1 /* Subroutines used for code generation on IA-32.
2 Copyright (C) 1988, 1992, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
3 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
30 #include "hard-reg-set.h"
32 #include "insn-config.h"
33 #include "conditions.h"
35 #include "insn-codes.h"
36 #include "insn-attr.h"
44 #include "basic-block.h"
47 #include "target-def.h"
48 #include "langhooks.h"
53 #include "tm-constrs.h"
57 static rtx legitimize_dllimport_symbol (rtx, bool);
59 #ifndef CHECK_STACK_LIMIT
60 #define CHECK_STACK_LIMIT (-1)
63 /* Return index of given mode in mult and division cost tables. */
64 #define MODE_INDEX(mode) \
65 ((mode) == QImode ? 0 \
66 : (mode) == HImode ? 1 \
67 : (mode) == SImode ? 2 \
68 : (mode) == DImode ? 3 \
71 /* Processor costs (relative to an add) */
72 /* We assume COSTS_N_INSNS is defined as (N)*4 and an addition is 2 bytes. */
73 #define COSTS_N_BYTES(N) ((N) * 2)
75 #define DUMMY_STRINGOP_ALGS {libcall, {{-1, libcall}}}
78 struct processor_costs ix86_size_cost = {/* costs for tuning for size */
79 COSTS_N_BYTES (2), /* cost of an add instruction */
80 COSTS_N_BYTES (3), /* cost of a lea instruction */
81 COSTS_N_BYTES (2), /* variable shift costs */
82 COSTS_N_BYTES (3), /* constant shift costs */
83 {COSTS_N_BYTES (3), /* cost of starting multiply for QI */
84 COSTS_N_BYTES (3), /* HI */
85 COSTS_N_BYTES (3), /* SI */
86 COSTS_N_BYTES (3), /* DI */
87 COSTS_N_BYTES (5)}, /* other */
88 0, /* cost of multiply per each bit set */
89 {COSTS_N_BYTES (3), /* cost of a divide/mod for QI */
90 COSTS_N_BYTES (3), /* HI */
91 COSTS_N_BYTES (3), /* SI */
92 COSTS_N_BYTES (3), /* DI */
93 COSTS_N_BYTES (5)}, /* other */
94 COSTS_N_BYTES (3), /* cost of movsx */
95 COSTS_N_BYTES (3), /* cost of movzx */
98 2, /* cost for loading QImode using movzbl */
99 {2, 2, 2}, /* cost of loading integer registers
100 in QImode, HImode and SImode.
101 Relative to reg-reg move (2). */
102 {2, 2, 2}, /* cost of storing integer registers */
103 2, /* cost of reg,reg fld/fst */
104 {2, 2, 2}, /* cost of loading fp registers
105 in SFmode, DFmode and XFmode */
106 {2, 2, 2}, /* cost of storing fp registers
107 in SFmode, DFmode and XFmode */
108 3, /* cost of moving MMX register */
109 {3, 3}, /* cost of loading MMX registers
110 in SImode and DImode */
111 {3, 3}, /* cost of storing MMX registers
112 in SImode and DImode */
113 3, /* cost of moving SSE register */
114 {3, 3, 3}, /* cost of loading SSE registers
115 in SImode, DImode and TImode */
116 {3, 3, 3}, /* cost of storing SSE registers
117 in SImode, DImode and TImode */
118 3, /* MMX or SSE register to integer */
119 0, /* size of l1 cache */
120 0, /* size of l2 cache */
121 0, /* size of prefetch block */
122 0, /* number of parallel prefetches */
124 COSTS_N_BYTES (2), /* cost of FADD and FSUB insns. */
125 COSTS_N_BYTES (2), /* cost of FMUL instruction. */
126 COSTS_N_BYTES (2), /* cost of FDIV instruction. */
127 COSTS_N_BYTES (2), /* cost of FABS instruction. */
128 COSTS_N_BYTES (2), /* cost of FCHS instruction. */
129 COSTS_N_BYTES (2), /* cost of FSQRT instruction. */
130 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
131 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
132 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
133 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
134 1, /* scalar_stmt_cost. */
135 1, /* scalar load_cost. */
136 1, /* scalar_store_cost. */
137 1, /* vec_stmt_cost. */
138 1, /* vec_to_scalar_cost. */
139 1, /* scalar_to_vec_cost. */
140 1, /* vec_align_load_cost. */
141 1, /* vec_unalign_load_cost. */
142 1, /* vec_store_cost. */
143 1, /* cond_taken_branch_cost. */
144 1, /* cond_not_taken_branch_cost. */
147 /* Processor costs (relative to an add) */
149 struct processor_costs i386_cost = { /* 386 specific costs */
150 COSTS_N_INSNS (1), /* cost of an add instruction */
151 COSTS_N_INSNS (1), /* cost of a lea instruction */
152 COSTS_N_INSNS (3), /* variable shift costs */
153 COSTS_N_INSNS (2), /* constant shift costs */
154 {COSTS_N_INSNS (6), /* cost of starting multiply for QI */
155 COSTS_N_INSNS (6), /* HI */
156 COSTS_N_INSNS (6), /* SI */
157 COSTS_N_INSNS (6), /* DI */
158 COSTS_N_INSNS (6)}, /* other */
159 COSTS_N_INSNS (1), /* cost of multiply per each bit set */
160 {COSTS_N_INSNS (23), /* cost of a divide/mod for QI */
161 COSTS_N_INSNS (23), /* HI */
162 COSTS_N_INSNS (23), /* SI */
163 COSTS_N_INSNS (23), /* DI */
164 COSTS_N_INSNS (23)}, /* other */
165 COSTS_N_INSNS (3), /* cost of movsx */
166 COSTS_N_INSNS (2), /* cost of movzx */
167 15, /* "large" insn */
169 4, /* cost for loading QImode using movzbl */
170 {2, 4, 2}, /* cost of loading integer registers
171 in QImode, HImode and SImode.
172 Relative to reg-reg move (2). */
173 {2, 4, 2}, /* cost of storing integer registers */
174 2, /* cost of reg,reg fld/fst */
175 {8, 8, 8}, /* cost of loading fp registers
176 in SFmode, DFmode and XFmode */
177 {8, 8, 8}, /* cost of storing fp registers
178 in SFmode, DFmode and XFmode */
179 2, /* cost of moving MMX register */
180 {4, 8}, /* cost of loading MMX registers
181 in SImode and DImode */
182 {4, 8}, /* cost of storing MMX registers
183 in SImode and DImode */
184 2, /* cost of moving SSE register */
185 {4, 8, 16}, /* cost of loading SSE registers
186 in SImode, DImode and TImode */
187 {4, 8, 16}, /* cost of storing SSE registers
188 in SImode, DImode and TImode */
189 3, /* MMX or SSE register to integer */
190 0, /* size of l1 cache */
191 0, /* size of l2 cache */
192 0, /* size of prefetch block */
193 0, /* number of parallel prefetches */
195 COSTS_N_INSNS (23), /* cost of FADD and FSUB insns. */
196 COSTS_N_INSNS (27), /* cost of FMUL instruction. */
197 COSTS_N_INSNS (88), /* cost of FDIV instruction. */
198 COSTS_N_INSNS (22), /* cost of FABS instruction. */
199 COSTS_N_INSNS (24), /* cost of FCHS instruction. */
200 COSTS_N_INSNS (122), /* cost of FSQRT instruction. */
201 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
202 DUMMY_STRINGOP_ALGS},
203 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
204 DUMMY_STRINGOP_ALGS},
205 1, /* scalar_stmt_cost. */
206 1, /* scalar load_cost. */
207 1, /* scalar_store_cost. */
208 1, /* vec_stmt_cost. */
209 1, /* vec_to_scalar_cost. */
210 1, /* scalar_to_vec_cost. */
211 1, /* vec_align_load_cost. */
212 2, /* vec_unalign_load_cost. */
213 1, /* vec_store_cost. */
214 3, /* cond_taken_branch_cost. */
215 1, /* cond_not_taken_branch_cost. */
219 struct processor_costs i486_cost = { /* 486 specific costs */
220 COSTS_N_INSNS (1), /* cost of an add instruction */
221 COSTS_N_INSNS (1), /* cost of a lea instruction */
222 COSTS_N_INSNS (3), /* variable shift costs */
223 COSTS_N_INSNS (2), /* constant shift costs */
224 {COSTS_N_INSNS (12), /* cost of starting multiply for QI */
225 COSTS_N_INSNS (12), /* HI */
226 COSTS_N_INSNS (12), /* SI */
227 COSTS_N_INSNS (12), /* DI */
228 COSTS_N_INSNS (12)}, /* other */
229 1, /* cost of multiply per each bit set */
230 {COSTS_N_INSNS (40), /* cost of a divide/mod for QI */
231 COSTS_N_INSNS (40), /* HI */
232 COSTS_N_INSNS (40), /* SI */
233 COSTS_N_INSNS (40), /* DI */
234 COSTS_N_INSNS (40)}, /* other */
235 COSTS_N_INSNS (3), /* cost of movsx */
236 COSTS_N_INSNS (2), /* cost of movzx */
237 15, /* "large" insn */
239 4, /* cost for loading QImode using movzbl */
240 {2, 4, 2}, /* cost of loading integer registers
241 in QImode, HImode and SImode.
242 Relative to reg-reg move (2). */
243 {2, 4, 2}, /* cost of storing integer registers */
244 2, /* cost of reg,reg fld/fst */
245 {8, 8, 8}, /* cost of loading fp registers
246 in SFmode, DFmode and XFmode */
247 {8, 8, 8}, /* cost of storing fp registers
248 in SFmode, DFmode and XFmode */
249 2, /* cost of moving MMX register */
250 {4, 8}, /* cost of loading MMX registers
251 in SImode and DImode */
252 {4, 8}, /* cost of storing MMX registers
253 in SImode and DImode */
254 2, /* cost of moving SSE register */
255 {4, 8, 16}, /* cost of loading SSE registers
256 in SImode, DImode and TImode */
257 {4, 8, 16}, /* cost of storing SSE registers
258 in SImode, DImode and TImode */
259 3, /* MMX or SSE register to integer */
260 4, /* size of l1 cache. 486 has 8kB cache
261 shared for code and data, so 4kB is
262 not really precise. */
263 4, /* size of l2 cache */
264 0, /* size of prefetch block */
265 0, /* number of parallel prefetches */
267 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
268 COSTS_N_INSNS (16), /* cost of FMUL instruction. */
269 COSTS_N_INSNS (73), /* cost of FDIV instruction. */
270 COSTS_N_INSNS (3), /* cost of FABS instruction. */
271 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
272 COSTS_N_INSNS (83), /* cost of FSQRT instruction. */
273 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
274 DUMMY_STRINGOP_ALGS},
275 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
276 DUMMY_STRINGOP_ALGS},
277 1, /* scalar_stmt_cost. */
278 1, /* scalar load_cost. */
279 1, /* scalar_store_cost. */
280 1, /* vec_stmt_cost. */
281 1, /* vec_to_scalar_cost. */
282 1, /* scalar_to_vec_cost. */
283 1, /* vec_align_load_cost. */
284 2, /* vec_unalign_load_cost. */
285 1, /* vec_store_cost. */
286 3, /* cond_taken_branch_cost. */
287 1, /* cond_not_taken_branch_cost. */
291 struct processor_costs pentium_cost = {
292 COSTS_N_INSNS (1), /* cost of an add instruction */
293 COSTS_N_INSNS (1), /* cost of a lea instruction */
294 COSTS_N_INSNS (4), /* variable shift costs */
295 COSTS_N_INSNS (1), /* constant shift costs */
296 {COSTS_N_INSNS (11), /* cost of starting multiply for QI */
297 COSTS_N_INSNS (11), /* HI */
298 COSTS_N_INSNS (11), /* SI */
299 COSTS_N_INSNS (11), /* DI */
300 COSTS_N_INSNS (11)}, /* other */
301 0, /* cost of multiply per each bit set */
302 {COSTS_N_INSNS (25), /* cost of a divide/mod for QI */
303 COSTS_N_INSNS (25), /* HI */
304 COSTS_N_INSNS (25), /* SI */
305 COSTS_N_INSNS (25), /* DI */
306 COSTS_N_INSNS (25)}, /* other */
307 COSTS_N_INSNS (3), /* cost of movsx */
308 COSTS_N_INSNS (2), /* cost of movzx */
309 8, /* "large" insn */
311 6, /* cost for loading QImode using movzbl */
312 {2, 4, 2}, /* cost of loading integer registers
313 in QImode, HImode and SImode.
314 Relative to reg-reg move (2). */
315 {2, 4, 2}, /* cost of storing integer registers */
316 2, /* cost of reg,reg fld/fst */
317 {2, 2, 6}, /* cost of loading fp registers
318 in SFmode, DFmode and XFmode */
319 {4, 4, 6}, /* cost of storing fp registers
320 in SFmode, DFmode and XFmode */
321 8, /* cost of moving MMX register */
322 {8, 8}, /* cost of loading MMX registers
323 in SImode and DImode */
324 {8, 8}, /* cost of storing MMX registers
325 in SImode and DImode */
326 2, /* cost of moving SSE register */
327 {4, 8, 16}, /* cost of loading SSE registers
328 in SImode, DImode and TImode */
329 {4, 8, 16}, /* cost of storing SSE registers
330 in SImode, DImode and TImode */
331 3, /* MMX or SSE register to integer */
332 8, /* size of l1 cache. */
333 8, /* size of l2 cache */
334 0, /* size of prefetch block */
335 0, /* number of parallel prefetches */
337 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
338 COSTS_N_INSNS (3), /* cost of FMUL instruction. */
339 COSTS_N_INSNS (39), /* cost of FDIV instruction. */
340 COSTS_N_INSNS (1), /* cost of FABS instruction. */
341 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
342 COSTS_N_INSNS (70), /* cost of FSQRT instruction. */
343 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
344 DUMMY_STRINGOP_ALGS},
345 {{libcall, {{-1, rep_prefix_4_byte}}},
346 DUMMY_STRINGOP_ALGS},
347 1, /* scalar_stmt_cost. */
348 1, /* scalar load_cost. */
349 1, /* scalar_store_cost. */
350 1, /* vec_stmt_cost. */
351 1, /* vec_to_scalar_cost. */
352 1, /* scalar_to_vec_cost. */
353 1, /* vec_align_load_cost. */
354 2, /* vec_unalign_load_cost. */
355 1, /* vec_store_cost. */
356 3, /* cond_taken_branch_cost. */
357 1, /* cond_not_taken_branch_cost. */
361 struct processor_costs pentiumpro_cost = {
362 COSTS_N_INSNS (1), /* cost of an add instruction */
363 COSTS_N_INSNS (1), /* cost of a lea instruction */
364 COSTS_N_INSNS (1), /* variable shift costs */
365 COSTS_N_INSNS (1), /* constant shift costs */
366 {COSTS_N_INSNS (4), /* cost of starting multiply for QI */
367 COSTS_N_INSNS (4), /* HI */
368 COSTS_N_INSNS (4), /* SI */
369 COSTS_N_INSNS (4), /* DI */
370 COSTS_N_INSNS (4)}, /* other */
371 0, /* cost of multiply per each bit set */
372 {COSTS_N_INSNS (17), /* cost of a divide/mod for QI */
373 COSTS_N_INSNS (17), /* HI */
374 COSTS_N_INSNS (17), /* SI */
375 COSTS_N_INSNS (17), /* DI */
376 COSTS_N_INSNS (17)}, /* other */
377 COSTS_N_INSNS (1), /* cost of movsx */
378 COSTS_N_INSNS (1), /* cost of movzx */
379 8, /* "large" insn */
381 2, /* cost for loading QImode using movzbl */
382 {4, 4, 4}, /* cost of loading integer registers
383 in QImode, HImode and SImode.
384 Relative to reg-reg move (2). */
385 {2, 2, 2}, /* cost of storing integer registers */
386 2, /* cost of reg,reg fld/fst */
387 {2, 2, 6}, /* cost of loading fp registers
388 in SFmode, DFmode and XFmode */
389 {4, 4, 6}, /* cost of storing fp registers
390 in SFmode, DFmode and XFmode */
391 2, /* cost of moving MMX register */
392 {2, 2}, /* cost of loading MMX registers
393 in SImode and DImode */
394 {2, 2}, /* cost of storing MMX registers
395 in SImode and DImode */
396 2, /* cost of moving SSE register */
397 {2, 2, 8}, /* cost of loading SSE registers
398 in SImode, DImode and TImode */
399 {2, 2, 8}, /* cost of storing SSE registers
400 in SImode, DImode and TImode */
401 3, /* MMX or SSE register to integer */
402 8, /* size of l1 cache. */
403 256, /* size of l2 cache */
404 32, /* size of prefetch block */
405 6, /* number of parallel prefetches */
407 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
408 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
409 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
410 COSTS_N_INSNS (2), /* cost of FABS instruction. */
411 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
412 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
413 /* PentiumPro has optimized rep instructions for blocks aligned by 8 bytes (we ensure
414 the alignment). For small blocks inline loop is still a noticeable win, for bigger
415 blocks either rep movsl or rep movsb is way to go. Rep movsb has apparently
416 more expensive startup time in CPU, but after 4K the difference is down in the noise.
418 {{rep_prefix_4_byte, {{128, loop}, {1024, unrolled_loop},
419 {8192, rep_prefix_4_byte}, {-1, rep_prefix_1_byte}}},
420 DUMMY_STRINGOP_ALGS},
421 {{rep_prefix_4_byte, {{1024, unrolled_loop},
422 {8192, rep_prefix_4_byte}, {-1, libcall}}},
423 DUMMY_STRINGOP_ALGS},
424 1, /* scalar_stmt_cost. */
425 1, /* scalar load_cost. */
426 1, /* scalar_store_cost. */
427 1, /* vec_stmt_cost. */
428 1, /* vec_to_scalar_cost. */
429 1, /* scalar_to_vec_cost. */
430 1, /* vec_align_load_cost. */
431 2, /* vec_unalign_load_cost. */
432 1, /* vec_store_cost. */
433 3, /* cond_taken_branch_cost. */
434 1, /* cond_not_taken_branch_cost. */
438 struct processor_costs geode_cost = {
439 COSTS_N_INSNS (1), /* cost of an add instruction */
440 COSTS_N_INSNS (1), /* cost of a lea instruction */
441 COSTS_N_INSNS (2), /* variable shift costs */
442 COSTS_N_INSNS (1), /* constant shift costs */
443 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
444 COSTS_N_INSNS (4), /* HI */
445 COSTS_N_INSNS (7), /* SI */
446 COSTS_N_INSNS (7), /* DI */
447 COSTS_N_INSNS (7)}, /* other */
448 0, /* cost of multiply per each bit set */
449 {COSTS_N_INSNS (15), /* cost of a divide/mod for QI */
450 COSTS_N_INSNS (23), /* HI */
451 COSTS_N_INSNS (39), /* SI */
452 COSTS_N_INSNS (39), /* DI */
453 COSTS_N_INSNS (39)}, /* other */
454 COSTS_N_INSNS (1), /* cost of movsx */
455 COSTS_N_INSNS (1), /* cost of movzx */
456 8, /* "large" insn */
458 1, /* cost for loading QImode using movzbl */
459 {1, 1, 1}, /* cost of loading integer registers
460 in QImode, HImode and SImode.
461 Relative to reg-reg move (2). */
462 {1, 1, 1}, /* cost of storing integer registers */
463 1, /* cost of reg,reg fld/fst */
464 {1, 1, 1}, /* cost of loading fp registers
465 in SFmode, DFmode and XFmode */
466 {4, 6, 6}, /* cost of storing fp registers
467 in SFmode, DFmode and XFmode */
469 1, /* cost of moving MMX register */
470 {1, 1}, /* cost of loading MMX registers
471 in SImode and DImode */
472 {1, 1}, /* cost of storing MMX registers
473 in SImode and DImode */
474 1, /* cost of moving SSE register */
475 {1, 1, 1}, /* cost of loading SSE registers
476 in SImode, DImode and TImode */
477 {1, 1, 1}, /* cost of storing SSE registers
478 in SImode, DImode and TImode */
479 1, /* MMX or SSE register to integer */
480 64, /* size of l1 cache. */
481 128, /* size of l2 cache. */
482 32, /* size of prefetch block */
483 1, /* number of parallel prefetches */
485 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
486 COSTS_N_INSNS (11), /* cost of FMUL instruction. */
487 COSTS_N_INSNS (47), /* cost of FDIV instruction. */
488 COSTS_N_INSNS (1), /* cost of FABS instruction. */
489 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
490 COSTS_N_INSNS (54), /* cost of FSQRT instruction. */
491 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
492 DUMMY_STRINGOP_ALGS},
493 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
494 DUMMY_STRINGOP_ALGS},
495 1, /* scalar_stmt_cost. */
496 1, /* scalar load_cost. */
497 1, /* scalar_store_cost. */
498 1, /* vec_stmt_cost. */
499 1, /* vec_to_scalar_cost. */
500 1, /* scalar_to_vec_cost. */
501 1, /* vec_align_load_cost. */
502 2, /* vec_unalign_load_cost. */
503 1, /* vec_store_cost. */
504 3, /* cond_taken_branch_cost. */
505 1, /* cond_not_taken_branch_cost. */
509 struct processor_costs k6_cost = {
510 COSTS_N_INSNS (1), /* cost of an add instruction */
511 COSTS_N_INSNS (2), /* cost of a lea instruction */
512 COSTS_N_INSNS (1), /* variable shift costs */
513 COSTS_N_INSNS (1), /* constant shift costs */
514 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
515 COSTS_N_INSNS (3), /* HI */
516 COSTS_N_INSNS (3), /* SI */
517 COSTS_N_INSNS (3), /* DI */
518 COSTS_N_INSNS (3)}, /* other */
519 0, /* cost of multiply per each bit set */
520 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
521 COSTS_N_INSNS (18), /* HI */
522 COSTS_N_INSNS (18), /* SI */
523 COSTS_N_INSNS (18), /* DI */
524 COSTS_N_INSNS (18)}, /* other */
525 COSTS_N_INSNS (2), /* cost of movsx */
526 COSTS_N_INSNS (2), /* cost of movzx */
527 8, /* "large" insn */
529 3, /* cost for loading QImode using movzbl */
530 {4, 5, 4}, /* cost of loading integer registers
531 in QImode, HImode and SImode.
532 Relative to reg-reg move (2). */
533 {2, 3, 2}, /* cost of storing integer registers */
534 4, /* cost of reg,reg fld/fst */
535 {6, 6, 6}, /* cost of loading fp registers
536 in SFmode, DFmode and XFmode */
537 {4, 4, 4}, /* cost of storing fp registers
538 in SFmode, DFmode and XFmode */
539 2, /* cost of moving MMX register */
540 {2, 2}, /* cost of loading MMX registers
541 in SImode and DImode */
542 {2, 2}, /* cost of storing MMX registers
543 in SImode and DImode */
544 2, /* cost of moving SSE register */
545 {2, 2, 8}, /* cost of loading SSE registers
546 in SImode, DImode and TImode */
547 {2, 2, 8}, /* cost of storing SSE registers
548 in SImode, DImode and TImode */
549 6, /* MMX or SSE register to integer */
550 32, /* size of l1 cache. */
551 32, /* size of l2 cache. Some models
552 have integrated l2 cache, but
553 optimizing for k6 is not important
554 enough to worry about that. */
555 32, /* size of prefetch block */
556 1, /* number of parallel prefetches */
558 COSTS_N_INSNS (2), /* cost of FADD and FSUB insns. */
559 COSTS_N_INSNS (2), /* cost of FMUL instruction. */
560 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
561 COSTS_N_INSNS (2), /* cost of FABS instruction. */
562 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
563 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
564 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
565 DUMMY_STRINGOP_ALGS},
566 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
567 DUMMY_STRINGOP_ALGS},
568 1, /* scalar_stmt_cost. */
569 1, /* scalar load_cost. */
570 1, /* scalar_store_cost. */
571 1, /* vec_stmt_cost. */
572 1, /* vec_to_scalar_cost. */
573 1, /* scalar_to_vec_cost. */
574 1, /* vec_align_load_cost. */
575 2, /* vec_unalign_load_cost. */
576 1, /* vec_store_cost. */
577 3, /* cond_taken_branch_cost. */
578 1, /* cond_not_taken_branch_cost. */
582 struct processor_costs athlon_cost = {
583 COSTS_N_INSNS (1), /* cost of an add instruction */
584 COSTS_N_INSNS (2), /* cost of a lea instruction */
585 COSTS_N_INSNS (1), /* variable shift costs */
586 COSTS_N_INSNS (1), /* constant shift costs */
587 {COSTS_N_INSNS (5), /* cost of starting multiply for QI */
588 COSTS_N_INSNS (5), /* HI */
589 COSTS_N_INSNS (5), /* SI */
590 COSTS_N_INSNS (5), /* DI */
591 COSTS_N_INSNS (5)}, /* other */
592 0, /* cost of multiply per each bit set */
593 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
594 COSTS_N_INSNS (26), /* HI */
595 COSTS_N_INSNS (42), /* SI */
596 COSTS_N_INSNS (74), /* DI */
597 COSTS_N_INSNS (74)}, /* other */
598 COSTS_N_INSNS (1), /* cost of movsx */
599 COSTS_N_INSNS (1), /* cost of movzx */
600 8, /* "large" insn */
602 4, /* cost for loading QImode using movzbl */
603 {3, 4, 3}, /* cost of loading integer registers
604 in QImode, HImode and SImode.
605 Relative to reg-reg move (2). */
606 {3, 4, 3}, /* cost of storing integer registers */
607 4, /* cost of reg,reg fld/fst */
608 {4, 4, 12}, /* cost of loading fp registers
609 in SFmode, DFmode and XFmode */
610 {6, 6, 8}, /* cost of storing fp registers
611 in SFmode, DFmode and XFmode */
612 2, /* cost of moving MMX register */
613 {4, 4}, /* cost of loading MMX registers
614 in SImode and DImode */
615 {4, 4}, /* cost of storing MMX registers
616 in SImode and DImode */
617 2, /* cost of moving SSE register */
618 {4, 4, 6}, /* cost of loading SSE registers
619 in SImode, DImode and TImode */
620 {4, 4, 5}, /* cost of storing SSE registers
621 in SImode, DImode and TImode */
622 5, /* MMX or SSE register to integer */
623 64, /* size of l1 cache. */
624 256, /* size of l2 cache. */
625 64, /* size of prefetch block */
626 6, /* number of parallel prefetches */
628 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
629 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
630 COSTS_N_INSNS (24), /* cost of FDIV instruction. */
631 COSTS_N_INSNS (2), /* cost of FABS instruction. */
632 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
633 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
634 /* For some reason, Athlon deals better with REP prefix (relative to loops)
635 compared to K8. Alignment becomes important after 8 bytes for memcpy and
636 128 bytes for memset. */
637 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
638 DUMMY_STRINGOP_ALGS},
639 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
640 DUMMY_STRINGOP_ALGS},
641 1, /* scalar_stmt_cost. */
642 1, /* scalar load_cost. */
643 1, /* scalar_store_cost. */
644 1, /* vec_stmt_cost. */
645 1, /* vec_to_scalar_cost. */
646 1, /* scalar_to_vec_cost. */
647 1, /* vec_align_load_cost. */
648 2, /* vec_unalign_load_cost. */
649 1, /* vec_store_cost. */
650 3, /* cond_taken_branch_cost. */
651 1, /* cond_not_taken_branch_cost. */
655 struct processor_costs k8_cost = {
656 COSTS_N_INSNS (1), /* cost of an add instruction */
657 COSTS_N_INSNS (2), /* cost of a lea instruction */
658 COSTS_N_INSNS (1), /* variable shift costs */
659 COSTS_N_INSNS (1), /* constant shift costs */
660 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
661 COSTS_N_INSNS (4), /* HI */
662 COSTS_N_INSNS (3), /* SI */
663 COSTS_N_INSNS (4), /* DI */
664 COSTS_N_INSNS (5)}, /* other */
665 0, /* cost of multiply per each bit set */
666 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
667 COSTS_N_INSNS (26), /* HI */
668 COSTS_N_INSNS (42), /* SI */
669 COSTS_N_INSNS (74), /* DI */
670 COSTS_N_INSNS (74)}, /* other */
671 COSTS_N_INSNS (1), /* cost of movsx */
672 COSTS_N_INSNS (1), /* cost of movzx */
673 8, /* "large" insn */
675 4, /* cost for loading QImode using movzbl */
676 {3, 4, 3}, /* cost of loading integer registers
677 in QImode, HImode and SImode.
678 Relative to reg-reg move (2). */
679 {3, 4, 3}, /* cost of storing integer registers */
680 4, /* cost of reg,reg fld/fst */
681 {4, 4, 12}, /* cost of loading fp registers
682 in SFmode, DFmode and XFmode */
683 {6, 6, 8}, /* cost of storing fp registers
684 in SFmode, DFmode and XFmode */
685 2, /* cost of moving MMX register */
686 {3, 3}, /* cost of loading MMX registers
687 in SImode and DImode */
688 {4, 4}, /* cost of storing MMX registers
689 in SImode and DImode */
690 2, /* cost of moving SSE register */
691 {4, 3, 6}, /* cost of loading SSE registers
692 in SImode, DImode and TImode */
693 {4, 4, 5}, /* cost of storing SSE registers
694 in SImode, DImode and TImode */
695 5, /* MMX or SSE register to integer */
696 64, /* size of l1 cache. */
697 512, /* size of l2 cache. */
698 64, /* size of prefetch block */
699 /* New AMD processors never drop prefetches; if they cannot be performed
700 immediately, they are queued. We set number of simultaneous prefetches
701 to a large constant to reflect this (it probably is not a good idea not
702 to limit number of prefetches at all, as their execution also takes some
704 100, /* number of parallel prefetches */
706 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
707 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
708 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
709 COSTS_N_INSNS (2), /* cost of FABS instruction. */
710 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
711 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
712 /* K8 has optimized REP instruction for medium sized blocks, but for very small
713 blocks it is better to use loop. For large blocks, libcall can do
714 nontemporary accesses and beat inline considerably. */
715 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
716 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
717 {{libcall, {{8, loop}, {24, unrolled_loop},
718 {2048, rep_prefix_4_byte}, {-1, libcall}}},
719 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
720 4, /* scalar_stmt_cost. */
721 2, /* scalar load_cost. */
722 2, /* scalar_store_cost. */
723 5, /* vec_stmt_cost. */
724 0, /* vec_to_scalar_cost. */
725 2, /* scalar_to_vec_cost. */
726 2, /* vec_align_load_cost. */
727 3, /* vec_unalign_load_cost. */
728 3, /* vec_store_cost. */
729 3, /* cond_taken_branch_cost. */
730 2, /* cond_not_taken_branch_cost. */
733 struct processor_costs amdfam10_cost = {
734 COSTS_N_INSNS (1), /* cost of an add instruction */
735 COSTS_N_INSNS (2), /* cost of a lea instruction */
736 COSTS_N_INSNS (1), /* variable shift costs */
737 COSTS_N_INSNS (1), /* constant shift costs */
738 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
739 COSTS_N_INSNS (4), /* HI */
740 COSTS_N_INSNS (3), /* SI */
741 COSTS_N_INSNS (4), /* DI */
742 COSTS_N_INSNS (5)}, /* other */
743 0, /* cost of multiply per each bit set */
744 {COSTS_N_INSNS (19), /* cost of a divide/mod for QI */
745 COSTS_N_INSNS (35), /* HI */
746 COSTS_N_INSNS (51), /* SI */
747 COSTS_N_INSNS (83), /* DI */
748 COSTS_N_INSNS (83)}, /* other */
749 COSTS_N_INSNS (1), /* cost of movsx */
750 COSTS_N_INSNS (1), /* cost of movzx */
751 8, /* "large" insn */
753 4, /* cost for loading QImode using movzbl */
754 {3, 4, 3}, /* cost of loading integer registers
755 in QImode, HImode and SImode.
756 Relative to reg-reg move (2). */
757 {3, 4, 3}, /* cost of storing integer registers */
758 4, /* cost of reg,reg fld/fst */
759 {4, 4, 12}, /* cost of loading fp registers
760 in SFmode, DFmode and XFmode */
761 {6, 6, 8}, /* cost of storing fp registers
762 in SFmode, DFmode and XFmode */
763 2, /* cost of moving MMX register */
764 {3, 3}, /* cost of loading MMX registers
765 in SImode and DImode */
766 {4, 4}, /* cost of storing MMX registers
767 in SImode and DImode */
768 2, /* cost of moving SSE register */
769 {4, 4, 3}, /* cost of loading SSE registers
770 in SImode, DImode and TImode */
771 {4, 4, 5}, /* cost of storing SSE registers
772 in SImode, DImode and TImode */
773 3, /* MMX or SSE register to integer */
775 MOVD reg64, xmmreg Double FSTORE 4
776 MOVD reg32, xmmreg Double FSTORE 4
778 MOVD reg64, xmmreg Double FADD 3
780 MOVD reg32, xmmreg Double FADD 3
782 64, /* size of l1 cache. */
783 512, /* size of l2 cache. */
784 64, /* size of prefetch block */
785 /* New AMD processors never drop prefetches; if they cannot be performed
786 immediately, they are queued. We set number of simultaneous prefetches
787 to a large constant to reflect this (it probably is not a good idea not
788 to limit number of prefetches at all, as their execution also takes some
790 100, /* number of parallel prefetches */
792 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
793 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
794 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
795 COSTS_N_INSNS (2), /* cost of FABS instruction. */
796 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
797 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
799 /* AMDFAM10 has optimized REP instruction for medium sized blocks, but for
800 very small blocks it is better to use loop. For large blocks, libcall can
801 do nontemporary accesses and beat inline considerably. */
802 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
803 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
804 {{libcall, {{8, loop}, {24, unrolled_loop},
805 {2048, rep_prefix_4_byte}, {-1, libcall}}},
806 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
807 4, /* scalar_stmt_cost. */
808 2, /* scalar load_cost. */
809 2, /* scalar_store_cost. */
810 6, /* vec_stmt_cost. */
811 0, /* vec_to_scalar_cost. */
812 2, /* scalar_to_vec_cost. */
813 2, /* vec_align_load_cost. */
814 2, /* vec_unalign_load_cost. */
815 2, /* vec_store_cost. */
816 2, /* cond_taken_branch_cost. */
817 1, /* cond_not_taken_branch_cost. */
821 struct processor_costs pentium4_cost = {
822 COSTS_N_INSNS (1), /* cost of an add instruction */
823 COSTS_N_INSNS (3), /* cost of a lea instruction */
824 COSTS_N_INSNS (4), /* variable shift costs */
825 COSTS_N_INSNS (4), /* constant shift costs */
826 {COSTS_N_INSNS (15), /* cost of starting multiply for QI */
827 COSTS_N_INSNS (15), /* HI */
828 COSTS_N_INSNS (15), /* SI */
829 COSTS_N_INSNS (15), /* DI */
830 COSTS_N_INSNS (15)}, /* other */
831 0, /* cost of multiply per each bit set */
832 {COSTS_N_INSNS (56), /* cost of a divide/mod for QI */
833 COSTS_N_INSNS (56), /* HI */
834 COSTS_N_INSNS (56), /* SI */
835 COSTS_N_INSNS (56), /* DI */
836 COSTS_N_INSNS (56)}, /* other */
837 COSTS_N_INSNS (1), /* cost of movsx */
838 COSTS_N_INSNS (1), /* cost of movzx */
839 16, /* "large" insn */
841 2, /* cost for loading QImode using movzbl */
842 {4, 5, 4}, /* cost of loading integer registers
843 in QImode, HImode and SImode.
844 Relative to reg-reg move (2). */
845 {2, 3, 2}, /* cost of storing integer registers */
846 2, /* cost of reg,reg fld/fst */
847 {2, 2, 6}, /* cost of loading fp registers
848 in SFmode, DFmode and XFmode */
849 {4, 4, 6}, /* cost of storing fp registers
850 in SFmode, DFmode and XFmode */
851 2, /* cost of moving MMX register */
852 {2, 2}, /* cost of loading MMX registers
853 in SImode and DImode */
854 {2, 2}, /* cost of storing MMX registers
855 in SImode and DImode */
856 12, /* cost of moving SSE register */
857 {12, 12, 12}, /* cost of loading SSE registers
858 in SImode, DImode and TImode */
859 {2, 2, 8}, /* cost of storing SSE registers
860 in SImode, DImode and TImode */
861 10, /* MMX or SSE register to integer */
862 8, /* size of l1 cache. */
863 256, /* size of l2 cache. */
864 64, /* size of prefetch block */
865 6, /* number of parallel prefetches */
867 COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */
868 COSTS_N_INSNS (7), /* cost of FMUL instruction. */
869 COSTS_N_INSNS (43), /* cost of FDIV instruction. */
870 COSTS_N_INSNS (2), /* cost of FABS instruction. */
871 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
872 COSTS_N_INSNS (43), /* cost of FSQRT instruction. */
873 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
874 DUMMY_STRINGOP_ALGS},
875 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
877 DUMMY_STRINGOP_ALGS},
878 1, /* scalar_stmt_cost. */
879 1, /* scalar load_cost. */
880 1, /* scalar_store_cost. */
881 1, /* vec_stmt_cost. */
882 1, /* vec_to_scalar_cost. */
883 1, /* scalar_to_vec_cost. */
884 1, /* vec_align_load_cost. */
885 2, /* vec_unalign_load_cost. */
886 1, /* vec_store_cost. */
887 3, /* cond_taken_branch_cost. */
888 1, /* cond_not_taken_branch_cost. */
892 struct processor_costs nocona_cost = {
893 COSTS_N_INSNS (1), /* cost of an add instruction */
894 COSTS_N_INSNS (1), /* cost of a lea instruction */
895 COSTS_N_INSNS (1), /* variable shift costs */
896 COSTS_N_INSNS (1), /* constant shift costs */
897 {COSTS_N_INSNS (10), /* cost of starting multiply for QI */
898 COSTS_N_INSNS (10), /* HI */
899 COSTS_N_INSNS (10), /* SI */
900 COSTS_N_INSNS (10), /* DI */
901 COSTS_N_INSNS (10)}, /* other */
902 0, /* cost of multiply per each bit set */
903 {COSTS_N_INSNS (66), /* cost of a divide/mod for QI */
904 COSTS_N_INSNS (66), /* HI */
905 COSTS_N_INSNS (66), /* SI */
906 COSTS_N_INSNS (66), /* DI */
907 COSTS_N_INSNS (66)}, /* other */
908 COSTS_N_INSNS (1), /* cost of movsx */
909 COSTS_N_INSNS (1), /* cost of movzx */
910 16, /* "large" insn */
912 4, /* cost for loading QImode using movzbl */
913 {4, 4, 4}, /* cost of loading integer registers
914 in QImode, HImode and SImode.
915 Relative to reg-reg move (2). */
916 {4, 4, 4}, /* cost of storing integer registers */
917 3, /* cost of reg,reg fld/fst */
918 {12, 12, 12}, /* cost of loading fp registers
919 in SFmode, DFmode and XFmode */
920 {4, 4, 4}, /* cost of storing fp registers
921 in SFmode, DFmode and XFmode */
922 6, /* cost of moving MMX register */
923 {12, 12}, /* cost of loading MMX registers
924 in SImode and DImode */
925 {12, 12}, /* cost of storing MMX registers
926 in SImode and DImode */
927 6, /* cost of moving SSE register */
928 {12, 12, 12}, /* cost of loading SSE registers
929 in SImode, DImode and TImode */
930 {12, 12, 12}, /* cost of storing SSE registers
931 in SImode, DImode and TImode */
932 8, /* MMX or SSE register to integer */
933 8, /* size of l1 cache. */
934 1024, /* size of l2 cache. */
935 128, /* size of prefetch block */
936 8, /* number of parallel prefetches */
938 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
939 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
940 COSTS_N_INSNS (40), /* cost of FDIV instruction. */
941 COSTS_N_INSNS (3), /* cost of FABS instruction. */
942 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
943 COSTS_N_INSNS (44), /* cost of FSQRT instruction. */
944 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
945 {libcall, {{32, loop}, {20000, rep_prefix_8_byte},
946 {100000, unrolled_loop}, {-1, libcall}}}},
947 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
949 {libcall, {{24, loop}, {64, unrolled_loop},
950 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
951 1, /* scalar_stmt_cost. */
952 1, /* scalar load_cost. */
953 1, /* scalar_store_cost. */
954 1, /* vec_stmt_cost. */
955 1, /* vec_to_scalar_cost. */
956 1, /* scalar_to_vec_cost. */
957 1, /* vec_align_load_cost. */
958 2, /* vec_unalign_load_cost. */
959 1, /* vec_store_cost. */
960 3, /* cond_taken_branch_cost. */
961 1, /* cond_not_taken_branch_cost. */
965 struct processor_costs core2_cost = {
966 COSTS_N_INSNS (1), /* cost of an add instruction */
967 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
968 COSTS_N_INSNS (1), /* variable shift costs */
969 COSTS_N_INSNS (1), /* constant shift costs */
970 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
971 COSTS_N_INSNS (3), /* HI */
972 COSTS_N_INSNS (3), /* SI */
973 COSTS_N_INSNS (3), /* DI */
974 COSTS_N_INSNS (3)}, /* other */
975 0, /* cost of multiply per each bit set */
976 {COSTS_N_INSNS (22), /* cost of a divide/mod for QI */
977 COSTS_N_INSNS (22), /* HI */
978 COSTS_N_INSNS (22), /* SI */
979 COSTS_N_INSNS (22), /* DI */
980 COSTS_N_INSNS (22)}, /* other */
981 COSTS_N_INSNS (1), /* cost of movsx */
982 COSTS_N_INSNS (1), /* cost of movzx */
983 8, /* "large" insn */
985 2, /* cost for loading QImode using movzbl */
986 {6, 6, 6}, /* cost of loading integer registers
987 in QImode, HImode and SImode.
988 Relative to reg-reg move (2). */
989 {4, 4, 4}, /* cost of storing integer registers */
990 2, /* cost of reg,reg fld/fst */
991 {6, 6, 6}, /* cost of loading fp registers
992 in SFmode, DFmode and XFmode */
993 {4, 4, 4}, /* cost of storing fp registers
994 in SFmode, DFmode and XFmode */
995 2, /* cost of moving MMX register */
996 {6, 6}, /* cost of loading MMX registers
997 in SImode and DImode */
998 {4, 4}, /* cost of storing MMX registers
999 in SImode and DImode */
1000 2, /* cost of moving SSE register */
1001 {6, 6, 6}, /* cost of loading SSE registers
1002 in SImode, DImode and TImode */
1003 {4, 4, 4}, /* cost of storing SSE registers
1004 in SImode, DImode and TImode */
1005 2, /* MMX or SSE register to integer */
1006 32, /* size of l1 cache. */
1007 2048, /* size of l2 cache. */
1008 128, /* size of prefetch block */
1009 8, /* number of parallel prefetches */
1010 3, /* Branch cost */
1011 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
1012 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
1013 COSTS_N_INSNS (32), /* cost of FDIV instruction. */
1014 COSTS_N_INSNS (1), /* cost of FABS instruction. */
1015 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
1016 COSTS_N_INSNS (58), /* cost of FSQRT instruction. */
1017 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1018 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1019 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1020 {{libcall, {{8, loop}, {15, unrolled_loop},
1021 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1022 {libcall, {{24, loop}, {32, unrolled_loop},
1023 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1024 1, /* scalar_stmt_cost. */
1025 1, /* scalar load_cost. */
1026 1, /* scalar_store_cost. */
1027 1, /* vec_stmt_cost. */
1028 1, /* vec_to_scalar_cost. */
1029 1, /* scalar_to_vec_cost. */
1030 1, /* vec_align_load_cost. */
1031 2, /* vec_unalign_load_cost. */
1032 1, /* vec_store_cost. */
1033 3, /* cond_taken_branch_cost. */
1034 1, /* cond_not_taken_branch_cost. */
1038 struct processor_costs atom_cost = {
1039 COSTS_N_INSNS (1), /* cost of an add instruction */
1040 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1041 COSTS_N_INSNS (1), /* variable shift costs */
1042 COSTS_N_INSNS (1), /* constant shift costs */
1043 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1044 COSTS_N_INSNS (4), /* HI */
1045 COSTS_N_INSNS (3), /* SI */
1046 COSTS_N_INSNS (4), /* DI */
1047 COSTS_N_INSNS (2)}, /* other */
1048 0, /* cost of multiply per each bit set */
1049 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1050 COSTS_N_INSNS (26), /* HI */
1051 COSTS_N_INSNS (42), /* SI */
1052 COSTS_N_INSNS (74), /* DI */
1053 COSTS_N_INSNS (74)}, /* other */
1054 COSTS_N_INSNS (1), /* cost of movsx */
1055 COSTS_N_INSNS (1), /* cost of movzx */
1056 8, /* "large" insn */
1057 17, /* MOVE_RATIO */
1058 2, /* cost for loading QImode using movzbl */
1059 {4, 4, 4}, /* cost of loading integer registers
1060 in QImode, HImode and SImode.
1061 Relative to reg-reg move (2). */
1062 {4, 4, 4}, /* cost of storing integer registers */
1063 4, /* cost of reg,reg fld/fst */
1064 {12, 12, 12}, /* cost of loading fp registers
1065 in SFmode, DFmode and XFmode */
1066 {6, 6, 8}, /* cost of storing fp registers
1067 in SFmode, DFmode and XFmode */
1068 2, /* cost of moving MMX register */
1069 {8, 8}, /* cost of loading MMX registers
1070 in SImode and DImode */
1071 {8, 8}, /* cost of storing MMX registers
1072 in SImode and DImode */
1073 2, /* cost of moving SSE register */
1074 {8, 8, 8}, /* cost of loading SSE registers
1075 in SImode, DImode and TImode */
1076 {8, 8, 8}, /* cost of storing SSE registers
1077 in SImode, DImode and TImode */
1078 5, /* MMX or SSE register to integer */
1079 32, /* size of l1 cache. */
1080 256, /* size of l2 cache. */
1081 64, /* size of prefetch block */
1082 6, /* number of parallel prefetches */
1083 3, /* Branch cost */
1084 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1085 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1086 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1087 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1088 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1089 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1090 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1091 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1092 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1093 {{libcall, {{8, loop}, {15, unrolled_loop},
1094 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1095 {libcall, {{24, loop}, {32, unrolled_loop},
1096 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1097 1, /* scalar_stmt_cost. */
1098 1, /* scalar load_cost. */
1099 1, /* scalar_store_cost. */
1100 1, /* vec_stmt_cost. */
1101 1, /* vec_to_scalar_cost. */
1102 1, /* scalar_to_vec_cost. */
1103 1, /* vec_align_load_cost. */
1104 2, /* vec_unalign_load_cost. */
1105 1, /* vec_store_cost. */
1106 3, /* cond_taken_branch_cost. */
1107 1, /* cond_not_taken_branch_cost. */
1110 /* Generic64 should produce code tuned for Nocona and K8. */
1112 struct processor_costs generic64_cost = {
1113 COSTS_N_INSNS (1), /* cost of an add instruction */
1114 /* On all chips taken into consideration lea is 2 cycles and more. With
1115 this cost however our current implementation of synth_mult results in
1116 use of unnecessary temporary registers causing regression on several
1117 SPECfp benchmarks. */
1118 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1119 COSTS_N_INSNS (1), /* variable shift costs */
1120 COSTS_N_INSNS (1), /* constant shift costs */
1121 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1122 COSTS_N_INSNS (4), /* HI */
1123 COSTS_N_INSNS (3), /* SI */
1124 COSTS_N_INSNS (4), /* DI */
1125 COSTS_N_INSNS (2)}, /* other */
1126 0, /* cost of multiply per each bit set */
1127 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1128 COSTS_N_INSNS (26), /* HI */
1129 COSTS_N_INSNS (42), /* SI */
1130 COSTS_N_INSNS (74), /* DI */
1131 COSTS_N_INSNS (74)}, /* other */
1132 COSTS_N_INSNS (1), /* cost of movsx */
1133 COSTS_N_INSNS (1), /* cost of movzx */
1134 8, /* "large" insn */
1135 17, /* MOVE_RATIO */
1136 4, /* cost for loading QImode using movzbl */
1137 {4, 4, 4}, /* cost of loading integer registers
1138 in QImode, HImode and SImode.
1139 Relative to reg-reg move (2). */
1140 {4, 4, 4}, /* cost of storing integer registers */
1141 4, /* cost of reg,reg fld/fst */
1142 {12, 12, 12}, /* cost of loading fp registers
1143 in SFmode, DFmode and XFmode */
1144 {6, 6, 8}, /* cost of storing fp registers
1145 in SFmode, DFmode and XFmode */
1146 2, /* cost of moving MMX register */
1147 {8, 8}, /* cost of loading MMX registers
1148 in SImode and DImode */
1149 {8, 8}, /* cost of storing MMX registers
1150 in SImode and DImode */
1151 2, /* cost of moving SSE register */
1152 {8, 8, 8}, /* cost of loading SSE registers
1153 in SImode, DImode and TImode */
1154 {8, 8, 8}, /* cost of storing SSE registers
1155 in SImode, DImode and TImode */
1156 5, /* MMX or SSE register to integer */
1157 32, /* size of l1 cache. */
1158 512, /* size of l2 cache. */
1159 64, /* size of prefetch block */
1160 6, /* number of parallel prefetches */
1161 /* Benchmarks shows large regressions on K8 sixtrack benchmark when this value
1162 is increased to perhaps more appropriate value of 5. */
1163 3, /* Branch cost */
1164 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1165 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1166 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1167 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1168 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1169 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1170 {DUMMY_STRINGOP_ALGS,
1171 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1172 {DUMMY_STRINGOP_ALGS,
1173 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1174 1, /* scalar_stmt_cost. */
1175 1, /* scalar load_cost. */
1176 1, /* scalar_store_cost. */
1177 1, /* vec_stmt_cost. */
1178 1, /* vec_to_scalar_cost. */
1179 1, /* scalar_to_vec_cost. */
1180 1, /* vec_align_load_cost. */
1181 2, /* vec_unalign_load_cost. */
1182 1, /* vec_store_cost. */
1183 3, /* cond_taken_branch_cost. */
1184 1, /* cond_not_taken_branch_cost. */
1187 /* Generic32 should produce code tuned for Athlon, PPro, Pentium4, Nocona and K8. */
1189 struct processor_costs generic32_cost = {
1190 COSTS_N_INSNS (1), /* cost of an add instruction */
1191 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1192 COSTS_N_INSNS (1), /* variable shift costs */
1193 COSTS_N_INSNS (1), /* constant shift costs */
1194 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1195 COSTS_N_INSNS (4), /* HI */
1196 COSTS_N_INSNS (3), /* SI */
1197 COSTS_N_INSNS (4), /* DI */
1198 COSTS_N_INSNS (2)}, /* other */
1199 0, /* cost of multiply per each bit set */
1200 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1201 COSTS_N_INSNS (26), /* HI */
1202 COSTS_N_INSNS (42), /* SI */
1203 COSTS_N_INSNS (74), /* DI */
1204 COSTS_N_INSNS (74)}, /* other */
1205 COSTS_N_INSNS (1), /* cost of movsx */
1206 COSTS_N_INSNS (1), /* cost of movzx */
1207 8, /* "large" insn */
1208 17, /* MOVE_RATIO */
1209 4, /* cost for loading QImode using movzbl */
1210 {4, 4, 4}, /* cost of loading integer registers
1211 in QImode, HImode and SImode.
1212 Relative to reg-reg move (2). */
1213 {4, 4, 4}, /* cost of storing integer registers */
1214 4, /* cost of reg,reg fld/fst */
1215 {12, 12, 12}, /* cost of loading fp registers
1216 in SFmode, DFmode and XFmode */
1217 {6, 6, 8}, /* cost of storing fp registers
1218 in SFmode, DFmode and XFmode */
1219 2, /* cost of moving MMX register */
1220 {8, 8}, /* cost of loading MMX registers
1221 in SImode and DImode */
1222 {8, 8}, /* cost of storing MMX registers
1223 in SImode and DImode */
1224 2, /* cost of moving SSE register */
1225 {8, 8, 8}, /* cost of loading SSE registers
1226 in SImode, DImode and TImode */
1227 {8, 8, 8}, /* cost of storing SSE registers
1228 in SImode, DImode and TImode */
1229 5, /* MMX or SSE register to integer */
1230 32, /* size of l1 cache. */
1231 256, /* size of l2 cache. */
1232 64, /* size of prefetch block */
1233 6, /* number of parallel prefetches */
1234 3, /* Branch cost */
1235 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1236 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1237 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1238 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1239 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1240 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1241 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1242 DUMMY_STRINGOP_ALGS},
1243 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1244 DUMMY_STRINGOP_ALGS},
1245 1, /* scalar_stmt_cost. */
1246 1, /* scalar load_cost. */
1247 1, /* scalar_store_cost. */
1248 1, /* vec_stmt_cost. */
1249 1, /* vec_to_scalar_cost. */
1250 1, /* scalar_to_vec_cost. */
1251 1, /* vec_align_load_cost. */
1252 2, /* vec_unalign_load_cost. */
1253 1, /* vec_store_cost. */
1254 3, /* cond_taken_branch_cost. */
1255 1, /* cond_not_taken_branch_cost. */
1258 const struct processor_costs *ix86_cost = &pentium_cost;
1260 /* Processor feature/optimization bitmasks. */
1261 #define m_386 (1<<PROCESSOR_I386)
1262 #define m_486 (1<<PROCESSOR_I486)
1263 #define m_PENT (1<<PROCESSOR_PENTIUM)
1264 #define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
1265 #define m_PENT4 (1<<PROCESSOR_PENTIUM4)
1266 #define m_NOCONA (1<<PROCESSOR_NOCONA)
1267 #define m_CORE2 (1<<PROCESSOR_CORE2)
1268 #define m_ATOM (1<<PROCESSOR_ATOM)
1270 #define m_GEODE (1<<PROCESSOR_GEODE)
1271 #define m_K6 (1<<PROCESSOR_K6)
1272 #define m_K6_GEODE (m_K6 | m_GEODE)
1273 #define m_K8 (1<<PROCESSOR_K8)
1274 #define m_ATHLON (1<<PROCESSOR_ATHLON)
1275 #define m_ATHLON_K8 (m_K8 | m_ATHLON)
1276 #define m_AMDFAM10 (1<<PROCESSOR_AMDFAM10)
1277 #define m_AMD_MULTIPLE (m_K8 | m_ATHLON | m_AMDFAM10)
1279 #define m_GENERIC32 (1<<PROCESSOR_GENERIC32)
1280 #define m_GENERIC64 (1<<PROCESSOR_GENERIC64)
1282 /* Generic instruction choice should be common subset of supported CPUs
1283 (PPro/PENT4/NOCONA/CORE2/Athlon/K8). */
1284 #define m_GENERIC (m_GENERIC32 | m_GENERIC64)
1286 /* Feature tests against the various tunings. */
1287 unsigned char ix86_tune_features[X86_TUNE_LAST];
1289 /* Feature tests against the various tunings used to create ix86_tune_features
1290 based on the processor mask. */
1291 static unsigned int initial_ix86_tune_features[X86_TUNE_LAST] = {
1292 /* X86_TUNE_USE_LEAVE: Leave does not affect Nocona SPEC2000 results
1293 negatively, so enabling for Generic64 seems like good code size
1294 tradeoff. We can't enable it for 32bit generic because it does not
1295 work well with PPro base chips. */
1296 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_CORE2 | m_GENERIC64,
1298 /* X86_TUNE_PUSH_MEMORY */
1299 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4
1300 | m_NOCONA | m_CORE2 | m_GENERIC,
1302 /* X86_TUNE_ZERO_EXTEND_WITH_AND */
1305 /* X86_TUNE_UNROLL_STRLEN */
1306 m_486 | m_PENT | m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_K6
1307 | m_CORE2 | m_GENERIC,
1309 /* X86_TUNE_DEEP_BRANCH_PREDICTION */
1310 m_ATOM | m_PPRO | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4 | m_GENERIC,
1312 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
1313 on simulation result. But after P4 was made, no performance benefit
1314 was observed with branch hints. It also increases the code size.
1315 As a result, icc never generates branch hints. */
1318 /* X86_TUNE_DOUBLE_WITH_ADD */
1321 /* X86_TUNE_USE_SAHF */
1322 m_ATOM | m_PPRO | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_PENT4
1323 | m_NOCONA | m_CORE2 | m_GENERIC,
1325 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
1326 partial dependencies. */
1327 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_PENT4 | m_NOCONA
1328 | m_CORE2 | m_GENERIC | m_GEODE /* m_386 | m_K6 */,
1330 /* X86_TUNE_PARTIAL_REG_STALL: We probably ought to watch for partial
1331 register stalls on Generic32 compilation setting as well. However
1332 in current implementation the partial register stalls are not eliminated
1333 very well - they can be introduced via subregs synthesized by combine
1334 and can happen in caller/callee saving sequences. Because this option
1335 pays back little on PPro based chips and is in conflict with partial reg
1336 dependencies used by Athlon/P4 based chips, it is better to leave it off
1337 for generic32 for now. */
1340 /* X86_TUNE_PARTIAL_FLAG_REG_STALL */
1341 m_CORE2 | m_GENERIC,
1343 /* X86_TUNE_USE_HIMODE_FIOP */
1344 m_386 | m_486 | m_K6_GEODE,
1346 /* X86_TUNE_USE_SIMODE_FIOP */
1347 ~(m_PPRO | m_AMD_MULTIPLE | m_PENT | m_ATOM | m_CORE2 | m_GENERIC),
1349 /* X86_TUNE_USE_MOV0 */
1352 /* X86_TUNE_USE_CLTD */
1353 ~(m_PENT | m_ATOM | m_K6 | m_CORE2 | m_GENERIC),
1355 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
1358 /* X86_TUNE_SPLIT_LONG_MOVES */
1361 /* X86_TUNE_READ_MODIFY_WRITE */
1364 /* X86_TUNE_READ_MODIFY */
1367 /* X86_TUNE_PROMOTE_QIMODE */
1368 m_K6_GEODE | m_PENT | m_ATOM | m_386 | m_486 | m_AMD_MULTIPLE
1369 | m_CORE2 | m_GENERIC /* | m_PENT4 ? */,
1371 /* X86_TUNE_FAST_PREFIX */
1372 ~(m_PENT | m_486 | m_386),
1374 /* X86_TUNE_SINGLE_STRINGOP */
1375 m_386 | m_PENT4 | m_NOCONA,
1377 /* X86_TUNE_QIMODE_MATH */
1380 /* X86_TUNE_HIMODE_MATH: On PPro this flag is meant to avoid partial
1381 register stalls. Just like X86_TUNE_PARTIAL_REG_STALL this option
1382 might be considered for Generic32 if our scheme for avoiding partial
1383 stalls was more effective. */
1386 /* X86_TUNE_PROMOTE_QI_REGS */
1389 /* X86_TUNE_PROMOTE_HI_REGS */
1392 /* X86_TUNE_ADD_ESP_4: Enable if add/sub is preferred over 1/2 push/pop. */
1393 m_ATOM | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT4 | m_NOCONA
1394 | m_CORE2 | m_GENERIC,
1396 /* X86_TUNE_ADD_ESP_8 */
1397 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_K6_GEODE | m_386
1398 | m_486 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1400 /* X86_TUNE_SUB_ESP_4 */
1401 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_PENT4 | m_NOCONA | m_CORE2
1404 /* X86_TUNE_SUB_ESP_8 */
1405 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_386 | m_486
1406 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1408 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
1409 for DFmode copies */
1410 ~(m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1411 | m_GENERIC | m_GEODE),
1413 /* X86_TUNE_PARTIAL_REG_DEPENDENCY */
1414 m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1416 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: In the Generic model we have a
1417 conflict here in between PPro/Pentium4 based chips that thread 128bit
1418 SSE registers as single units versus K8 based chips that divide SSE
1419 registers to two 64bit halves. This knob promotes all store destinations
1420 to be 128bit to allow register renaming on 128bit SSE units, but usually
1421 results in one extra microop on 64bit SSE units. Experimental results
1422 shows that disabling this option on P4 brings over 20% SPECfp regression,
1423 while enabling it on K8 brings roughly 2.4% regression that can be partly
1424 masked by careful scheduling of moves. */
1425 m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC
1428 /* X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL */
1431 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
1432 are resolved on SSE register parts instead of whole registers, so we may
1433 maintain just lower part of scalar values in proper format leaving the
1434 upper part undefined. */
1437 /* X86_TUNE_SSE_TYPELESS_STORES */
1440 /* X86_TUNE_SSE_LOAD0_BY_PXOR */
1441 m_PPRO | m_PENT4 | m_NOCONA,
1443 /* X86_TUNE_MEMORY_MISMATCH_STALL */
1444 m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1446 /* X86_TUNE_PROLOGUE_USING_MOVE */
1447 m_ATHLON_K8 | m_ATOM | m_PPRO | m_CORE2 | m_GENERIC,
1449 /* X86_TUNE_EPILOGUE_USING_MOVE */
1450 m_ATHLON_K8 | m_ATOM | m_PPRO | m_CORE2 | m_GENERIC,
1452 /* X86_TUNE_SHIFT1 */
1455 /* X86_TUNE_USE_FFREEP */
1458 /* X86_TUNE_INTER_UNIT_MOVES */
1459 ~(m_AMD_MULTIPLE | m_ATOM | m_GENERIC),
1461 /* X86_TUNE_INTER_UNIT_CONVERSIONS */
1464 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
1465 than 4 branch instructions in the 16 byte window. */
1466 m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2
1469 /* X86_TUNE_SCHEDULE */
1470 m_PPRO | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT | m_ATOM | m_CORE2
1473 /* X86_TUNE_USE_BT */
1474 m_AMD_MULTIPLE | m_ATOM | m_CORE2 | m_GENERIC,
1476 /* X86_TUNE_USE_INCDEC */
1477 ~(m_PENT4 | m_NOCONA | m_GENERIC | m_ATOM),
1479 /* X86_TUNE_PAD_RETURNS */
1480 m_AMD_MULTIPLE | m_CORE2 | m_GENERIC,
1482 /* X86_TUNE_EXT_80387_CONSTANTS */
1483 m_K6_GEODE | m_ATHLON_K8 | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO
1484 | m_CORE2 | m_GENERIC,
1486 /* X86_TUNE_SHORTEN_X87_SSE */
1489 /* X86_TUNE_AVOID_VECTOR_DECODE */
1492 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
1493 and SImode multiply, but 386 and 486 do HImode multiply faster. */
1496 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
1497 vector path on AMD machines. */
1498 m_K8 | m_GENERIC64 | m_AMDFAM10,
1500 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
1502 m_K8 | m_GENERIC64 | m_AMDFAM10,
1504 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
1508 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
1509 but one byte longer. */
1512 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
1513 operand that cannot be represented using a modRM byte. The XOR
1514 replacement is long decoded, so this split helps here as well. */
1517 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
1519 m_AMDFAM10 | m_GENERIC,
1521 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
1522 from integer to FP. */
1525 /* X86_TUNE_FUSE_CMP_AND_BRANCH: Fuse a compare or test instruction
1526 with a subsequent conditional jump instruction into a single
1527 compare-and-branch uop. */
1530 /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
1531 will impact LEA instruction selection. */
1535 /* Feature tests against the various architecture variations. */
1536 unsigned char ix86_arch_features[X86_ARCH_LAST];
1538 /* Feature tests against the various architecture variations, used to create
1539 ix86_arch_features based on the processor mask. */
1540 static unsigned int initial_ix86_arch_features[X86_ARCH_LAST] = {
1541 /* X86_ARCH_CMOVE: Conditional move was added for pentiumpro. */
1542 ~(m_386 | m_486 | m_PENT | m_K6),
1544 /* X86_ARCH_CMPXCHG: Compare and exchange was added for 80486. */
1547 /* X86_ARCH_CMPXCHG8B: Compare and exchange 8 bytes was added for pentium. */
1550 /* X86_ARCH_XADD: Exchange and add was added for 80486. */
1553 /* X86_ARCH_BSWAP: Byteswap was added for 80486. */
1557 static const unsigned int x86_accumulate_outgoing_args
1558 = m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1561 static const unsigned int x86_arch_always_fancy_math_387
1562 = m_PENT | m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_PENT4
1563 | m_NOCONA | m_CORE2 | m_GENERIC;
1565 static enum stringop_alg stringop_alg = no_stringop;
1567 /* In case the average insn count for single function invocation is
1568 lower than this constant, emit fast (but longer) prologue and
1570 #define FAST_PROLOGUE_INSN_COUNT 20
1572 /* Names for 8 (low), 8 (high), and 16-bit registers, respectively. */
1573 static const char *const qi_reg_name[] = QI_REGISTER_NAMES;
1574 static const char *const qi_high_reg_name[] = QI_HIGH_REGISTER_NAMES;
1575 static const char *const hi_reg_name[] = HI_REGISTER_NAMES;
1577 /* Array of the smallest class containing reg number REGNO, indexed by
1578 REGNO. Used by REGNO_REG_CLASS in i386.h. */
1580 enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER] =
1582 /* ax, dx, cx, bx */
1583 AREG, DREG, CREG, BREG,
1584 /* si, di, bp, sp */
1585 SIREG, DIREG, NON_Q_REGS, NON_Q_REGS,
1587 FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS,
1588 FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
1591 /* flags, fpsr, fpcr, frame */
1592 NO_REGS, NO_REGS, NO_REGS, NON_Q_REGS,
1594 SSE_FIRST_REG, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1597 MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS,
1600 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1601 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1602 /* SSE REX registers */
1603 SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1607 /* The "default" register map used in 32bit mode. */
1609 int const dbx_register_map[FIRST_PSEUDO_REGISTER] =
1611 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */
1612 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */
1613 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1614 21, 22, 23, 24, 25, 26, 27, 28, /* SSE */
1615 29, 30, 31, 32, 33, 34, 35, 36, /* MMX */
1616 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1617 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1620 /* The "default" register map used in 64bit mode. */
1622 int const dbx64_register_map[FIRST_PSEUDO_REGISTER] =
1624 0, 1, 2, 3, 4, 5, 6, 7, /* general regs */
1625 33, 34, 35, 36, 37, 38, 39, 40, /* fp regs */
1626 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1627 17, 18, 19, 20, 21, 22, 23, 24, /* SSE */
1628 41, 42, 43, 44, 45, 46, 47, 48, /* MMX */
1629 8,9,10,11,12,13,14,15, /* extended integer registers */
1630 25, 26, 27, 28, 29, 30, 31, 32, /* extended SSE registers */
1633 /* Define the register numbers to be used in Dwarf debugging information.
1634 The SVR4 reference port C compiler uses the following register numbers
1635 in its Dwarf output code:
1636 0 for %eax (gcc regno = 0)
1637 1 for %ecx (gcc regno = 2)
1638 2 for %edx (gcc regno = 1)
1639 3 for %ebx (gcc regno = 3)
1640 4 for %esp (gcc regno = 7)
1641 5 for %ebp (gcc regno = 6)
1642 6 for %esi (gcc regno = 4)
1643 7 for %edi (gcc regno = 5)
1644 The following three DWARF register numbers are never generated by
1645 the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4
1646 believes these numbers have these meanings.
1647 8 for %eip (no gcc equivalent)
1648 9 for %eflags (gcc regno = 17)
1649 10 for %trapno (no gcc equivalent)
1650 It is not at all clear how we should number the FP stack registers
1651 for the x86 architecture. If the version of SDB on x86/svr4 were
1652 a bit less brain dead with respect to floating-point then we would
1653 have a precedent to follow with respect to DWARF register numbers
1654 for x86 FP registers, but the SDB on x86/svr4 is so completely
1655 broken with respect to FP registers that it is hardly worth thinking
1656 of it as something to strive for compatibility with.
1657 The version of x86/svr4 SDB I have at the moment does (partially)
1658 seem to believe that DWARF register number 11 is associated with
1659 the x86 register %st(0), but that's about all. Higher DWARF
1660 register numbers don't seem to be associated with anything in
1661 particular, and even for DWARF regno 11, SDB only seems to under-
1662 stand that it should say that a variable lives in %st(0) (when
1663 asked via an `=' command) if we said it was in DWARF regno 11,
1664 but SDB still prints garbage when asked for the value of the
1665 variable in question (via a `/' command).
1666 (Also note that the labels SDB prints for various FP stack regs
1667 when doing an `x' command are all wrong.)
1668 Note that these problems generally don't affect the native SVR4
1669 C compiler because it doesn't allow the use of -O with -g and
1670 because when it is *not* optimizing, it allocates a memory
1671 location for each floating-point variable, and the memory
1672 location is what gets described in the DWARF AT_location
1673 attribute for the variable in question.
1674 Regardless of the severe mental illness of the x86/svr4 SDB, we
1675 do something sensible here and we use the following DWARF
1676 register numbers. Note that these are all stack-top-relative
1678 11 for %st(0) (gcc regno = 8)
1679 12 for %st(1) (gcc regno = 9)
1680 13 for %st(2) (gcc regno = 10)
1681 14 for %st(3) (gcc regno = 11)
1682 15 for %st(4) (gcc regno = 12)
1683 16 for %st(5) (gcc regno = 13)
1684 17 for %st(6) (gcc regno = 14)
1685 18 for %st(7) (gcc regno = 15)
1687 int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER] =
1689 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */
1690 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */
1691 -1, 9, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1692 21, 22, 23, 24, 25, 26, 27, 28, /* SSE registers */
1693 29, 30, 31, 32, 33, 34, 35, 36, /* MMX registers */
1694 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1695 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1698 /* Test and compare insns in i386.md store the information needed to
1699 generate branch and scc insns here. */
1701 rtx ix86_compare_op0 = NULL_RTX;
1702 rtx ix86_compare_op1 = NULL_RTX;
1704 /* Define parameter passing and return registers. */
1706 static int const x86_64_int_parameter_registers[6] =
1708 DI_REG, SI_REG, DX_REG, CX_REG, R8_REG, R9_REG
1711 static int const x86_64_ms_abi_int_parameter_registers[4] =
1713 CX_REG, DX_REG, R8_REG, R9_REG
1716 static int const x86_64_int_return_registers[4] =
1718 AX_REG, DX_REG, DI_REG, SI_REG
1721 /* Define the structure for the machine field in struct function. */
1723 struct GTY(()) stack_local_entry {
1724 unsigned short mode;
1727 struct stack_local_entry *next;
1730 /* Structure describing stack frame layout.
1731 Stack grows downward:
1737 saved frame pointer if frame_pointer_needed
1738 <- HARD_FRAME_POINTER
1747 [va_arg registers] (
1748 > to_allocate <- FRAME_POINTER
1760 HOST_WIDE_INT frame;
1762 int outgoing_arguments_size;
1765 HOST_WIDE_INT to_allocate;
1766 /* The offsets relative to ARG_POINTER. */
1767 HOST_WIDE_INT frame_pointer_offset;
1768 HOST_WIDE_INT hard_frame_pointer_offset;
1769 HOST_WIDE_INT stack_pointer_offset;
1771 /* When save_regs_using_mov is set, emit prologue using
1772 move instead of push instructions. */
1773 bool save_regs_using_mov;
1776 /* Code model option. */
1777 enum cmodel ix86_cmodel;
1779 enum asm_dialect ix86_asm_dialect = ASM_ATT;
1781 enum tls_dialect ix86_tls_dialect = TLS_DIALECT_GNU;
1783 /* Which unit we are generating floating point math for. */
1784 enum fpmath_unit ix86_fpmath;
1786 /* Which cpu are we scheduling for. */
1787 enum attr_cpu ix86_schedule;
1789 /* Which cpu are we optimizing for. */
1790 enum processor_type ix86_tune;
1792 /* Which instruction set architecture to use. */
1793 enum processor_type ix86_arch;
1795 /* true if sse prefetch instruction is not NOOP. */
1796 int x86_prefetch_sse;
1798 /* ix86_regparm_string as a number */
1799 static int ix86_regparm;
1801 /* -mstackrealign option */
1802 extern int ix86_force_align_arg_pointer;
1803 static const char ix86_force_align_arg_pointer_string[]
1804 = "force_align_arg_pointer";
1806 static rtx (*ix86_gen_leave) (void);
1807 static rtx (*ix86_gen_pop1) (rtx);
1808 static rtx (*ix86_gen_add3) (rtx, rtx, rtx);
1809 static rtx (*ix86_gen_sub3) (rtx, rtx, rtx);
1810 static rtx (*ix86_gen_sub3_carry) (rtx, rtx, rtx, rtx, rtx);
1811 static rtx (*ix86_gen_one_cmpl2) (rtx, rtx);
1812 static rtx (*ix86_gen_monitor) (rtx, rtx, rtx);
1813 static rtx (*ix86_gen_andsp) (rtx, rtx, rtx);
1815 /* Preferred alignment for stack boundary in bits. */
1816 unsigned int ix86_preferred_stack_boundary;
1818 /* Alignment for incoming stack boundary in bits specified at
1820 static unsigned int ix86_user_incoming_stack_boundary;
1822 /* Default alignment for incoming stack boundary in bits. */
1823 static unsigned int ix86_default_incoming_stack_boundary;
1825 /* Alignment for incoming stack boundary in bits. */
1826 unsigned int ix86_incoming_stack_boundary;
1828 /* The abi used by target. */
1829 enum calling_abi ix86_abi;
1831 /* Values 1-5: see jump.c */
1832 int ix86_branch_cost;
1834 /* Calling abi specific va_list type nodes. */
1835 static GTY(()) tree sysv_va_list_type_node;
1836 static GTY(()) tree ms_va_list_type_node;
1838 /* Variables which are this size or smaller are put in the data/bss
1839 or ldata/lbss sections. */
1841 int ix86_section_threshold = 65536;
1843 /* Prefix built by ASM_GENERATE_INTERNAL_LABEL. */
1844 char internal_label_prefix[16];
1845 int internal_label_prefix_len;
1847 /* Fence to use after loop using movnt. */
1850 /* Register class used for passing given 64bit part of the argument.
1851 These represent classes as documented by the PS ABI, with the exception
1852 of SSESF, SSEDF classes, that are basically SSE class, just gcc will
1853 use SF or DFmode move instead of DImode to avoid reformatting penalties.
1855 Similarly we play games with INTEGERSI_CLASS to use cheaper SImode moves
1856 whenever possible (upper half does contain padding). */
1857 enum x86_64_reg_class
1860 X86_64_INTEGER_CLASS,
1861 X86_64_INTEGERSI_CLASS,
1868 X86_64_COMPLEX_X87_CLASS,
1872 #define MAX_CLASSES 4
1874 /* Table of constants used by fldpi, fldln2, etc.... */
1875 static REAL_VALUE_TYPE ext_80387_constants_table [5];
1876 static bool ext_80387_constants_init = 0;
1879 static struct machine_function * ix86_init_machine_status (void);
1880 static rtx ix86_function_value (const_tree, const_tree, bool);
1881 static rtx ix86_static_chain (const_tree, bool);
1882 static int ix86_function_regparm (const_tree, const_tree);
1883 static void ix86_compute_frame_layout (struct ix86_frame *);
1884 static bool ix86_expand_vector_init_one_nonzero (bool, enum machine_mode,
1886 static void ix86_add_new_builtins (int);
1887 static rtx ix86_expand_vec_perm_builtin (tree);
1889 enum ix86_function_specific_strings
1891 IX86_FUNCTION_SPECIFIC_ARCH,
1892 IX86_FUNCTION_SPECIFIC_TUNE,
1893 IX86_FUNCTION_SPECIFIC_FPMATH,
1894 IX86_FUNCTION_SPECIFIC_MAX
1897 static char *ix86_target_string (int, int, const char *, const char *,
1898 const char *, bool);
1899 static void ix86_debug_options (void) ATTRIBUTE_UNUSED;
1900 static void ix86_function_specific_save (struct cl_target_option *);
1901 static void ix86_function_specific_restore (struct cl_target_option *);
1902 static void ix86_function_specific_print (FILE *, int,
1903 struct cl_target_option *);
1904 static bool ix86_valid_target_attribute_p (tree, tree, tree, int);
1905 static bool ix86_valid_target_attribute_inner_p (tree, char *[]);
1906 static bool ix86_can_inline_p (tree, tree);
1907 static void ix86_set_current_function (tree);
1908 static unsigned int ix86_minimum_incoming_stack_boundary (bool);
1910 static enum calling_abi ix86_function_abi (const_tree);
1913 /* The svr4 ABI for the i386 says that records and unions are returned
1915 #ifndef DEFAULT_PCC_STRUCT_RETURN
1916 #define DEFAULT_PCC_STRUCT_RETURN 1
1919 /* Whether -mtune= or -march= were specified */
1920 static int ix86_tune_defaulted;
1921 static int ix86_arch_specified;
1923 /* Bit flags that specify the ISA we are compiling for. */
1924 int ix86_isa_flags = TARGET_64BIT_DEFAULT | TARGET_SUBTARGET_ISA_DEFAULT;
1926 /* A mask of ix86_isa_flags that includes bit X if X
1927 was set or cleared on the command line. */
1928 static int ix86_isa_flags_explicit;
1930 /* Define a set of ISAs which are available when a given ISA is
1931 enabled. MMX and SSE ISAs are handled separately. */
1933 #define OPTION_MASK_ISA_MMX_SET OPTION_MASK_ISA_MMX
1934 #define OPTION_MASK_ISA_3DNOW_SET \
1935 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_MMX_SET)
1937 #define OPTION_MASK_ISA_SSE_SET OPTION_MASK_ISA_SSE
1938 #define OPTION_MASK_ISA_SSE2_SET \
1939 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE_SET)
1940 #define OPTION_MASK_ISA_SSE3_SET \
1941 (OPTION_MASK_ISA_SSE3 | OPTION_MASK_ISA_SSE2_SET)
1942 #define OPTION_MASK_ISA_SSSE3_SET \
1943 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE3_SET)
1944 #define OPTION_MASK_ISA_SSE4_1_SET \
1945 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSSE3_SET)
1946 #define OPTION_MASK_ISA_SSE4_2_SET \
1947 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_SSE4_1_SET)
1948 #define OPTION_MASK_ISA_AVX_SET \
1949 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_SSE4_2_SET)
1950 #define OPTION_MASK_ISA_FMA_SET \
1951 (OPTION_MASK_ISA_FMA | OPTION_MASK_ISA_AVX_SET)
1953 /* SSE4 includes both SSE4.1 and SSE4.2. -msse4 should be the same
1955 #define OPTION_MASK_ISA_SSE4_SET OPTION_MASK_ISA_SSE4_2_SET
1957 #define OPTION_MASK_ISA_SSE4A_SET \
1958 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE3_SET)
1959 #define OPTION_MASK_ISA_FMA4_SET \
1960 (OPTION_MASK_ISA_FMA4 | OPTION_MASK_ISA_SSE4A_SET \
1961 | OPTION_MASK_ISA_AVX_SET)
1962 #define OPTION_MASK_ISA_XOP_SET \
1963 (OPTION_MASK_ISA_XOP | OPTION_MASK_ISA_FMA4_SET)
1964 #define OPTION_MASK_ISA_LWP_SET \
1967 /* AES and PCLMUL need SSE2 because they use xmm registers */
1968 #define OPTION_MASK_ISA_AES_SET \
1969 (OPTION_MASK_ISA_AES | OPTION_MASK_ISA_SSE2_SET)
1970 #define OPTION_MASK_ISA_PCLMUL_SET \
1971 (OPTION_MASK_ISA_PCLMUL | OPTION_MASK_ISA_SSE2_SET)
1973 #define OPTION_MASK_ISA_ABM_SET \
1974 (OPTION_MASK_ISA_ABM | OPTION_MASK_ISA_POPCNT)
1976 #define OPTION_MASK_ISA_POPCNT_SET OPTION_MASK_ISA_POPCNT
1977 #define OPTION_MASK_ISA_CX16_SET OPTION_MASK_ISA_CX16
1978 #define OPTION_MASK_ISA_SAHF_SET OPTION_MASK_ISA_SAHF
1979 #define OPTION_MASK_ISA_MOVBE_SET OPTION_MASK_ISA_MOVBE
1980 #define OPTION_MASK_ISA_CRC32_SET OPTION_MASK_ISA_CRC32
1982 /* Define a set of ISAs which aren't available when a given ISA is
1983 disabled. MMX and SSE ISAs are handled separately. */
1985 #define OPTION_MASK_ISA_MMX_UNSET \
1986 (OPTION_MASK_ISA_MMX | OPTION_MASK_ISA_3DNOW_UNSET)
1987 #define OPTION_MASK_ISA_3DNOW_UNSET \
1988 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_3DNOW_A_UNSET)
1989 #define OPTION_MASK_ISA_3DNOW_A_UNSET OPTION_MASK_ISA_3DNOW_A
1991 #define OPTION_MASK_ISA_SSE_UNSET \
1992 (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_SSE2_UNSET)
1993 #define OPTION_MASK_ISA_SSE2_UNSET \
1994 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE3_UNSET)
1995 #define OPTION_MASK_ISA_SSE3_UNSET \
1996 (OPTION_MASK_ISA_SSE3 \
1997 | OPTION_MASK_ISA_SSSE3_UNSET \
1998 | OPTION_MASK_ISA_SSE4A_UNSET )
1999 #define OPTION_MASK_ISA_SSSE3_UNSET \
2000 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE4_1_UNSET)
2001 #define OPTION_MASK_ISA_SSE4_1_UNSET \
2002 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSE4_2_UNSET)
2003 #define OPTION_MASK_ISA_SSE4_2_UNSET \
2004 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_AVX_UNSET )
2005 #define OPTION_MASK_ISA_AVX_UNSET \
2006 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_FMA_UNSET \
2007 | OPTION_MASK_ISA_FMA4_UNSET)
2008 #define OPTION_MASK_ISA_FMA_UNSET OPTION_MASK_ISA_FMA
2010 /* SSE4 includes both SSE4.1 and SSE4.2. -mno-sse4 should the same
2012 #define OPTION_MASK_ISA_SSE4_UNSET OPTION_MASK_ISA_SSE4_1_UNSET
2014 #define OPTION_MASK_ISA_SSE4A_UNSET \
2015 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_FMA4_UNSET)
2017 #define OPTION_MASK_ISA_FMA4_UNSET \
2018 (OPTION_MASK_ISA_FMA4 | OPTION_MASK_ISA_XOP_UNSET)
2019 #define OPTION_MASK_ISA_XOP_UNSET OPTION_MASK_ISA_XOP
2020 #define OPTION_MASK_ISA_LWP_UNSET OPTION_MASK_ISA_LWP
2022 #define OPTION_MASK_ISA_AES_UNSET OPTION_MASK_ISA_AES
2023 #define OPTION_MASK_ISA_PCLMUL_UNSET OPTION_MASK_ISA_PCLMUL
2024 #define OPTION_MASK_ISA_ABM_UNSET OPTION_MASK_ISA_ABM
2025 #define OPTION_MASK_ISA_POPCNT_UNSET OPTION_MASK_ISA_POPCNT
2026 #define OPTION_MASK_ISA_CX16_UNSET OPTION_MASK_ISA_CX16
2027 #define OPTION_MASK_ISA_SAHF_UNSET OPTION_MASK_ISA_SAHF
2028 #define OPTION_MASK_ISA_MOVBE_UNSET OPTION_MASK_ISA_MOVBE
2029 #define OPTION_MASK_ISA_CRC32_UNSET OPTION_MASK_ISA_CRC32
2031 /* Vectorization library interface and handlers. */
2032 tree (*ix86_veclib_handler)(enum built_in_function, tree, tree) = NULL;
2033 static tree ix86_veclibabi_svml (enum built_in_function, tree, tree);
2034 static tree ix86_veclibabi_acml (enum built_in_function, tree, tree);
2036 /* Processor target table, indexed by processor number */
2039 const struct processor_costs *cost; /* Processor costs */
2040 const int align_loop; /* Default alignments. */
2041 const int align_loop_max_skip;
2042 const int align_jump;
2043 const int align_jump_max_skip;
2044 const int align_func;
2047 static const struct ptt processor_target_table[PROCESSOR_max] =
2049 {&i386_cost, 4, 3, 4, 3, 4},
2050 {&i486_cost, 16, 15, 16, 15, 16},
2051 {&pentium_cost, 16, 7, 16, 7, 16},
2052 {&pentiumpro_cost, 16, 15, 16, 10, 16},
2053 {&geode_cost, 0, 0, 0, 0, 0},
2054 {&k6_cost, 32, 7, 32, 7, 32},
2055 {&athlon_cost, 16, 7, 16, 7, 16},
2056 {&pentium4_cost, 0, 0, 0, 0, 0},
2057 {&k8_cost, 16, 7, 16, 7, 16},
2058 {&nocona_cost, 0, 0, 0, 0, 0},
2059 {&core2_cost, 16, 10, 16, 10, 16},
2060 {&generic32_cost, 16, 7, 16, 7, 16},
2061 {&generic64_cost, 16, 10, 16, 10, 16},
2062 {&amdfam10_cost, 32, 24, 32, 7, 32},
2063 {&atom_cost, 16, 7, 16, 7, 16}
2066 static const char *const cpu_names[TARGET_CPU_DEFAULT_max] =
2092 /* Implement TARGET_HANDLE_OPTION. */
2095 ix86_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, int value)
2102 ix86_isa_flags |= OPTION_MASK_ISA_MMX_SET;
2103 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_SET;
2107 ix86_isa_flags &= ~OPTION_MASK_ISA_MMX_UNSET;
2108 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_UNSET;
2115 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_SET;
2116 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_SET;
2120 ix86_isa_flags &= ~OPTION_MASK_ISA_3DNOW_UNSET;
2121 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_UNSET;
2131 ix86_isa_flags |= OPTION_MASK_ISA_SSE_SET;
2132 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_SET;
2136 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE_UNSET;
2137 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_UNSET;
2144 ix86_isa_flags |= OPTION_MASK_ISA_SSE2_SET;
2145 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_SET;
2149 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE2_UNSET;
2150 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_UNSET;
2157 ix86_isa_flags |= OPTION_MASK_ISA_SSE3_SET;
2158 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_SET;
2162 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE3_UNSET;
2163 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_UNSET;
2170 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3_SET;
2171 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_SET;
2175 ix86_isa_flags &= ~OPTION_MASK_ISA_SSSE3_UNSET;
2176 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_UNSET;
2183 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1_SET;
2184 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_SET;
2188 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_1_UNSET;
2189 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_UNSET;
2196 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2_SET;
2197 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_SET;
2201 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_2_UNSET;
2202 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_UNSET;
2209 ix86_isa_flags |= OPTION_MASK_ISA_AVX_SET;
2210 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_SET;
2214 ix86_isa_flags &= ~OPTION_MASK_ISA_AVX_UNSET;
2215 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_UNSET;
2222 ix86_isa_flags |= OPTION_MASK_ISA_FMA_SET;
2223 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_SET;
2227 ix86_isa_flags &= ~OPTION_MASK_ISA_FMA_UNSET;
2228 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_UNSET;
2233 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
2234 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
2238 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
2239 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
2245 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A_SET;
2246 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_SET;
2250 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4A_UNSET;
2251 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_UNSET;
2258 ix86_isa_flags |= OPTION_MASK_ISA_FMA4_SET;
2259 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA4_SET;
2263 ix86_isa_flags &= ~OPTION_MASK_ISA_FMA4_UNSET;
2264 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA4_UNSET;
2271 ix86_isa_flags |= OPTION_MASK_ISA_XOP_SET;
2272 ix86_isa_flags_explicit |= OPTION_MASK_ISA_XOP_SET;
2276 ix86_isa_flags &= ~OPTION_MASK_ISA_XOP_UNSET;
2277 ix86_isa_flags_explicit |= OPTION_MASK_ISA_XOP_UNSET;
2284 ix86_isa_flags |= OPTION_MASK_ISA_LWP_SET;
2285 ix86_isa_flags_explicit |= OPTION_MASK_ISA_LWP_SET;
2289 ix86_isa_flags &= ~OPTION_MASK_ISA_LWP_UNSET;
2290 ix86_isa_flags_explicit |= OPTION_MASK_ISA_LWP_UNSET;
2297 ix86_isa_flags |= OPTION_MASK_ISA_ABM_SET;
2298 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_SET;
2302 ix86_isa_flags &= ~OPTION_MASK_ISA_ABM_UNSET;
2303 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_UNSET;
2310 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT_SET;
2311 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_SET;
2315 ix86_isa_flags &= ~OPTION_MASK_ISA_POPCNT_UNSET;
2316 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_UNSET;
2323 ix86_isa_flags |= OPTION_MASK_ISA_SAHF_SET;
2324 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_SET;
2328 ix86_isa_flags &= ~OPTION_MASK_ISA_SAHF_UNSET;
2329 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_UNSET;
2336 ix86_isa_flags |= OPTION_MASK_ISA_CX16_SET;
2337 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_SET;
2341 ix86_isa_flags &= ~OPTION_MASK_ISA_CX16_UNSET;
2342 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_UNSET;
2349 ix86_isa_flags |= OPTION_MASK_ISA_MOVBE_SET;
2350 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MOVBE_SET;
2354 ix86_isa_flags &= ~OPTION_MASK_ISA_MOVBE_UNSET;
2355 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MOVBE_UNSET;
2362 ix86_isa_flags |= OPTION_MASK_ISA_CRC32_SET;
2363 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CRC32_SET;
2367 ix86_isa_flags &= ~OPTION_MASK_ISA_CRC32_UNSET;
2368 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CRC32_UNSET;
2375 ix86_isa_flags |= OPTION_MASK_ISA_AES_SET;
2376 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_SET;
2380 ix86_isa_flags &= ~OPTION_MASK_ISA_AES_UNSET;
2381 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_UNSET;
2388 ix86_isa_flags |= OPTION_MASK_ISA_PCLMUL_SET;
2389 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_SET;
2393 ix86_isa_flags &= ~OPTION_MASK_ISA_PCLMUL_UNSET;
2394 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_UNSET;
2403 /* Return a string the documents the current -m options. The caller is
2404 responsible for freeing the string. */
2407 ix86_target_string (int isa, int flags, const char *arch, const char *tune,
2408 const char *fpmath, bool add_nl_p)
2410 struct ix86_target_opts
2412 const char *option; /* option string */
2413 int mask; /* isa mask options */
2416 /* This table is ordered so that options like -msse4.2 that imply
2417 preceding options while match those first. */
2418 static struct ix86_target_opts isa_opts[] =
2420 { "-m64", OPTION_MASK_ISA_64BIT },
2421 { "-mfma4", OPTION_MASK_ISA_FMA4 },
2422 { "-mxop", OPTION_MASK_ISA_XOP },
2423 { "-mlwp", OPTION_MASK_ISA_LWP },
2424 { "-msse4a", OPTION_MASK_ISA_SSE4A },
2425 { "-msse4.2", OPTION_MASK_ISA_SSE4_2 },
2426 { "-msse4.1", OPTION_MASK_ISA_SSE4_1 },
2427 { "-mssse3", OPTION_MASK_ISA_SSSE3 },
2428 { "-msse3", OPTION_MASK_ISA_SSE3 },
2429 { "-msse2", OPTION_MASK_ISA_SSE2 },
2430 { "-msse", OPTION_MASK_ISA_SSE },
2431 { "-m3dnow", OPTION_MASK_ISA_3DNOW },
2432 { "-m3dnowa", OPTION_MASK_ISA_3DNOW_A },
2433 { "-mmmx", OPTION_MASK_ISA_MMX },
2434 { "-mabm", OPTION_MASK_ISA_ABM },
2435 { "-mpopcnt", OPTION_MASK_ISA_POPCNT },
2436 { "-mmovbe", OPTION_MASK_ISA_MOVBE },
2437 { "-mcrc32", OPTION_MASK_ISA_CRC32 },
2438 { "-maes", OPTION_MASK_ISA_AES },
2439 { "-mpclmul", OPTION_MASK_ISA_PCLMUL },
2443 static struct ix86_target_opts flag_opts[] =
2445 { "-m128bit-long-double", MASK_128BIT_LONG_DOUBLE },
2446 { "-m80387", MASK_80387 },
2447 { "-maccumulate-outgoing-args", MASK_ACCUMULATE_OUTGOING_ARGS },
2448 { "-malign-double", MASK_ALIGN_DOUBLE },
2449 { "-mcld", MASK_CLD },
2450 { "-mfp-ret-in-387", MASK_FLOAT_RETURNS },
2451 { "-mieee-fp", MASK_IEEE_FP },
2452 { "-minline-all-stringops", MASK_INLINE_ALL_STRINGOPS },
2453 { "-minline-stringops-dynamically", MASK_INLINE_STRINGOPS_DYNAMICALLY },
2454 { "-mms-bitfields", MASK_MS_BITFIELD_LAYOUT },
2455 { "-mno-align-stringops", MASK_NO_ALIGN_STRINGOPS },
2456 { "-mno-fancy-math-387", MASK_NO_FANCY_MATH_387 },
2457 { "-mno-push-args", MASK_NO_PUSH_ARGS },
2458 { "-mno-red-zone", MASK_NO_RED_ZONE },
2459 { "-momit-leaf-frame-pointer", MASK_OMIT_LEAF_FRAME_POINTER },
2460 { "-mrecip", MASK_RECIP },
2461 { "-mrtd", MASK_RTD },
2462 { "-msseregparm", MASK_SSEREGPARM },
2463 { "-mstack-arg-probe", MASK_STACK_PROBE },
2464 { "-mtls-direct-seg-refs", MASK_TLS_DIRECT_SEG_REFS },
2467 const char *opts[ARRAY_SIZE (isa_opts) + ARRAY_SIZE (flag_opts) + 6][2];
2470 char target_other[40];
2479 memset (opts, '\0', sizeof (opts));
2481 /* Add -march= option. */
2484 opts[num][0] = "-march=";
2485 opts[num++][1] = arch;
2488 /* Add -mtune= option. */
2491 opts[num][0] = "-mtune=";
2492 opts[num++][1] = tune;
2495 /* Pick out the options in isa options. */
2496 for (i = 0; i < ARRAY_SIZE (isa_opts); i++)
2498 if ((isa & isa_opts[i].mask) != 0)
2500 opts[num++][0] = isa_opts[i].option;
2501 isa &= ~ isa_opts[i].mask;
2505 if (isa && add_nl_p)
2507 opts[num++][0] = isa_other;
2508 sprintf (isa_other, "(other isa: 0x%x)", isa);
2511 /* Add flag options. */
2512 for (i = 0; i < ARRAY_SIZE (flag_opts); i++)
2514 if ((flags & flag_opts[i].mask) != 0)
2516 opts[num++][0] = flag_opts[i].option;
2517 flags &= ~ flag_opts[i].mask;
2521 if (flags && add_nl_p)
2523 opts[num++][0] = target_other;
2524 sprintf (target_other, "(other flags: 0x%x)", isa);
2527 /* Add -fpmath= option. */
2530 opts[num][0] = "-mfpmath=";
2531 opts[num++][1] = fpmath;
2538 gcc_assert (num < ARRAY_SIZE (opts));
2540 /* Size the string. */
2542 sep_len = (add_nl_p) ? 3 : 1;
2543 for (i = 0; i < num; i++)
2546 for (j = 0; j < 2; j++)
2548 len += strlen (opts[i][j]);
2551 /* Build the string. */
2552 ret = ptr = (char *) xmalloc (len);
2555 for (i = 0; i < num; i++)
2559 for (j = 0; j < 2; j++)
2560 len2[j] = (opts[i][j]) ? strlen (opts[i][j]) : 0;
2567 if (add_nl_p && line_len + len2[0] + len2[1] > 70)
2575 for (j = 0; j < 2; j++)
2578 memcpy (ptr, opts[i][j], len2[j]);
2580 line_len += len2[j];
2585 gcc_assert (ret + len >= ptr);
2590 /* Function that is callable from the debugger to print the current
2593 ix86_debug_options (void)
2595 char *opts = ix86_target_string (ix86_isa_flags, target_flags,
2596 ix86_arch_string, ix86_tune_string,
2597 ix86_fpmath_string, true);
2601 fprintf (stderr, "%s\n\n", opts);
2605 fputs ("<no options>\n\n", stderr);
2610 /* Sometimes certain combinations of command options do not make
2611 sense on a particular target machine. You can define a macro
2612 `OVERRIDE_OPTIONS' to take account of this. This macro, if
2613 defined, is executed once just after all the command options have
2616 Don't use this macro to turn on various extra optimizations for
2617 `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */
2620 override_options (bool main_args_p)
2623 unsigned int ix86_arch_mask, ix86_tune_mask;
2628 /* Comes from final.c -- no real reason to change it. */
2629 #define MAX_CODE_ALIGN 16
2637 PTA_PREFETCH_SSE = 1 << 4,
2639 PTA_3DNOW_A = 1 << 6,
2643 PTA_POPCNT = 1 << 10,
2645 PTA_SSE4A = 1 << 12,
2646 PTA_NO_SAHF = 1 << 13,
2647 PTA_SSE4_1 = 1 << 14,
2648 PTA_SSE4_2 = 1 << 15,
2650 PTA_PCLMUL = 1 << 17,
2653 PTA_MOVBE = 1 << 20,
2661 const char *const name; /* processor name or nickname. */
2662 const enum processor_type processor;
2663 const enum attr_cpu schedule;
2664 const unsigned /*enum pta_flags*/ flags;
2666 const processor_alias_table[] =
2668 {"i386", PROCESSOR_I386, CPU_NONE, 0},
2669 {"i486", PROCESSOR_I486, CPU_NONE, 0},
2670 {"i586", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2671 {"pentium", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2672 {"pentium-mmx", PROCESSOR_PENTIUM, CPU_PENTIUM, PTA_MMX},
2673 {"winchip-c6", PROCESSOR_I486, CPU_NONE, PTA_MMX},
2674 {"winchip2", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2675 {"c3", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2676 {"c3-2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX | PTA_SSE},
2677 {"i686", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2678 {"pentiumpro", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2679 {"pentium2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX},
2680 {"pentium3", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2682 {"pentium3m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2684 {"pentium-m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2685 PTA_MMX | PTA_SSE | PTA_SSE2},
2686 {"pentium4", PROCESSOR_PENTIUM4, CPU_NONE,
2687 PTA_MMX |PTA_SSE | PTA_SSE2},
2688 {"pentium4m", PROCESSOR_PENTIUM4, CPU_NONE,
2689 PTA_MMX | PTA_SSE | PTA_SSE2},
2690 {"prescott", PROCESSOR_NOCONA, CPU_NONE,
2691 PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3},
2692 {"nocona", PROCESSOR_NOCONA, CPU_NONE,
2693 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2694 | PTA_CX16 | PTA_NO_SAHF},
2695 {"core2", PROCESSOR_CORE2, CPU_CORE2,
2696 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2697 | PTA_SSSE3 | PTA_CX16},
2698 {"atom", PROCESSOR_ATOM, CPU_ATOM,
2699 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2700 | PTA_SSSE3 | PTA_CX16 | PTA_MOVBE},
2701 {"geode", PROCESSOR_GEODE, CPU_GEODE,
2702 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A |PTA_PREFETCH_SSE},
2703 {"k6", PROCESSOR_K6, CPU_K6, PTA_MMX},
2704 {"k6-2", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2705 {"k6-3", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2706 {"athlon", PROCESSOR_ATHLON, CPU_ATHLON,
2707 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2708 {"athlon-tbird", PROCESSOR_ATHLON, CPU_ATHLON,
2709 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2710 {"athlon-4", PROCESSOR_ATHLON, CPU_ATHLON,
2711 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2712 {"athlon-xp", PROCESSOR_ATHLON, CPU_ATHLON,
2713 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2714 {"athlon-mp", PROCESSOR_ATHLON, CPU_ATHLON,
2715 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2716 {"x86-64", PROCESSOR_K8, CPU_K8,
2717 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_NO_SAHF},
2718 {"k8", PROCESSOR_K8, CPU_K8,
2719 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2720 | PTA_SSE2 | PTA_NO_SAHF},
2721 {"k8-sse3", PROCESSOR_K8, CPU_K8,
2722 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2723 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2724 {"opteron", PROCESSOR_K8, CPU_K8,
2725 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2726 | PTA_SSE2 | PTA_NO_SAHF},
2727 {"opteron-sse3", PROCESSOR_K8, CPU_K8,
2728 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2729 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2730 {"athlon64", PROCESSOR_K8, CPU_K8,
2731 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2732 | PTA_SSE2 | PTA_NO_SAHF},
2733 {"athlon64-sse3", PROCESSOR_K8, CPU_K8,
2734 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2735 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2736 {"athlon-fx", PROCESSOR_K8, CPU_K8,
2737 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2738 | PTA_SSE2 | PTA_NO_SAHF},
2739 {"amdfam10", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2740 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2741 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2742 {"barcelona", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2743 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2744 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2745 {"generic32", PROCESSOR_GENERIC32, CPU_PENTIUMPRO,
2746 0 /* flags are only used for -march switch. */ },
2747 {"generic64", PROCESSOR_GENERIC64, CPU_GENERIC64,
2748 PTA_64BIT /* flags are only used for -march switch. */ },
2751 int const pta_size = ARRAY_SIZE (processor_alias_table);
2753 /* Set up prefix/suffix so the error messages refer to either the command
2754 line argument, or the attribute(target). */
2763 prefix = "option(\"";
2768 #ifdef SUBTARGET_OVERRIDE_OPTIONS
2769 SUBTARGET_OVERRIDE_OPTIONS;
2772 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
2773 SUBSUBTARGET_OVERRIDE_OPTIONS;
2776 /* -fPIC is the default for x86_64. */
2777 if (TARGET_MACHO && TARGET_64BIT)
2780 /* Set the default values for switches whose default depends on TARGET_64BIT
2781 in case they weren't overwritten by command line options. */
2784 /* Mach-O doesn't support omitting the frame pointer for now. */
2785 if (flag_omit_frame_pointer == 2)
2786 flag_omit_frame_pointer = (TARGET_MACHO ? 0 : 1);
2787 if (flag_asynchronous_unwind_tables == 2)
2788 flag_asynchronous_unwind_tables = 1;
2789 if (flag_pcc_struct_return == 2)
2790 flag_pcc_struct_return = 0;
2794 if (flag_omit_frame_pointer == 2)
2795 flag_omit_frame_pointer = 0;
2796 if (flag_asynchronous_unwind_tables == 2)
2797 flag_asynchronous_unwind_tables = 0;
2798 if (flag_pcc_struct_return == 2)
2799 flag_pcc_struct_return = DEFAULT_PCC_STRUCT_RETURN;
2802 /* Need to check -mtune=generic first. */
2803 if (ix86_tune_string)
2805 if (!strcmp (ix86_tune_string, "generic")
2806 || !strcmp (ix86_tune_string, "i686")
2807 /* As special support for cross compilers we read -mtune=native
2808 as -mtune=generic. With native compilers we won't see the
2809 -mtune=native, as it was changed by the driver. */
2810 || !strcmp (ix86_tune_string, "native"))
2813 ix86_tune_string = "generic64";
2815 ix86_tune_string = "generic32";
2817 /* If this call is for setting the option attribute, allow the
2818 generic32/generic64 that was previously set. */
2819 else if (!main_args_p
2820 && (!strcmp (ix86_tune_string, "generic32")
2821 || !strcmp (ix86_tune_string, "generic64")))
2823 else if (!strncmp (ix86_tune_string, "generic", 7))
2824 error ("bad value (%s) for %stune=%s %s",
2825 ix86_tune_string, prefix, suffix, sw);
2829 if (ix86_arch_string)
2830 ix86_tune_string = ix86_arch_string;
2831 if (!ix86_tune_string)
2833 ix86_tune_string = cpu_names[TARGET_CPU_DEFAULT];
2834 ix86_tune_defaulted = 1;
2837 /* ix86_tune_string is set to ix86_arch_string or defaulted. We
2838 need to use a sensible tune option. */
2839 if (!strcmp (ix86_tune_string, "generic")
2840 || !strcmp (ix86_tune_string, "x86-64")
2841 || !strcmp (ix86_tune_string, "i686"))
2844 ix86_tune_string = "generic64";
2846 ix86_tune_string = "generic32";
2849 if (ix86_stringop_string)
2851 if (!strcmp (ix86_stringop_string, "rep_byte"))
2852 stringop_alg = rep_prefix_1_byte;
2853 else if (!strcmp (ix86_stringop_string, "libcall"))
2854 stringop_alg = libcall;
2855 else if (!strcmp (ix86_stringop_string, "rep_4byte"))
2856 stringop_alg = rep_prefix_4_byte;
2857 else if (!strcmp (ix86_stringop_string, "rep_8byte")
2859 /* rep; movq isn't available in 32-bit code. */
2860 stringop_alg = rep_prefix_8_byte;
2861 else if (!strcmp (ix86_stringop_string, "byte_loop"))
2862 stringop_alg = loop_1_byte;
2863 else if (!strcmp (ix86_stringop_string, "loop"))
2864 stringop_alg = loop;
2865 else if (!strcmp (ix86_stringop_string, "unrolled_loop"))
2866 stringop_alg = unrolled_loop;
2868 error ("bad value (%s) for %sstringop-strategy=%s %s",
2869 ix86_stringop_string, prefix, suffix, sw);
2871 if (!strcmp (ix86_tune_string, "x86-64"))
2872 warning (OPT_Wdeprecated, "%stune=x86-64%s is deprecated. Use "
2873 "%stune=k8%s or %stune=generic%s instead as appropriate.",
2874 prefix, suffix, prefix, suffix, prefix, suffix);
2876 if (!ix86_arch_string)
2877 ix86_arch_string = TARGET_64BIT ? "x86-64" : "i386";
2879 ix86_arch_specified = 1;
2881 if (!strcmp (ix86_arch_string, "generic"))
2882 error ("generic CPU can be used only for %stune=%s %s",
2883 prefix, suffix, sw);
2884 if (!strncmp (ix86_arch_string, "generic", 7))
2885 error ("bad value (%s) for %sarch=%s %s",
2886 ix86_arch_string, prefix, suffix, sw);
2888 /* Validate -mabi= value. */
2889 if (ix86_abi_string)
2891 if (strcmp (ix86_abi_string, "sysv") == 0)
2892 ix86_abi = SYSV_ABI;
2893 else if (strcmp (ix86_abi_string, "ms") == 0)
2896 error ("unknown ABI (%s) for %sabi=%s %s",
2897 ix86_abi_string, prefix, suffix, sw);
2900 ix86_abi = DEFAULT_ABI;
2902 if (ix86_cmodel_string != 0)
2904 if (!strcmp (ix86_cmodel_string, "small"))
2905 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2906 else if (!strcmp (ix86_cmodel_string, "medium"))
2907 ix86_cmodel = flag_pic ? CM_MEDIUM_PIC : CM_MEDIUM;
2908 else if (!strcmp (ix86_cmodel_string, "large"))
2909 ix86_cmodel = flag_pic ? CM_LARGE_PIC : CM_LARGE;
2911 error ("code model %s does not support PIC mode", ix86_cmodel_string);
2912 else if (!strcmp (ix86_cmodel_string, "32"))
2913 ix86_cmodel = CM_32;
2914 else if (!strcmp (ix86_cmodel_string, "kernel") && !flag_pic)
2915 ix86_cmodel = CM_KERNEL;
2917 error ("bad value (%s) for %scmodel=%s %s",
2918 ix86_cmodel_string, prefix, suffix, sw);
2922 /* For TARGET_64BIT and MS_ABI, force pic on, in order to enable the
2923 use of rip-relative addressing. This eliminates fixups that
2924 would otherwise be needed if this object is to be placed in a
2925 DLL, and is essentially just as efficient as direct addressing. */
2926 if (TARGET_64BIT && DEFAULT_ABI == MS_ABI)
2927 ix86_cmodel = CM_SMALL_PIC, flag_pic = 1;
2928 else if (TARGET_64BIT)
2929 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2931 ix86_cmodel = CM_32;
2933 if (ix86_asm_string != 0)
2936 && !strcmp (ix86_asm_string, "intel"))
2937 ix86_asm_dialect = ASM_INTEL;
2938 else if (!strcmp (ix86_asm_string, "att"))
2939 ix86_asm_dialect = ASM_ATT;
2941 error ("bad value (%s) for %sasm=%s %s",
2942 ix86_asm_string, prefix, suffix, sw);
2944 if ((TARGET_64BIT == 0) != (ix86_cmodel == CM_32))
2945 error ("code model %qs not supported in the %s bit mode",
2946 ix86_cmodel_string, TARGET_64BIT ? "64" : "32");
2947 if ((TARGET_64BIT != 0) != ((ix86_isa_flags & OPTION_MASK_ISA_64BIT) != 0))
2948 sorry ("%i-bit mode not compiled in",
2949 (ix86_isa_flags & OPTION_MASK_ISA_64BIT) ? 64 : 32);
2951 for (i = 0; i < pta_size; i++)
2952 if (! strcmp (ix86_arch_string, processor_alias_table[i].name))
2954 ix86_schedule = processor_alias_table[i].schedule;
2955 ix86_arch = processor_alias_table[i].processor;
2956 /* Default cpu tuning to the architecture. */
2957 ix86_tune = ix86_arch;
2959 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
2960 error ("CPU you selected does not support x86-64 "
2963 if (processor_alias_table[i].flags & PTA_MMX
2964 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_MMX))
2965 ix86_isa_flags |= OPTION_MASK_ISA_MMX;
2966 if (processor_alias_table[i].flags & PTA_3DNOW
2967 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW))
2968 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW;
2969 if (processor_alias_table[i].flags & PTA_3DNOW_A
2970 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW_A))
2971 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_A;
2972 if (processor_alias_table[i].flags & PTA_SSE
2973 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE))
2974 ix86_isa_flags |= OPTION_MASK_ISA_SSE;
2975 if (processor_alias_table[i].flags & PTA_SSE2
2976 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE2))
2977 ix86_isa_flags |= OPTION_MASK_ISA_SSE2;
2978 if (processor_alias_table[i].flags & PTA_SSE3
2979 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE3))
2980 ix86_isa_flags |= OPTION_MASK_ISA_SSE3;
2981 if (processor_alias_table[i].flags & PTA_SSSE3
2982 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSSE3))
2983 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3;
2984 if (processor_alias_table[i].flags & PTA_SSE4_1
2985 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_1))
2986 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1;
2987 if (processor_alias_table[i].flags & PTA_SSE4_2
2988 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_2))
2989 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2;
2990 if (processor_alias_table[i].flags & PTA_AVX
2991 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_AVX))
2992 ix86_isa_flags |= OPTION_MASK_ISA_AVX;
2993 if (processor_alias_table[i].flags & PTA_FMA
2994 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_FMA))
2995 ix86_isa_flags |= OPTION_MASK_ISA_FMA;
2996 if (processor_alias_table[i].flags & PTA_SSE4A
2997 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4A))
2998 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A;
2999 if (processor_alias_table[i].flags & PTA_FMA4
3000 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_FMA4))
3001 ix86_isa_flags |= OPTION_MASK_ISA_FMA4;
3002 if (processor_alias_table[i].flags & PTA_XOP
3003 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_XOP))
3004 ix86_isa_flags |= OPTION_MASK_ISA_XOP;
3005 if (processor_alias_table[i].flags & PTA_LWP
3006 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_LWP))
3007 ix86_isa_flags |= OPTION_MASK_ISA_LWP;
3008 if (processor_alias_table[i].flags & PTA_ABM
3009 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_ABM))
3010 ix86_isa_flags |= OPTION_MASK_ISA_ABM;
3011 if (processor_alias_table[i].flags & PTA_CX16
3012 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_CX16))
3013 ix86_isa_flags |= OPTION_MASK_ISA_CX16;
3014 if (processor_alias_table[i].flags & (PTA_POPCNT | PTA_ABM)
3015 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_POPCNT))
3016 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT;
3017 if (!(TARGET_64BIT && (processor_alias_table[i].flags & PTA_NO_SAHF))
3018 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SAHF))
3019 ix86_isa_flags |= OPTION_MASK_ISA_SAHF;
3020 if (processor_alias_table[i].flags & PTA_MOVBE
3021 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_MOVBE))
3022 ix86_isa_flags |= OPTION_MASK_ISA_MOVBE;
3023 if (processor_alias_table[i].flags & PTA_AES
3024 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_AES))
3025 ix86_isa_flags |= OPTION_MASK_ISA_AES;
3026 if (processor_alias_table[i].flags & PTA_PCLMUL
3027 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_PCLMUL))
3028 ix86_isa_flags |= OPTION_MASK_ISA_PCLMUL;
3029 if (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE))
3030 x86_prefetch_sse = true;
3036 error ("bad value (%s) for %sarch=%s %s",
3037 ix86_arch_string, prefix, suffix, sw);
3039 ix86_arch_mask = 1u << ix86_arch;
3040 for (i = 0; i < X86_ARCH_LAST; ++i)
3041 ix86_arch_features[i] = !!(initial_ix86_arch_features[i] & ix86_arch_mask);
3043 for (i = 0; i < pta_size; i++)
3044 if (! strcmp (ix86_tune_string, processor_alias_table[i].name))
3046 ix86_schedule = processor_alias_table[i].schedule;
3047 ix86_tune = processor_alias_table[i].processor;
3048 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
3050 if (ix86_tune_defaulted)
3052 ix86_tune_string = "x86-64";
3053 for (i = 0; i < pta_size; i++)
3054 if (! strcmp (ix86_tune_string,
3055 processor_alias_table[i].name))
3057 ix86_schedule = processor_alias_table[i].schedule;
3058 ix86_tune = processor_alias_table[i].processor;
3061 error ("CPU you selected does not support x86-64 "
3064 /* Intel CPUs have always interpreted SSE prefetch instructions as
3065 NOPs; so, we can enable SSE prefetch instructions even when
3066 -mtune (rather than -march) points us to a processor that has them.
3067 However, the VIA C3 gives a SIGILL, so we only do that for i686 and
3068 higher processors. */
3070 && (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE)))
3071 x86_prefetch_sse = true;
3075 error ("bad value (%s) for %stune=%s %s",
3076 ix86_tune_string, prefix, suffix, sw);
3078 ix86_tune_mask = 1u << ix86_tune;
3079 for (i = 0; i < X86_TUNE_LAST; ++i)
3080 ix86_tune_features[i] = !!(initial_ix86_tune_features[i] & ix86_tune_mask);
3083 ix86_cost = &ix86_size_cost;
3085 ix86_cost = processor_target_table[ix86_tune].cost;
3087 /* Arrange to set up i386_stack_locals for all functions. */
3088 init_machine_status = ix86_init_machine_status;
3090 /* Validate -mregparm= value. */
3091 if (ix86_regparm_string)
3094 warning (0, "%sregparm%s is ignored in 64-bit mode", prefix, suffix);
3095 i = atoi (ix86_regparm_string);
3096 if (i < 0 || i > REGPARM_MAX)
3097 error ("%sregparm=%d%s is not between 0 and %d",
3098 prefix, i, suffix, REGPARM_MAX);
3103 ix86_regparm = REGPARM_MAX;
3105 /* If the user has provided any of the -malign-* options,
3106 warn and use that value only if -falign-* is not set.
3107 Remove this code in GCC 3.2 or later. */
3108 if (ix86_align_loops_string)
3110 warning (0, "%salign-loops%s is obsolete, use -falign-loops%s",
3111 prefix, suffix, suffix);
3112 if (align_loops == 0)
3114 i = atoi (ix86_align_loops_string);
3115 if (i < 0 || i > MAX_CODE_ALIGN)
3116 error ("%salign-loops=%d%s is not between 0 and %d",
3117 prefix, i, suffix, MAX_CODE_ALIGN);
3119 align_loops = 1 << i;
3123 if (ix86_align_jumps_string)
3125 warning (0, "%salign-jumps%s is obsolete, use -falign-jumps%s",
3126 prefix, suffix, suffix);
3127 if (align_jumps == 0)
3129 i = atoi (ix86_align_jumps_string);
3130 if (i < 0 || i > MAX_CODE_ALIGN)
3131 error ("%salign-loops=%d%s is not between 0 and %d",
3132 prefix, i, suffix, MAX_CODE_ALIGN);
3134 align_jumps = 1 << i;
3138 if (ix86_align_funcs_string)
3140 warning (0, "%salign-functions%s is obsolete, use -falign-functions%s",
3141 prefix, suffix, suffix);
3142 if (align_functions == 0)
3144 i = atoi (ix86_align_funcs_string);
3145 if (i < 0 || i > MAX_CODE_ALIGN)
3146 error ("%salign-loops=%d%s is not between 0 and %d",
3147 prefix, i, suffix, MAX_CODE_ALIGN);
3149 align_functions = 1 << i;
3153 /* Default align_* from the processor table. */
3154 if (align_loops == 0)
3156 align_loops = processor_target_table[ix86_tune].align_loop;
3157 align_loops_max_skip = processor_target_table[ix86_tune].align_loop_max_skip;
3159 if (align_jumps == 0)
3161 align_jumps = processor_target_table[ix86_tune].align_jump;
3162 align_jumps_max_skip = processor_target_table[ix86_tune].align_jump_max_skip;
3164 if (align_functions == 0)
3166 align_functions = processor_target_table[ix86_tune].align_func;
3169 /* Validate -mbranch-cost= value, or provide default. */
3170 ix86_branch_cost = ix86_cost->branch_cost;
3171 if (ix86_branch_cost_string)
3173 i = atoi (ix86_branch_cost_string);
3175 error ("%sbranch-cost=%d%s is not between 0 and 5", prefix, i, suffix);
3177 ix86_branch_cost = i;
3179 if (ix86_section_threshold_string)
3181 i = atoi (ix86_section_threshold_string);
3183 error ("%slarge-data-threshold=%d%s is negative", prefix, i, suffix);
3185 ix86_section_threshold = i;
3188 if (ix86_tls_dialect_string)
3190 if (strcmp (ix86_tls_dialect_string, "gnu") == 0)
3191 ix86_tls_dialect = TLS_DIALECT_GNU;
3192 else if (strcmp (ix86_tls_dialect_string, "gnu2") == 0)
3193 ix86_tls_dialect = TLS_DIALECT_GNU2;
3194 else if (strcmp (ix86_tls_dialect_string, "sun") == 0)
3195 ix86_tls_dialect = TLS_DIALECT_SUN;
3197 error ("bad value (%s) for %stls-dialect=%s %s",
3198 ix86_tls_dialect_string, prefix, suffix, sw);
3201 if (ix87_precision_string)
3203 i = atoi (ix87_precision_string);
3204 if (i != 32 && i != 64 && i != 80)
3205 error ("pc%d is not valid precision setting (32, 64 or 80)", i);
3210 target_flags |= TARGET_SUBTARGET64_DEFAULT & ~target_flags_explicit;
3212 /* Enable by default the SSE and MMX builtins. Do allow the user to
3213 explicitly disable any of these. In particular, disabling SSE and
3214 MMX for kernel code is extremely useful. */
3215 if (!ix86_arch_specified)
3217 |= ((OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_MMX
3218 | TARGET_SUBTARGET64_ISA_DEFAULT) & ~ix86_isa_flags_explicit);
3221 warning (0, "%srtd%s is ignored in 64bit mode", prefix, suffix);
3225 target_flags |= TARGET_SUBTARGET32_DEFAULT & ~target_flags_explicit;
3227 if (!ix86_arch_specified)
3229 |= TARGET_SUBTARGET32_ISA_DEFAULT & ~ix86_isa_flags_explicit;
3231 /* i386 ABI does not specify red zone. It still makes sense to use it
3232 when programmer takes care to stack from being destroyed. */
3233 if (!(target_flags_explicit & MASK_NO_RED_ZONE))
3234 target_flags |= MASK_NO_RED_ZONE;
3237 /* Keep nonleaf frame pointers. */
3238 if (flag_omit_frame_pointer)
3239 target_flags &= ~MASK_OMIT_LEAF_FRAME_POINTER;
3240 else if (TARGET_OMIT_LEAF_FRAME_POINTER)
3241 flag_omit_frame_pointer = 1;
3243 /* If we're doing fast math, we don't care about comparison order
3244 wrt NaNs. This lets us use a shorter comparison sequence. */
3245 if (flag_finite_math_only)
3246 target_flags &= ~MASK_IEEE_FP;
3248 /* If the architecture always has an FPU, turn off NO_FANCY_MATH_387,
3249 since the insns won't need emulation. */
3250 if (x86_arch_always_fancy_math_387 & ix86_arch_mask)
3251 target_flags &= ~MASK_NO_FANCY_MATH_387;
3253 /* Likewise, if the target doesn't have a 387, or we've specified
3254 software floating point, don't use 387 inline intrinsics. */
3256 target_flags |= MASK_NO_FANCY_MATH_387;
3258 /* Turn on MMX builtins for -msse. */
3261 ix86_isa_flags |= OPTION_MASK_ISA_MMX & ~ix86_isa_flags_explicit;
3262 x86_prefetch_sse = true;
3265 /* Turn on popcnt instruction for -msse4.2 or -mabm. */
3266 if (TARGET_SSE4_2 || TARGET_ABM)
3267 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT & ~ix86_isa_flags_explicit;
3269 /* Validate -mpreferred-stack-boundary= value or default it to
3270 PREFERRED_STACK_BOUNDARY_DEFAULT. */
3271 ix86_preferred_stack_boundary = PREFERRED_STACK_BOUNDARY_DEFAULT;
3272 if (ix86_preferred_stack_boundary_string)
3274 i = atoi (ix86_preferred_stack_boundary_string);
3275 if (i < (TARGET_64BIT ? 4 : 2) || i > 12)
3276 error ("%spreferred-stack-boundary=%d%s is not between %d and 12",
3277 prefix, i, suffix, TARGET_64BIT ? 4 : 2);
3279 ix86_preferred_stack_boundary = (1 << i) * BITS_PER_UNIT;
3282 /* Set the default value for -mstackrealign. */
3283 if (ix86_force_align_arg_pointer == -1)
3284 ix86_force_align_arg_pointer = STACK_REALIGN_DEFAULT;
3286 ix86_default_incoming_stack_boundary = PREFERRED_STACK_BOUNDARY;
3288 /* Validate -mincoming-stack-boundary= value or default it to
3289 MIN_STACK_BOUNDARY/PREFERRED_STACK_BOUNDARY. */
3290 ix86_incoming_stack_boundary = ix86_default_incoming_stack_boundary;
3291 if (ix86_incoming_stack_boundary_string)
3293 i = atoi (ix86_incoming_stack_boundary_string);
3294 if (i < (TARGET_64BIT ? 4 : 2) || i > 12)
3295 error ("-mincoming-stack-boundary=%d is not between %d and 12",
3296 i, TARGET_64BIT ? 4 : 2);
3299 ix86_user_incoming_stack_boundary = (1 << i) * BITS_PER_UNIT;
3300 ix86_incoming_stack_boundary
3301 = ix86_user_incoming_stack_boundary;
3305 /* Accept -msseregparm only if at least SSE support is enabled. */
3306 if (TARGET_SSEREGPARM
3308 error ("%ssseregparm%s used without SSE enabled", prefix, suffix);
3310 ix86_fpmath = TARGET_FPMATH_DEFAULT;
3311 if (ix86_fpmath_string != 0)
3313 if (! strcmp (ix86_fpmath_string, "387"))
3314 ix86_fpmath = FPMATH_387;
3315 else if (! strcmp (ix86_fpmath_string, "sse"))
3319 warning (0, "SSE instruction set disabled, using 387 arithmetics");
3320 ix86_fpmath = FPMATH_387;
3323 ix86_fpmath = FPMATH_SSE;
3325 else if (! strcmp (ix86_fpmath_string, "387,sse")
3326 || ! strcmp (ix86_fpmath_string, "387+sse")
3327 || ! strcmp (ix86_fpmath_string, "sse,387")
3328 || ! strcmp (ix86_fpmath_string, "sse+387")
3329 || ! strcmp (ix86_fpmath_string, "both"))
3333 warning (0, "SSE instruction set disabled, using 387 arithmetics");
3334 ix86_fpmath = FPMATH_387;
3336 else if (!TARGET_80387)
3338 warning (0, "387 instruction set disabled, using SSE arithmetics");
3339 ix86_fpmath = FPMATH_SSE;
3342 ix86_fpmath = (enum fpmath_unit) (FPMATH_SSE | FPMATH_387);
3345 error ("bad value (%s) for %sfpmath=%s %s",
3346 ix86_fpmath_string, prefix, suffix, sw);
3349 /* If the i387 is disabled, then do not return values in it. */
3351 target_flags &= ~MASK_FLOAT_RETURNS;
3353 /* Use external vectorized library in vectorizing intrinsics. */
3354 if (ix86_veclibabi_string)
3356 if (strcmp (ix86_veclibabi_string, "svml") == 0)
3357 ix86_veclib_handler = ix86_veclibabi_svml;
3358 else if (strcmp (ix86_veclibabi_string, "acml") == 0)
3359 ix86_veclib_handler = ix86_veclibabi_acml;
3361 error ("unknown vectorization library ABI type (%s) for "
3362 "%sveclibabi=%s %s", ix86_veclibabi_string,
3363 prefix, suffix, sw);
3366 if ((x86_accumulate_outgoing_args & ix86_tune_mask)
3367 && !(target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3369 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3371 /* ??? Unwind info is not correct around the CFG unless either a frame
3372 pointer is present or M_A_O_A is set. Fixing this requires rewriting
3373 unwind info generation to be aware of the CFG and propagating states
3375 if ((flag_unwind_tables || flag_asynchronous_unwind_tables
3376 || flag_exceptions || flag_non_call_exceptions)
3377 && flag_omit_frame_pointer
3378 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
3380 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3381 warning (0, "unwind tables currently require either a frame pointer "
3382 "or %saccumulate-outgoing-args%s for correctness",
3384 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3387 /* If stack probes are required, the space used for large function
3388 arguments on the stack must also be probed, so enable
3389 -maccumulate-outgoing-args so this happens in the prologue. */
3390 if (TARGET_STACK_PROBE
3391 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
3393 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3394 warning (0, "stack probing requires %saccumulate-outgoing-args%s "
3395 "for correctness", prefix, suffix);
3396 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3399 /* For sane SSE instruction set generation we need fcomi instruction.
3400 It is safe to enable all CMOVE instructions. */
3404 /* Figure out what ASM_GENERATE_INTERNAL_LABEL builds as a prefix. */
3407 ASM_GENERATE_INTERNAL_LABEL (internal_label_prefix, "LX", 0);
3408 p = strchr (internal_label_prefix, 'X');
3409 internal_label_prefix_len = p - internal_label_prefix;
3413 /* When scheduling description is not available, disable scheduler pass
3414 so it won't slow down the compilation and make x87 code slower. */
3415 if (!TARGET_SCHEDULE)
3416 flag_schedule_insns_after_reload = flag_schedule_insns = 0;
3418 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES))
3419 set_param_value ("simultaneous-prefetches",
3420 ix86_cost->simultaneous_prefetches);
3421 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE))
3422 set_param_value ("l1-cache-line-size", ix86_cost->prefetch_block);
3423 if (!PARAM_SET_P (PARAM_L1_CACHE_SIZE))
3424 set_param_value ("l1-cache-size", ix86_cost->l1_cache_size);
3425 if (!PARAM_SET_P (PARAM_L2_CACHE_SIZE))
3426 set_param_value ("l2-cache-size", ix86_cost->l2_cache_size);
3428 /* If using typedef char *va_list, signal that __builtin_va_start (&ap, 0)
3429 can be optimized to ap = __builtin_next_arg (0). */
3431 targetm.expand_builtin_va_start = NULL;
3435 ix86_gen_leave = gen_leave_rex64;
3436 ix86_gen_pop1 = gen_popdi1;
3437 ix86_gen_add3 = gen_adddi3;
3438 ix86_gen_sub3 = gen_subdi3;
3439 ix86_gen_sub3_carry = gen_subdi3_carry;
3440 ix86_gen_one_cmpl2 = gen_one_cmpldi2;
3441 ix86_gen_monitor = gen_sse3_monitor64;
3442 ix86_gen_andsp = gen_anddi3;
3446 ix86_gen_leave = gen_leave;
3447 ix86_gen_pop1 = gen_popsi1;
3448 ix86_gen_add3 = gen_addsi3;
3449 ix86_gen_sub3 = gen_subsi3;
3450 ix86_gen_sub3_carry = gen_subsi3_carry;
3451 ix86_gen_one_cmpl2 = gen_one_cmplsi2;
3452 ix86_gen_monitor = gen_sse3_monitor;
3453 ix86_gen_andsp = gen_andsi3;
3457 /* Use -mcld by default for 32-bit code if configured with --enable-cld. */
3459 target_flags |= MASK_CLD & ~target_flags_explicit;
3462 /* Save the initial options in case the user does function specific options */
3464 target_option_default_node = target_option_current_node
3465 = build_target_option_node ();
3468 /* Update register usage after having seen the compiler flags. */
3471 ix86_conditional_register_usage (void)
3476 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3478 if (fixed_regs[i] > 1)
3479 fixed_regs[i] = (fixed_regs[i] == (TARGET_64BIT ? 3 : 2));
3480 if (call_used_regs[i] > 1)
3481 call_used_regs[i] = (call_used_regs[i] == (TARGET_64BIT ? 3 : 2));
3484 /* The PIC register, if it exists, is fixed. */
3485 j = PIC_OFFSET_TABLE_REGNUM;
3486 if (j != INVALID_REGNUM)
3487 fixed_regs[j] = call_used_regs[j] = 1;
3489 /* The MS_ABI changes the set of call-used registers. */
3490 if (TARGET_64BIT && ix86_cfun_abi () == MS_ABI)
3492 call_used_regs[SI_REG] = 0;
3493 call_used_regs[DI_REG] = 0;
3494 call_used_regs[XMM6_REG] = 0;
3495 call_used_regs[XMM7_REG] = 0;
3496 for (i = FIRST_REX_SSE_REG; i <= LAST_REX_SSE_REG; i++)
3497 call_used_regs[i] = 0;
3500 /* The default setting of CLOBBERED_REGS is for 32-bit; add in the
3501 other call-clobbered regs for 64-bit. */
3504 CLEAR_HARD_REG_SET (reg_class_contents[(int)CLOBBERED_REGS]);
3506 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3507 if (TEST_HARD_REG_BIT (reg_class_contents[(int)GENERAL_REGS], i)
3508 && call_used_regs[i])
3509 SET_HARD_REG_BIT (reg_class_contents[(int)CLOBBERED_REGS], i);
3512 /* If MMX is disabled, squash the registers. */
3514 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3515 if (TEST_HARD_REG_BIT (reg_class_contents[(int)MMX_REGS], i))
3516 fixed_regs[i] = call_used_regs[i] = 1, reg_names[i] = "";
3518 /* If SSE is disabled, squash the registers. */
3520 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3521 if (TEST_HARD_REG_BIT (reg_class_contents[(int)SSE_REGS], i))
3522 fixed_regs[i] = call_used_regs[i] = 1, reg_names[i] = "";
3524 /* If the FPU is disabled, squash the registers. */
3525 if (! (TARGET_80387 || TARGET_FLOAT_RETURNS_IN_80387))
3526 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
3527 if (TEST_HARD_REG_BIT (reg_class_contents[(int)FLOAT_REGS], i))
3528 fixed_regs[i] = call_used_regs[i] = 1, reg_names[i] = "";
3530 /* If 32-bit, squash the 64-bit registers. */
3533 for (i = FIRST_REX_INT_REG; i <= LAST_REX_INT_REG; i++)
3535 for (i = FIRST_REX_SSE_REG; i <= LAST_REX_SSE_REG; i++)
3541 /* Save the current options */
3544 ix86_function_specific_save (struct cl_target_option *ptr)
3546 ptr->arch = ix86_arch;
3547 ptr->schedule = ix86_schedule;
3548 ptr->tune = ix86_tune;
3549 ptr->fpmath = ix86_fpmath;
3550 ptr->branch_cost = ix86_branch_cost;
3551 ptr->tune_defaulted = ix86_tune_defaulted;
3552 ptr->arch_specified = ix86_arch_specified;
3553 ptr->ix86_isa_flags_explicit = ix86_isa_flags_explicit;
3554 ptr->target_flags_explicit = target_flags_explicit;
3556 /* The fields are char but the variables are not; make sure the
3557 values fit in the fields. */
3558 gcc_assert (ptr->arch == ix86_arch);
3559 gcc_assert (ptr->schedule == ix86_schedule);
3560 gcc_assert (ptr->tune == ix86_tune);
3561 gcc_assert (ptr->fpmath == ix86_fpmath);
3562 gcc_assert (ptr->branch_cost == ix86_branch_cost);
3565 /* Restore the current options */
3568 ix86_function_specific_restore (struct cl_target_option *ptr)
3570 enum processor_type old_tune = ix86_tune;
3571 enum processor_type old_arch = ix86_arch;
3572 unsigned int ix86_arch_mask, ix86_tune_mask;
3575 ix86_arch = (enum processor_type) ptr->arch;
3576 ix86_schedule = (enum attr_cpu) ptr->schedule;
3577 ix86_tune = (enum processor_type) ptr->tune;
3578 ix86_fpmath = (enum fpmath_unit) ptr->fpmath;
3579 ix86_branch_cost = ptr->branch_cost;
3580 ix86_tune_defaulted = ptr->tune_defaulted;
3581 ix86_arch_specified = ptr->arch_specified;
3582 ix86_isa_flags_explicit = ptr->ix86_isa_flags_explicit;
3583 target_flags_explicit = ptr->target_flags_explicit;
3585 /* Recreate the arch feature tests if the arch changed */
3586 if (old_arch != ix86_arch)
3588 ix86_arch_mask = 1u << ix86_arch;
3589 for (i = 0; i < X86_ARCH_LAST; ++i)
3590 ix86_arch_features[i]
3591 = !!(initial_ix86_arch_features[i] & ix86_arch_mask);
3594 /* Recreate the tune optimization tests */
3595 if (old_tune != ix86_tune)
3597 ix86_tune_mask = 1u << ix86_tune;
3598 for (i = 0; i < X86_TUNE_LAST; ++i)
3599 ix86_tune_features[i]
3600 = !!(initial_ix86_tune_features[i] & ix86_tune_mask);
3604 /* Print the current options */
3607 ix86_function_specific_print (FILE *file, int indent,
3608 struct cl_target_option *ptr)
3611 = ix86_target_string (ptr->ix86_isa_flags, ptr->target_flags,
3612 NULL, NULL, NULL, false);
3614 fprintf (file, "%*sarch = %d (%s)\n",
3617 ((ptr->arch < TARGET_CPU_DEFAULT_max)
3618 ? cpu_names[ptr->arch]
3621 fprintf (file, "%*stune = %d (%s)\n",
3624 ((ptr->tune < TARGET_CPU_DEFAULT_max)
3625 ? cpu_names[ptr->tune]
3628 fprintf (file, "%*sfpmath = %d%s%s\n", indent, "", ptr->fpmath,
3629 (ptr->fpmath & FPMATH_387) ? ", 387" : "",
3630 (ptr->fpmath & FPMATH_SSE) ? ", sse" : "");
3631 fprintf (file, "%*sbranch_cost = %d\n", indent, "", ptr->branch_cost);
3635 fprintf (file, "%*s%s\n", indent, "", target_string);
3636 free (target_string);
3641 /* Inner function to process the attribute((target(...))), take an argument and
3642 set the current options from the argument. If we have a list, recursively go
3646 ix86_valid_target_attribute_inner_p (tree args, char *p_strings[])
3651 #define IX86_ATTR_ISA(S,O) { S, sizeof (S)-1, ix86_opt_isa, O, 0 }
3652 #define IX86_ATTR_STR(S,O) { S, sizeof (S)-1, ix86_opt_str, O, 0 }
3653 #define IX86_ATTR_YES(S,O,M) { S, sizeof (S)-1, ix86_opt_yes, O, M }
3654 #define IX86_ATTR_NO(S,O,M) { S, sizeof (S)-1, ix86_opt_no, O, M }
3669 enum ix86_opt_type type;
3674 IX86_ATTR_ISA ("3dnow", OPT_m3dnow),
3675 IX86_ATTR_ISA ("abm", OPT_mabm),
3676 IX86_ATTR_ISA ("aes", OPT_maes),
3677 IX86_ATTR_ISA ("avx", OPT_mavx),
3678 IX86_ATTR_ISA ("mmx", OPT_mmmx),
3679 IX86_ATTR_ISA ("pclmul", OPT_mpclmul),
3680 IX86_ATTR_ISA ("popcnt", OPT_mpopcnt),
3681 IX86_ATTR_ISA ("sse", OPT_msse),
3682 IX86_ATTR_ISA ("sse2", OPT_msse2),
3683 IX86_ATTR_ISA ("sse3", OPT_msse3),
3684 IX86_ATTR_ISA ("sse4", OPT_msse4),
3685 IX86_ATTR_ISA ("sse4.1", OPT_msse4_1),
3686 IX86_ATTR_ISA ("sse4.2", OPT_msse4_2),
3687 IX86_ATTR_ISA ("sse4a", OPT_msse4a),
3688 IX86_ATTR_ISA ("ssse3", OPT_mssse3),
3689 IX86_ATTR_ISA ("fma4", OPT_mfma4),
3690 IX86_ATTR_ISA ("xop", OPT_mxop),
3691 IX86_ATTR_ISA ("lwp", OPT_mlwp),
3693 /* string options */
3694 IX86_ATTR_STR ("arch=", IX86_FUNCTION_SPECIFIC_ARCH),
3695 IX86_ATTR_STR ("fpmath=", IX86_FUNCTION_SPECIFIC_FPMATH),
3696 IX86_ATTR_STR ("tune=", IX86_FUNCTION_SPECIFIC_TUNE),
3699 IX86_ATTR_YES ("cld",
3703 IX86_ATTR_NO ("fancy-math-387",
3704 OPT_mfancy_math_387,
3705 MASK_NO_FANCY_MATH_387),
3707 IX86_ATTR_YES ("ieee-fp",
3711 IX86_ATTR_YES ("inline-all-stringops",
3712 OPT_minline_all_stringops,
3713 MASK_INLINE_ALL_STRINGOPS),
3715 IX86_ATTR_YES ("inline-stringops-dynamically",
3716 OPT_minline_stringops_dynamically,
3717 MASK_INLINE_STRINGOPS_DYNAMICALLY),
3719 IX86_ATTR_NO ("align-stringops",
3720 OPT_mno_align_stringops,
3721 MASK_NO_ALIGN_STRINGOPS),
3723 IX86_ATTR_YES ("recip",
3729 /* If this is a list, recurse to get the options. */
3730 if (TREE_CODE (args) == TREE_LIST)
3734 for (; args; args = TREE_CHAIN (args))
3735 if (TREE_VALUE (args)
3736 && !ix86_valid_target_attribute_inner_p (TREE_VALUE (args), p_strings))
3742 else if (TREE_CODE (args) != STRING_CST)
3745 /* Handle multiple arguments separated by commas. */
3746 next_optstr = ASTRDUP (TREE_STRING_POINTER (args));
3748 while (next_optstr && *next_optstr != '\0')
3750 char *p = next_optstr;
3752 char *comma = strchr (next_optstr, ',');
3753 const char *opt_string;
3754 size_t len, opt_len;
3759 enum ix86_opt_type type = ix86_opt_unknown;
3765 len = comma - next_optstr;
3766 next_optstr = comma + 1;
3774 /* Recognize no-xxx. */
3775 if (len > 3 && p[0] == 'n' && p[1] == 'o' && p[2] == '-')
3784 /* Find the option. */
3787 for (i = 0; i < ARRAY_SIZE (attrs); i++)
3789 type = attrs[i].type;
3790 opt_len = attrs[i].len;
3791 if (ch == attrs[i].string[0]
3792 && ((type != ix86_opt_str) ? len == opt_len : len > opt_len)
3793 && memcmp (p, attrs[i].string, opt_len) == 0)
3796 mask = attrs[i].mask;
3797 opt_string = attrs[i].string;
3802 /* Process the option. */
3805 error ("attribute(target(\"%s\")) is unknown", orig_p);
3809 else if (type == ix86_opt_isa)
3810 ix86_handle_option (opt, p, opt_set_p);
3812 else if (type == ix86_opt_yes || type == ix86_opt_no)
3814 if (type == ix86_opt_no)
3815 opt_set_p = !opt_set_p;
3818 target_flags |= mask;
3820 target_flags &= ~mask;
3823 else if (type == ix86_opt_str)
3827 error ("option(\"%s\") was already specified", opt_string);
3831 p_strings[opt] = xstrdup (p + opt_len);
3841 /* Return a TARGET_OPTION_NODE tree of the target options listed or NULL. */
3844 ix86_valid_target_attribute_tree (tree args)
3846 const char *orig_arch_string = ix86_arch_string;
3847 const char *orig_tune_string = ix86_tune_string;
3848 const char *orig_fpmath_string = ix86_fpmath_string;
3849 int orig_tune_defaulted = ix86_tune_defaulted;
3850 int orig_arch_specified = ix86_arch_specified;
3851 char *option_strings[IX86_FUNCTION_SPECIFIC_MAX] = { NULL, NULL, NULL };
3854 struct cl_target_option *def
3855 = TREE_TARGET_OPTION (target_option_default_node);
3857 /* Process each of the options on the chain. */
3858 if (! ix86_valid_target_attribute_inner_p (args, option_strings))
3861 /* If the changed options are different from the default, rerun override_options,
3862 and then save the options away. The string options are are attribute options,
3863 and will be undone when we copy the save structure. */
3864 if (ix86_isa_flags != def->ix86_isa_flags
3865 || target_flags != def->target_flags
3866 || option_strings[IX86_FUNCTION_SPECIFIC_ARCH]
3867 || option_strings[IX86_FUNCTION_SPECIFIC_TUNE]
3868 || option_strings[IX86_FUNCTION_SPECIFIC_FPMATH])
3870 /* If we are using the default tune= or arch=, undo the string assigned,
3871 and use the default. */
3872 if (option_strings[IX86_FUNCTION_SPECIFIC_ARCH])
3873 ix86_arch_string = option_strings[IX86_FUNCTION_SPECIFIC_ARCH];
3874 else if (!orig_arch_specified)
3875 ix86_arch_string = NULL;
3877 if (option_strings[IX86_FUNCTION_SPECIFIC_TUNE])
3878 ix86_tune_string = option_strings[IX86_FUNCTION_SPECIFIC_TUNE];
3879 else if (orig_tune_defaulted)
3880 ix86_tune_string = NULL;
3882 /* If fpmath= is not set, and we now have sse2 on 32-bit, use it. */
3883 if (option_strings[IX86_FUNCTION_SPECIFIC_FPMATH])
3884 ix86_fpmath_string = option_strings[IX86_FUNCTION_SPECIFIC_FPMATH];
3885 else if (!TARGET_64BIT && TARGET_SSE)
3886 ix86_fpmath_string = "sse,387";
3888 /* Do any overrides, such as arch=xxx, or tune=xxx support. */
3889 override_options (false);
3891 /* Add any builtin functions with the new isa if any. */
3892 ix86_add_new_builtins (ix86_isa_flags);
3894 /* Save the current options unless we are validating options for
3896 t = build_target_option_node ();
3898 ix86_arch_string = orig_arch_string;
3899 ix86_tune_string = orig_tune_string;
3900 ix86_fpmath_string = orig_fpmath_string;
3902 /* Free up memory allocated to hold the strings */
3903 for (i = 0; i < IX86_FUNCTION_SPECIFIC_MAX; i++)
3904 if (option_strings[i])
3905 free (option_strings[i]);
3911 /* Hook to validate attribute((target("string"))). */
3914 ix86_valid_target_attribute_p (tree fndecl,
3915 tree ARG_UNUSED (name),
3917 int ARG_UNUSED (flags))
3919 struct cl_target_option cur_target;
3921 tree old_optimize = build_optimization_node ();
3922 tree new_target, new_optimize;
3923 tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl);
3925 /* If the function changed the optimization levels as well as setting target
3926 options, start with the optimizations specified. */
3927 if (func_optimize && func_optimize != old_optimize)
3928 cl_optimization_restore (TREE_OPTIMIZATION (func_optimize));
3930 /* The target attributes may also change some optimization flags, so update
3931 the optimization options if necessary. */
3932 cl_target_option_save (&cur_target);
3933 new_target = ix86_valid_target_attribute_tree (args);
3934 new_optimize = build_optimization_node ();
3941 DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = new_target;
3943 if (old_optimize != new_optimize)
3944 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize;
3947 cl_target_option_restore (&cur_target);
3949 if (old_optimize != new_optimize)
3950 cl_optimization_restore (TREE_OPTIMIZATION (old_optimize));
3956 /* Hook to determine if one function can safely inline another. */
3959 ix86_can_inline_p (tree caller, tree callee)
3962 tree caller_tree = DECL_FUNCTION_SPECIFIC_TARGET (caller);
3963 tree callee_tree = DECL_FUNCTION_SPECIFIC_TARGET (callee);
3965 /* If callee has no option attributes, then it is ok to inline. */
3969 /* If caller has no option attributes, but callee does then it is not ok to
3971 else if (!caller_tree)
3976 struct cl_target_option *caller_opts = TREE_TARGET_OPTION (caller_tree);
3977 struct cl_target_option *callee_opts = TREE_TARGET_OPTION (callee_tree);
3979 /* Callee's isa options should a subset of the caller's, i.e. a SSE4 function
3980 can inline a SSE2 function but a SSE2 function can't inline a SSE4
3982 if ((caller_opts->ix86_isa_flags & callee_opts->ix86_isa_flags)
3983 != callee_opts->ix86_isa_flags)
3986 /* See if we have the same non-isa options. */
3987 else if (caller_opts->target_flags != callee_opts->target_flags)
3990 /* See if arch, tune, etc. are the same. */
3991 else if (caller_opts->arch != callee_opts->arch)
3994 else if (caller_opts->tune != callee_opts->tune)
3997 else if (caller_opts->fpmath != callee_opts->fpmath)
4000 else if (caller_opts->branch_cost != callee_opts->branch_cost)
4011 /* Remember the last target of ix86_set_current_function. */
4012 static GTY(()) tree ix86_previous_fndecl;
4014 /* Establish appropriate back-end context for processing the function
4015 FNDECL. The argument might be NULL to indicate processing at top
4016 level, outside of any function scope. */
4018 ix86_set_current_function (tree fndecl)
4020 /* Only change the context if the function changes. This hook is called
4021 several times in the course of compiling a function, and we don't want to
4022 slow things down too much or call target_reinit when it isn't safe. */
4023 if (fndecl && fndecl != ix86_previous_fndecl)
4025 tree old_tree = (ix86_previous_fndecl
4026 ? DECL_FUNCTION_SPECIFIC_TARGET (ix86_previous_fndecl)
4029 tree new_tree = (fndecl
4030 ? DECL_FUNCTION_SPECIFIC_TARGET (fndecl)
4033 ix86_previous_fndecl = fndecl;
4034 if (old_tree == new_tree)
4039 cl_target_option_restore (TREE_TARGET_OPTION (new_tree));
4045 struct cl_target_option *def
4046 = TREE_TARGET_OPTION (target_option_current_node);
4048 cl_target_option_restore (def);
4055 /* Return true if this goes in large data/bss. */
4058 ix86_in_large_data_p (tree exp)
4060 if (ix86_cmodel != CM_MEDIUM && ix86_cmodel != CM_MEDIUM_PIC)
4063 /* Functions are never large data. */
4064 if (TREE_CODE (exp) == FUNCTION_DECL)
4067 if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
4069 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));
4070 if (strcmp (section, ".ldata") == 0
4071 || strcmp (section, ".lbss") == 0)
4077 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
4079 /* If this is an incomplete type with size 0, then we can't put it
4080 in data because it might be too big when completed. */
4081 if (!size || size > ix86_section_threshold)
4088 /* Switch to the appropriate section for output of DECL.
4089 DECL is either a `VAR_DECL' node or a constant of some sort.
4090 RELOC indicates whether forming the initial value of DECL requires
4091 link-time relocations. */
4093 static section * x86_64_elf_select_section (tree, int, unsigned HOST_WIDE_INT)
4097 x86_64_elf_select_section (tree decl, int reloc,
4098 unsigned HOST_WIDE_INT align)
4100 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
4101 && ix86_in_large_data_p (decl))
4103 const char *sname = NULL;
4104 unsigned int flags = SECTION_WRITE;
4105 switch (categorize_decl_for_section (decl, reloc))
4110 case SECCAT_DATA_REL:
4111 sname = ".ldata.rel";
4113 case SECCAT_DATA_REL_LOCAL:
4114 sname = ".ldata.rel.local";
4116 case SECCAT_DATA_REL_RO:
4117 sname = ".ldata.rel.ro";
4119 case SECCAT_DATA_REL_RO_LOCAL:
4120 sname = ".ldata.rel.ro.local";
4124 flags |= SECTION_BSS;
4127 case SECCAT_RODATA_MERGE_STR:
4128 case SECCAT_RODATA_MERGE_STR_INIT:
4129 case SECCAT_RODATA_MERGE_CONST:
4133 case SECCAT_SRODATA:
4140 /* We don't split these for medium model. Place them into
4141 default sections and hope for best. */
4143 case SECCAT_EMUTLS_VAR:
4144 case SECCAT_EMUTLS_TMPL:
4149 /* We might get called with string constants, but get_named_section
4150 doesn't like them as they are not DECLs. Also, we need to set
4151 flags in that case. */
4153 return get_section (sname, flags, NULL);
4154 return get_named_section (decl, sname, reloc);
4157 return default_elf_select_section (decl, reloc, align);
4160 /* Build up a unique section name, expressed as a
4161 STRING_CST node, and assign it to DECL_SECTION_NAME (decl).
4162 RELOC indicates whether the initial value of EXP requires
4163 link-time relocations. */
4165 static void ATTRIBUTE_UNUSED
4166 x86_64_elf_unique_section (tree decl, int reloc)
4168 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
4169 && ix86_in_large_data_p (decl))
4171 const char *prefix = NULL;
4172 /* We only need to use .gnu.linkonce if we don't have COMDAT groups. */
4173 bool one_only = DECL_ONE_ONLY (decl) && !HAVE_COMDAT_GROUP;
4175 switch (categorize_decl_for_section (decl, reloc))
4178 case SECCAT_DATA_REL:
4179 case SECCAT_DATA_REL_LOCAL:
4180 case SECCAT_DATA_REL_RO:
4181 case SECCAT_DATA_REL_RO_LOCAL:
4182 prefix = one_only ? ".ld" : ".ldata";
4185 prefix = one_only ? ".lb" : ".lbss";
4188 case SECCAT_RODATA_MERGE_STR:
4189 case SECCAT_RODATA_MERGE_STR_INIT:
4190 case SECCAT_RODATA_MERGE_CONST:
4191 prefix = one_only ? ".lr" : ".lrodata";
4193 case SECCAT_SRODATA:
4200 /* We don't split these for medium model. Place them into
4201 default sections and hope for best. */
4203 case SECCAT_EMUTLS_VAR:
4204 prefix = targetm.emutls.var_section;
4206 case SECCAT_EMUTLS_TMPL:
4207 prefix = targetm.emutls.tmpl_section;
4212 const char *name, *linkonce;
4215 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
4216 name = targetm.strip_name_encoding (name);
4218 /* If we're using one_only, then there needs to be a .gnu.linkonce
4219 prefix to the section name. */
4220 linkonce = one_only ? ".gnu.linkonce" : "";
4222 string = ACONCAT ((linkonce, prefix, ".", name, NULL));
4224 DECL_SECTION_NAME (decl) = build_string (strlen (string), string);
4228 default_unique_section (decl, reloc);
4231 #ifdef COMMON_ASM_OP
4232 /* This says how to output assembler code to declare an
4233 uninitialized external linkage data object.
4235 For medium model x86-64 we need to use .largecomm opcode for
4238 x86_elf_aligned_common (FILE *file,
4239 const char *name, unsigned HOST_WIDE_INT size,
4242 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
4243 && size > (unsigned int)ix86_section_threshold)
4244 fputs (".largecomm\t", file);
4246 fputs (COMMON_ASM_OP, file);
4247 assemble_name (file, name);
4248 fprintf (file, "," HOST_WIDE_INT_PRINT_UNSIGNED ",%u\n",
4249 size, align / BITS_PER_UNIT);
4253 /* Utility function for targets to use in implementing
4254 ASM_OUTPUT_ALIGNED_BSS. */
4257 x86_output_aligned_bss (FILE *file, tree decl ATTRIBUTE_UNUSED,
4258 const char *name, unsigned HOST_WIDE_INT size,
4261 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
4262 && size > (unsigned int)ix86_section_threshold)
4263 switch_to_section (get_named_section (decl, ".lbss", 0));
4265 switch_to_section (bss_section);
4266 ASM_OUTPUT_ALIGN (file, floor_log2 (align / BITS_PER_UNIT));
4267 #ifdef ASM_DECLARE_OBJECT_NAME
4268 last_assemble_variable_decl = decl;
4269 ASM_DECLARE_OBJECT_NAME (file, name, decl);
4271 /* Standard thing is just output label for the object. */
4272 ASM_OUTPUT_LABEL (file, name);
4273 #endif /* ASM_DECLARE_OBJECT_NAME */
4274 ASM_OUTPUT_SKIP (file, size ? size : 1);
4278 optimization_options (int level, int size ATTRIBUTE_UNUSED)
4280 /* For -O2 and beyond, turn off -fschedule-insns by default. It tends to
4281 make the problem with not enough registers even worse. */
4282 #ifdef INSN_SCHEDULING
4284 flag_schedule_insns = 0;
4288 /* The Darwin libraries never set errno, so we might as well
4289 avoid calling them when that's the only reason we would. */
4290 flag_errno_math = 0;
4292 /* The default values of these switches depend on the TARGET_64BIT
4293 that is not known at this moment. Mark these values with 2 and
4294 let user the to override these. In case there is no command line option
4295 specifying them, we will set the defaults in override_options. */
4297 flag_omit_frame_pointer = 2;
4298 flag_pcc_struct_return = 2;
4299 flag_asynchronous_unwind_tables = 2;
4300 flag_vect_cost_model = 1;
4301 #ifdef SUBTARGET_OPTIMIZATION_OPTIONS
4302 SUBTARGET_OPTIMIZATION_OPTIONS;
4306 /* Decide whether we can make a sibling call to a function. DECL is the
4307 declaration of the function being targeted by the call and EXP is the
4308 CALL_EXPR representing the call. */
4311 ix86_function_ok_for_sibcall (tree decl, tree exp)
4313 tree type, decl_or_type;
4316 /* If we are generating position-independent code, we cannot sibcall
4317 optimize any indirect call, or a direct call to a global function,
4318 as the PLT requires %ebx be live. */
4319 if (!TARGET_64BIT && flag_pic && (!decl || !targetm.binds_local_p (decl)))
4322 /* If we need to align the outgoing stack, then sibcalling would
4323 unalign the stack, which may break the called function. */
4324 if (ix86_minimum_incoming_stack_boundary (true)
4325 < PREFERRED_STACK_BOUNDARY)
4330 decl_or_type = decl;
4331 type = TREE_TYPE (decl);
4335 /* We're looking at the CALL_EXPR, we need the type of the function. */
4336 type = CALL_EXPR_FN (exp); /* pointer expression */
4337 type = TREE_TYPE (type); /* pointer type */
4338 type = TREE_TYPE (type); /* function type */
4339 decl_or_type = type;
4342 /* Check that the return value locations are the same. Like
4343 if we are returning floats on the 80387 register stack, we cannot
4344 make a sibcall from a function that doesn't return a float to a
4345 function that does or, conversely, from a function that does return
4346 a float to a function that doesn't; the necessary stack adjustment
4347 would not be executed. This is also the place we notice
4348 differences in the return value ABI. Note that it is ok for one
4349 of the functions to have void return type as long as the return
4350 value of the other is passed in a register. */
4351 a = ix86_function_value (TREE_TYPE (exp), decl_or_type, false);
4352 b = ix86_function_value (TREE_TYPE (DECL_RESULT (cfun->decl)),
4354 if (STACK_REG_P (a) || STACK_REG_P (b))
4356 if (!rtx_equal_p (a, b))
4359 else if (VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun->decl))))
4361 else if (!rtx_equal_p (a, b))
4366 /* The SYSV ABI has more call-clobbered registers;
4367 disallow sibcalls from MS to SYSV. */
4368 if (cfun->machine->call_abi == MS_ABI
4369 && ix86_function_type_abi (type) == SYSV_ABI)
4374 /* If this call is indirect, we'll need to be able to use a
4375 call-clobbered register for the address of the target function.
4376 Make sure that all such registers are not used for passing
4377 parameters. Note that DLLIMPORT functions are indirect. */
4379 || (TARGET_DLLIMPORT_DECL_ATTRIBUTES && DECL_DLLIMPORT_P (decl)))
4381 if (ix86_function_regparm (type, NULL) >= 3)
4383 /* ??? Need to count the actual number of registers to be used,
4384 not the possible number of registers. Fix later. */
4390 /* Otherwise okay. That also includes certain types of indirect calls. */
4394 /* Handle "cdecl", "stdcall", "fastcall", "regparm" and "sseregparm"
4395 calling convention attributes;
4396 arguments as in struct attribute_spec.handler. */
4399 ix86_handle_cconv_attribute (tree *node, tree name,
4401 int flags ATTRIBUTE_UNUSED,
4404 if (TREE_CODE (*node) != FUNCTION_TYPE
4405 && TREE_CODE (*node) != METHOD_TYPE
4406 && TREE_CODE (*node) != FIELD_DECL
4407 && TREE_CODE (*node) != TYPE_DECL)
4409 warning (OPT_Wattributes, "%qE attribute only applies to functions",
4411 *no_add_attrs = true;
4415 /* Can combine regparm with all attributes but fastcall. */
4416 if (is_attribute_p ("regparm", name))
4420 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4422 error ("fastcall and regparm attributes are not compatible");
4425 cst = TREE_VALUE (args);
4426 if (TREE_CODE (cst) != INTEGER_CST)
4428 warning (OPT_Wattributes,
4429 "%qE attribute requires an integer constant argument",
4431 *no_add_attrs = true;
4433 else if (compare_tree_int (cst, REGPARM_MAX) > 0)
4435 warning (OPT_Wattributes, "argument to %qE attribute larger than %d",
4437 *no_add_attrs = true;
4445 /* Do not warn when emulating the MS ABI. */
4446 if (TREE_CODE (*node) != FUNCTION_TYPE
4447 || ix86_function_type_abi (*node) != MS_ABI)
4448 warning (OPT_Wattributes, "%qE attribute ignored",
4450 *no_add_attrs = true;
4454 /* Can combine fastcall with stdcall (redundant) and sseregparm. */
4455 if (is_attribute_p ("fastcall", name))
4457 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
4459 error ("fastcall and cdecl attributes are not compatible");
4461 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
4463 error ("fastcall and stdcall attributes are not compatible");
4465 if (lookup_attribute ("regparm", TYPE_ATTRIBUTES (*node)))
4467 error ("fastcall and regparm attributes are not compatible");
4471 /* Can combine stdcall with fastcall (redundant), regparm and
4473 else if (is_attribute_p ("stdcall", name))
4475 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
4477 error ("stdcall and cdecl attributes are not compatible");
4479 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4481 error ("stdcall and fastcall attributes are not compatible");
4485 /* Can combine cdecl with regparm and sseregparm. */
4486 else if (is_attribute_p ("cdecl", name))
4488 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
4490 error ("stdcall and cdecl attributes are not compatible");
4492 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4494 error ("fastcall and cdecl attributes are not compatible");
4498 /* Can combine sseregparm with all attributes. */
4503 /* Return 0 if the attributes for two types are incompatible, 1 if they
4504 are compatible, and 2 if they are nearly compatible (which causes a
4505 warning to be generated). */
4508 ix86_comp_type_attributes (const_tree type1, const_tree type2)
4510 /* Check for mismatch of non-default calling convention. */
4511 const char *const rtdstr = TARGET_RTD ? "cdecl" : "stdcall";
4513 if (TREE_CODE (type1) != FUNCTION_TYPE
4514 && TREE_CODE (type1) != METHOD_TYPE)
4517 /* Check for mismatched fastcall/regparm types. */
4518 if ((!lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type1))
4519 != !lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type2)))
4520 || (ix86_function_regparm (type1, NULL)
4521 != ix86_function_regparm (type2, NULL)))
4524 /* Check for mismatched sseregparm types. */
4525 if (!lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type1))
4526 != !lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type2)))
4529 /* Check for mismatched return types (cdecl vs stdcall). */
4530 if (!lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type1))
4531 != !lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type2)))
4537 /* Return the regparm value for a function with the indicated TYPE and DECL.
4538 DECL may be NULL when calling function indirectly
4539 or considering a libcall. */
4542 ix86_function_regparm (const_tree type, const_tree decl)
4548 return (ix86_function_type_abi (type) == SYSV_ABI
4549 ? X86_64_REGPARM_MAX : X86_64_MS_REGPARM_MAX);
4551 regparm = ix86_regparm;
4552 attr = lookup_attribute ("regparm", TYPE_ATTRIBUTES (type));
4555 regparm = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attr)));
4559 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type)))
4562 /* Use register calling convention for local functions when possible. */
4564 && TREE_CODE (decl) == FUNCTION_DECL
4568 /* FIXME: remove this CONST_CAST when cgraph.[ch] is constified. */
4569 struct cgraph_local_info *i = cgraph_local_info (CONST_CAST_TREE (decl));
4572 int local_regparm, globals = 0, regno;
4574 /* Make sure no regparm register is taken by a
4575 fixed register variable. */
4576 for (local_regparm = 0; local_regparm < REGPARM_MAX; local_regparm++)
4577 if (fixed_regs[local_regparm])
4580 /* We don't want to use regparm(3) for nested functions as
4581 these use a static chain pointer in the third argument. */
4582 if (local_regparm == 3 && DECL_STATIC_CHAIN (decl))
4585 /* Each fixed register usage increases register pressure,
4586 so less registers should be used for argument passing.
4587 This functionality can be overriden by an explicit
4589 for (regno = 0; regno <= DI_REG; regno++)
4590 if (fixed_regs[regno])
4594 = globals < local_regparm ? local_regparm - globals : 0;
4596 if (local_regparm > regparm)
4597 regparm = local_regparm;
4604 /* Return 1 or 2, if we can pass up to SSE_REGPARM_MAX SFmode (1) and
4605 DFmode (2) arguments in SSE registers for a function with the
4606 indicated TYPE and DECL. DECL may be NULL when calling function
4607 indirectly or considering a libcall. Otherwise return 0. */
4610 ix86_function_sseregparm (const_tree type, const_tree decl, bool warn)
4612 gcc_assert (!TARGET_64BIT);
4614 /* Use SSE registers to pass SFmode and DFmode arguments if requested
4615 by the sseregparm attribute. */
4616 if (TARGET_SSEREGPARM
4617 || (type && lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type))))
4624 error ("Calling %qD with attribute sseregparm without "
4625 "SSE/SSE2 enabled", decl);
4627 error ("Calling %qT with attribute sseregparm without "
4628 "SSE/SSE2 enabled", type);
4636 /* For local functions, pass up to SSE_REGPARM_MAX SFmode
4637 (and DFmode for SSE2) arguments in SSE registers. */
4638 if (decl && TARGET_SSE_MATH && optimize && !profile_flag)
4640 /* FIXME: remove this CONST_CAST when cgraph.[ch] is constified. */
4641 struct cgraph_local_info *i = cgraph_local_info (CONST_CAST_TREE(decl));
4643 return TARGET_SSE2 ? 2 : 1;
4649 /* Return true if EAX is live at the start of the function. Used by
4650 ix86_expand_prologue to determine if we need special help before
4651 calling allocate_stack_worker. */
4654 ix86_eax_live_at_start_p (void)
4656 /* Cheat. Don't bother working forward from ix86_function_regparm
4657 to the function type to whether an actual argument is located in
4658 eax. Instead just look at cfg info, which is still close enough
4659 to correct at this point. This gives false positives for broken
4660 functions that might use uninitialized data that happens to be
4661 allocated in eax, but who cares? */
4662 return REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR), 0);
4665 /* Value is the number of bytes of arguments automatically
4666 popped when returning from a subroutine call.
4667 FUNDECL is the declaration node of the function (as a tree),
4668 FUNTYPE is the data type of the function (as a tree),
4669 or for a library call it is an identifier node for the subroutine name.
4670 SIZE is the number of bytes of arguments passed on the stack.
4672 On the 80386, the RTD insn may be used to pop them if the number
4673 of args is fixed, but if the number is variable then the caller
4674 must pop them all. RTD can't be used for library calls now
4675 because the library is compiled with the Unix compiler.
4676 Use of RTD is a selectable option, since it is incompatible with
4677 standard Unix calling sequences. If the option is not selected,
4678 the caller must always pop the args.
4680 The attribute stdcall is equivalent to RTD on a per module basis. */
4683 ix86_return_pops_args (tree fundecl, tree funtype, int size)
4687 /* None of the 64-bit ABIs pop arguments. */
4691 rtd = TARGET_RTD && (!fundecl || TREE_CODE (fundecl) != IDENTIFIER_NODE);
4693 /* Cdecl functions override -mrtd, and never pop the stack. */
4694 if (! lookup_attribute ("cdecl", TYPE_ATTRIBUTES (funtype)))
4696 /* Stdcall and fastcall functions will pop the stack if not
4698 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (funtype))
4699 || lookup_attribute ("fastcall", TYPE_ATTRIBUTES (funtype)))
4702 if (rtd && ! stdarg_p (funtype))
4706 /* Lose any fake structure return argument if it is passed on the stack. */
4707 if (aggregate_value_p (TREE_TYPE (funtype), fundecl)
4708 && !KEEP_AGGREGATE_RETURN_POINTER)
4710 int nregs = ix86_function_regparm (funtype, fundecl);
4712 return GET_MODE_SIZE (Pmode);
4718 /* Argument support functions. */
4720 /* Return true when register may be used to pass function parameters. */
4722 ix86_function_arg_regno_p (int regno)
4725 const int *parm_regs;
4730 return (regno < REGPARM_MAX
4731 || (TARGET_SSE && SSE_REGNO_P (regno) && !fixed_regs[regno]));
4733 return (regno < REGPARM_MAX
4734 || (TARGET_MMX && MMX_REGNO_P (regno)
4735 && (regno < FIRST_MMX_REG + MMX_REGPARM_MAX))
4736 || (TARGET_SSE && SSE_REGNO_P (regno)
4737 && (regno < FIRST_SSE_REG + SSE_REGPARM_MAX)));
4742 if (SSE_REGNO_P (regno) && TARGET_SSE)
4747 if (TARGET_SSE && SSE_REGNO_P (regno)
4748 && (regno < FIRST_SSE_REG + SSE_REGPARM_MAX))
4752 /* TODO: The function should depend on current function ABI but
4753 builtins.c would need updating then. Therefore we use the
4756 /* RAX is used as hidden argument to va_arg functions. */
4757 if (ix86_abi == SYSV_ABI && regno == AX_REG)
4760 if (ix86_abi == MS_ABI)
4761 parm_regs = x86_64_ms_abi_int_parameter_registers;
4763 parm_regs = x86_64_int_parameter_registers;
4764 for (i = 0; i < (ix86_abi == MS_ABI
4765 ? X86_64_MS_REGPARM_MAX : X86_64_REGPARM_MAX); i++)
4766 if (regno == parm_regs[i])
4771 /* Return if we do not know how to pass TYPE solely in registers. */
4774 ix86_must_pass_in_stack (enum machine_mode mode, const_tree type)
4776 if (must_pass_in_stack_var_size_or_pad (mode, type))
4779 /* For 32-bit, we want TImode aggregates to go on the stack. But watch out!
4780 The layout_type routine is crafty and tries to trick us into passing
4781 currently unsupported vector types on the stack by using TImode. */
4782 return (!TARGET_64BIT && mode == TImode
4783 && type && TREE_CODE (type) != VECTOR_TYPE);
4786 /* It returns the size, in bytes, of the area reserved for arguments passed
4787 in registers for the function represented by fndecl dependent to the used
4790 ix86_reg_parm_stack_space (const_tree fndecl)
4792 enum calling_abi call_abi = SYSV_ABI;
4793 if (fndecl != NULL_TREE && TREE_CODE (fndecl) == FUNCTION_DECL)
4794 call_abi = ix86_function_abi (fndecl);
4796 call_abi = ix86_function_type_abi (fndecl);
4797 if (call_abi == MS_ABI)
4802 /* Returns value SYSV_ABI, MS_ABI dependent on fntype, specifying the
4805 ix86_function_type_abi (const_tree fntype)
4807 if (TARGET_64BIT && fntype != NULL)
4809 enum calling_abi abi = ix86_abi;
4810 if (abi == SYSV_ABI)
4812 if (lookup_attribute ("ms_abi", TYPE_ATTRIBUTES (fntype)))
4815 else if (lookup_attribute ("sysv_abi", TYPE_ATTRIBUTES (fntype)))
4823 ix86_function_ms_hook_prologue (const_tree fntype)
4827 if (lookup_attribute ("ms_hook_prologue", DECL_ATTRIBUTES (fntype)))
4829 if (decl_function_context (fntype) != NULL_TREE)
4831 error_at (DECL_SOURCE_LOCATION (fntype),
4832 "ms_hook_prologue is not compatible with nested function");
4841 static enum calling_abi
4842 ix86_function_abi (const_tree fndecl)
4846 return ix86_function_type_abi (TREE_TYPE (fndecl));
4849 /* Returns value SYSV_ABI, MS_ABI dependent on cfun, specifying the
4852 ix86_cfun_abi (void)
4854 if (! cfun || ! TARGET_64BIT)
4856 return cfun->machine->call_abi;
4860 extern void init_regs (void);
4862 /* Implementation of call abi switching target hook. Specific to FNDECL
4863 the specific call register sets are set. See also CONDITIONAL_REGISTER_USAGE
4864 for more details. */
4866 ix86_call_abi_override (const_tree fndecl)
4868 if (fndecl == NULL_TREE)
4869 cfun->machine->call_abi = ix86_abi;
4871 cfun->machine->call_abi = ix86_function_type_abi (TREE_TYPE (fndecl));
4874 /* MS and SYSV ABI have different set of call used registers. Avoid expensive
4875 re-initialization of init_regs each time we switch function context since
4876 this is needed only during RTL expansion. */
4878 ix86_maybe_switch_abi (void)
4881 call_used_regs[SI_REG] == (cfun->machine->call_abi == MS_ABI))
4885 /* Initialize a variable CUM of type CUMULATIVE_ARGS
4886 for a call to a function whose data type is FNTYPE.
4887 For a library call, FNTYPE is 0. */
4890 init_cumulative_args (CUMULATIVE_ARGS *cum, /* Argument info to initialize */
4891 tree fntype, /* tree ptr for function decl */
4892 rtx libname, /* SYMBOL_REF of library name or 0 */
4895 struct cgraph_local_info *i = fndecl ? cgraph_local_info (fndecl) : NULL;
4896 memset (cum, 0, sizeof (*cum));
4899 cum->call_abi = ix86_function_abi (fndecl);
4901 cum->call_abi = ix86_function_type_abi (fntype);
4902 /* Set up the number of registers to use for passing arguments. */
4904 if (cum->call_abi == MS_ABI && !ACCUMULATE_OUTGOING_ARGS)
4905 sorry ("ms_abi attribute requires -maccumulate-outgoing-args "
4906 "or subtarget optimization implying it");
4907 cum->nregs = ix86_regparm;
4910 if (cum->call_abi != ix86_abi)
4911 cum->nregs = (ix86_abi != SYSV_ABI
4912 ? X86_64_REGPARM_MAX : X86_64_MS_REGPARM_MAX);
4916 cum->sse_nregs = SSE_REGPARM_MAX;
4919 if (cum->call_abi != ix86_abi)
4920 cum->sse_nregs = (ix86_abi != SYSV_ABI
4921 ? X86_64_SSE_REGPARM_MAX
4922 : X86_64_MS_SSE_REGPARM_MAX);
4926 cum->mmx_nregs = MMX_REGPARM_MAX;
4927 cum->warn_avx = true;
4928 cum->warn_sse = true;
4929 cum->warn_mmx = true;
4931 /* Because type might mismatch in between caller and callee, we need to
4932 use actual type of function for local calls.
4933 FIXME: cgraph_analyze can be told to actually record if function uses
4934 va_start so for local functions maybe_vaarg can be made aggressive
4936 FIXME: once typesytem is fixed, we won't need this code anymore. */
4938 fntype = TREE_TYPE (fndecl);
4939 cum->maybe_vaarg = (fntype
4940 ? (!prototype_p (fntype) || stdarg_p (fntype))
4945 /* If there are variable arguments, then we won't pass anything
4946 in registers in 32-bit mode. */
4947 if (stdarg_p (fntype))
4958 /* Use ecx and edx registers if function has fastcall attribute,
4959 else look for regparm information. */
4962 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (fntype)))
4968 cum->nregs = ix86_function_regparm (fntype, fndecl);
4971 /* Set up the number of SSE registers used for passing SFmode
4972 and DFmode arguments. Warn for mismatching ABI. */
4973 cum->float_in_sse = ix86_function_sseregparm (fntype, fndecl, true);
4977 /* Return the "natural" mode for TYPE. In most cases, this is just TYPE_MODE.
4978 But in the case of vector types, it is some vector mode.
4980 When we have only some of our vector isa extensions enabled, then there
4981 are some modes for which vector_mode_supported_p is false. For these
4982 modes, the generic vector support in gcc will choose some non-vector mode
4983 in order to implement the type. By computing the natural mode, we'll
4984 select the proper ABI location for the operand and not depend on whatever
4985 the middle-end decides to do with these vector types.
4987 The midde-end can't deal with the vector types > 16 bytes. In this
4988 case, we return the original mode and warn ABI change if CUM isn't
4991 static enum machine_mode
4992 type_natural_mode (const_tree type, CUMULATIVE_ARGS *cum)
4994 enum machine_mode mode = TYPE_MODE (type);
4996 if (TREE_CODE (type) == VECTOR_TYPE && !VECTOR_MODE_P (mode))
4998 HOST_WIDE_INT size = int_size_in_bytes (type);
4999 if ((size == 8 || size == 16 || size == 32)
5000 /* ??? Generic code allows us to create width 1 vectors. Ignore. */
5001 && TYPE_VECTOR_SUBPARTS (type) > 1)
5003 enum machine_mode innermode = TYPE_MODE (TREE_TYPE (type));
5005 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
5006 mode = MIN_MODE_VECTOR_FLOAT;
5008 mode = MIN_MODE_VECTOR_INT;
5010 /* Get the mode which has this inner mode and number of units. */
5011 for (; mode != VOIDmode; mode = GET_MODE_WIDER_MODE (mode))
5012 if (GET_MODE_NUNITS (mode) == TYPE_VECTOR_SUBPARTS (type)
5013 && GET_MODE_INNER (mode) == innermode)
5015 if (size == 32 && !TARGET_AVX)
5017 static bool warnedavx;
5024 warning (0, "AVX vector argument without AVX "
5025 "enabled changes the ABI");
5027 return TYPE_MODE (type);
5040 /* We want to pass a value in REGNO whose "natural" mode is MODE. However,
5041 this may not agree with the mode that the type system has chosen for the
5042 register, which is ORIG_MODE. If ORIG_MODE is not BLKmode, then we can
5043 go ahead and use it. Otherwise we have to build a PARALLEL instead. */
5046 gen_reg_or_parallel (enum machine_mode mode, enum machine_mode orig_mode,
5051 if (orig_mode != BLKmode)
5052 tmp = gen_rtx_REG (orig_mode, regno);
5055 tmp = gen_rtx_REG (mode, regno);
5056 tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, const0_rtx);
5057 tmp = gen_rtx_PARALLEL (orig_mode, gen_rtvec (1, tmp));
5063 /* x86-64 register passing implementation. See x86-64 ABI for details. Goal
5064 of this code is to classify each 8bytes of incoming argument by the register
5065 class and assign registers accordingly. */
5067 /* Return the union class of CLASS1 and CLASS2.
5068 See the x86-64 PS ABI for details. */
5070 static enum x86_64_reg_class
5071 merge_classes (enum x86_64_reg_class class1, enum x86_64_reg_class class2)
5073 /* Rule #1: If both classes are equal, this is the resulting class. */
5074 if (class1 == class2)
5077 /* Rule #2: If one of the classes is NO_CLASS, the resulting class is
5079 if (class1 == X86_64_NO_CLASS)
5081 if (class2 == X86_64_NO_CLASS)
5084 /* Rule #3: If one of the classes is MEMORY, the result is MEMORY. */
5085 if (class1 == X86_64_MEMORY_CLASS || class2 == X86_64_MEMORY_CLASS)
5086 return X86_64_MEMORY_CLASS;
5088 /* Rule #4: If one of the classes is INTEGER, the result is INTEGER. */
5089 if ((class1 == X86_64_INTEGERSI_CLASS && class2 == X86_64_SSESF_CLASS)
5090 || (class2 == X86_64_INTEGERSI_CLASS && class1 == X86_64_SSESF_CLASS))
5091 return X86_64_INTEGERSI_CLASS;
5092 if (class1 == X86_64_INTEGER_CLASS || class1 == X86_64_INTEGERSI_CLASS
5093 || class2 == X86_64_INTEGER_CLASS || class2 == X86_64_INTEGERSI_CLASS)
5094 return X86_64_INTEGER_CLASS;
5096 /* Rule #5: If one of the classes is X87, X87UP, or COMPLEX_X87 class,
5098 if (class1 == X86_64_X87_CLASS
5099 || class1 == X86_64_X87UP_CLASS
5100 || class1 == X86_64_COMPLEX_X87_CLASS
5101 || class2 == X86_64_X87_CLASS
5102 || class2 == X86_64_X87UP_CLASS
5103 || class2 == X86_64_COMPLEX_X87_CLASS)
5104 return X86_64_MEMORY_CLASS;
5106 /* Rule #6: Otherwise class SSE is used. */
5107 return X86_64_SSE_CLASS;
5110 /* Classify the argument of type TYPE and mode MODE.
5111 CLASSES will be filled by the register class used to pass each word
5112 of the operand. The number of words is returned. In case the parameter
5113 should be passed in memory, 0 is returned. As a special case for zero
5114 sized containers, classes[0] will be NO_CLASS and 1 is returned.
5116 BIT_OFFSET is used internally for handling records and specifies offset
5117 of the offset in bits modulo 256 to avoid overflow cases.
5119 See the x86-64 PS ABI for details.
5123 classify_argument (enum machine_mode mode, const_tree type,
5124 enum x86_64_reg_class classes[MAX_CLASSES], int bit_offset)
5126 HOST_WIDE_INT bytes =
5127 (mode == BLKmode) ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode);
5128 int words = (bytes + (bit_offset % 64) / 8 + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5130 /* Variable sized entities are always passed/returned in memory. */
5134 if (mode != VOIDmode
5135 && targetm.calls.must_pass_in_stack (mode, type))
5138 if (type && AGGREGATE_TYPE_P (type))
5142 enum x86_64_reg_class subclasses[MAX_CLASSES];
5144 /* On x86-64 we pass structures larger than 32 bytes on the stack. */
5148 for (i = 0; i < words; i++)
5149 classes[i] = X86_64_NO_CLASS;
5151 /* Zero sized arrays or structures are NO_CLASS. We return 0 to
5152 signalize memory class, so handle it as special case. */
5155 classes[0] = X86_64_NO_CLASS;
5159 /* Classify each field of record and merge classes. */
5160 switch (TREE_CODE (type))
5163 /* And now merge the fields of structure. */
5164 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5166 if (TREE_CODE (field) == FIELD_DECL)
5170 if (TREE_TYPE (field) == error_mark_node)
5173 /* Bitfields are always classified as integer. Handle them
5174 early, since later code would consider them to be
5175 misaligned integers. */
5176 if (DECL_BIT_FIELD (field))
5178 for (i = (int_bit_position (field) + (bit_offset % 64)) / 8 / 8;
5179 i < ((int_bit_position (field) + (bit_offset % 64))
5180 + tree_low_cst (DECL_SIZE (field), 0)
5183 merge_classes (X86_64_INTEGER_CLASS,
5190 type = TREE_TYPE (field);
5192 /* Flexible array member is ignored. */
5193 if (TYPE_MODE (type) == BLKmode
5194 && TREE_CODE (type) == ARRAY_TYPE
5195 && TYPE_SIZE (type) == NULL_TREE
5196 && TYPE_DOMAIN (type) != NULL_TREE
5197 && (TYPE_MAX_VALUE (TYPE_DOMAIN (type))
5202 if (!warned && warn_psabi)
5205 inform (input_location,
5206 "The ABI of passing struct with"
5207 " a flexible array member has"
5208 " changed in GCC 4.4");
5212 num = classify_argument (TYPE_MODE (type), type,
5214 (int_bit_position (field)
5215 + bit_offset) % 256);
5218 pos = (int_bit_position (field) + (bit_offset % 64)) / 8 / 8;
5219 for (i = 0; i < num && (i + pos) < words; i++)
5221 merge_classes (subclasses[i], classes[i + pos]);
5228 /* Arrays are handled as small records. */
5231 num = classify_argument (TYPE_MODE (TREE_TYPE (type)),
5232 TREE_TYPE (type), subclasses, bit_offset);
5236 /* The partial classes are now full classes. */
5237 if (subclasses[0] == X86_64_SSESF_CLASS && bytes != 4)
5238 subclasses[0] = X86_64_SSE_CLASS;
5239 if (subclasses[0] == X86_64_INTEGERSI_CLASS
5240 && !((bit_offset % 64) == 0 && bytes == 4))
5241 subclasses[0] = X86_64_INTEGER_CLASS;
5243 for (i = 0; i < words; i++)
5244 classes[i] = subclasses[i % num];
5249 case QUAL_UNION_TYPE:
5250 /* Unions are similar to RECORD_TYPE but offset is always 0.
5252 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5254 if (TREE_CODE (field) == FIELD_DECL)
5258 if (TREE_TYPE (field) == error_mark_node)
5261 num = classify_argument (TYPE_MODE (TREE_TYPE (field)),
5262 TREE_TYPE (field), subclasses,
5266 for (i = 0; i < num; i++)
5267 classes[i] = merge_classes (subclasses[i], classes[i]);
5278 /* When size > 16 bytes, if the first one isn't
5279 X86_64_SSE_CLASS or any other ones aren't
5280 X86_64_SSEUP_CLASS, everything should be passed in
5282 if (classes[0] != X86_64_SSE_CLASS)
5285 for (i = 1; i < words; i++)
5286 if (classes[i] != X86_64_SSEUP_CLASS)
5290 /* Final merger cleanup. */
5291 for (i = 0; i < words; i++)
5293 /* If one class is MEMORY, everything should be passed in
5295 if (classes[i] == X86_64_MEMORY_CLASS)
5298 /* The X86_64_SSEUP_CLASS should be always preceded by
5299 X86_64_SSE_CLASS or X86_64_SSEUP_CLASS. */
5300 if (classes[i] == X86_64_SSEUP_CLASS
5301 && classes[i - 1] != X86_64_SSE_CLASS
5302 && classes[i - 1] != X86_64_SSEUP_CLASS)
5304 /* The first one should never be X86_64_SSEUP_CLASS. */
5305 gcc_assert (i != 0);
5306 classes[i] = X86_64_SSE_CLASS;
5309 /* If X86_64_X87UP_CLASS isn't preceded by X86_64_X87_CLASS,
5310 everything should be passed in memory. */
5311 if (classes[i] == X86_64_X87UP_CLASS
5312 && (classes[i - 1] != X86_64_X87_CLASS))
5316 /* The first one should never be X86_64_X87UP_CLASS. */
5317 gcc_assert (i != 0);
5318 if (!warned && warn_psabi)
5321 inform (input_location,
5322 "The ABI of passing union with long double"
5323 " has changed in GCC 4.4");
5331 /* Compute alignment needed. We align all types to natural boundaries with
5332 exception of XFmode that is aligned to 64bits. */
5333 if (mode != VOIDmode && mode != BLKmode)
5335 int mode_alignment = GET_MODE_BITSIZE (mode);
5338 mode_alignment = 128;
5339 else if (mode == XCmode)
5340 mode_alignment = 256;
5341 if (COMPLEX_MODE_P (mode))
5342 mode_alignment /= 2;
5343 /* Misaligned fields are always returned in memory. */
5344 if (bit_offset % mode_alignment)
5348 /* for V1xx modes, just use the base mode */
5349 if (VECTOR_MODE_P (mode) && mode != V1DImode && mode != V1TImode
5350 && GET_MODE_SIZE (GET_MODE_INNER (mode)) == bytes)
5351 mode = GET_MODE_INNER (mode);
5353 /* Classification of atomic types. */
5358 classes[0] = X86_64_SSE_CLASS;
5361 classes[0] = X86_64_SSE_CLASS;
5362 classes[1] = X86_64_SSEUP_CLASS;
5372 int size = (bit_offset % 64)+ (int) GET_MODE_BITSIZE (mode);
5376 classes[0] = X86_64_INTEGERSI_CLASS;
5379 else if (size <= 64)
5381 classes[0] = X86_64_INTEGER_CLASS;
5384 else if (size <= 64+32)
5386 classes[0] = X86_64_INTEGER_CLASS;
5387 classes[1] = X86_64_INTEGERSI_CLASS;
5390 else if (size <= 64+64)
5392 classes[0] = classes[1] = X86_64_INTEGER_CLASS;
5400 classes[0] = classes[1] = X86_64_INTEGER_CLASS;
5404 /* OImode shouldn't be used directly. */
5409 if (!(bit_offset % 64))
5410 classes[0] = X86_64_SSESF_CLASS;
5412 classes[0] = X86_64_SSE_CLASS;
5415 classes[0] = X86_64_SSEDF_CLASS;
5418 classes[0] = X86_64_X87_CLASS;
5419 classes[1] = X86_64_X87UP_CLASS;
5422 classes[0] = X86_64_SSE_CLASS;
5423 classes[1] = X86_64_SSEUP_CLASS;
5426 classes[0] = X86_64_SSE_CLASS;
5427 if (!(bit_offset % 64))
5433 if (!warned && warn_psabi)
5436 inform (input_location,
5437 "The ABI of passing structure with complex float"
5438 " member has changed in GCC 4.4");
5440 classes[1] = X86_64_SSESF_CLASS;
5444 classes[0] = X86_64_SSEDF_CLASS;
5445 classes[1] = X86_64_SSEDF_CLASS;
5448 classes[0] = X86_64_COMPLEX_X87_CLASS;
5451 /* This modes is larger than 16 bytes. */
5459 classes[0] = X86_64_SSE_CLASS;
5460 classes[1] = X86_64_SSEUP_CLASS;
5461 classes[2] = X86_64_SSEUP_CLASS;
5462 classes[3] = X86_64_SSEUP_CLASS;
5470 classes[0] = X86_64_SSE_CLASS;
5471 classes[1] = X86_64_SSEUP_CLASS;
5479 classes[0] = X86_64_SSE_CLASS;
5485 gcc_assert (VECTOR_MODE_P (mode));
5490 gcc_assert (GET_MODE_CLASS (GET_MODE_INNER (mode)) == MODE_INT);
5492 if (bit_offset + GET_MODE_BITSIZE (mode) <= 32)
5493 classes[0] = X86_64_INTEGERSI_CLASS;
5495 classes[0] = X86_64_INTEGER_CLASS;
5496 classes[1] = X86_64_INTEGER_CLASS;
5497 return 1 + (bytes > 8);
5501 /* Examine the argument and return set number of register required in each
5502 class. Return 0 iff parameter should be passed in memory. */
5504 examine_argument (enum machine_mode mode, const_tree type, int in_return,
5505 int *int_nregs, int *sse_nregs)
5507 enum x86_64_reg_class regclass[MAX_CLASSES];
5508 int n = classify_argument (mode, type, regclass, 0);
5514 for (n--; n >= 0; n--)
5515 switch (regclass[n])
5517 case X86_64_INTEGER_CLASS:
5518 case X86_64_INTEGERSI_CLASS:
5521 case X86_64_SSE_CLASS:
5522 case X86_64_SSESF_CLASS:
5523 case X86_64_SSEDF_CLASS:
5526 case X86_64_NO_CLASS:
5527 case X86_64_SSEUP_CLASS:
5529 case X86_64_X87_CLASS:
5530 case X86_64_X87UP_CLASS:
5534 case X86_64_COMPLEX_X87_CLASS:
5535 return in_return ? 2 : 0;
5536 case X86_64_MEMORY_CLASS:
5542 /* Construct container for the argument used by GCC interface. See
5543 FUNCTION_ARG for the detailed description. */
5546 construct_container (enum machine_mode mode, enum machine_mode orig_mode,
5547 const_tree type, int in_return, int nintregs, int nsseregs,
5548 const int *intreg, int sse_regno)
5550 /* The following variables hold the static issued_error state. */
5551 static bool issued_sse_arg_error;
5552 static bool issued_sse_ret_error;
5553 static bool issued_x87_ret_error;
5555 enum machine_mode tmpmode;
5557 (mode == BLKmode) ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode);
5558 enum x86_64_reg_class regclass[MAX_CLASSES];
5562 int needed_sseregs, needed_intregs;
5563 rtx exp[MAX_CLASSES];
5566 n = classify_argument (mode, type, regclass, 0);
5569 if (!examine_argument (mode, type, in_return, &needed_intregs,
5572 if (needed_intregs > nintregs || needed_sseregs > nsseregs)
5575 /* We allowed the user to turn off SSE for kernel mode. Don't crash if
5576 some less clueful developer tries to use floating-point anyway. */
5577 if (needed_sseregs && !TARGET_SSE)
5581 if (!issued_sse_ret_error)
5583 error ("SSE register return with SSE disabled");
5584 issued_sse_ret_error = true;
5587 else if (!issued_sse_arg_error)
5589 error ("SSE register argument with SSE disabled");
5590 issued_sse_arg_error = true;
5595 /* Likewise, error if the ABI requires us to return values in the
5596 x87 registers and the user specified -mno-80387. */
5597 if (!TARGET_80387 && in_return)
5598 for (i = 0; i < n; i++)
5599 if (regclass[i] == X86_64_X87_CLASS
5600 || regclass[i] == X86_64_X87UP_CLASS
5601 || regclass[i] == X86_64_COMPLEX_X87_CLASS)
5603 if (!issued_x87_ret_error)
5605 error ("x87 register return with x87 disabled");
5606 issued_x87_ret_error = true;
5611 /* First construct simple cases. Avoid SCmode, since we want to use
5612 single register to pass this type. */
5613 if (n == 1 && mode != SCmode)
5614 switch (regclass[0])
5616 case X86_64_INTEGER_CLASS:
5617 case X86_64_INTEGERSI_CLASS:
5618 return gen_rtx_REG (mode, intreg[0]);
5619 case X86_64_SSE_CLASS:
5620 case X86_64_SSESF_CLASS:
5621 case X86_64_SSEDF_CLASS:
5622 if (mode != BLKmode)
5623 return gen_reg_or_parallel (mode, orig_mode,
5624 SSE_REGNO (sse_regno));
5626 case X86_64_X87_CLASS:
5627 case X86_64_COMPLEX_X87_CLASS:
5628 return gen_rtx_REG (mode, FIRST_STACK_REG);
5629 case X86_64_NO_CLASS:
5630 /* Zero sized array, struct or class. */
5635 if (n == 2 && regclass[0] == X86_64_SSE_CLASS
5636 && regclass[1] == X86_64_SSEUP_CLASS && mode != BLKmode)
5637 return gen_rtx_REG (mode, SSE_REGNO (sse_regno));
5639 && regclass[0] == X86_64_SSE_CLASS
5640 && regclass[1] == X86_64_SSEUP_CLASS
5641 && regclass[2] == X86_64_SSEUP_CLASS
5642 && regclass[3] == X86_64_SSEUP_CLASS
5644 return gen_rtx_REG (mode, SSE_REGNO (sse_regno));
5647 && regclass[0] == X86_64_X87_CLASS && regclass[1] == X86_64_X87UP_CLASS)
5648 return gen_rtx_REG (XFmode, FIRST_STACK_REG);
5649 if (n == 2 && regclass[0] == X86_64_INTEGER_CLASS
5650 && regclass[1] == X86_64_INTEGER_CLASS
5651 && (mode == CDImode || mode == TImode || mode == TFmode)
5652 && intreg[0] + 1 == intreg[1])
5653 return gen_rtx_REG (mode, intreg[0]);
5655 /* Otherwise figure out the entries of the PARALLEL. */
5656 for (i = 0; i < n; i++)
5660 switch (regclass[i])
5662 case X86_64_NO_CLASS:
5664 case X86_64_INTEGER_CLASS:
5665 case X86_64_INTEGERSI_CLASS:
5666 /* Merge TImodes on aligned occasions here too. */
5667 if (i * 8 + 8 > bytes)
5668 tmpmode = mode_for_size ((bytes - i * 8) * BITS_PER_UNIT, MODE_INT, 0);
5669 else if (regclass[i] == X86_64_INTEGERSI_CLASS)
5673 /* We've requested 24 bytes we don't have mode for. Use DImode. */
5674 if (tmpmode == BLKmode)
5676 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5677 gen_rtx_REG (tmpmode, *intreg),
5681 case X86_64_SSESF_CLASS:
5682 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5683 gen_rtx_REG (SFmode,
5684 SSE_REGNO (sse_regno)),
5688 case X86_64_SSEDF_CLASS:
5689 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5690 gen_rtx_REG (DFmode,
5691 SSE_REGNO (sse_regno)),
5695 case X86_64_SSE_CLASS:
5703 if (i == 0 && regclass[1] == X86_64_SSEUP_CLASS)
5713 && regclass[1] == X86_64_SSEUP_CLASS
5714 && regclass[2] == X86_64_SSEUP_CLASS
5715 && regclass[3] == X86_64_SSEUP_CLASS);
5722 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5723 gen_rtx_REG (tmpmode,
5724 SSE_REGNO (sse_regno)),
5733 /* Empty aligned struct, union or class. */
5737 ret = gen_rtx_PARALLEL (mode, rtvec_alloc (nexps));
5738 for (i = 0; i < nexps; i++)
5739 XVECEXP (ret, 0, i) = exp [i];
5743 /* Update the data in CUM to advance over an argument of mode MODE
5744 and data type TYPE. (TYPE is null for libcalls where that information
5745 may not be available.) */
5748 function_arg_advance_32 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5749 tree type, HOST_WIDE_INT bytes, HOST_WIDE_INT words)
5765 cum->words += words;
5766 cum->nregs -= words;
5767 cum->regno += words;
5769 if (cum->nregs <= 0)
5777 /* OImode shouldn't be used directly. */
5781 if (cum->float_in_sse < 2)
5784 if (cum->float_in_sse < 1)
5801 if (!type || !AGGREGATE_TYPE_P (type))
5803 cum->sse_words += words;
5804 cum->sse_nregs -= 1;
5805 cum->sse_regno += 1;
5806 if (cum->sse_nregs <= 0)
5820 if (!type || !AGGREGATE_TYPE_P (type))
5822 cum->mmx_words += words;
5823 cum->mmx_nregs -= 1;
5824 cum->mmx_regno += 1;
5825 if (cum->mmx_nregs <= 0)
5836 function_arg_advance_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5837 tree type, HOST_WIDE_INT words, int named)
5839 int int_nregs, sse_nregs;
5841 /* Unnamed 256bit vector mode parameters are passed on stack. */
5842 if (!named && VALID_AVX256_REG_MODE (mode))
5845 if (!examine_argument (mode, type, 0, &int_nregs, &sse_nregs))
5846 cum->words += words;
5847 else if (sse_nregs <= cum->sse_nregs && int_nregs <= cum->nregs)
5849 cum->nregs -= int_nregs;
5850 cum->sse_nregs -= sse_nregs;
5851 cum->regno += int_nregs;
5852 cum->sse_regno += sse_nregs;
5855 cum->words += words;
5859 function_arg_advance_ms_64 (CUMULATIVE_ARGS *cum, HOST_WIDE_INT bytes,
5860 HOST_WIDE_INT words)
5862 /* Otherwise, this should be passed indirect. */
5863 gcc_assert (bytes == 1 || bytes == 2 || bytes == 4 || bytes == 8);
5865 cum->words += words;
5874 function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5875 tree type, int named)
5877 HOST_WIDE_INT bytes, words;
5879 if (mode == BLKmode)
5880 bytes = int_size_in_bytes (type);
5882 bytes = GET_MODE_SIZE (mode);
5883 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5886 mode = type_natural_mode (type, NULL);
5888 if (TARGET_64BIT && (cum ? cum->call_abi : ix86_abi) == MS_ABI)
5889 function_arg_advance_ms_64 (cum, bytes, words);
5890 else if (TARGET_64BIT)
5891 function_arg_advance_64 (cum, mode, type, words, named);
5893 function_arg_advance_32 (cum, mode, type, bytes, words);
5896 /* Define where to put the arguments to a function.
5897 Value is zero to push the argument on the stack,
5898 or a hard register in which to store the argument.
5900 MODE is the argument's machine mode.
5901 TYPE is the data type of the argument (as a tree).
5902 This is null for libcalls where that information may
5904 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5905 the preceding args and about the function being called.
5906 NAMED is nonzero if this argument is a named parameter
5907 (otherwise it is an extra parameter matching an ellipsis). */
5910 function_arg_32 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5911 enum machine_mode orig_mode, tree type,
5912 HOST_WIDE_INT bytes, HOST_WIDE_INT words)
5914 static bool warnedsse, warnedmmx;
5916 /* Avoid the AL settings for the Unix64 ABI. */
5917 if (mode == VOIDmode)
5933 if (words <= cum->nregs)
5935 int regno = cum->regno;
5937 /* Fastcall allocates the first two DWORD (SImode) or
5938 smaller arguments to ECX and EDX if it isn't an
5944 || (type && AGGREGATE_TYPE_P (type)))
5947 /* ECX not EAX is the first allocated register. */
5948 if (regno == AX_REG)
5951 return gen_rtx_REG (mode, regno);
5956 if (cum->float_in_sse < 2)
5959 if (cum->float_in_sse < 1)
5963 /* In 32bit, we pass TImode in xmm registers. */
5970 if (!type || !AGGREGATE_TYPE_P (type))
5972 if (!TARGET_SSE && !warnedsse && cum->warn_sse)
5975 warning (0, "SSE vector argument without SSE enabled "
5979 return gen_reg_or_parallel (mode, orig_mode,
5980 cum->sse_regno + FIRST_SSE_REG);
5985 /* OImode shouldn't be used directly. */
5994 if (!type || !AGGREGATE_TYPE_P (type))
5997 return gen_reg_or_parallel (mode, orig_mode,
5998 cum->sse_regno + FIRST_SSE_REG);
6008 if (!type || !AGGREGATE_TYPE_P (type))
6010 if (!TARGET_MMX && !warnedmmx && cum->warn_mmx)
6013 warning (0, "MMX vector argument without MMX enabled "
6017 return gen_reg_or_parallel (mode, orig_mode,
6018 cum->mmx_regno + FIRST_MMX_REG);
6027 function_arg_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
6028 enum machine_mode orig_mode, tree type, int named)
6030 /* Handle a hidden AL argument containing number of registers
6031 for varargs x86-64 functions. */
6032 if (mode == VOIDmode)
6033 return GEN_INT (cum->maybe_vaarg
6034 ? (cum->sse_nregs < 0
6035 ? (cum->call_abi == ix86_abi
6037 : (ix86_abi != SYSV_ABI
6038 ? X86_64_SSE_REGPARM_MAX
6039 : X86_64_MS_SSE_REGPARM_MAX))
6054 /* Unnamed 256bit vector mode parameters are passed on stack. */
6060 return construct_container (mode, orig_mode, type, 0, cum->nregs,
6062 &x86_64_int_parameter_registers [cum->regno],
6067 function_arg_ms_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
6068 enum machine_mode orig_mode, int named,
6069 HOST_WIDE_INT bytes)
6073 /* We need to add clobber for MS_ABI->SYSV ABI calls in expand_call.
6074 We use value of -2 to specify that current function call is MSABI. */
6075 if (mode == VOIDmode)
6076 return GEN_INT (-2);
6078 /* If we've run out of registers, it goes on the stack. */
6079 if (cum->nregs == 0)
6082 regno = x86_64_ms_abi_int_parameter_registers[cum->regno];
6084 /* Only floating point modes are passed in anything but integer regs. */
6085 if (TARGET_SSE && (mode == SFmode || mode == DFmode))
6088 regno = cum->regno + FIRST_SSE_REG;
6093 /* Unnamed floating parameters are passed in both the
6094 SSE and integer registers. */
6095 t1 = gen_rtx_REG (mode, cum->regno + FIRST_SSE_REG);
6096 t2 = gen_rtx_REG (mode, regno);
6097 t1 = gen_rtx_EXPR_LIST (VOIDmode, t1, const0_rtx);
6098 t2 = gen_rtx_EXPR_LIST (VOIDmode, t2, const0_rtx);
6099 return gen_rtx_PARALLEL (mode, gen_rtvec (2, t1, t2));
6102 /* Handle aggregated types passed in register. */
6103 if (orig_mode == BLKmode)
6105 if (bytes > 0 && bytes <= 8)
6106 mode = (bytes > 4 ? DImode : SImode);
6107 if (mode == BLKmode)
6111 return gen_reg_or_parallel (mode, orig_mode, regno);
6115 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode omode,
6116 tree type, int named)
6118 enum machine_mode mode = omode;
6119 HOST_WIDE_INT bytes, words;
6121 if (mode == BLKmode)
6122 bytes = int_size_in_bytes (type);
6124 bytes = GET_MODE_SIZE (mode);
6125 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6127 /* To simplify the code below, represent vector types with a vector mode
6128 even if MMX/SSE are not active. */
6129 if (type && TREE_CODE (type) == VECTOR_TYPE)
6130 mode = type_natural_mode (type, cum);
6132 if (TARGET_64BIT && (cum ? cum->call_abi : ix86_abi) == MS_ABI)
6133 return function_arg_ms_64 (cum, mode, omode, named, bytes);
6134 else if (TARGET_64BIT)
6135 return function_arg_64 (cum, mode, omode, type, named);
6137 return function_arg_32 (cum, mode, omode, type, bytes, words);
6140 /* A C expression that indicates when an argument must be passed by
6141 reference. If nonzero for an argument, a copy of that argument is
6142 made in memory and a pointer to the argument is passed instead of
6143 the argument itself. The pointer is passed in whatever way is
6144 appropriate for passing a pointer to that type. */
6147 ix86_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
6148 enum machine_mode mode ATTRIBUTE_UNUSED,
6149 const_tree type, bool named ATTRIBUTE_UNUSED)
6151 /* See Windows x64 Software Convention. */
6152 if (TARGET_64BIT && (cum ? cum->call_abi : ix86_abi) == MS_ABI)
6154 int msize = (int) GET_MODE_SIZE (mode);
6157 /* Arrays are passed by reference. */
6158 if (TREE_CODE (type) == ARRAY_TYPE)
6161 if (AGGREGATE_TYPE_P (type))
6163 /* Structs/unions of sizes other than 8, 16, 32, or 64 bits
6164 are passed by reference. */
6165 msize = int_size_in_bytes (type);
6169 /* __m128 is passed by reference. */
6171 case 1: case 2: case 4: case 8:
6177 else if (TARGET_64BIT && type && int_size_in_bytes (type) == -1)
6183 /* Return true when TYPE should be 128bit aligned for 32bit argument passing
6186 contains_aligned_value_p (tree type)
6188 enum machine_mode mode = TYPE_MODE (type);
6189 if (((TARGET_SSE && SSE_REG_MODE_P (mode))
6193 && (!TYPE_USER_ALIGN (type) || TYPE_ALIGN (type) > 128))
6195 if (TYPE_ALIGN (type) < 128)
6198 if (AGGREGATE_TYPE_P (type))
6200 /* Walk the aggregates recursively. */
6201 switch (TREE_CODE (type))
6205 case QUAL_UNION_TYPE:
6209 /* Walk all the structure fields. */
6210 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
6212 if (TREE_CODE (field) == FIELD_DECL
6213 && contains_aligned_value_p (TREE_TYPE (field)))
6220 /* Just for use if some languages passes arrays by value. */
6221 if (contains_aligned_value_p (TREE_TYPE (type)))
6232 /* Gives the alignment boundary, in bits, of an argument with the
6233 specified mode and type. */
6236 ix86_function_arg_boundary (enum machine_mode mode, tree type)
6241 /* Since canonical type is used for call, we convert it to
6242 canonical type if needed. */
6243 if (!TYPE_STRUCTURAL_EQUALITY_P (type))
6244 type = TYPE_CANONICAL (type);
6245 align = TYPE_ALIGN (type);
6248 align = GET_MODE_ALIGNMENT (mode);
6249 if (align < PARM_BOUNDARY)
6250 align = PARM_BOUNDARY;
6251 /* In 32bit, only _Decimal128 and __float128 are aligned to their
6252 natural boundaries. */
6253 if (!TARGET_64BIT && mode != TDmode && mode != TFmode)
6255 /* i386 ABI defines all arguments to be 4 byte aligned. We have to
6256 make an exception for SSE modes since these require 128bit
6259 The handling here differs from field_alignment. ICC aligns MMX
6260 arguments to 4 byte boundaries, while structure fields are aligned
6261 to 8 byte boundaries. */
6264 if (!(TARGET_SSE && SSE_REG_MODE_P (mode)))
6265 align = PARM_BOUNDARY;
6269 if (!contains_aligned_value_p (type))
6270 align = PARM_BOUNDARY;
6273 if (align > BIGGEST_ALIGNMENT)
6274 align = BIGGEST_ALIGNMENT;
6278 /* Return true if N is a possible register number of function value. */
6281 ix86_function_value_regno_p (int regno)
6288 case FIRST_FLOAT_REG:
6289 /* TODO: The function should depend on current function ABI but
6290 builtins.c would need updating then. Therefore we use the
6292 if (TARGET_64BIT && ix86_abi == MS_ABI)
6294 return TARGET_FLOAT_RETURNS_IN_80387;
6300 if (TARGET_MACHO || TARGET_64BIT)
6308 /* Define how to find the value returned by a function.
6309 VALTYPE is the data type of the value (as a tree).
6310 If the precise function being called is known, FUNC is its FUNCTION_DECL;
6311 otherwise, FUNC is 0. */
6314 function_value_32 (enum machine_mode orig_mode, enum machine_mode mode,
6315 const_tree fntype, const_tree fn)
6319 /* 8-byte vector modes in %mm0. See ix86_return_in_memory for where
6320 we normally prevent this case when mmx is not available. However
6321 some ABIs may require the result to be returned like DImode. */
6322 if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 8)
6323 regno = TARGET_MMX ? FIRST_MMX_REG : 0;
6325 /* 16-byte vector modes in %xmm0. See ix86_return_in_memory for where
6326 we prevent this case when sse is not available. However some ABIs
6327 may require the result to be returned like integer TImode. */
6328 else if (mode == TImode
6329 || (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 16))
6330 regno = TARGET_SSE ? FIRST_SSE_REG : 0;
6332 /* 32-byte vector modes in %ymm0. */
6333 else if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 32)
6334 regno = TARGET_AVX ? FIRST_SSE_REG : 0;
6336 /* Floating point return values in %st(0) (unless -mno-fp-ret-in-387). */
6337 else if (X87_FLOAT_MODE_P (mode) && TARGET_FLOAT_RETURNS_IN_80387)
6338 regno = FIRST_FLOAT_REG;
6340 /* Most things go in %eax. */
6343 /* Override FP return register with %xmm0 for local functions when
6344 SSE math is enabled or for functions with sseregparm attribute. */
6345 if ((fn || fntype) && (mode == SFmode || mode == DFmode))
6347 int sse_level = ix86_function_sseregparm (fntype, fn, false);
6348 if ((sse_level >= 1 && mode == SFmode)
6349 || (sse_level == 2 && mode == DFmode))
6350 regno = FIRST_SSE_REG;
6353 /* OImode shouldn't be used directly. */
6354 gcc_assert (mode != OImode);
6356 return gen_rtx_REG (orig_mode, regno);
6360 function_value_64 (enum machine_mode orig_mode, enum machine_mode mode,
6365 /* Handle libcalls, which don't provide a type node. */
6366 if (valtype == NULL)
6378 return gen_rtx_REG (mode, FIRST_SSE_REG);
6381 return gen_rtx_REG (mode, FIRST_FLOAT_REG);
6385 return gen_rtx_REG (mode, AX_REG);
6389 ret = construct_container (mode, orig_mode, valtype, 1,
6390 X86_64_REGPARM_MAX, X86_64_SSE_REGPARM_MAX,
6391 x86_64_int_return_registers, 0);
6393 /* For zero sized structures, construct_container returns NULL, but we
6394 need to keep rest of compiler happy by returning meaningful value. */
6396 ret = gen_rtx_REG (orig_mode, AX_REG);
6402 function_value_ms_64 (enum machine_mode orig_mode, enum machine_mode mode)
6404 unsigned int regno = AX_REG;
6408 switch (GET_MODE_SIZE (mode))
6411 if((SCALAR_INT_MODE_P (mode) || VECTOR_MODE_P (mode))
6412 && !COMPLEX_MODE_P (mode))
6413 regno = FIRST_SSE_REG;
6417 if (mode == SFmode || mode == DFmode)
6418 regno = FIRST_SSE_REG;
6424 return gen_rtx_REG (orig_mode, regno);
6428 ix86_function_value_1 (const_tree valtype, const_tree fntype_or_decl,
6429 enum machine_mode orig_mode, enum machine_mode mode)
6431 const_tree fn, fntype;
6434 if (fntype_or_decl && DECL_P (fntype_or_decl))
6435 fn = fntype_or_decl;
6436 fntype = fn ? TREE_TYPE (fn) : fntype_or_decl;
6438 if (TARGET_64BIT && ix86_function_type_abi (fntype) == MS_ABI)
6439 return function_value_ms_64 (orig_mode, mode);
6440 else if (TARGET_64BIT)
6441 return function_value_64 (orig_mode, mode, valtype);
6443 return function_value_32 (orig_mode, mode, fntype, fn);
6447 ix86_function_value (const_tree valtype, const_tree fntype_or_decl,
6448 bool outgoing ATTRIBUTE_UNUSED)
6450 enum machine_mode mode, orig_mode;
6452 orig_mode = TYPE_MODE (valtype);
6453 mode = type_natural_mode (valtype, NULL);
6454 return ix86_function_value_1 (valtype, fntype_or_decl, orig_mode, mode);
6458 ix86_libcall_value (enum machine_mode mode)
6460 return ix86_function_value_1 (NULL, NULL, mode, mode);
6463 /* Return true iff type is returned in memory. */
6465 static int ATTRIBUTE_UNUSED
6466 return_in_memory_32 (const_tree type, enum machine_mode mode)
6470 if (mode == BLKmode)
6473 size = int_size_in_bytes (type);
6475 if (MS_AGGREGATE_RETURN && AGGREGATE_TYPE_P (type) && size <= 8)
6478 if (VECTOR_MODE_P (mode) || mode == TImode)
6480 /* User-created vectors small enough to fit in EAX. */
6484 /* MMX/3dNow values are returned in MM0,
6485 except when it doesn't exits. */
6487 return (TARGET_MMX ? 0 : 1);
6489 /* SSE values are returned in XMM0, except when it doesn't exist. */
6491 return (TARGET_SSE ? 0 : 1);
6493 /* AVX values are returned in YMM0, except when it doesn't exist. */
6495 return TARGET_AVX ? 0 : 1;
6504 /* OImode shouldn't be used directly. */
6505 gcc_assert (mode != OImode);
6510 static int ATTRIBUTE_UNUSED
6511 return_in_memory_64 (const_tree type, enum machine_mode mode)
6513 int needed_intregs, needed_sseregs;
6514 return !examine_argument (mode, type, 1, &needed_intregs, &needed_sseregs);
6517 static int ATTRIBUTE_UNUSED
6518 return_in_memory_ms_64 (const_tree type, enum machine_mode mode)
6520 HOST_WIDE_INT size = int_size_in_bytes (type);
6522 /* __m128 is returned in xmm0. */
6523 if ((SCALAR_INT_MODE_P (mode) || VECTOR_MODE_P (mode))
6524 && !COMPLEX_MODE_P (mode) && (GET_MODE_SIZE (mode) == 16 || size == 16))
6527 /* Otherwise, the size must be exactly in [1248]. */
6528 return (size != 1 && size != 2 && size != 4 && size != 8);
6532 ix86_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6534 #ifdef SUBTARGET_RETURN_IN_MEMORY
6535 return SUBTARGET_RETURN_IN_MEMORY (type, fntype);
6537 const enum machine_mode mode = type_natural_mode (type, NULL);
6541 if (ix86_function_type_abi (fntype) == MS_ABI)
6542 return return_in_memory_ms_64 (type, mode);
6544 return return_in_memory_64 (type, mode);
6547 return return_in_memory_32 (type, mode);
6551 /* Return false iff TYPE is returned in memory. This version is used
6552 on Solaris 10. It is similar to the generic ix86_return_in_memory,
6553 but differs notably in that when MMX is available, 8-byte vectors
6554 are returned in memory, rather than in MMX registers. */
6557 ix86_sol10_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6560 enum machine_mode mode = type_natural_mode (type, NULL);
6563 return return_in_memory_64 (type, mode);
6565 if (mode == BLKmode)
6568 size = int_size_in_bytes (type);
6570 if (VECTOR_MODE_P (mode))
6572 /* Return in memory only if MMX registers *are* available. This
6573 seems backwards, but it is consistent with the existing
6580 else if (mode == TImode)
6582 else if (mode == XFmode)
6588 /* When returning SSE vector types, we have a choice of either
6589 (1) being abi incompatible with a -march switch, or
6590 (2) generating an error.
6591 Given no good solution, I think the safest thing is one warning.
6592 The user won't be able to use -Werror, but....
6594 Choose the STRUCT_VALUE_RTX hook because that's (at present) only
6595 called in response to actually generating a caller or callee that
6596 uses such a type. As opposed to TARGET_RETURN_IN_MEMORY, which is called
6597 via aggregate_value_p for general type probing from tree-ssa. */
6600 ix86_struct_value_rtx (tree type, int incoming ATTRIBUTE_UNUSED)
6602 static bool warnedsse, warnedmmx;
6604 if (!TARGET_64BIT && type)
6606 /* Look at the return type of the function, not the function type. */
6607 enum machine_mode mode = TYPE_MODE (TREE_TYPE (type));
6609 if (!TARGET_SSE && !warnedsse)
6612 || (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 16))
6615 warning (0, "SSE vector return without SSE enabled "
6620 if (!TARGET_MMX && !warnedmmx)
6622 if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 8)
6625 warning (0, "MMX vector return without MMX enabled "
6635 /* Create the va_list data type. */
6637 /* Returns the calling convention specific va_list date type.
6638 The argument ABI can be DEFAULT_ABI, MS_ABI, or SYSV_ABI. */
6641 ix86_build_builtin_va_list_abi (enum calling_abi abi)
6643 tree f_gpr, f_fpr, f_ovf, f_sav, record, type_decl;
6645 /* For i386 we use plain pointer to argument area. */
6646 if (!TARGET_64BIT || abi == MS_ABI)
6647 return build_pointer_type (char_type_node);
6649 record = (*lang_hooks.types.make_type) (RECORD_TYPE);
6650 type_decl = build_decl (BUILTINS_LOCATION,
6651 TYPE_DECL, get_identifier ("__va_list_tag"), record);
6653 f_gpr = build_decl (BUILTINS_LOCATION,
6654 FIELD_DECL, get_identifier ("gp_offset"),
6655 unsigned_type_node);
6656 f_fpr = build_decl (BUILTINS_LOCATION,
6657 FIELD_DECL, get_identifier ("fp_offset"),
6658 unsigned_type_node);
6659 f_ovf = build_decl (BUILTINS_LOCATION,
6660 FIELD_DECL, get_identifier ("overflow_arg_area"),
6662 f_sav = build_decl (BUILTINS_LOCATION,
6663 FIELD_DECL, get_identifier ("reg_save_area"),
6666 va_list_gpr_counter_field = f_gpr;
6667 va_list_fpr_counter_field = f_fpr;
6669 DECL_FIELD_CONTEXT (f_gpr) = record;
6670 DECL_FIELD_CONTEXT (f_fpr) = record;
6671 DECL_FIELD_CONTEXT (f_ovf) = record;
6672 DECL_FIELD_CONTEXT (f_sav) = record;
6674 TREE_CHAIN (record) = type_decl;
6675 TYPE_NAME (record) = type_decl;
6676 TYPE_FIELDS (record) = f_gpr;
6677 TREE_CHAIN (f_gpr) = f_fpr;
6678 TREE_CHAIN (f_fpr) = f_ovf;
6679 TREE_CHAIN (f_ovf) = f_sav;
6681 layout_type (record);
6683 /* The correct type is an array type of one element. */
6684 return build_array_type (record, build_index_type (size_zero_node));
6687 /* Setup the builtin va_list data type and for 64-bit the additional
6688 calling convention specific va_list data types. */
6691 ix86_build_builtin_va_list (void)
6693 tree ret = ix86_build_builtin_va_list_abi (ix86_abi);
6695 /* Initialize abi specific va_list builtin types. */
6699 if (ix86_abi == MS_ABI)
6701 t = ix86_build_builtin_va_list_abi (SYSV_ABI);
6702 if (TREE_CODE (t) != RECORD_TYPE)
6703 t = build_variant_type_copy (t);
6704 sysv_va_list_type_node = t;
6709 if (TREE_CODE (t) != RECORD_TYPE)
6710 t = build_variant_type_copy (t);
6711 sysv_va_list_type_node = t;
6713 if (ix86_abi != MS_ABI)
6715 t = ix86_build_builtin_va_list_abi (MS_ABI);
6716 if (TREE_CODE (t) != RECORD_TYPE)
6717 t = build_variant_type_copy (t);
6718 ms_va_list_type_node = t;
6723 if (TREE_CODE (t) != RECORD_TYPE)
6724 t = build_variant_type_copy (t);
6725 ms_va_list_type_node = t;
6732 /* Worker function for TARGET_SETUP_INCOMING_VARARGS. */
6735 setup_incoming_varargs_64 (CUMULATIVE_ARGS *cum)
6744 int regparm = ix86_regparm;
6746 if (cum->call_abi != ix86_abi)
6747 regparm = (ix86_abi != SYSV_ABI
6748 ? X86_64_REGPARM_MAX : X86_64_MS_REGPARM_MAX);
6750 /* GPR size of varargs save area. */
6751 if (cfun->va_list_gpr_size)
6752 ix86_varargs_gpr_size = X86_64_REGPARM_MAX * UNITS_PER_WORD;
6754 ix86_varargs_gpr_size = 0;
6756 /* FPR size of varargs save area. We don't need it if we don't pass
6757 anything in SSE registers. */
6758 if (cum->sse_nregs && cfun->va_list_fpr_size)
6759 ix86_varargs_fpr_size = X86_64_SSE_REGPARM_MAX * 16;
6761 ix86_varargs_fpr_size = 0;
6763 if (! ix86_varargs_gpr_size && ! ix86_varargs_fpr_size)
6766 save_area = frame_pointer_rtx;
6767 set = get_varargs_alias_set ();
6769 for (i = cum->regno;
6771 && i < cum->regno + cfun->va_list_gpr_size / UNITS_PER_WORD;
6774 mem = gen_rtx_MEM (Pmode,
6775 plus_constant (save_area, i * UNITS_PER_WORD));
6776 MEM_NOTRAP_P (mem) = 1;
6777 set_mem_alias_set (mem, set);
6778 emit_move_insn (mem, gen_rtx_REG (Pmode,
6779 x86_64_int_parameter_registers[i]));
6782 if (ix86_varargs_fpr_size)
6784 /* Now emit code to save SSE registers. The AX parameter contains number
6785 of SSE parameter registers used to call this function. We use
6786 sse_prologue_save insn template that produces computed jump across
6787 SSE saves. We need some preparation work to get this working. */
6789 label = gen_label_rtx ();
6790 label_ref = gen_rtx_LABEL_REF (Pmode, label);
6792 /* Compute address to jump to :
6793 label - eax*4 + nnamed_sse_arguments*4 Or
6794 label - eax*5 + nnamed_sse_arguments*5 for AVX. */
6795 tmp_reg = gen_reg_rtx (Pmode);
6796 nsse_reg = gen_reg_rtx (Pmode);
6797 emit_insn (gen_zero_extendqidi2 (nsse_reg, gen_rtx_REG (QImode, AX_REG)));
6798 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6799 gen_rtx_MULT (Pmode, nsse_reg,
6802 /* vmovaps is one byte longer than movaps. */
6804 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6805 gen_rtx_PLUS (Pmode, tmp_reg,
6811 gen_rtx_CONST (DImode,
6812 gen_rtx_PLUS (DImode,
6814 GEN_INT (cum->sse_regno
6815 * (TARGET_AVX ? 5 : 4)))));
6817 emit_move_insn (nsse_reg, label_ref);
6818 emit_insn (gen_subdi3 (nsse_reg, nsse_reg, tmp_reg));
6820 /* Compute address of memory block we save into. We always use pointer
6821 pointing 127 bytes after first byte to store - this is needed to keep
6822 instruction size limited by 4 bytes (5 bytes for AVX) with one
6823 byte displacement. */
6824 tmp_reg = gen_reg_rtx (Pmode);
6825 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6826 plus_constant (save_area,
6827 ix86_varargs_gpr_size + 127)));
6828 mem = gen_rtx_MEM (BLKmode, plus_constant (tmp_reg, -127));
6829 MEM_NOTRAP_P (mem) = 1;
6830 set_mem_alias_set (mem, set);
6831 set_mem_align (mem, BITS_PER_WORD);
6833 /* And finally do the dirty job! */
6834 emit_insn (gen_sse_prologue_save (mem, nsse_reg,
6835 GEN_INT (cum->sse_regno), label));
6840 setup_incoming_varargs_ms_64 (CUMULATIVE_ARGS *cum)
6842 alias_set_type set = get_varargs_alias_set ();
6845 for (i = cum->regno; i < X86_64_MS_REGPARM_MAX; i++)
6849 mem = gen_rtx_MEM (Pmode,
6850 plus_constant (virtual_incoming_args_rtx,
6851 i * UNITS_PER_WORD));
6852 MEM_NOTRAP_P (mem) = 1;
6853 set_mem_alias_set (mem, set);
6855 reg = gen_rtx_REG (Pmode, x86_64_ms_abi_int_parameter_registers[i]);
6856 emit_move_insn (mem, reg);
6861 ix86_setup_incoming_varargs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
6862 tree type, int *pretend_size ATTRIBUTE_UNUSED,
6865 CUMULATIVE_ARGS next_cum;
6868 /* This argument doesn't appear to be used anymore. Which is good,
6869 because the old code here didn't suppress rtl generation. */
6870 gcc_assert (!no_rtl);
6875 fntype = TREE_TYPE (current_function_decl);
6877 /* For varargs, we do not want to skip the dummy va_dcl argument.
6878 For stdargs, we do want to skip the last named argument. */
6880 if (stdarg_p (fntype))
6881 function_arg_advance (&next_cum, mode, type, 1);
6883 if (cum->call_abi == MS_ABI)
6884 setup_incoming_varargs_ms_64 (&next_cum);
6886 setup_incoming_varargs_64 (&next_cum);
6889 /* Checks if TYPE is of kind va_list char *. */
6892 is_va_list_char_pointer (tree type)
6896 /* For 32-bit it is always true. */
6899 canonic = ix86_canonical_va_list_type (type);
6900 return (canonic == ms_va_list_type_node
6901 || (ix86_abi == MS_ABI && canonic == va_list_type_node));
6904 /* Implement va_start. */
6907 ix86_va_start (tree valist, rtx nextarg)
6909 HOST_WIDE_INT words, n_gpr, n_fpr;
6910 tree f_gpr, f_fpr, f_ovf, f_sav;
6911 tree gpr, fpr, ovf, sav, t;
6914 /* Only 64bit target needs something special. */
6915 if (!TARGET_64BIT || is_va_list_char_pointer (TREE_TYPE (valist)))
6917 std_expand_builtin_va_start (valist, nextarg);
6921 f_gpr = TYPE_FIELDS (TREE_TYPE (sysv_va_list_type_node));
6922 f_fpr = TREE_CHAIN (f_gpr);
6923 f_ovf = TREE_CHAIN (f_fpr);
6924 f_sav = TREE_CHAIN (f_ovf);
6926 valist = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (valist)), valist);
6927 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr), valist, f_gpr, NULL_TREE);
6928 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
6929 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
6930 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
6932 /* Count number of gp and fp argument registers used. */
6933 words = crtl->args.info.words;
6934 n_gpr = crtl->args.info.regno;
6935 n_fpr = crtl->args.info.sse_regno;
6937 if (cfun->va_list_gpr_size)
6939 type = TREE_TYPE (gpr);
6940 t = build2 (MODIFY_EXPR, type,
6941 gpr, build_int_cst (type, n_gpr * 8));
6942 TREE_SIDE_EFFECTS (t) = 1;
6943 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6946 if (TARGET_SSE && cfun->va_list_fpr_size)
6948 type = TREE_TYPE (fpr);
6949 t = build2 (MODIFY_EXPR, type, fpr,
6950 build_int_cst (type, n_fpr * 16 + 8*X86_64_REGPARM_MAX));
6951 TREE_SIDE_EFFECTS (t) = 1;
6952 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6955 /* Find the overflow area. */
6956 type = TREE_TYPE (ovf);
6957 t = make_tree (type, crtl->args.internal_arg_pointer);
6959 t = build2 (POINTER_PLUS_EXPR, type, t,
6960 size_int (words * UNITS_PER_WORD));
6961 t = build2 (MODIFY_EXPR, type, ovf, t);
6962 TREE_SIDE_EFFECTS (t) = 1;
6963 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6965 if (ix86_varargs_gpr_size || ix86_varargs_fpr_size)
6967 /* Find the register save area.
6968 Prologue of the function save it right above stack frame. */
6969 type = TREE_TYPE (sav);
6970 t = make_tree (type, frame_pointer_rtx);
6971 if (!ix86_varargs_gpr_size)
6972 t = build2 (POINTER_PLUS_EXPR, type, t,
6973 size_int (-8 * X86_64_REGPARM_MAX));
6974 t = build2 (MODIFY_EXPR, type, sav, t);
6975 TREE_SIDE_EFFECTS (t) = 1;
6976 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6980 /* Implement va_arg. */
6983 ix86_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
6986 static const int intreg[6] = { 0, 1, 2, 3, 4, 5 };
6987 tree f_gpr, f_fpr, f_ovf, f_sav;
6988 tree gpr, fpr, ovf, sav, t;
6990 tree lab_false, lab_over = NULL_TREE;
6995 enum machine_mode nat_mode;
6998 /* Only 64bit target needs something special. */
6999 if (!TARGET_64BIT || is_va_list_char_pointer (TREE_TYPE (valist)))
7000 return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
7002 f_gpr = TYPE_FIELDS (TREE_TYPE (sysv_va_list_type_node));
7003 f_fpr = TREE_CHAIN (f_gpr);
7004 f_ovf = TREE_CHAIN (f_fpr);
7005 f_sav = TREE_CHAIN (f_ovf);
7007 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr),
7008 build_va_arg_indirect_ref (valist), f_gpr, NULL_TREE);
7009 valist = build_va_arg_indirect_ref (valist);
7010 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
7011 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
7012 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
7014 indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false);
7016 type = build_pointer_type (type);
7017 size = int_size_in_bytes (type);
7018 rsize = (size + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
7020 nat_mode = type_natural_mode (type, NULL);
7029 /* Unnamed 256bit vector mode parameters are passed on stack. */
7030 if (ix86_cfun_abi () == SYSV_ABI)
7037 container = construct_container (nat_mode, TYPE_MODE (type),
7038 type, 0, X86_64_REGPARM_MAX,
7039 X86_64_SSE_REGPARM_MAX, intreg,
7044 /* Pull the value out of the saved registers. */
7046 addr = create_tmp_var (ptr_type_node, "addr");
7050 int needed_intregs, needed_sseregs;
7052 tree int_addr, sse_addr;
7054 lab_false = create_artificial_label (UNKNOWN_LOCATION);
7055 lab_over = create_artificial_label (UNKNOWN_LOCATION);
7057 examine_argument (nat_mode, type, 0, &needed_intregs, &needed_sseregs);
7059 need_temp = (!REG_P (container)
7060 && ((needed_intregs && TYPE_ALIGN (type) > 64)
7061 || TYPE_ALIGN (type) > 128));
7063 /* In case we are passing structure, verify that it is consecutive block
7064 on the register save area. If not we need to do moves. */
7065 if (!need_temp && !REG_P (container))
7067 /* Verify that all registers are strictly consecutive */
7068 if (SSE_REGNO_P (REGNO (XEXP (XVECEXP (container, 0, 0), 0))))
7072 for (i = 0; i < XVECLEN (container, 0) && !need_temp; i++)
7074 rtx slot = XVECEXP (container, 0, i);
7075 if (REGNO (XEXP (slot, 0)) != FIRST_SSE_REG + (unsigned int) i
7076 || INTVAL (XEXP (slot, 1)) != i * 16)
7084 for (i = 0; i < XVECLEN (container, 0) && !need_temp; i++)
7086 rtx slot = XVECEXP (container, 0, i);
7087 if (REGNO (XEXP (slot, 0)) != (unsigned int) i
7088 || INTVAL (XEXP (slot, 1)) != i * 8)
7100 int_addr = create_tmp_var (ptr_type_node, "int_addr");
7101 sse_addr = create_tmp_var (ptr_type_node, "sse_addr");
7104 /* First ensure that we fit completely in registers. */
7107 t = build_int_cst (TREE_TYPE (gpr),
7108 (X86_64_REGPARM_MAX - needed_intregs + 1) * 8);
7109 t = build2 (GE_EXPR, boolean_type_node, gpr, t);
7110 t2 = build1 (GOTO_EXPR, void_type_node, lab_false);
7111 t = build3 (COND_EXPR, void_type_node, t, t2, NULL_TREE);
7112 gimplify_and_add (t, pre_p);
7116 t = build_int_cst (TREE_TYPE (fpr),
7117 (X86_64_SSE_REGPARM_MAX - needed_sseregs + 1) * 16
7118 + X86_64_REGPARM_MAX * 8);
7119 t = build2 (GE_EXPR, boolean_type_node, fpr, t);
7120 t2 = build1 (GOTO_EXPR, void_type_node, lab_false);
7121 t = build3 (COND_EXPR, void_type_node, t, t2, NULL_TREE);
7122 gimplify_and_add (t, pre_p);
7125 /* Compute index to start of area used for integer regs. */
7128 /* int_addr = gpr + sav; */
7129 t = fold_convert (sizetype, gpr);
7130 t = build2 (POINTER_PLUS_EXPR, ptr_type_node, sav, t);
7131 gimplify_assign (int_addr, t, pre_p);
7135 /* sse_addr = fpr + sav; */
7136 t = fold_convert (sizetype, fpr);
7137 t = build2 (POINTER_PLUS_EXPR, ptr_type_node, sav, t);
7138 gimplify_assign (sse_addr, t, pre_p);
7143 tree temp = create_tmp_var (type, "va_arg_tmp");
7146 t = build1 (ADDR_EXPR, build_pointer_type (type), temp);
7147 gimplify_assign (addr, t, pre_p);
7149 for (i = 0; i < XVECLEN (container, 0); i++)
7151 rtx slot = XVECEXP (container, 0, i);
7152 rtx reg = XEXP (slot, 0);
7153 enum machine_mode mode = GET_MODE (reg);
7154 tree piece_type = lang_hooks.types.type_for_mode (mode, 1);
7155 tree addr_type = build_pointer_type (piece_type);
7156 tree daddr_type = build_pointer_type_for_mode (piece_type,
7160 tree dest_addr, dest;
7162 if (SSE_REGNO_P (REGNO (reg)))
7164 src_addr = sse_addr;
7165 src_offset = (REGNO (reg) - FIRST_SSE_REG) * 16;
7169 src_addr = int_addr;
7170 src_offset = REGNO (reg) * 8;
7172 src_addr = fold_convert (addr_type, src_addr);
7173 src_addr = fold_build2 (POINTER_PLUS_EXPR, addr_type, src_addr,
7174 size_int (src_offset));
7175 src = build_va_arg_indirect_ref (src_addr);
7177 dest_addr = fold_convert (daddr_type, addr);
7178 dest_addr = fold_build2 (POINTER_PLUS_EXPR, daddr_type, dest_addr,
7179 size_int (INTVAL (XEXP (slot, 1))));
7180 dest = build_va_arg_indirect_ref (dest_addr);
7182 gimplify_assign (dest, src, pre_p);
7188 t = build2 (PLUS_EXPR, TREE_TYPE (gpr), gpr,
7189 build_int_cst (TREE_TYPE (gpr), needed_intregs * 8));
7190 gimplify_assign (gpr, t, pre_p);
7195 t = build2 (PLUS_EXPR, TREE_TYPE (fpr), fpr,
7196 build_int_cst (TREE_TYPE (fpr), needed_sseregs * 16));
7197 gimplify_assign (fpr, t, pre_p);
7200 gimple_seq_add_stmt (pre_p, gimple_build_goto (lab_over));
7202 gimple_seq_add_stmt (pre_p, gimple_build_label (lab_false));
7205 /* ... otherwise out of the overflow area. */
7207 /* When we align parameter on stack for caller, if the parameter
7208 alignment is beyond MAX_SUPPORTED_STACK_ALIGNMENT, it will be
7209 aligned at MAX_SUPPORTED_STACK_ALIGNMENT. We will match callee
7210 here with caller. */
7211 arg_boundary = FUNCTION_ARG_BOUNDARY (VOIDmode, type);
7212 if ((unsigned int) arg_boundary > MAX_SUPPORTED_STACK_ALIGNMENT)
7213 arg_boundary = MAX_SUPPORTED_STACK_ALIGNMENT;
7215 /* Care for on-stack alignment if needed. */
7216 if (arg_boundary <= 64
7217 || integer_zerop (TYPE_SIZE (type)))
7221 HOST_WIDE_INT align = arg_boundary / 8;
7222 t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (ovf), ovf,
7223 size_int (align - 1));
7224 t = fold_convert (sizetype, t);
7225 t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t,
7227 t = fold_convert (TREE_TYPE (ovf), t);
7229 gimplify_expr (&t, pre_p, NULL, is_gimple_val, fb_rvalue);
7230 gimplify_assign (addr, t, pre_p);
7232 t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (t), t,
7233 size_int (rsize * UNITS_PER_WORD));
7234 gimplify_assign (unshare_expr (ovf), t, pre_p);
7237 gimple_seq_add_stmt (pre_p, gimple_build_label (lab_over));
7239 ptrtype = build_pointer_type_for_mode (type, ptr_mode, true);
7240 addr = fold_convert (ptrtype, addr);
7243 addr = build_va_arg_indirect_ref (addr);
7244 return build_va_arg_indirect_ref (addr);
7247 /* Return nonzero if OPNUM's MEM should be matched
7248 in movabs* patterns. */
7251 ix86_check_movabs (rtx insn, int opnum)
7255 set = PATTERN (insn);
7256 if (GET_CODE (set) == PARALLEL)
7257 set = XVECEXP (set, 0, 0);
7258 gcc_assert (GET_CODE (set) == SET);
7259 mem = XEXP (set, opnum);
7260 while (GET_CODE (mem) == SUBREG)
7261 mem = SUBREG_REG (mem);
7262 gcc_assert (MEM_P (mem));
7263 return (volatile_ok || !MEM_VOLATILE_P (mem));
7266 /* Initialize the table of extra 80387 mathematical constants. */
7269 init_ext_80387_constants (void)
7271 static const char * cst[5] =
7273 "0.3010299956639811952256464283594894482", /* 0: fldlg2 */
7274 "0.6931471805599453094286904741849753009", /* 1: fldln2 */
7275 "1.4426950408889634073876517827983434472", /* 2: fldl2e */
7276 "3.3219280948873623478083405569094566090", /* 3: fldl2t */
7277 "3.1415926535897932385128089594061862044", /* 4: fldpi */
7281 for (i = 0; i < 5; i++)
7283 real_from_string (&ext_80387_constants_table[i], cst[i]);
7284 /* Ensure each constant is rounded to XFmode precision. */
7285 real_convert (&ext_80387_constants_table[i],
7286 XFmode, &ext_80387_constants_table[i]);
7289 ext_80387_constants_init = 1;
7292 /* Return true if the constant is something that can be loaded with
7293 a special instruction. */
7296 standard_80387_constant_p (rtx x)
7298 enum machine_mode mode = GET_MODE (x);
7302 if (!(X87_FLOAT_MODE_P (mode) && (GET_CODE (x) == CONST_DOUBLE)))
7305 if (x == CONST0_RTX (mode))
7307 if (x == CONST1_RTX (mode))
7310 REAL_VALUE_FROM_CONST_DOUBLE (r, x);
7312 /* For XFmode constants, try to find a special 80387 instruction when
7313 optimizing for size or on those CPUs that benefit from them. */
7315 && (optimize_function_for_size_p (cfun) || TARGET_EXT_80387_CONSTANTS))
7319 if (! ext_80387_constants_init)
7320 init_ext_80387_constants ();
7322 for (i = 0; i < 5; i++)
7323 if (real_identical (&r, &ext_80387_constants_table[i]))
7327 /* Load of the constant -0.0 or -1.0 will be split as
7328 fldz;fchs or fld1;fchs sequence. */
7329 if (real_isnegzero (&r))
7331 if (real_identical (&r, &dconstm1))
7337 /* Return the opcode of the special instruction to be used to load
7341 standard_80387_constant_opcode (rtx x)
7343 switch (standard_80387_constant_p (x))
7367 /* Return the CONST_DOUBLE representing the 80387 constant that is
7368 loaded by the specified special instruction. The argument IDX
7369 matches the return value from standard_80387_constant_p. */
7372 standard_80387_constant_rtx (int idx)
7376 if (! ext_80387_constants_init)
7377 init_ext_80387_constants ();
7393 return CONST_DOUBLE_FROM_REAL_VALUE (ext_80387_constants_table[i],
7397 /* Return 1 if X is all 0s and 2 if x is all 1s
7398 in supported SSE vector mode. */
7401 standard_sse_constant_p (rtx x)
7403 enum machine_mode mode = GET_MODE (x);
7405 if (x == const0_rtx || x == CONST0_RTX (GET_MODE (x)))
7407 if (vector_all_ones_operand (x, mode))
7423 /* Return the opcode of the special instruction to be used to load
7427 standard_sse_constant_opcode (rtx insn, rtx x)
7429 switch (standard_sse_constant_p (x))
7432 switch (get_attr_mode (insn))
7435 return TARGET_AVX ? "vxorps\t%0, %0, %0" : "xorps\t%0, %0";
7437 return TARGET_AVX ? "vxorpd\t%0, %0, %0" : "xorpd\t%0, %0";
7439 return TARGET_AVX ? "vpxor\t%0, %0, %0" : "pxor\t%0, %0";
7441 return "vxorps\t%x0, %x0, %x0";
7443 return "vxorpd\t%x0, %x0, %x0";
7445 return "vpxor\t%x0, %x0, %x0";
7450 return TARGET_AVX ? "vpcmpeqd\t%0, %0, %0" : "pcmpeqd\t%0, %0";
7457 /* Returns 1 if OP contains a symbol reference */
7460 symbolic_reference_mentioned_p (rtx op)
7465 if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
7468 fmt = GET_RTX_FORMAT (GET_CODE (op));
7469 for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--)
7475 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
7476 if (symbolic_reference_mentioned_p (XVECEXP (op, i, j)))
7480 else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i)))
7487 /* Return 1 if it is appropriate to emit `ret' instructions in the
7488 body of a function. Do this only if the epilogue is simple, needing a
7489 couple of insns. Prior to reloading, we can't tell how many registers
7490 must be saved, so return 0 then. Return 0 if there is no frame
7491 marker to de-allocate. */
7494 ix86_can_use_return_insn_p (void)
7496 struct ix86_frame frame;
7498 if (! reload_completed || frame_pointer_needed)
7501 /* Don't allow more than 32 pop, since that's all we can do
7502 with one instruction. */
7503 if (crtl->args.pops_args
7504 && crtl->args.size >= 32768)
7507 ix86_compute_frame_layout (&frame);
7508 return frame.to_allocate == 0 && frame.padding0 == 0
7509 && (frame.nregs + frame.nsseregs) == 0;
7512 /* Value should be nonzero if functions must have frame pointers.
7513 Zero means the frame pointer need not be set up (and parms may
7514 be accessed via the stack pointer) in functions that seem suitable. */
7517 ix86_frame_pointer_required (void)
7519 /* If we accessed previous frames, then the generated code expects
7520 to be able to access the saved ebp value in our frame. */
7521 if (cfun->machine->accesses_prev_frame)
7524 /* Several x86 os'es need a frame pointer for other reasons,
7525 usually pertaining to setjmp. */
7526 if (SUBTARGET_FRAME_POINTER_REQUIRED)
7529 /* In override_options, TARGET_OMIT_LEAF_FRAME_POINTER turns off
7530 the frame pointer by default. Turn it back on now if we've not
7531 got a leaf function. */
7532 if (TARGET_OMIT_LEAF_FRAME_POINTER
7533 && (!current_function_is_leaf
7534 || ix86_current_function_calls_tls_descriptor))
7543 /* Record that the current function accesses previous call frames. */
7546 ix86_setup_frame_addresses (void)
7548 cfun->machine->accesses_prev_frame = 1;
7551 #ifndef USE_HIDDEN_LINKONCE
7552 # if (defined(HAVE_GAS_HIDDEN) && (SUPPORTS_ONE_ONLY - 0)) || TARGET_MACHO
7553 # define USE_HIDDEN_LINKONCE 1
7555 # define USE_HIDDEN_LINKONCE 0
7559 static int pic_labels_used;
7561 /* Fills in the label name that should be used for a pc thunk for
7562 the given register. */
7565 get_pc_thunk_name (char name[32], unsigned int regno)
7567 gcc_assert (!TARGET_64BIT);
7569 if (USE_HIDDEN_LINKONCE)
7570 sprintf (name, "__i686.get_pc_thunk.%s", reg_names[regno]);
7572 ASM_GENERATE_INTERNAL_LABEL (name, "LPR", regno);
7576 /* This function generates code for -fpic that loads %ebx with
7577 the return address of the caller and then returns. */
7580 ix86_file_end (void)
7585 for (regno = 0; regno < 8; ++regno)
7589 if (! ((pic_labels_used >> regno) & 1))
7592 get_pc_thunk_name (name, regno);
7597 switch_to_section (darwin_sections[text_coal_section]);
7598 fputs ("\t.weak_definition\t", asm_out_file);
7599 assemble_name (asm_out_file, name);
7600 fputs ("\n\t.private_extern\t", asm_out_file);
7601 assemble_name (asm_out_file, name);
7602 fputs ("\n", asm_out_file);
7603 ASM_OUTPUT_LABEL (asm_out_file, name);
7607 if (USE_HIDDEN_LINKONCE)
7611 decl = build_decl (BUILTINS_LOCATION,
7612 FUNCTION_DECL, get_identifier (name),
7614 TREE_PUBLIC (decl) = 1;
7615 TREE_STATIC (decl) = 1;
7616 DECL_COMDAT_GROUP (decl) = DECL_ASSEMBLER_NAME (decl);
7618 (*targetm.asm_out.unique_section) (decl, 0);
7619 switch_to_section (get_named_section (decl, NULL, 0));
7621 (*targetm.asm_out.globalize_label) (asm_out_file, name);
7622 fputs ("\t.hidden\t", asm_out_file);
7623 assemble_name (asm_out_file, name);
7624 putc ('\n', asm_out_file);
7625 ASM_DECLARE_FUNCTION_NAME (asm_out_file, name, decl);
7629 switch_to_section (text_section);
7630 ASM_OUTPUT_LABEL (asm_out_file, name);
7633 xops[0] = gen_rtx_REG (Pmode, regno);
7634 xops[1] = gen_rtx_MEM (Pmode, stack_pointer_rtx);
7635 output_asm_insn ("mov%z0\t{%1, %0|%0, %1}", xops);
7636 output_asm_insn ("ret", xops);
7639 if (NEED_INDICATE_EXEC_STACK)
7640 file_end_indicate_exec_stack ();
7643 /* Emit code for the SET_GOT patterns. */
7646 output_set_got (rtx dest, rtx label ATTRIBUTE_UNUSED)
7652 if (TARGET_VXWORKS_RTP && flag_pic)
7654 /* Load (*VXWORKS_GOTT_BASE) into the PIC register. */
7655 xops[2] = gen_rtx_MEM (Pmode,
7656 gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE));
7657 output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops);
7659 /* Load (*VXWORKS_GOTT_BASE)[VXWORKS_GOTT_INDEX] into the PIC register.
7660 Use %P and a local symbol in order to print VXWORKS_GOTT_INDEX as
7661 an unadorned address. */
7662 xops[2] = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_INDEX);
7663 SYMBOL_REF_FLAGS (xops[2]) |= SYMBOL_FLAG_LOCAL;
7664 output_asm_insn ("mov{l}\t{%P2(%0), %0|%0, DWORD PTR %P2[%0]}", xops);
7668 xops[1] = gen_rtx_SYMBOL_REF (Pmode, GOT_SYMBOL_NAME);
7670 if (! TARGET_DEEP_BRANCH_PREDICTION || !flag_pic)
7672 xops[2] = gen_rtx_LABEL_REF (Pmode, label ? label : gen_label_rtx ());
7675 output_asm_insn ("mov%z0\t{%2, %0|%0, %2}", xops);
7677 output_asm_insn ("call\t%a2", xops);
7680 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
7681 is what will be referenced by the Mach-O PIC subsystem. */
7683 ASM_OUTPUT_LABEL (asm_out_file, MACHOPIC_FUNCTION_BASE_NAME);
7686 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7687 CODE_LABEL_NUMBER (XEXP (xops[2], 0)));
7690 output_asm_insn ("pop%z0\t%0", xops);
7695 get_pc_thunk_name (name, REGNO (dest));
7696 pic_labels_used |= 1 << REGNO (dest);
7698 xops[2] = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
7699 xops[2] = gen_rtx_MEM (QImode, xops[2]);
7700 output_asm_insn ("call\t%X2", xops);
7701 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
7702 is what will be referenced by the Mach-O PIC subsystem. */
7705 ASM_OUTPUT_LABEL (asm_out_file, MACHOPIC_FUNCTION_BASE_NAME);
7707 targetm.asm_out.internal_label (asm_out_file, "L",
7708 CODE_LABEL_NUMBER (label));
7715 if (!flag_pic || TARGET_DEEP_BRANCH_PREDICTION)
7716 output_asm_insn ("add%z0\t{%1, %0|%0, %1}", xops);
7718 output_asm_insn ("add%z0\t{%1+[.-%a2], %0|%0, %1+(.-%a2)}", xops);
7723 /* Generate an "push" pattern for input ARG. */
7728 if (ix86_cfa_state->reg == stack_pointer_rtx)
7729 ix86_cfa_state->offset += UNITS_PER_WORD;
7731 return gen_rtx_SET (VOIDmode,
7733 gen_rtx_PRE_DEC (Pmode,
7734 stack_pointer_rtx)),
7738 /* Return >= 0 if there is an unused call-clobbered register available
7739 for the entire function. */
7742 ix86_select_alt_pic_regnum (void)
7744 if (current_function_is_leaf && !crtl->profile
7745 && !ix86_current_function_calls_tls_descriptor)
7748 /* Can't use the same register for both PIC and DRAP. */
7750 drap = REGNO (crtl->drap_reg);
7753 for (i = 2; i >= 0; --i)
7754 if (i != drap && !df_regs_ever_live_p (i))
7758 return INVALID_REGNUM;
7761 /* Return 1 if we need to save REGNO. */
7763 ix86_save_reg (unsigned int regno, int maybe_eh_return)
7765 if (pic_offset_table_rtx
7766 && regno == REAL_PIC_OFFSET_TABLE_REGNUM
7767 && (df_regs_ever_live_p (REAL_PIC_OFFSET_TABLE_REGNUM)
7769 || crtl->calls_eh_return
7770 || crtl->uses_const_pool))
7772 if (ix86_select_alt_pic_regnum () != INVALID_REGNUM)
7777 if (crtl->calls_eh_return && maybe_eh_return)
7782 unsigned test = EH_RETURN_DATA_REGNO (i);
7783 if (test == INVALID_REGNUM)
7790 if (crtl->drap_reg && regno == REGNO (crtl->drap_reg))
7793 return (df_regs_ever_live_p (regno)
7794 && !call_used_regs[regno]
7795 && !fixed_regs[regno]
7796 && (regno != HARD_FRAME_POINTER_REGNUM || !frame_pointer_needed));
7799 /* Return number of saved general prupose registers. */
7802 ix86_nsaved_regs (void)
7807 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7808 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7813 /* Return number of saved SSE registrers. */
7816 ix86_nsaved_sseregs (void)
7821 if (ix86_cfun_abi () != MS_ABI)
7823 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7824 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7829 /* Given FROM and TO register numbers, say whether this elimination is
7830 allowed. If stack alignment is needed, we can only replace argument
7831 pointer with hard frame pointer, or replace frame pointer with stack
7832 pointer. Otherwise, frame pointer elimination is automatically
7833 handled and all other eliminations are valid. */
7836 ix86_can_eliminate (const int from, const int to)
7838 if (stack_realign_fp)
7839 return ((from == ARG_POINTER_REGNUM
7840 && to == HARD_FRAME_POINTER_REGNUM)
7841 || (from == FRAME_POINTER_REGNUM
7842 && to == STACK_POINTER_REGNUM));
7844 return to == STACK_POINTER_REGNUM ? !frame_pointer_needed : true;
7847 /* Return the offset between two registers, one to be eliminated, and the other
7848 its replacement, at the start of a routine. */
7851 ix86_initial_elimination_offset (int from, int to)
7853 struct ix86_frame frame;
7854 ix86_compute_frame_layout (&frame);
7856 if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
7857 return frame.hard_frame_pointer_offset;
7858 else if (from == FRAME_POINTER_REGNUM
7859 && to == HARD_FRAME_POINTER_REGNUM)
7860 return frame.hard_frame_pointer_offset - frame.frame_pointer_offset;
7863 gcc_assert (to == STACK_POINTER_REGNUM);
7865 if (from == ARG_POINTER_REGNUM)
7866 return frame.stack_pointer_offset;
7868 gcc_assert (from == FRAME_POINTER_REGNUM);
7869 return frame.stack_pointer_offset - frame.frame_pointer_offset;
7873 /* In a dynamically-aligned function, we can't know the offset from
7874 stack pointer to frame pointer, so we must ensure that setjmp
7875 eliminates fp against the hard fp (%ebp) rather than trying to
7876 index from %esp up to the top of the frame across a gap that is
7877 of unknown (at compile-time) size. */
7879 ix86_builtin_setjmp_frame_value (void)
7881 return stack_realign_fp ? hard_frame_pointer_rtx : virtual_stack_vars_rtx;
7884 /* Fill structure ix86_frame about frame of currently computed function. */
7887 ix86_compute_frame_layout (struct ix86_frame *frame)
7889 unsigned int stack_alignment_needed;
7890 HOST_WIDE_INT offset;
7891 unsigned int preferred_alignment;
7892 HOST_WIDE_INT size = get_frame_size ();
7894 frame->nregs = ix86_nsaved_regs ();
7895 frame->nsseregs = ix86_nsaved_sseregs ();
7897 stack_alignment_needed = crtl->stack_alignment_needed / BITS_PER_UNIT;
7898 preferred_alignment = crtl->preferred_stack_boundary / BITS_PER_UNIT;
7900 /* MS ABI seem to require stack alignment to be always 16 except for function
7902 if (ix86_cfun_abi () == MS_ABI && preferred_alignment < 16)
7904 preferred_alignment = 16;
7905 stack_alignment_needed = 16;
7906 crtl->preferred_stack_boundary = 128;
7907 crtl->stack_alignment_needed = 128;
7910 gcc_assert (!size || stack_alignment_needed);
7911 gcc_assert (preferred_alignment >= STACK_BOUNDARY / BITS_PER_UNIT);
7912 gcc_assert (preferred_alignment <= stack_alignment_needed);
7914 /* During reload iteration the amount of registers saved can change.
7915 Recompute the value as needed. Do not recompute when amount of registers
7916 didn't change as reload does multiple calls to the function and does not
7917 expect the decision to change within single iteration. */
7918 if (!optimize_function_for_size_p (cfun)
7919 && cfun->machine->use_fast_prologue_epilogue_nregs != frame->nregs)
7921 int count = frame->nregs;
7923 cfun->machine->use_fast_prologue_epilogue_nregs = count;
7924 /* The fast prologue uses move instead of push to save registers. This
7925 is significantly longer, but also executes faster as modern hardware
7926 can execute the moves in parallel, but can't do that for push/pop.
7928 Be careful about choosing what prologue to emit: When function takes
7929 many instructions to execute we may use slow version as well as in
7930 case function is known to be outside hot spot (this is known with
7931 feedback only). Weight the size of function by number of registers
7932 to save as it is cheap to use one or two push instructions but very
7933 slow to use many of them. */
7935 count = (count - 1) * FAST_PROLOGUE_INSN_COUNT;
7936 if (cfun->function_frequency < FUNCTION_FREQUENCY_NORMAL
7937 || (flag_branch_probabilities
7938 && cfun->function_frequency < FUNCTION_FREQUENCY_HOT))
7939 cfun->machine->use_fast_prologue_epilogue = false;
7941 cfun->machine->use_fast_prologue_epilogue
7942 = !expensive_function_p (count);
7944 if (TARGET_PROLOGUE_USING_MOVE
7945 && cfun->machine->use_fast_prologue_epilogue)
7946 frame->save_regs_using_mov = true;
7948 frame->save_regs_using_mov = false;
7950 /* Skip return address. */
7951 offset = UNITS_PER_WORD;
7953 /* Skip pushed static chain. */
7954 if (ix86_static_chain_on_stack)
7955 offset += UNITS_PER_WORD;
7957 /* Skip saved base pointer. */
7958 if (frame_pointer_needed)
7959 offset += UNITS_PER_WORD;
7961 frame->hard_frame_pointer_offset = offset;
7963 /* Set offset to aligned because the realigned frame starts from
7965 if (stack_realign_fp)
7966 offset = (offset + stack_alignment_needed -1) & -stack_alignment_needed;
7968 /* Register save area */
7969 offset += frame->nregs * UNITS_PER_WORD;
7971 /* Align SSE reg save area. */
7972 if (frame->nsseregs)
7973 frame->padding0 = ((offset + 16 - 1) & -16) - offset;
7975 frame->padding0 = 0;
7977 /* SSE register save area. */
7978 offset += frame->padding0 + frame->nsseregs * 16;
7981 frame->va_arg_size = ix86_varargs_gpr_size + ix86_varargs_fpr_size;
7982 offset += frame->va_arg_size;
7984 /* Align start of frame for local function. */
7985 frame->padding1 = ((offset + stack_alignment_needed - 1)
7986 & -stack_alignment_needed) - offset;
7988 offset += frame->padding1;
7990 /* Frame pointer points here. */
7991 frame->frame_pointer_offset = offset;
7995 /* Add outgoing arguments area. Can be skipped if we eliminated
7996 all the function calls as dead code.
7997 Skipping is however impossible when function calls alloca. Alloca
7998 expander assumes that last crtl->outgoing_args_size
7999 of stack frame are unused. */
8000 if (ACCUMULATE_OUTGOING_ARGS
8001 && (!current_function_is_leaf || cfun->calls_alloca
8002 || ix86_current_function_calls_tls_descriptor))
8004 offset += crtl->outgoing_args_size;
8005 frame->outgoing_arguments_size = crtl->outgoing_args_size;
8008 frame->outgoing_arguments_size = 0;
8010 /* Align stack boundary. Only needed if we're calling another function
8012 if (!current_function_is_leaf || cfun->calls_alloca
8013 || ix86_current_function_calls_tls_descriptor)
8014 frame->padding2 = ((offset + preferred_alignment - 1)
8015 & -preferred_alignment) - offset;
8017 frame->padding2 = 0;
8019 offset += frame->padding2;
8021 /* We've reached end of stack frame. */
8022 frame->stack_pointer_offset = offset;
8024 /* Size prologue needs to allocate. */
8025 frame->to_allocate =
8026 (size + frame->padding1 + frame->padding2
8027 + frame->outgoing_arguments_size + frame->va_arg_size);
8029 if ((!frame->to_allocate && frame->nregs <= 1)
8030 || (TARGET_64BIT && frame->to_allocate >= (HOST_WIDE_INT) 0x80000000))
8031 frame->save_regs_using_mov = false;
8033 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE
8034 && current_function_sp_is_unchanging
8035 && current_function_is_leaf
8036 && !ix86_current_function_calls_tls_descriptor)
8038 frame->red_zone_size = frame->to_allocate;
8039 if (frame->save_regs_using_mov)
8040 frame->red_zone_size += frame->nregs * UNITS_PER_WORD;
8041 if (frame->red_zone_size > RED_ZONE_SIZE - RED_ZONE_RESERVE)
8042 frame->red_zone_size = RED_ZONE_SIZE - RED_ZONE_RESERVE;
8045 frame->red_zone_size = 0;
8046 frame->to_allocate -= frame->red_zone_size;
8047 frame->stack_pointer_offset -= frame->red_zone_size;
8050 /* Emit code to save registers in the prologue. */
8053 ix86_emit_save_regs (void)
8058 for (regno = FIRST_PSEUDO_REGISTER - 1; regno-- > 0; )
8059 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
8061 insn = emit_insn (gen_push (gen_rtx_REG (Pmode, regno)));
8062 RTX_FRAME_RELATED_P (insn) = 1;
8066 /* Emit code to save registers using MOV insns. First register
8067 is restored from POINTER + OFFSET. */
8069 ix86_emit_save_regs_using_mov (rtx pointer, HOST_WIDE_INT offset)
8074 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8075 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
8077 insn = emit_move_insn (adjust_address (gen_rtx_MEM (Pmode, pointer),
8079 gen_rtx_REG (Pmode, regno));
8080 RTX_FRAME_RELATED_P (insn) = 1;
8081 offset += UNITS_PER_WORD;
8085 /* Emit code to save registers using MOV insns. First register
8086 is restored from POINTER + OFFSET. */
8088 ix86_emit_save_sse_regs_using_mov (rtx pointer, HOST_WIDE_INT offset)
8094 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8095 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
8097 mem = adjust_address (gen_rtx_MEM (TImode, pointer), TImode, offset);
8098 set_mem_align (mem, 128);
8099 insn = emit_move_insn (mem, gen_rtx_REG (TImode, regno));
8100 RTX_FRAME_RELATED_P (insn) = 1;
8105 static GTY(()) rtx queued_cfa_restores;
8107 /* Add a REG_CFA_RESTORE REG note to INSN or queue them until next stack
8108 manipulation insn. Don't add it if the previously
8109 saved value will be left untouched within stack red-zone till return,
8110 as unwinders can find the same value in the register and
8114 ix86_add_cfa_restore_note (rtx insn, rtx reg, HOST_WIDE_INT red_offset)
8117 && !TARGET_64BIT_MS_ABI
8118 && red_offset + RED_ZONE_SIZE >= 0
8119 && crtl->args.pops_args < 65536)
8124 add_reg_note (insn, REG_CFA_RESTORE, reg);
8125 RTX_FRAME_RELATED_P (insn) = 1;
8129 = alloc_reg_note (REG_CFA_RESTORE, reg, queued_cfa_restores);
8132 /* Add queued REG_CFA_RESTORE notes if any to INSN. */
8135 ix86_add_queued_cfa_restore_notes (rtx insn)
8138 if (!queued_cfa_restores)
8140 for (last = queued_cfa_restores; XEXP (last, 1); last = XEXP (last, 1))
8142 XEXP (last, 1) = REG_NOTES (insn);
8143 REG_NOTES (insn) = queued_cfa_restores;
8144 queued_cfa_restores = NULL_RTX;
8145 RTX_FRAME_RELATED_P (insn) = 1;
8148 /* Expand prologue or epilogue stack adjustment.
8149 The pattern exist to put a dependency on all ebp-based memory accesses.
8150 STYLE should be negative if instructions should be marked as frame related,
8151 zero if %r11 register is live and cannot be freely used and positive
8155 pro_epilogue_adjust_stack (rtx dest, rtx src, rtx offset,
8156 int style, bool set_cfa)
8161 insn = emit_insn (gen_pro_epilogue_adjust_stack_1 (dest, src, offset));
8162 else if (x86_64_immediate_operand (offset, DImode))
8163 insn = emit_insn (gen_pro_epilogue_adjust_stack_rex64 (dest, src, offset));
8167 /* r11 is used by indirect sibcall return as well, set before the
8168 epilogue and used after the epilogue. ATM indirect sibcall
8169 shouldn't be used together with huge frame sizes in one
8170 function because of the frame_size check in sibcall.c. */
8172 r11 = gen_rtx_REG (DImode, R11_REG);
8173 insn = emit_insn (gen_rtx_SET (DImode, r11, offset));
8175 RTX_FRAME_RELATED_P (insn) = 1;
8176 insn = emit_insn (gen_pro_epilogue_adjust_stack_rex64_2 (dest, src, r11,
8181 ix86_add_queued_cfa_restore_notes (insn);
8187 gcc_assert (ix86_cfa_state->reg == src);
8188 ix86_cfa_state->offset += INTVAL (offset);
8189 ix86_cfa_state->reg = dest;
8191 r = gen_rtx_PLUS (Pmode, src, offset);
8192 r = gen_rtx_SET (VOIDmode, dest, r);
8193 add_reg_note (insn, REG_CFA_ADJUST_CFA, r);
8194 RTX_FRAME_RELATED_P (insn) = 1;
8197 RTX_FRAME_RELATED_P (insn) = 1;
8200 /* Find an available register to be used as dynamic realign argument
8201 pointer regsiter. Such a register will be written in prologue and
8202 used in begin of body, so it must not be
8203 1. parameter passing register.
8205 We reuse static-chain register if it is available. Otherwise, we
8206 use DI for i386 and R13 for x86-64. We chose R13 since it has
8209 Return: the regno of chosen register. */
8212 find_drap_reg (void)
8214 tree decl = cfun->decl;
8218 /* Use R13 for nested function or function need static chain.
8219 Since function with tail call may use any caller-saved
8220 registers in epilogue, DRAP must not use caller-saved
8221 register in such case. */
8222 if (DECL_STATIC_CHAIN (decl) || crtl->tail_call_emit)
8229 /* Use DI for nested function or function need static chain.
8230 Since function with tail call may use any caller-saved
8231 registers in epilogue, DRAP must not use caller-saved
8232 register in such case. */
8233 if (DECL_STATIC_CHAIN (decl) || crtl->tail_call_emit)
8236 /* Reuse static chain register if it isn't used for parameter
8238 if (ix86_function_regparm (TREE_TYPE (decl), decl) <= 2
8239 && !lookup_attribute ("fastcall",
8240 TYPE_ATTRIBUTES (TREE_TYPE (decl))))
8247 /* Return minimum incoming stack alignment. */
8250 ix86_minimum_incoming_stack_boundary (bool sibcall)
8252 unsigned int incoming_stack_boundary;
8254 /* Prefer the one specified at command line. */
8255 if (ix86_user_incoming_stack_boundary)
8256 incoming_stack_boundary = ix86_user_incoming_stack_boundary;
8257 /* In 32bit, use MIN_STACK_BOUNDARY for incoming stack boundary
8258 if -mstackrealign is used, it isn't used for sibcall check and
8259 estimated stack alignment is 128bit. */
8262 && ix86_force_align_arg_pointer
8263 && crtl->stack_alignment_estimated == 128)
8264 incoming_stack_boundary = MIN_STACK_BOUNDARY;
8266 incoming_stack_boundary = ix86_default_incoming_stack_boundary;
8268 /* Incoming stack alignment can be changed on individual functions
8269 via force_align_arg_pointer attribute. We use the smallest
8270 incoming stack boundary. */
8271 if (incoming_stack_boundary > MIN_STACK_BOUNDARY
8272 && lookup_attribute (ix86_force_align_arg_pointer_string,
8273 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
8274 incoming_stack_boundary = MIN_STACK_BOUNDARY;
8276 /* The incoming stack frame has to be aligned at least at
8277 parm_stack_boundary. */
8278 if (incoming_stack_boundary < crtl->parm_stack_boundary)
8279 incoming_stack_boundary = crtl->parm_stack_boundary;
8281 /* Stack at entrance of main is aligned by runtime. We use the
8282 smallest incoming stack boundary. */
8283 if (incoming_stack_boundary > MAIN_STACK_BOUNDARY
8284 && DECL_NAME (current_function_decl)
8285 && MAIN_NAME_P (DECL_NAME (current_function_decl))
8286 && DECL_FILE_SCOPE_P (current_function_decl))
8287 incoming_stack_boundary = MAIN_STACK_BOUNDARY;
8289 return incoming_stack_boundary;
8292 /* Update incoming stack boundary and estimated stack alignment. */
8295 ix86_update_stack_boundary (void)
8297 ix86_incoming_stack_boundary
8298 = ix86_minimum_incoming_stack_boundary (false);
8300 /* x86_64 vararg needs 16byte stack alignment for register save
8304 && crtl->stack_alignment_estimated < 128)
8305 crtl->stack_alignment_estimated = 128;
8308 /* Handle the TARGET_GET_DRAP_RTX hook. Return NULL if no DRAP is
8309 needed or an rtx for DRAP otherwise. */
8312 ix86_get_drap_rtx (void)
8314 if (ix86_force_drap || !ACCUMULATE_OUTGOING_ARGS)
8315 crtl->need_drap = true;
8317 if (stack_realign_drap)
8319 /* Assign DRAP to vDRAP and returns vDRAP */
8320 unsigned int regno = find_drap_reg ();
8325 arg_ptr = gen_rtx_REG (Pmode, regno);
8326 crtl->drap_reg = arg_ptr;
8329 drap_vreg = copy_to_reg (arg_ptr);
8333 insn = emit_insn_before (seq, NEXT_INSN (entry_of_function ()));
8334 RTX_FRAME_RELATED_P (insn) = 1;
8341 /* Handle the TARGET_INTERNAL_ARG_POINTER hook. */
8344 ix86_internal_arg_pointer (void)
8346 return virtual_incoming_args_rtx;
8349 /* Finalize stack_realign_needed flag, which will guide prologue/epilogue
8350 to be generated in correct form. */
8352 ix86_finalize_stack_realign_flags (void)
8354 /* Check if stack realign is really needed after reload, and
8355 stores result in cfun */
8356 unsigned int incoming_stack_boundary
8357 = (crtl->parm_stack_boundary > ix86_incoming_stack_boundary
8358 ? crtl->parm_stack_boundary : ix86_incoming_stack_boundary);
8359 unsigned int stack_realign = (incoming_stack_boundary
8360 < (current_function_is_leaf
8361 ? crtl->max_used_stack_slot_alignment
8362 : crtl->stack_alignment_needed));
8364 if (crtl->stack_realign_finalized)
8366 /* After stack_realign_needed is finalized, we can't no longer
8368 gcc_assert (crtl->stack_realign_needed == stack_realign);
8372 crtl->stack_realign_needed = stack_realign;
8373 crtl->stack_realign_finalized = true;
8377 /* Expand the prologue into a bunch of separate insns. */
8380 ix86_expand_prologue (void)
8384 struct ix86_frame frame;
8385 HOST_WIDE_INT allocate;
8386 int gen_frame_pointer = frame_pointer_needed;
8388 ix86_finalize_stack_realign_flags ();
8390 /* DRAP should not coexist with stack_realign_fp */
8391 gcc_assert (!(crtl->drap_reg && stack_realign_fp));
8393 /* Initialize CFA state for before the prologue. */
8394 ix86_cfa_state->reg = stack_pointer_rtx;
8395 ix86_cfa_state->offset = INCOMING_FRAME_SP_OFFSET;
8397 ix86_compute_frame_layout (&frame);
8399 if (ix86_function_ms_hook_prologue (current_function_decl))
8403 /* Make sure the function starts with
8404 8b ff movl.s %edi,%edi
8406 8b ec movl.s %esp,%ebp
8408 This matches the hookable function prologue in Win32 API
8409 functions in Microsoft Windows XP Service Pack 2 and newer.
8410 Wine uses this to enable Windows apps to hook the Win32 API
8411 functions provided by Wine. */
8412 insn = emit_insn (gen_vswapmov (gen_rtx_REG (SImode, DI_REG),
8413 gen_rtx_REG (SImode, DI_REG)));
8414 push = emit_insn (gen_push (hard_frame_pointer_rtx));
8415 mov = emit_insn (gen_vswapmov (hard_frame_pointer_rtx,
8416 stack_pointer_rtx));
8418 if (frame_pointer_needed && !(crtl->drap_reg
8419 && crtl->stack_realign_needed))
8421 /* The push %ebp and movl.s %esp, %ebp already set up
8422 the frame pointer. No need to do this again. */
8423 gen_frame_pointer = 0;
8424 RTX_FRAME_RELATED_P (push) = 1;
8425 RTX_FRAME_RELATED_P (mov) = 1;
8426 if (ix86_cfa_state->reg == stack_pointer_rtx)
8427 ix86_cfa_state->reg = hard_frame_pointer_rtx;
8430 /* If the frame pointer is not needed, pop %ebp again. This
8431 could be optimized for cases where ebp needs to be backed up
8432 for some other reason. If stack realignment is needed, pop
8433 the base pointer again, align the stack, and later regenerate
8434 the frame pointer setup. The frame pointer generated by the
8435 hook prologue is not aligned, so it can't be used. */
8436 insn = emit_insn ((*ix86_gen_pop1) (hard_frame_pointer_rtx));
8439 /* The first insn of a function that accepts its static chain on the
8440 stack is to push the register that would be filled in by a direct
8441 call. This insn will be skipped by the trampoline. */
8442 if (ix86_static_chain_on_stack)
8446 insn = emit_insn (gen_push (ix86_static_chain (cfun->decl, false)));
8447 emit_insn (gen_blockage ());
8449 /* We don't want to interpret this push insn as a register save,
8450 only as a stack adjustment. The real copy of the register as
8451 a save will be done later, if needed. */
8452 t = plus_constant (stack_pointer_rtx, -UNITS_PER_WORD);
8453 t = gen_rtx_SET (VOIDmode, stack_pointer_rtx, t);
8454 add_reg_note (insn, REG_CFA_ADJUST_CFA, t);
8455 RTX_FRAME_RELATED_P (insn) = 1;
8458 /* Emit prologue code to adjust stack alignment and setup DRAP, in case
8459 of DRAP is needed and stack realignment is really needed after reload */
8460 if (crtl->drap_reg && crtl->stack_realign_needed)
8463 int align_bytes = crtl->stack_alignment_needed / BITS_PER_UNIT;
8464 int param_ptr_offset = UNITS_PER_WORD;
8466 if (ix86_static_chain_on_stack)
8467 param_ptr_offset += UNITS_PER_WORD;
8468 if (!call_used_regs[REGNO (crtl->drap_reg)])
8469 param_ptr_offset += UNITS_PER_WORD;
8471 gcc_assert (stack_realign_drap);
8473 /* Grab the argument pointer. */
8474 x = plus_constant (stack_pointer_rtx, param_ptr_offset);
8477 /* Only need to push parameter pointer reg if it is caller
8479 if (!call_used_regs[REGNO (crtl->drap_reg)])
8481 /* Push arg pointer reg */
8482 insn = emit_insn (gen_push (y));
8483 RTX_FRAME_RELATED_P (insn) = 1;
8486 insn = emit_insn (gen_rtx_SET (VOIDmode, y, x));
8487 RTX_FRAME_RELATED_P (insn) = 1;
8488 ix86_cfa_state->reg = crtl->drap_reg;
8490 /* Align the stack. */
8491 insn = emit_insn ((*ix86_gen_andsp) (stack_pointer_rtx,
8493 GEN_INT (-align_bytes)));
8494 RTX_FRAME_RELATED_P (insn) = 1;
8496 /* Replicate the return address on the stack so that return
8497 address can be reached via (argp - 1) slot. This is needed
8498 to implement macro RETURN_ADDR_RTX and intrinsic function
8499 expand_builtin_return_addr etc. */
8501 x = gen_frame_mem (Pmode,
8502 plus_constant (x, -UNITS_PER_WORD));
8503 insn = emit_insn (gen_push (x));
8504 RTX_FRAME_RELATED_P (insn) = 1;
8507 /* Note: AT&T enter does NOT have reversed args. Enter is probably
8508 slower on all targets. Also sdb doesn't like it. */
8510 if (gen_frame_pointer)
8512 insn = emit_insn (gen_push (hard_frame_pointer_rtx));
8513 RTX_FRAME_RELATED_P (insn) = 1;
8515 insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);
8516 RTX_FRAME_RELATED_P (insn) = 1;
8518 if (ix86_cfa_state->reg == stack_pointer_rtx)
8519 ix86_cfa_state->reg = hard_frame_pointer_rtx;
8522 if (stack_realign_fp)
8524 int align_bytes = crtl->stack_alignment_needed / BITS_PER_UNIT;
8525 gcc_assert (align_bytes > MIN_STACK_BOUNDARY / BITS_PER_UNIT);
8527 /* Align the stack. */
8528 insn = emit_insn ((*ix86_gen_andsp) (stack_pointer_rtx,
8530 GEN_INT (-align_bytes)));
8531 RTX_FRAME_RELATED_P (insn) = 1;
8534 allocate = frame.to_allocate + frame.nsseregs * 16 + frame.padding0;
8536 if (!frame.save_regs_using_mov)
8537 ix86_emit_save_regs ();
8539 allocate += frame.nregs * UNITS_PER_WORD;
8541 /* When using red zone we may start register saving before allocating
8542 the stack frame saving one cycle of the prologue. However I will
8543 avoid doing this if I am going to have to probe the stack since
8544 at least on x86_64 the stack probe can turn into a call that clobbers
8545 a red zone location */
8546 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE && frame.save_regs_using_mov
8547 && (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT))
8548 ix86_emit_save_regs_using_mov ((frame_pointer_needed
8549 && !crtl->stack_realign_needed)
8550 ? hard_frame_pointer_rtx
8551 : stack_pointer_rtx,
8552 -frame.nregs * UNITS_PER_WORD);
8556 else if (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT)
8557 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
8558 GEN_INT (-allocate), -1,
8559 ix86_cfa_state->reg == stack_pointer_rtx);
8562 /* Only valid for Win32. */
8563 rtx eax = gen_rtx_REG (Pmode, AX_REG);
8567 gcc_assert (!TARGET_64BIT || cfun->machine->call_abi == MS_ABI);
8569 if (cfun->machine->call_abi == MS_ABI)
8572 eax_live = ix86_eax_live_at_start_p ();
8576 emit_insn (gen_push (eax));
8577 allocate -= UNITS_PER_WORD;
8580 emit_move_insn (eax, GEN_INT (allocate));
8583 insn = gen_allocate_stack_worker_64 (eax, eax);
8585 insn = gen_allocate_stack_worker_32 (eax, eax);
8586 insn = emit_insn (insn);
8588 if (ix86_cfa_state->reg == stack_pointer_rtx)
8590 ix86_cfa_state->offset += allocate;
8591 t = gen_rtx_PLUS (Pmode, stack_pointer_rtx, GEN_INT (-allocate));
8592 t = gen_rtx_SET (VOIDmode, stack_pointer_rtx, t);
8593 add_reg_note (insn, REG_CFA_ADJUST_CFA, t);
8594 RTX_FRAME_RELATED_P (insn) = 1;
8599 if (frame_pointer_needed)
8600 t = plus_constant (hard_frame_pointer_rtx,
8603 - frame.nregs * UNITS_PER_WORD);
8605 t = plus_constant (stack_pointer_rtx, allocate);
8606 emit_move_insn (eax, gen_rtx_MEM (Pmode, t));
8610 if (frame.save_regs_using_mov
8611 && !(!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE
8612 && (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT)))
8614 if (!frame_pointer_needed
8615 || !(frame.to_allocate + frame.padding0)
8616 || crtl->stack_realign_needed)
8617 ix86_emit_save_regs_using_mov (stack_pointer_rtx,
8619 + frame.nsseregs * 16 + frame.padding0);
8621 ix86_emit_save_regs_using_mov (hard_frame_pointer_rtx,
8622 -frame.nregs * UNITS_PER_WORD);
8624 if (!frame_pointer_needed
8625 || !(frame.to_allocate + frame.padding0)
8626 || crtl->stack_realign_needed)
8627 ix86_emit_save_sse_regs_using_mov (stack_pointer_rtx,
8630 ix86_emit_save_sse_regs_using_mov (hard_frame_pointer_rtx,
8631 - frame.nregs * UNITS_PER_WORD
8632 - frame.nsseregs * 16
8635 pic_reg_used = false;
8636 if (pic_offset_table_rtx
8637 && (df_regs_ever_live_p (REAL_PIC_OFFSET_TABLE_REGNUM)
8640 unsigned int alt_pic_reg_used = ix86_select_alt_pic_regnum ();
8642 if (alt_pic_reg_used != INVALID_REGNUM)
8643 SET_REGNO (pic_offset_table_rtx, alt_pic_reg_used);
8645 pic_reg_used = true;
8652 if (ix86_cmodel == CM_LARGE_PIC)
8654 rtx tmp_reg = gen_rtx_REG (DImode, R11_REG);
8655 rtx label = gen_label_rtx ();
8657 LABEL_PRESERVE_P (label) = 1;
8658 gcc_assert (REGNO (pic_offset_table_rtx) != REGNO (tmp_reg));
8659 insn = emit_insn (gen_set_rip_rex64 (pic_offset_table_rtx, label));
8660 insn = emit_insn (gen_set_got_offset_rex64 (tmp_reg, label));
8661 insn = emit_insn (gen_adddi3 (pic_offset_table_rtx,
8662 pic_offset_table_rtx, tmp_reg));
8665 insn = emit_insn (gen_set_got_rex64 (pic_offset_table_rtx));
8668 insn = emit_insn (gen_set_got (pic_offset_table_rtx));
8671 /* In the pic_reg_used case, make sure that the got load isn't deleted
8672 when mcount needs it. Blockage to avoid call movement across mcount
8673 call is emitted in generic code after the NOTE_INSN_PROLOGUE_END
8675 if (crtl->profile && pic_reg_used)
8676 emit_insn (gen_prologue_use (pic_offset_table_rtx));
8678 if (crtl->drap_reg && !crtl->stack_realign_needed)
8680 /* vDRAP is setup but after reload it turns out stack realign
8681 isn't necessary, here we will emit prologue to setup DRAP
8682 without stack realign adjustment */
8684 int drap_bp_offset = UNITS_PER_WORD * 2;
8686 if (ix86_static_chain_on_stack)
8687 drap_bp_offset += UNITS_PER_WORD;
8688 x = plus_constant (hard_frame_pointer_rtx, drap_bp_offset);
8689 insn = emit_insn (gen_rtx_SET (VOIDmode, crtl->drap_reg, x));
8692 /* Prevent instructions from being scheduled into register save push
8693 sequence when access to the redzone area is done through frame pointer.
8694 The offset between the frame pointer and the stack pointer is calculated
8695 relative to the value of the stack pointer at the end of the function
8696 prologue, and moving instructions that access redzone area via frame
8697 pointer inside push sequence violates this assumption. */
8698 if (frame_pointer_needed && frame.red_zone_size)
8699 emit_insn (gen_memory_blockage ());
8701 /* Emit cld instruction if stringops are used in the function. */
8702 if (TARGET_CLD && ix86_current_function_needs_cld)
8703 emit_insn (gen_cld ());
8706 /* Emit code to restore REG using a POP insn. */
8709 ix86_emit_restore_reg_using_pop (rtx reg, HOST_WIDE_INT red_offset)
8711 rtx insn = emit_insn (ix86_gen_pop1 (reg));
8713 if (ix86_cfa_state->reg == crtl->drap_reg
8714 && REGNO (reg) == REGNO (crtl->drap_reg))
8716 /* Previously we'd represented the CFA as an expression
8717 like *(%ebp - 8). We've just popped that value from
8718 the stack, which means we need to reset the CFA to
8719 the drap register. This will remain until we restore
8720 the stack pointer. */
8721 add_reg_note (insn, REG_CFA_DEF_CFA, reg);
8722 RTX_FRAME_RELATED_P (insn) = 1;
8726 if (ix86_cfa_state->reg == stack_pointer_rtx)
8728 ix86_cfa_state->offset -= UNITS_PER_WORD;
8729 add_reg_note (insn, REG_CFA_ADJUST_CFA,
8730 copy_rtx (XVECEXP (PATTERN (insn), 0, 1)));
8731 RTX_FRAME_RELATED_P (insn) = 1;
8734 /* When the frame pointer is the CFA, and we pop it, we are
8735 swapping back to the stack pointer as the CFA. This happens
8736 for stack frames that don't allocate other data, so we assume
8737 the stack pointer is now pointing at the return address, i.e.
8738 the function entry state, which makes the offset be 1 word. */
8739 else if (ix86_cfa_state->reg == hard_frame_pointer_rtx
8740 && reg == hard_frame_pointer_rtx)
8742 ix86_cfa_state->reg = stack_pointer_rtx;
8743 ix86_cfa_state->offset -= UNITS_PER_WORD;
8745 add_reg_note (insn, REG_CFA_DEF_CFA,
8746 gen_rtx_PLUS (Pmode, stack_pointer_rtx,
8747 GEN_INT (ix86_cfa_state->offset)));
8748 RTX_FRAME_RELATED_P (insn) = 1;
8751 ix86_add_cfa_restore_note (insn, reg, red_offset);
8754 /* Emit code to restore saved registers using POP insns. */
8757 ix86_emit_restore_regs_using_pop (HOST_WIDE_INT red_offset)
8761 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8762 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, false))
8764 ix86_emit_restore_reg_using_pop (gen_rtx_REG (Pmode, regno),
8766 red_offset += UNITS_PER_WORD;
8770 /* Emit code and notes for the LEAVE instruction. */
8773 ix86_emit_leave (HOST_WIDE_INT red_offset)
8775 rtx insn = emit_insn (ix86_gen_leave ());
8777 ix86_add_queued_cfa_restore_notes (insn);
8779 if (ix86_cfa_state->reg == hard_frame_pointer_rtx)
8781 ix86_cfa_state->reg = stack_pointer_rtx;
8782 ix86_cfa_state->offset -= UNITS_PER_WORD;
8784 add_reg_note (insn, REG_CFA_ADJUST_CFA,
8785 copy_rtx (XVECEXP (PATTERN (insn), 0, 0)));
8786 RTX_FRAME_RELATED_P (insn) = 1;
8787 ix86_add_cfa_restore_note (insn, hard_frame_pointer_rtx, red_offset);
8791 /* Emit code to restore saved registers using MOV insns. First register
8792 is restored from POINTER + OFFSET. */
8794 ix86_emit_restore_regs_using_mov (rtx pointer, HOST_WIDE_INT offset,
8795 HOST_WIDE_INT red_offset,
8796 int maybe_eh_return)
8799 rtx base_address = gen_rtx_MEM (Pmode, pointer);
8802 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8803 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, maybe_eh_return))
8805 rtx reg = gen_rtx_REG (Pmode, regno);
8807 /* Ensure that adjust_address won't be forced to produce pointer
8808 out of range allowed by x86-64 instruction set. */
8809 if (TARGET_64BIT && offset != trunc_int_for_mode (offset, SImode))
8813 r11 = gen_rtx_REG (DImode, R11_REG);
8814 emit_move_insn (r11, GEN_INT (offset));
8815 emit_insn (gen_adddi3 (r11, r11, pointer));
8816 base_address = gen_rtx_MEM (Pmode, r11);
8819 insn = emit_move_insn (reg,
8820 adjust_address (base_address, Pmode, offset));
8821 offset += UNITS_PER_WORD;
8823 if (ix86_cfa_state->reg == crtl->drap_reg
8824 && regno == REGNO (crtl->drap_reg))
8826 /* Previously we'd represented the CFA as an expression
8827 like *(%ebp - 8). We've just popped that value from
8828 the stack, which means we need to reset the CFA to
8829 the drap register. This will remain until we restore
8830 the stack pointer. */
8831 add_reg_note (insn, REG_CFA_DEF_CFA, reg);
8832 RTX_FRAME_RELATED_P (insn) = 1;
8835 ix86_add_cfa_restore_note (NULL_RTX, reg, red_offset);
8837 red_offset += UNITS_PER_WORD;
8841 /* Emit code to restore saved registers using MOV insns. First register
8842 is restored from POINTER + OFFSET. */
8844 ix86_emit_restore_sse_regs_using_mov (rtx pointer, HOST_WIDE_INT offset,
8845 HOST_WIDE_INT red_offset,
8846 int maybe_eh_return)
8849 rtx base_address = gen_rtx_MEM (TImode, pointer);
8852 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8853 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, maybe_eh_return))
8855 rtx reg = gen_rtx_REG (TImode, regno);
8857 /* Ensure that adjust_address won't be forced to produce pointer
8858 out of range allowed by x86-64 instruction set. */
8859 if (TARGET_64BIT && offset != trunc_int_for_mode (offset, SImode))
8863 r11 = gen_rtx_REG (DImode, R11_REG);
8864 emit_move_insn (r11, GEN_INT (offset));
8865 emit_insn (gen_adddi3 (r11, r11, pointer));
8866 base_address = gen_rtx_MEM (TImode, r11);
8869 mem = adjust_address (base_address, TImode, offset);
8870 set_mem_align (mem, 128);
8871 emit_move_insn (reg, mem);
8874 ix86_add_cfa_restore_note (NULL_RTX, reg, red_offset);
8880 /* Restore function stack, frame, and registers. */
8883 ix86_expand_epilogue (int style)
8886 struct ix86_frame frame;
8887 HOST_WIDE_INT offset, red_offset;
8888 struct machine_cfa_state cfa_state_save = *ix86_cfa_state;
8891 ix86_finalize_stack_realign_flags ();
8893 /* When stack is realigned, SP must be valid. */
8894 sp_valid = (!frame_pointer_needed
8895 || current_function_sp_is_unchanging
8896 || stack_realign_fp);
8898 ix86_compute_frame_layout (&frame);
8900 /* See the comment about red zone and frame
8901 pointer usage in ix86_expand_prologue. */
8902 if (frame_pointer_needed && frame.red_zone_size)
8903 emit_insn (gen_memory_blockage ());
8905 using_drap = crtl->drap_reg && crtl->stack_realign_needed;
8906 gcc_assert (!using_drap || ix86_cfa_state->reg == crtl->drap_reg);
8908 /* Calculate start of saved registers relative to ebp. Special care
8909 must be taken for the normal return case of a function using
8910 eh_return: the eax and edx registers are marked as saved, but not
8911 restored along this path. */
8912 offset = frame.nregs;
8913 if (crtl->calls_eh_return && style != 2)
8915 offset *= -UNITS_PER_WORD;
8916 offset -= frame.nsseregs * 16 + frame.padding0;
8918 /* Calculate start of saved registers relative to esp on entry of the
8919 function. When realigning stack, this needs to be the most negative
8920 value possible at runtime. */
8921 red_offset = offset;
8923 red_offset -= crtl->stack_alignment_needed / BITS_PER_UNIT
8925 else if (stack_realign_fp)
8926 red_offset -= crtl->stack_alignment_needed / BITS_PER_UNIT
8928 if (ix86_static_chain_on_stack)
8929 red_offset -= UNITS_PER_WORD;
8930 if (frame_pointer_needed)
8931 red_offset -= UNITS_PER_WORD;
8933 /* If we're only restoring one register and sp is not valid then
8934 using a move instruction to restore the register since it's
8935 less work than reloading sp and popping the register.
8937 The default code result in stack adjustment using add/lea instruction,
8938 while this code results in LEAVE instruction (or discrete equivalent),
8939 so it is profitable in some other cases as well. Especially when there
8940 are no registers to restore. We also use this code when TARGET_USE_LEAVE
8941 and there is exactly one register to pop. This heuristic may need some
8942 tuning in future. */
8943 if ((!sp_valid && (frame.nregs + frame.nsseregs) <= 1)
8944 || (TARGET_EPILOGUE_USING_MOVE
8945 && cfun->machine->use_fast_prologue_epilogue
8946 && ((frame.nregs + frame.nsseregs) > 1
8947 || (frame.to_allocate + frame.padding0) != 0))
8948 || (frame_pointer_needed && !(frame.nregs + frame.nsseregs)
8949 && (frame.to_allocate + frame.padding0) != 0)
8950 || (frame_pointer_needed && TARGET_USE_LEAVE
8951 && cfun->machine->use_fast_prologue_epilogue
8952 && (frame.nregs + frame.nsseregs) == 1)
8953 || crtl->calls_eh_return)
8955 /* Restore registers. We can use ebp or esp to address the memory
8956 locations. If both are available, default to ebp, since offsets
8957 are known to be small. Only exception is esp pointing directly
8958 to the end of block of saved registers, where we may simplify
8961 If we are realigning stack with bp and sp, regs restore can't
8962 be addressed by bp. sp must be used instead. */
8964 if (!frame_pointer_needed
8965 || (sp_valid && !(frame.to_allocate + frame.padding0))
8966 || stack_realign_fp)
8968 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
8969 frame.to_allocate, red_offset,
8971 ix86_emit_restore_regs_using_mov (stack_pointer_rtx,
8973 + frame.nsseregs * 16
8976 + frame.nsseregs * 16
8977 + frame.padding0, style == 2);
8981 ix86_emit_restore_sse_regs_using_mov (hard_frame_pointer_rtx,
8984 ix86_emit_restore_regs_using_mov (hard_frame_pointer_rtx,
8986 + frame.nsseregs * 16
8989 + frame.nsseregs * 16
8990 + frame.padding0, style == 2);
8993 red_offset -= offset;
8995 /* eh_return epilogues need %ecx added to the stack pointer. */
8998 rtx tmp, sa = EH_RETURN_STACKADJ_RTX;
9000 /* Stack align doesn't work with eh_return. */
9001 gcc_assert (!crtl->stack_realign_needed);
9002 /* Neither does regparm nested functions. */
9003 gcc_assert (!ix86_static_chain_on_stack);
9005 if (frame_pointer_needed)
9007 tmp = gen_rtx_PLUS (Pmode, hard_frame_pointer_rtx, sa);
9008 tmp = plus_constant (tmp, UNITS_PER_WORD);
9009 tmp = emit_insn (gen_rtx_SET (VOIDmode, sa, tmp));
9011 tmp = gen_rtx_MEM (Pmode, hard_frame_pointer_rtx);
9012 tmp = emit_move_insn (hard_frame_pointer_rtx, tmp);
9014 /* Note that we use SA as a temporary CFA, as the return
9015 address is at the proper place relative to it. We
9016 pretend this happens at the FP restore insn because
9017 prior to this insn the FP would be stored at the wrong
9018 offset relative to SA, and after this insn we have no
9019 other reasonable register to use for the CFA. We don't
9020 bother resetting the CFA to the SP for the duration of
9022 add_reg_note (tmp, REG_CFA_DEF_CFA,
9023 plus_constant (sa, UNITS_PER_WORD));
9024 ix86_add_queued_cfa_restore_notes (tmp);
9025 add_reg_note (tmp, REG_CFA_RESTORE, hard_frame_pointer_rtx);
9026 RTX_FRAME_RELATED_P (tmp) = 1;
9027 ix86_cfa_state->reg = sa;
9028 ix86_cfa_state->offset = UNITS_PER_WORD;
9030 pro_epilogue_adjust_stack (stack_pointer_rtx, sa,
9031 const0_rtx, style, false);
9035 tmp = gen_rtx_PLUS (Pmode, stack_pointer_rtx, sa);
9036 tmp = plus_constant (tmp, (frame.to_allocate
9037 + frame.nregs * UNITS_PER_WORD
9038 + frame.nsseregs * 16
9040 tmp = emit_insn (gen_rtx_SET (VOIDmode, stack_pointer_rtx, tmp));
9041 ix86_add_queued_cfa_restore_notes (tmp);
9043 gcc_assert (ix86_cfa_state->reg == stack_pointer_rtx);
9044 if (ix86_cfa_state->offset != UNITS_PER_WORD)
9046 ix86_cfa_state->offset = UNITS_PER_WORD;
9047 add_reg_note (tmp, REG_CFA_DEF_CFA,
9048 plus_constant (stack_pointer_rtx,
9050 RTX_FRAME_RELATED_P (tmp) = 1;
9054 else if (!frame_pointer_needed)
9055 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
9056 GEN_INT (frame.to_allocate
9057 + frame.nregs * UNITS_PER_WORD
9058 + frame.nsseregs * 16
9060 style, !using_drap);
9061 /* If not an i386, mov & pop is faster than "leave". */
9062 else if (TARGET_USE_LEAVE || optimize_function_for_size_p (cfun)
9063 || !cfun->machine->use_fast_prologue_epilogue)
9064 ix86_emit_leave (red_offset);
9067 pro_epilogue_adjust_stack (stack_pointer_rtx,
9068 hard_frame_pointer_rtx,
9069 const0_rtx, style, !using_drap);
9071 ix86_emit_restore_reg_using_pop (hard_frame_pointer_rtx, red_offset);
9076 /* First step is to deallocate the stack frame so that we can
9079 If we realign stack with frame pointer, then stack pointer
9080 won't be able to recover via lea $offset(%bp), %sp, because
9081 there is a padding area between bp and sp for realign.
9082 "add $to_allocate, %sp" must be used instead. */
9085 gcc_assert (frame_pointer_needed);
9086 gcc_assert (!stack_realign_fp);
9087 pro_epilogue_adjust_stack (stack_pointer_rtx,
9088 hard_frame_pointer_rtx,
9089 GEN_INT (offset), style, false);
9090 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
9093 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
9094 GEN_INT (frame.nsseregs * 16
9098 else if (frame.to_allocate || frame.padding0 || frame.nsseregs)
9100 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
9101 frame.to_allocate, red_offset,
9103 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
9104 GEN_INT (frame.to_allocate
9105 + frame.nsseregs * 16
9106 + frame.padding0), style,
9107 !using_drap && !frame_pointer_needed);
9110 ix86_emit_restore_regs_using_pop (red_offset + frame.nsseregs * 16
9112 red_offset -= offset;
9114 if (frame_pointer_needed)
9116 /* Leave results in shorter dependency chains on CPUs that are
9117 able to grok it fast. */
9118 if (TARGET_USE_LEAVE)
9119 ix86_emit_leave (red_offset);
9122 /* For stack realigned really happens, recover stack
9123 pointer to hard frame pointer is a must, if not using
9125 if (stack_realign_fp)
9126 pro_epilogue_adjust_stack (stack_pointer_rtx,
9127 hard_frame_pointer_rtx,
9128 const0_rtx, style, !using_drap);
9129 ix86_emit_restore_reg_using_pop (hard_frame_pointer_rtx,
9137 int param_ptr_offset = UNITS_PER_WORD;
9140 gcc_assert (stack_realign_drap);
9142 if (ix86_static_chain_on_stack)
9143 param_ptr_offset += UNITS_PER_WORD;
9144 if (!call_used_regs[REGNO (crtl->drap_reg)])
9145 param_ptr_offset += UNITS_PER_WORD;
9147 insn = emit_insn ((*ix86_gen_add3) (stack_pointer_rtx,
9149 GEN_INT (-param_ptr_offset)));
9151 ix86_cfa_state->reg = stack_pointer_rtx;
9152 ix86_cfa_state->offset = param_ptr_offset;
9154 add_reg_note (insn, REG_CFA_DEF_CFA,
9155 gen_rtx_PLUS (Pmode, ix86_cfa_state->reg,
9156 GEN_INT (ix86_cfa_state->offset)));
9157 RTX_FRAME_RELATED_P (insn) = 1;
9159 if (!call_used_regs[REGNO (crtl->drap_reg)])
9160 ix86_emit_restore_reg_using_pop (crtl->drap_reg, -UNITS_PER_WORD);
9163 /* Remove the saved static chain from the stack. The use of ECX is
9164 merely as a scratch register, not as the actual static chain. */
9165 if (ix86_static_chain_on_stack)
9169 gcc_assert (ix86_cfa_state->reg == stack_pointer_rtx);
9170 ix86_cfa_state->offset += UNITS_PER_WORD;
9172 r = gen_rtx_REG (Pmode, CX_REG);
9173 insn = emit_insn (ix86_gen_pop1 (r));
9175 r = plus_constant (stack_pointer_rtx, UNITS_PER_WORD);
9176 r = gen_rtx_SET (VOIDmode, stack_pointer_rtx, r);
9177 add_reg_note (insn, REG_CFA_ADJUST_CFA, r);
9178 RTX_FRAME_RELATED_P (insn) = 1;
9181 /* Sibcall epilogues don't want a return instruction. */
9184 *ix86_cfa_state = cfa_state_save;
9188 if (crtl->args.pops_args && crtl->args.size)
9190 rtx popc = GEN_INT (crtl->args.pops_args);
9192 /* i386 can only pop 64K bytes. If asked to pop more, pop return
9193 address, do explicit add, and jump indirectly to the caller. */
9195 if (crtl->args.pops_args >= 65536)
9197 rtx ecx = gen_rtx_REG (SImode, CX_REG);
9200 /* There is no "pascal" calling convention in any 64bit ABI. */
9201 gcc_assert (!TARGET_64BIT);
9203 insn = emit_insn (gen_popsi1 (ecx));
9204 ix86_cfa_state->offset -= UNITS_PER_WORD;
9206 add_reg_note (insn, REG_CFA_ADJUST_CFA,
9207 copy_rtx (XVECEXP (PATTERN (insn), 0, 1)));
9208 add_reg_note (insn, REG_CFA_REGISTER,
9209 gen_rtx_SET (VOIDmode, ecx, pc_rtx));
9210 RTX_FRAME_RELATED_P (insn) = 1;
9212 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
9214 emit_jump_insn (gen_return_indirect_internal (ecx));
9217 emit_jump_insn (gen_return_pop_internal (popc));
9220 emit_jump_insn (gen_return_internal ());
9222 /* Restore the state back to the state from the prologue,
9223 so that it's correct for the next epilogue. */
9224 *ix86_cfa_state = cfa_state_save;
9227 /* Reset from the function's potential modifications. */
9230 ix86_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
9231 HOST_WIDE_INT size ATTRIBUTE_UNUSED)
9233 if (pic_offset_table_rtx)
9234 SET_REGNO (pic_offset_table_rtx, REAL_PIC_OFFSET_TABLE_REGNUM);
9236 /* Mach-O doesn't support labels at the end of objects, so if
9237 it looks like we might want one, insert a NOP. */
9239 rtx insn = get_last_insn ();
9242 && NOTE_KIND (insn) != NOTE_INSN_DELETED_LABEL)
9243 insn = PREV_INSN (insn);
9247 && NOTE_KIND (insn) == NOTE_INSN_DELETED_LABEL)))
9248 fputs ("\tnop\n", file);
9254 /* Extract the parts of an RTL expression that is a valid memory address
9255 for an instruction. Return 0 if the structure of the address is
9256 grossly off. Return -1 if the address contains ASHIFT, so it is not
9257 strictly valid, but still used for computing length of lea instruction. */
9260 ix86_decompose_address (rtx addr, struct ix86_address *out)
9262 rtx base = NULL_RTX, index = NULL_RTX, disp = NULL_RTX;
9263 rtx base_reg, index_reg;
9264 HOST_WIDE_INT scale = 1;
9265 rtx scale_rtx = NULL_RTX;
9267 enum ix86_address_seg seg = SEG_DEFAULT;
9269 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
9271 else if (GET_CODE (addr) == PLUS)
9281 addends[n++] = XEXP (op, 1);
9284 while (GET_CODE (op) == PLUS);
9289 for (i = n; i >= 0; --i)
9292 switch (GET_CODE (op))
9297 index = XEXP (op, 0);
9298 scale_rtx = XEXP (op, 1);
9302 if (XINT (op, 1) == UNSPEC_TP
9303 && TARGET_TLS_DIRECT_SEG_REFS
9304 && seg == SEG_DEFAULT)
9305 seg = TARGET_64BIT ? SEG_FS : SEG_GS;
9334 else if (GET_CODE (addr) == MULT)
9336 index = XEXP (addr, 0); /* index*scale */
9337 scale_rtx = XEXP (addr, 1);
9339 else if (GET_CODE (addr) == ASHIFT)
9343 /* We're called for lea too, which implements ashift on occasion. */
9344 index = XEXP (addr, 0);
9345 tmp = XEXP (addr, 1);
9346 if (!CONST_INT_P (tmp))
9348 scale = INTVAL (tmp);
9349 if ((unsigned HOST_WIDE_INT) scale > 3)
9355 disp = addr; /* displacement */
9357 /* Extract the integral value of scale. */
9360 if (!CONST_INT_P (scale_rtx))
9362 scale = INTVAL (scale_rtx);
9365 base_reg = base && GET_CODE (base) == SUBREG ? SUBREG_REG (base) : base;
9366 index_reg = index && GET_CODE (index) == SUBREG ? SUBREG_REG (index) : index;
9368 /* Avoid useless 0 displacement. */
9369 if (disp == const0_rtx && (base || index))
9372 /* Allow arg pointer and stack pointer as index if there is not scaling. */
9373 if (base_reg && index_reg && scale == 1
9374 && (index_reg == arg_pointer_rtx
9375 || index_reg == frame_pointer_rtx
9376 || (REG_P (index_reg) && REGNO (index_reg) == STACK_POINTER_REGNUM)))
9379 tmp = base, base = index, index = tmp;
9380 tmp = base_reg, base_reg = index_reg, index_reg = tmp;
9383 /* Special case: %ebp cannot be encoded as a base without a displacement.
9387 && (base_reg == hard_frame_pointer_rtx
9388 || base_reg == frame_pointer_rtx
9389 || base_reg == arg_pointer_rtx
9390 || (REG_P (base_reg)
9391 && (REGNO (base_reg) == HARD_FRAME_POINTER_REGNUM
9392 || REGNO (base_reg) == R13_REG))))
9395 /* Special case: on K6, [%esi] makes the instruction vector decoded.
9396 Avoid this by transforming to [%esi+0].
9397 Reload calls address legitimization without cfun defined, so we need
9398 to test cfun for being non-NULL. */
9399 if (TARGET_K6 && cfun && optimize_function_for_speed_p (cfun)
9400 && base_reg && !index_reg && !disp
9402 && REGNO_REG_CLASS (REGNO (base_reg)) == SIREG)
9405 /* Special case: encode reg+reg instead of reg*2. */
9406 if (!base && index && scale == 2)
9407 base = index, base_reg = index_reg, scale = 1;
9409 /* Special case: scaling cannot be encoded without base or displacement. */
9410 if (!base && !disp && index && scale != 1)
9422 /* Return cost of the memory address x.
9423 For i386, it is better to use a complex address than let gcc copy
9424 the address into a reg and make a new pseudo. But not if the address
9425 requires to two regs - that would mean more pseudos with longer
9428 ix86_address_cost (rtx x, bool speed ATTRIBUTE_UNUSED)
9430 struct ix86_address parts;
9432 int ok = ix86_decompose_address (x, &parts);
9436 if (parts.base && GET_CODE (parts.base) == SUBREG)
9437 parts.base = SUBREG_REG (parts.base);
9438 if (parts.index && GET_CODE (parts.index) == SUBREG)
9439 parts.index = SUBREG_REG (parts.index);
9441 /* Attempt to minimize number of registers in the address. */
9443 && (!REG_P (parts.base) || REGNO (parts.base) >= FIRST_PSEUDO_REGISTER))
9445 && (!REG_P (parts.index)
9446 || REGNO (parts.index) >= FIRST_PSEUDO_REGISTER)))
9450 && (!REG_P (parts.base) || REGNO (parts.base) >= FIRST_PSEUDO_REGISTER)
9452 && (!REG_P (parts.index) || REGNO (parts.index) >= FIRST_PSEUDO_REGISTER)
9453 && parts.base != parts.index)
9456 /* AMD-K6 don't like addresses with ModR/M set to 00_xxx_100b,
9457 since it's predecode logic can't detect the length of instructions
9458 and it degenerates to vector decoded. Increase cost of such
9459 addresses here. The penalty is minimally 2 cycles. It may be worthwhile
9460 to split such addresses or even refuse such addresses at all.
9462 Following addressing modes are affected:
9467 The first and last case may be avoidable by explicitly coding the zero in
9468 memory address, but I don't have AMD-K6 machine handy to check this
9472 && ((!parts.disp && parts.base && parts.index && parts.scale != 1)
9473 || (parts.disp && !parts.base && parts.index && parts.scale != 1)
9474 || (!parts.disp && parts.base && parts.index && parts.scale == 1)))
9480 /* Allow {LABEL | SYMBOL}_REF - SYMBOL_REF-FOR-PICBASE for Mach-O as
9481 this is used for to form addresses to local data when -fPIC is in
9485 darwin_local_data_pic (rtx disp)
9487 return (GET_CODE (disp) == UNSPEC
9488 && XINT (disp, 1) == UNSPEC_MACHOPIC_OFFSET);
9491 /* Determine if a given RTX is a valid constant. We already know this
9492 satisfies CONSTANT_P. */
9495 legitimate_constant_p (rtx x)
9497 switch (GET_CODE (x))
9502 if (GET_CODE (x) == PLUS)
9504 if (!CONST_INT_P (XEXP (x, 1)))
9509 if (TARGET_MACHO && darwin_local_data_pic (x))
9512 /* Only some unspecs are valid as "constants". */
9513 if (GET_CODE (x) == UNSPEC)
9514 switch (XINT (x, 1))
9519 return TARGET_64BIT;
9522 x = XVECEXP (x, 0, 0);
9523 return (GET_CODE (x) == SYMBOL_REF
9524 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_EXEC);
9526 x = XVECEXP (x, 0, 0);
9527 return (GET_CODE (x) == SYMBOL_REF
9528 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC);
9533 /* We must have drilled down to a symbol. */
9534 if (GET_CODE (x) == LABEL_REF)
9536 if (GET_CODE (x) != SYMBOL_REF)
9541 /* TLS symbols are never valid. */
9542 if (SYMBOL_REF_TLS_MODEL (x))
9545 /* DLLIMPORT symbols are never valid. */
9546 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
9547 && SYMBOL_REF_DLLIMPORT_P (x))
9552 if (GET_MODE (x) == TImode
9553 && x != CONST0_RTX (TImode)
9559 if (!standard_sse_constant_p (x))
9566 /* Otherwise we handle everything else in the move patterns. */
9570 /* Determine if it's legal to put X into the constant pool. This
9571 is not possible for the address of thread-local symbols, which
9572 is checked above. */
9575 ix86_cannot_force_const_mem (rtx x)
9577 /* We can always put integral constants and vectors in memory. */
9578 switch (GET_CODE (x))
9588 return !legitimate_constant_p (x);
9592 /* Nonzero if the constant value X is a legitimate general operand
9593 when generating PIC code. It is given that flag_pic is on and
9594 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
9597 legitimate_pic_operand_p (rtx x)
9601 switch (GET_CODE (x))
9604 inner = XEXP (x, 0);
9605 if (GET_CODE (inner) == PLUS
9606 && CONST_INT_P (XEXP (inner, 1)))
9607 inner = XEXP (inner, 0);
9609 /* Only some unspecs are valid as "constants". */
9610 if (GET_CODE (inner) == UNSPEC)
9611 switch (XINT (inner, 1))
9616 return TARGET_64BIT;
9618 x = XVECEXP (inner, 0, 0);
9619 return (GET_CODE (x) == SYMBOL_REF
9620 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_EXEC);
9621 case UNSPEC_MACHOPIC_OFFSET:
9622 return legitimate_pic_address_disp_p (x);
9630 return legitimate_pic_address_disp_p (x);
9637 /* Determine if a given CONST RTX is a valid memory displacement
9641 legitimate_pic_address_disp_p (rtx disp)
9645 /* In 64bit mode we can allow direct addresses of symbols and labels
9646 when they are not dynamic symbols. */
9649 rtx op0 = disp, op1;
9651 switch (GET_CODE (disp))
9657 if (GET_CODE (XEXP (disp, 0)) != PLUS)
9659 op0 = XEXP (XEXP (disp, 0), 0);
9660 op1 = XEXP (XEXP (disp, 0), 1);
9661 if (!CONST_INT_P (op1)
9662 || INTVAL (op1) >= 16*1024*1024
9663 || INTVAL (op1) < -16*1024*1024)
9665 if (GET_CODE (op0) == LABEL_REF)
9667 if (GET_CODE (op0) != SYMBOL_REF)
9672 /* TLS references should always be enclosed in UNSPEC. */
9673 if (SYMBOL_REF_TLS_MODEL (op0))
9675 if (!SYMBOL_REF_FAR_ADDR_P (op0) && SYMBOL_REF_LOCAL_P (op0)
9676 && ix86_cmodel != CM_LARGE_PIC)
9684 if (GET_CODE (disp) != CONST)
9686 disp = XEXP (disp, 0);
9690 /* We are unsafe to allow PLUS expressions. This limit allowed distance
9691 of GOT tables. We should not need these anyway. */
9692 if (GET_CODE (disp) != UNSPEC
9693 || (XINT (disp, 1) != UNSPEC_GOTPCREL
9694 && XINT (disp, 1) != UNSPEC_GOTOFF
9695 && XINT (disp, 1) != UNSPEC_PLTOFF))
9698 if (GET_CODE (XVECEXP (disp, 0, 0)) != SYMBOL_REF
9699 && GET_CODE (XVECEXP (disp, 0, 0)) != LABEL_REF)
9705 if (GET_CODE (disp) == PLUS)
9707 if (!CONST_INT_P (XEXP (disp, 1)))
9709 disp = XEXP (disp, 0);
9713 if (TARGET_MACHO && darwin_local_data_pic (disp))
9716 if (GET_CODE (disp) != UNSPEC)
9719 switch (XINT (disp, 1))
9724 /* We need to check for both symbols and labels because VxWorks loads
9725 text labels with @GOT rather than @GOTOFF. See gotoff_operand for
9727 return (GET_CODE (XVECEXP (disp, 0, 0)) == SYMBOL_REF
9728 || GET_CODE (XVECEXP (disp, 0, 0)) == LABEL_REF);
9730 /* Refuse GOTOFF in 64bit mode since it is always 64bit when used.
9731 While ABI specify also 32bit relocation but we don't produce it in
9732 small PIC model at all. */
9733 if ((GET_CODE (XVECEXP (disp, 0, 0)) == SYMBOL_REF
9734 || GET_CODE (XVECEXP (disp, 0, 0)) == LABEL_REF)
9736 return gotoff_operand (XVECEXP (disp, 0, 0), Pmode);
9738 case UNSPEC_GOTTPOFF:
9739 case UNSPEC_GOTNTPOFF:
9740 case UNSPEC_INDNTPOFF:
9743 disp = XVECEXP (disp, 0, 0);
9744 return (GET_CODE (disp) == SYMBOL_REF
9745 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_INITIAL_EXEC);
9747 disp = XVECEXP (disp, 0, 0);
9748 return (GET_CODE (disp) == SYMBOL_REF
9749 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_LOCAL_EXEC);
9751 disp = XVECEXP (disp, 0, 0);
9752 return (GET_CODE (disp) == SYMBOL_REF
9753 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_LOCAL_DYNAMIC);
9759 /* Recognizes RTL expressions that are valid memory addresses for an
9760 instruction. The MODE argument is the machine mode for the MEM
9761 expression that wants to use this address.
9763 It only recognizes address in canonical form. LEGITIMIZE_ADDRESS should
9764 convert common non-canonical forms to canonical form so that they will
9768 ix86_legitimate_address_p (enum machine_mode mode ATTRIBUTE_UNUSED,
9769 rtx addr, bool strict)
9771 struct ix86_address parts;
9772 rtx base, index, disp;
9773 HOST_WIDE_INT scale;
9775 if (ix86_decompose_address (addr, &parts) <= 0)
9776 /* Decomposition failed. */
9780 index = parts.index;
9782 scale = parts.scale;
9784 /* Validate base register.
9786 Don't allow SUBREG's that span more than a word here. It can lead to spill
9787 failures when the base is one word out of a two word structure, which is
9788 represented internally as a DImode int. */
9796 else if (GET_CODE (base) == SUBREG
9797 && REG_P (SUBREG_REG (base))
9798 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (base)))
9800 reg = SUBREG_REG (base);
9802 /* Base is not a register. */
9805 if (GET_MODE (base) != Pmode)
9806 /* Base is not in Pmode. */
9809 if ((strict && ! REG_OK_FOR_BASE_STRICT_P (reg))
9810 || (! strict && ! REG_OK_FOR_BASE_NONSTRICT_P (reg)))
9811 /* Base is not valid. */
9815 /* Validate index register.
9817 Don't allow SUBREG's that span more than a word here -- same as above. */
9825 else if (GET_CODE (index) == SUBREG
9826 && REG_P (SUBREG_REG (index))
9827 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (index)))
9829 reg = SUBREG_REG (index);
9831 /* Index is not a register. */
9834 if (GET_MODE (index) != Pmode)
9835 /* Index is not in Pmode. */
9838 if ((strict && ! REG_OK_FOR_INDEX_STRICT_P (reg))
9839 || (! strict && ! REG_OK_FOR_INDEX_NONSTRICT_P (reg)))
9840 /* Index is not valid. */
9844 /* Validate scale factor. */
9848 /* Scale without index. */
9851 if (scale != 2 && scale != 4 && scale != 8)
9852 /* Scale is not a valid multiplier. */
9856 /* Validate displacement. */
9859 if (GET_CODE (disp) == CONST
9860 && GET_CODE (XEXP (disp, 0)) == UNSPEC
9861 && XINT (XEXP (disp, 0), 1) != UNSPEC_MACHOPIC_OFFSET)
9862 switch (XINT (XEXP (disp, 0), 1))
9864 /* Refuse GOTOFF and GOT in 64bit mode since it is always 64bit when
9865 used. While ABI specify also 32bit relocations, we don't produce
9866 them at all and use IP relative instead. */
9869 gcc_assert (flag_pic);
9871 goto is_legitimate_pic;
9873 /* 64bit address unspec. */
9876 case UNSPEC_GOTPCREL:
9877 gcc_assert (flag_pic);
9878 goto is_legitimate_pic;
9880 case UNSPEC_GOTTPOFF:
9881 case UNSPEC_GOTNTPOFF:
9882 case UNSPEC_INDNTPOFF:
9888 /* Invalid address unspec. */
9892 else if (SYMBOLIC_CONST (disp)
9896 && MACHOPIC_INDIRECT
9897 && !machopic_operand_p (disp)
9903 if (TARGET_64BIT && (index || base))
9905 /* foo@dtpoff(%rX) is ok. */
9906 if (GET_CODE (disp) != CONST
9907 || GET_CODE (XEXP (disp, 0)) != PLUS
9908 || GET_CODE (XEXP (XEXP (disp, 0), 0)) != UNSPEC
9909 || !CONST_INT_P (XEXP (XEXP (disp, 0), 1))
9910 || (XINT (XEXP (XEXP (disp, 0), 0), 1) != UNSPEC_DTPOFF
9911 && XINT (XEXP (XEXP (disp, 0), 0), 1) != UNSPEC_NTPOFF))
9912 /* Non-constant pic memory reference. */
9915 else if (! legitimate_pic_address_disp_p (disp))
9916 /* Displacement is an invalid pic construct. */
9919 /* This code used to verify that a symbolic pic displacement
9920 includes the pic_offset_table_rtx register.
9922 While this is good idea, unfortunately these constructs may
9923 be created by "adds using lea" optimization for incorrect
9932 This code is nonsensical, but results in addressing
9933 GOT table with pic_offset_table_rtx base. We can't
9934 just refuse it easily, since it gets matched by
9935 "addsi3" pattern, that later gets split to lea in the
9936 case output register differs from input. While this
9937 can be handled by separate addsi pattern for this case
9938 that never results in lea, this seems to be easier and
9939 correct fix for crash to disable this test. */
9941 else if (GET_CODE (disp) != LABEL_REF
9942 && !CONST_INT_P (disp)
9943 && (GET_CODE (disp) != CONST
9944 || !legitimate_constant_p (disp))
9945 && (GET_CODE (disp) != SYMBOL_REF
9946 || !legitimate_constant_p (disp)))
9947 /* Displacement is not constant. */
9949 else if (TARGET_64BIT
9950 && !x86_64_immediate_operand (disp, VOIDmode))
9951 /* Displacement is out of range. */
9955 /* Everything looks valid. */
9959 /* Determine if a given RTX is a valid constant address. */
9962 constant_address_p (rtx x)
9964 return CONSTANT_P (x) && ix86_legitimate_address_p (Pmode, x, 1);
9967 /* Return a unique alias set for the GOT. */
9969 static alias_set_type
9970 ix86_GOT_alias_set (void)
9972 static alias_set_type set = -1;
9974 set = new_alias_set ();
9978 /* Return a legitimate reference for ORIG (an address) using the
9979 register REG. If REG is 0, a new pseudo is generated.
9981 There are two types of references that must be handled:
9983 1. Global data references must load the address from the GOT, via
9984 the PIC reg. An insn is emitted to do this load, and the reg is
9987 2. Static data references, constant pool addresses, and code labels
9988 compute the address as an offset from the GOT, whose base is in
9989 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
9990 differentiate them from global data objects. The returned
9991 address is the PIC reg + an unspec constant.
9993 TARGET_LEGITIMATE_ADDRESS_P rejects symbolic references unless the PIC
9994 reg also appears in the address. */
9997 legitimize_pic_address (rtx orig, rtx reg)
10000 rtx new_rtx = orig;
10004 if (TARGET_MACHO && !TARGET_64BIT)
10007 reg = gen_reg_rtx (Pmode);
10008 /* Use the generic Mach-O PIC machinery. */
10009 return machopic_legitimize_pic_address (orig, GET_MODE (orig), reg);
10013 if (TARGET_64BIT && legitimate_pic_address_disp_p (addr))
10015 else if (TARGET_64BIT
10016 && ix86_cmodel != CM_SMALL_PIC
10017 && gotoff_operand (addr, Pmode))
10020 /* This symbol may be referenced via a displacement from the PIC
10021 base address (@GOTOFF). */
10023 if (reload_in_progress)
10024 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
10025 if (GET_CODE (addr) == CONST)
10026 addr = XEXP (addr, 0);
10027 if (GET_CODE (addr) == PLUS)
10029 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, XEXP (addr, 0)),
10031 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, XEXP (addr, 1));
10034 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTOFF);
10035 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
10037 tmpreg = gen_reg_rtx (Pmode);
10040 emit_move_insn (tmpreg, new_rtx);
10044 new_rtx = expand_simple_binop (Pmode, PLUS, reg, pic_offset_table_rtx,
10045 tmpreg, 1, OPTAB_DIRECT);
10048 else new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, tmpreg);
10050 else if (!TARGET_64BIT && gotoff_operand (addr, Pmode))
10052 /* This symbol may be referenced via a displacement from the PIC
10053 base address (@GOTOFF). */
10055 if (reload_in_progress)
10056 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
10057 if (GET_CODE (addr) == CONST)
10058 addr = XEXP (addr, 0);
10059 if (GET_CODE (addr) == PLUS)
10061 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, XEXP (addr, 0)),
10063 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, XEXP (addr, 1));
10066 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTOFF);
10067 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
10068 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
10072 emit_move_insn (reg, new_rtx);
10076 else if ((GET_CODE (addr) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (addr) == 0)
10077 /* We can't use @GOTOFF for text labels on VxWorks;
10078 see gotoff_operand. */
10079 || (TARGET_VXWORKS_RTP && GET_CODE (addr) == LABEL_REF))
10081 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES)
10083 if (GET_CODE (addr) == SYMBOL_REF && SYMBOL_REF_DLLIMPORT_P (addr))
10084 return legitimize_dllimport_symbol (addr, true);
10085 if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
10086 && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF
10087 && SYMBOL_REF_DLLIMPORT_P (XEXP (XEXP (addr, 0), 0)))
10089 rtx t = legitimize_dllimport_symbol (XEXP (XEXP (addr, 0), 0), true);
10090 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (addr, 0), 1));
10094 if (TARGET_64BIT && ix86_cmodel != CM_LARGE_PIC)
10096 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTPCREL);
10097 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
10098 new_rtx = gen_const_mem (Pmode, new_rtx);
10099 set_mem_alias_set (new_rtx, ix86_GOT_alias_set ());
10102 reg = gen_reg_rtx (Pmode);
10103 /* Use directly gen_movsi, otherwise the address is loaded
10104 into register for CSE. We don't want to CSE this addresses,
10105 instead we CSE addresses from the GOT table, so skip this. */
10106 emit_insn (gen_movsi (reg, new_rtx));
10111 /* This symbol must be referenced via a load from the
10112 Global Offset Table (@GOT). */
10114 if (reload_in_progress)
10115 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
10116 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOT);
10117 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
10119 new_rtx = force_reg (Pmode, new_rtx);
10120 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
10121 new_rtx = gen_const_mem (Pmode, new_rtx);
10122 set_mem_alias_set (new_rtx, ix86_GOT_alias_set ());
10125 reg = gen_reg_rtx (Pmode);
10126 emit_move_insn (reg, new_rtx);
10132 if (CONST_INT_P (addr)
10133 && !x86_64_immediate_operand (addr, VOIDmode))
10137 emit_move_insn (reg, addr);
10141 new_rtx = force_reg (Pmode, addr);
10143 else if (GET_CODE (addr) == CONST)
10145 addr = XEXP (addr, 0);
10147 /* We must match stuff we generate before. Assume the only
10148 unspecs that can get here are ours. Not that we could do
10149 anything with them anyway.... */
10150 if (GET_CODE (addr) == UNSPEC
10151 || (GET_CODE (addr) == PLUS
10152 && GET_CODE (XEXP (addr, 0)) == UNSPEC))
10154 gcc_assert (GET_CODE (addr) == PLUS);
10156 if (GET_CODE (addr) == PLUS)
10158 rtx op0 = XEXP (addr, 0), op1 = XEXP (addr, 1);
10160 /* Check first to see if this is a constant offset from a @GOTOFF
10161 symbol reference. */
10162 if (gotoff_operand (op0, Pmode)
10163 && CONST_INT_P (op1))
10167 if (reload_in_progress)
10168 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
10169 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, op0),
10171 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, op1);
10172 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
10173 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
10177 emit_move_insn (reg, new_rtx);
10183 if (INTVAL (op1) < -16*1024*1024
10184 || INTVAL (op1) >= 16*1024*1024)
10186 if (!x86_64_immediate_operand (op1, Pmode))
10187 op1 = force_reg (Pmode, op1);
10188 new_rtx = gen_rtx_PLUS (Pmode, force_reg (Pmode, op0), op1);
10194 base = legitimize_pic_address (XEXP (addr, 0), reg);
10195 new_rtx = legitimize_pic_address (XEXP (addr, 1),
10196 base == reg ? NULL_RTX : reg);
10198 if (CONST_INT_P (new_rtx))
10199 new_rtx = plus_constant (base, INTVAL (new_rtx));
10202 if (GET_CODE (new_rtx) == PLUS && CONSTANT_P (XEXP (new_rtx, 1)))
10204 base = gen_rtx_PLUS (Pmode, base, XEXP (new_rtx, 0));
10205 new_rtx = XEXP (new_rtx, 1);
10207 new_rtx = gen_rtx_PLUS (Pmode, base, new_rtx);
10215 /* Load the thread pointer. If TO_REG is true, force it into a register. */
10218 get_thread_pointer (int to_reg)
10222 tp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), UNSPEC_TP);
10226 reg = gen_reg_rtx (Pmode);
10227 insn = gen_rtx_SET (VOIDmode, reg, tp);
10228 insn = emit_insn (insn);
10233 /* A subroutine of ix86_legitimize_address and ix86_expand_move. FOR_MOV is
10234 false if we expect this to be used for a memory address and true if
10235 we expect to load the address into a register. */
10238 legitimize_tls_address (rtx x, enum tls_model model, int for_mov)
10240 rtx dest, base, off, pic, tp;
10245 case TLS_MODEL_GLOBAL_DYNAMIC:
10246 dest = gen_reg_rtx (Pmode);
10247 tp = TARGET_GNU2_TLS ? get_thread_pointer (1) : 0;
10249 if (TARGET_64BIT && ! TARGET_GNU2_TLS)
10251 rtx rax = gen_rtx_REG (Pmode, AX_REG), insns;
10254 emit_call_insn (gen_tls_global_dynamic_64 (rax, x));
10255 insns = get_insns ();
10258 RTL_CONST_CALL_P (insns) = 1;
10259 emit_libcall_block (insns, dest, rax, x);
10261 else if (TARGET_64BIT && TARGET_GNU2_TLS)
10262 emit_insn (gen_tls_global_dynamic_64 (dest, x));
10264 emit_insn (gen_tls_global_dynamic_32 (dest, x));
10266 if (TARGET_GNU2_TLS)
10268 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, tp, dest));
10270 set_unique_reg_note (get_last_insn (), REG_EQUIV, x);
10274 case TLS_MODEL_LOCAL_DYNAMIC:
10275 base = gen_reg_rtx (Pmode);
10276 tp = TARGET_GNU2_TLS ? get_thread_pointer (1) : 0;
10278 if (TARGET_64BIT && ! TARGET_GNU2_TLS)
10280 rtx rax = gen_rtx_REG (Pmode, AX_REG), insns, note;
10283 emit_call_insn (gen_tls_local_dynamic_base_64 (rax));
10284 insns = get_insns ();
10287 note = gen_rtx_EXPR_LIST (VOIDmode, const0_rtx, NULL);
10288 note = gen_rtx_EXPR_LIST (VOIDmode, ix86_tls_get_addr (), note);
10289 RTL_CONST_CALL_P (insns) = 1;
10290 emit_libcall_block (insns, base, rax, note);
10292 else if (TARGET_64BIT && TARGET_GNU2_TLS)
10293 emit_insn (gen_tls_local_dynamic_base_64 (base));
10295 emit_insn (gen_tls_local_dynamic_base_32 (base));
10297 if (TARGET_GNU2_TLS)
10299 rtx x = ix86_tls_module_base ();
10301 set_unique_reg_note (get_last_insn (), REG_EQUIV,
10302 gen_rtx_MINUS (Pmode, x, tp));
10305 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), UNSPEC_DTPOFF);
10306 off = gen_rtx_CONST (Pmode, off);
10308 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, base, off));
10310 if (TARGET_GNU2_TLS)
10312 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, dest, tp));
10314 set_unique_reg_note (get_last_insn (), REG_EQUIV, x);
10319 case TLS_MODEL_INITIAL_EXEC:
10323 type = UNSPEC_GOTNTPOFF;
10327 if (reload_in_progress)
10328 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
10329 pic = pic_offset_table_rtx;
10330 type = TARGET_ANY_GNU_TLS ? UNSPEC_GOTNTPOFF : UNSPEC_GOTTPOFF;
10332 else if (!TARGET_ANY_GNU_TLS)
10334 pic = gen_reg_rtx (Pmode);
10335 emit_insn (gen_set_got (pic));
10336 type = UNSPEC_GOTTPOFF;
10341 type = UNSPEC_INDNTPOFF;
10344 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), type);
10345 off = gen_rtx_CONST (Pmode, off);
10347 off = gen_rtx_PLUS (Pmode, pic, off);
10348 off = gen_const_mem (Pmode, off);
10349 set_mem_alias_set (off, ix86_GOT_alias_set ());
10351 if (TARGET_64BIT || TARGET_ANY_GNU_TLS)
10353 base = get_thread_pointer (for_mov || !TARGET_TLS_DIRECT_SEG_REFS);
10354 off = force_reg (Pmode, off);
10355 return gen_rtx_PLUS (Pmode, base, off);
10359 base = get_thread_pointer (true);
10360 dest = gen_reg_rtx (Pmode);
10361 emit_insn (gen_subsi3 (dest, base, off));
10365 case TLS_MODEL_LOCAL_EXEC:
10366 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x),
10367 (TARGET_64BIT || TARGET_ANY_GNU_TLS)
10368 ? UNSPEC_NTPOFF : UNSPEC_TPOFF);
10369 off = gen_rtx_CONST (Pmode, off);
10371 if (TARGET_64BIT || TARGET_ANY_GNU_TLS)
10373 base = get_thread_pointer (for_mov || !TARGET_TLS_DIRECT_SEG_REFS);
10374 return gen_rtx_PLUS (Pmode, base, off);
10378 base = get_thread_pointer (true);
10379 dest = gen_reg_rtx (Pmode);
10380 emit_insn (gen_subsi3 (dest, base, off));
10385 gcc_unreachable ();
10391 /* Create or return the unique __imp_DECL dllimport symbol corresponding
10394 static GTY((if_marked ("tree_map_marked_p"), param_is (struct tree_map)))
10395 htab_t dllimport_map;
10398 get_dllimport_decl (tree decl)
10400 struct tree_map *h, in;
10403 const char *prefix;
10404 size_t namelen, prefixlen;
10409 if (!dllimport_map)
10410 dllimport_map = htab_create_ggc (512, tree_map_hash, tree_map_eq, 0);
10412 in.hash = htab_hash_pointer (decl);
10413 in.base.from = decl;
10414 loc = htab_find_slot_with_hash (dllimport_map, &in, in.hash, INSERT);
10415 h = (struct tree_map *) *loc;
10419 *loc = h = GGC_NEW (struct tree_map);
10421 h->base.from = decl;
10422 h->to = to = build_decl (DECL_SOURCE_LOCATION (decl),
10423 VAR_DECL, NULL, ptr_type_node);
10424 DECL_ARTIFICIAL (to) = 1;
10425 DECL_IGNORED_P (to) = 1;
10426 DECL_EXTERNAL (to) = 1;
10427 TREE_READONLY (to) = 1;
10429 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
10430 name = targetm.strip_name_encoding (name);
10431 prefix = name[0] == FASTCALL_PREFIX || user_label_prefix[0] == 0
10432 ? "*__imp_" : "*__imp__";
10433 namelen = strlen (name);
10434 prefixlen = strlen (prefix);
10435 imp_name = (char *) alloca (namelen + prefixlen + 1);
10436 memcpy (imp_name, prefix, prefixlen);
10437 memcpy (imp_name + prefixlen, name, namelen + 1);
10439 name = ggc_alloc_string (imp_name, namelen + prefixlen);
10440 rtl = gen_rtx_SYMBOL_REF (Pmode, name);
10441 SET_SYMBOL_REF_DECL (rtl, to);
10442 SYMBOL_REF_FLAGS (rtl) = SYMBOL_FLAG_LOCAL;
10444 rtl = gen_const_mem (Pmode, rtl);
10445 set_mem_alias_set (rtl, ix86_GOT_alias_set ());
10447 SET_DECL_RTL (to, rtl);
10448 SET_DECL_ASSEMBLER_NAME (to, get_identifier (name));
10453 /* Expand SYMBOL into its corresponding dllimport symbol. WANT_REG is
10454 true if we require the result be a register. */
10457 legitimize_dllimport_symbol (rtx symbol, bool want_reg)
10462 gcc_assert (SYMBOL_REF_DECL (symbol));
10463 imp_decl = get_dllimport_decl (SYMBOL_REF_DECL (symbol));
10465 x = DECL_RTL (imp_decl);
10467 x = force_reg (Pmode, x);
10471 /* Try machine-dependent ways of modifying an illegitimate address
10472 to be legitimate. If we find one, return the new, valid address.
10473 This macro is used in only one place: `memory_address' in explow.c.
10475 OLDX is the address as it was before break_out_memory_refs was called.
10476 In some cases it is useful to look at this to decide what needs to be done.
10478 It is always safe for this macro to do nothing. It exists to recognize
10479 opportunities to optimize the output.
10481 For the 80386, we handle X+REG by loading X into a register R and
10482 using R+REG. R will go in a general reg and indexing will be used.
10483 However, if REG is a broken-out memory address or multiplication,
10484 nothing needs to be done because REG can certainly go in a general reg.
10486 When -fpic is used, special handling is needed for symbolic references.
10487 See comments by legitimize_pic_address in i386.c for details. */
10490 ix86_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED,
10491 enum machine_mode mode)
10496 log = GET_CODE (x) == SYMBOL_REF ? SYMBOL_REF_TLS_MODEL (x) : 0;
10498 return legitimize_tls_address (x, (enum tls_model) log, false);
10499 if (GET_CODE (x) == CONST
10500 && GET_CODE (XEXP (x, 0)) == PLUS
10501 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
10502 && (log = SYMBOL_REF_TLS_MODEL (XEXP (XEXP (x, 0), 0))))
10504 rtx t = legitimize_tls_address (XEXP (XEXP (x, 0), 0),
10505 (enum tls_model) log, false);
10506 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (x, 0), 1));
10509 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES)
10511 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_DLLIMPORT_P (x))
10512 return legitimize_dllimport_symbol (x, true);
10513 if (GET_CODE (x) == CONST
10514 && GET_CODE (XEXP (x, 0)) == PLUS
10515 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
10516 && SYMBOL_REF_DLLIMPORT_P (XEXP (XEXP (x, 0), 0)))
10518 rtx t = legitimize_dllimport_symbol (XEXP (XEXP (x, 0), 0), true);
10519 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (x, 0), 1));
10523 if (flag_pic && SYMBOLIC_CONST (x))
10524 return legitimize_pic_address (x, 0);
10526 /* Canonicalize shifts by 0, 1, 2, 3 into multiply */
10527 if (GET_CODE (x) == ASHIFT
10528 && CONST_INT_P (XEXP (x, 1))
10529 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (x, 1)) < 4)
10532 log = INTVAL (XEXP (x, 1));
10533 x = gen_rtx_MULT (Pmode, force_reg (Pmode, XEXP (x, 0)),
10534 GEN_INT (1 << log));
10537 if (GET_CODE (x) == PLUS)
10539 /* Canonicalize shifts by 0, 1, 2, 3 into multiply. */
10541 if (GET_CODE (XEXP (x, 0)) == ASHIFT
10542 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
10543 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (x, 0), 1)) < 4)
10546 log = INTVAL (XEXP (XEXP (x, 0), 1));
10547 XEXP (x, 0) = gen_rtx_MULT (Pmode,
10548 force_reg (Pmode, XEXP (XEXP (x, 0), 0)),
10549 GEN_INT (1 << log));
10552 if (GET_CODE (XEXP (x, 1)) == ASHIFT
10553 && CONST_INT_P (XEXP (XEXP (x, 1), 1))
10554 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (x, 1), 1)) < 4)
10557 log = INTVAL (XEXP (XEXP (x, 1), 1));
10558 XEXP (x, 1) = gen_rtx_MULT (Pmode,
10559 force_reg (Pmode, XEXP (XEXP (x, 1), 0)),
10560 GEN_INT (1 << log));
10563 /* Put multiply first if it isn't already. */
10564 if (GET_CODE (XEXP (x, 1)) == MULT)
10566 rtx tmp = XEXP (x, 0);
10567 XEXP (x, 0) = XEXP (x, 1);
10572 /* Canonicalize (plus (mult (reg) (const)) (plus (reg) (const)))
10573 into (plus (plus (mult (reg) (const)) (reg)) (const)). This can be
10574 created by virtual register instantiation, register elimination, and
10575 similar optimizations. */
10576 if (GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == PLUS)
10579 x = gen_rtx_PLUS (Pmode,
10580 gen_rtx_PLUS (Pmode, XEXP (x, 0),
10581 XEXP (XEXP (x, 1), 0)),
10582 XEXP (XEXP (x, 1), 1));
10586 (plus (plus (mult (reg) (const)) (plus (reg) (const))) const)
10587 into (plus (plus (mult (reg) (const)) (reg)) (const)). */
10588 else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS
10589 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
10590 && GET_CODE (XEXP (XEXP (x, 0), 1)) == PLUS
10591 && CONSTANT_P (XEXP (x, 1)))
10594 rtx other = NULL_RTX;
10596 if (CONST_INT_P (XEXP (x, 1)))
10598 constant = XEXP (x, 1);
10599 other = XEXP (XEXP (XEXP (x, 0), 1), 1);
10601 else if (CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 1), 1)))
10603 constant = XEXP (XEXP (XEXP (x, 0), 1), 1);
10604 other = XEXP (x, 1);
10612 x = gen_rtx_PLUS (Pmode,
10613 gen_rtx_PLUS (Pmode, XEXP (XEXP (x, 0), 0),
10614 XEXP (XEXP (XEXP (x, 0), 1), 0)),
10615 plus_constant (other, INTVAL (constant)));
10619 if (changed && ix86_legitimate_address_p (mode, x, FALSE))
10622 if (GET_CODE (XEXP (x, 0)) == MULT)
10625 XEXP (x, 0) = force_operand (XEXP (x, 0), 0);
10628 if (GET_CODE (XEXP (x, 1)) == MULT)
10631 XEXP (x, 1) = force_operand (XEXP (x, 1), 0);
10635 && REG_P (XEXP (x, 1))
10636 && REG_P (XEXP (x, 0)))
10639 if (flag_pic && SYMBOLIC_CONST (XEXP (x, 1)))
10642 x = legitimize_pic_address (x, 0);
10645 if (changed && ix86_legitimate_address_p (mode, x, FALSE))
10648 if (REG_P (XEXP (x, 0)))
10650 rtx temp = gen_reg_rtx (Pmode);
10651 rtx val = force_operand (XEXP (x, 1), temp);
10653 emit_move_insn (temp, val);
10655 XEXP (x, 1) = temp;
10659 else if (REG_P (XEXP (x, 1)))
10661 rtx temp = gen_reg_rtx (Pmode);
10662 rtx val = force_operand (XEXP (x, 0), temp);
10664 emit_move_insn (temp, val);
10666 XEXP (x, 0) = temp;
10674 /* Print an integer constant expression in assembler syntax. Addition
10675 and subtraction are the only arithmetic that may appear in these
10676 expressions. FILE is the stdio stream to write to, X is the rtx, and
10677 CODE is the operand print code from the output string. */
10680 output_pic_addr_const (FILE *file, rtx x, int code)
10684 switch (GET_CODE (x))
10687 gcc_assert (flag_pic);
10692 if (! TARGET_MACHO || TARGET_64BIT)
10693 output_addr_const (file, x);
10696 const char *name = XSTR (x, 0);
10698 /* Mark the decl as referenced so that cgraph will
10699 output the function. */
10700 if (SYMBOL_REF_DECL (x))
10701 mark_decl_referenced (SYMBOL_REF_DECL (x));
10704 if (MACHOPIC_INDIRECT
10705 && machopic_classify_symbol (x) == MACHOPIC_UNDEFINED_FUNCTION)
10706 name = machopic_indirection_name (x, /*stub_p=*/true);
10708 assemble_name (file, name);
10710 if (!TARGET_MACHO && !(TARGET_64BIT && DEFAULT_ABI == MS_ABI)
10711 && code == 'P' && ! SYMBOL_REF_LOCAL_P (x))
10712 fputs ("@PLT", file);
10719 ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x));
10720 assemble_name (asm_out_file, buf);
10724 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x));
10728 /* This used to output parentheses around the expression,
10729 but that does not work on the 386 (either ATT or BSD assembler). */
10730 output_pic_addr_const (file, XEXP (x, 0), code);
10734 if (GET_MODE (x) == VOIDmode)
10736 /* We can use %d if the number is <32 bits and positive. */
10737 if (CONST_DOUBLE_HIGH (x) || CONST_DOUBLE_LOW (x) < 0)
10738 fprintf (file, "0x%lx%08lx",
10739 (unsigned long) CONST_DOUBLE_HIGH (x),
10740 (unsigned long) CONST_DOUBLE_LOW (x));
10742 fprintf (file, HOST_WIDE_INT_PRINT_DEC, CONST_DOUBLE_LOW (x));
10745 /* We can't handle floating point constants;
10746 PRINT_OPERAND must handle them. */
10747 output_operand_lossage ("floating constant misused");
10751 /* Some assemblers need integer constants to appear first. */
10752 if (CONST_INT_P (XEXP (x, 0)))
10754 output_pic_addr_const (file, XEXP (x, 0), code);
10756 output_pic_addr_const (file, XEXP (x, 1), code);
10760 gcc_assert (CONST_INT_P (XEXP (x, 1)));
10761 output_pic_addr_const (file, XEXP (x, 1), code);
10763 output_pic_addr_const (file, XEXP (x, 0), code);
10769 putc (ASSEMBLER_DIALECT == ASM_INTEL ? '(' : '[', file);
10770 output_pic_addr_const (file, XEXP (x, 0), code);
10772 output_pic_addr_const (file, XEXP (x, 1), code);
10774 putc (ASSEMBLER_DIALECT == ASM_INTEL ? ')' : ']', file);
10778 gcc_assert (XVECLEN (x, 0) == 1);
10779 output_pic_addr_const (file, XVECEXP (x, 0, 0), code);
10780 switch (XINT (x, 1))
10783 fputs ("@GOT", file);
10785 case UNSPEC_GOTOFF:
10786 fputs ("@GOTOFF", file);
10788 case UNSPEC_PLTOFF:
10789 fputs ("@PLTOFF", file);
10791 case UNSPEC_GOTPCREL:
10792 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
10793 "@GOTPCREL(%rip)" : "@GOTPCREL[rip]", file);
10795 case UNSPEC_GOTTPOFF:
10796 /* FIXME: This might be @TPOFF in Sun ld too. */
10797 fputs ("@GOTTPOFF", file);
10800 fputs ("@TPOFF", file);
10802 case UNSPEC_NTPOFF:
10804 fputs ("@TPOFF", file);
10806 fputs ("@NTPOFF", file);
10808 case UNSPEC_DTPOFF:
10809 fputs ("@DTPOFF", file);
10811 case UNSPEC_GOTNTPOFF:
10813 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
10814 "@GOTTPOFF(%rip)": "@GOTTPOFF[rip]", file);
10816 fputs ("@GOTNTPOFF", file);
10818 case UNSPEC_INDNTPOFF:
10819 fputs ("@INDNTPOFF", file);
10822 case UNSPEC_MACHOPIC_OFFSET:
10824 machopic_output_function_base_name (file);
10828 output_operand_lossage ("invalid UNSPEC as operand");
10834 output_operand_lossage ("invalid expression as operand");
10838 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10839 We need to emit DTP-relative relocations. */
10841 static void ATTRIBUTE_UNUSED
10842 i386_output_dwarf_dtprel (FILE *file, int size, rtx x)
10844 fputs (ASM_LONG, file);
10845 output_addr_const (file, x);
10846 fputs ("@DTPOFF", file);
10852 fputs (", 0", file);
10855 gcc_unreachable ();
10859 /* Return true if X is a representation of the PIC register. This copes
10860 with calls from ix86_find_base_term, where the register might have
10861 been replaced by a cselib value. */
10864 ix86_pic_register_p (rtx x)
10866 if (GET_CODE (x) == VALUE && CSELIB_VAL_PTR (x))
10867 return (pic_offset_table_rtx
10868 && rtx_equal_for_cselib_p (x, pic_offset_table_rtx));
10870 return REG_P (x) && REGNO (x) == PIC_OFFSET_TABLE_REGNUM;
10873 /* In the name of slightly smaller debug output, and to cater to
10874 general assembler lossage, recognize PIC+GOTOFF and turn it back
10875 into a direct symbol reference.
10877 On Darwin, this is necessary to avoid a crash, because Darwin
10878 has a different PIC label for each routine but the DWARF debugging
10879 information is not associated with any particular routine, so it's
10880 necessary to remove references to the PIC label from RTL stored by
10881 the DWARF output code. */
10884 ix86_delegitimize_address (rtx x)
10886 rtx orig_x = delegitimize_mem_from_attrs (x);
10887 /* reg_addend is NULL or a multiple of some register. */
10888 rtx reg_addend = NULL_RTX;
10889 /* const_addend is NULL or a const_int. */
10890 rtx const_addend = NULL_RTX;
10891 /* This is the result, or NULL. */
10892 rtx result = NULL_RTX;
10901 if (GET_CODE (x) != CONST
10902 || GET_CODE (XEXP (x, 0)) != UNSPEC
10903 || XINT (XEXP (x, 0), 1) != UNSPEC_GOTPCREL
10904 || !MEM_P (orig_x))
10906 return XVECEXP (XEXP (x, 0), 0, 0);
10909 if (GET_CODE (x) != PLUS
10910 || GET_CODE (XEXP (x, 1)) != CONST)
10913 if (ix86_pic_register_p (XEXP (x, 0)))
10914 /* %ebx + GOT/GOTOFF */
10916 else if (GET_CODE (XEXP (x, 0)) == PLUS)
10918 /* %ebx + %reg * scale + GOT/GOTOFF */
10919 reg_addend = XEXP (x, 0);
10920 if (ix86_pic_register_p (XEXP (reg_addend, 0)))
10921 reg_addend = XEXP (reg_addend, 1);
10922 else if (ix86_pic_register_p (XEXP (reg_addend, 1)))
10923 reg_addend = XEXP (reg_addend, 0);
10926 if (!REG_P (reg_addend)
10927 && GET_CODE (reg_addend) != MULT
10928 && GET_CODE (reg_addend) != ASHIFT)
10934 x = XEXP (XEXP (x, 1), 0);
10935 if (GET_CODE (x) == PLUS
10936 && CONST_INT_P (XEXP (x, 1)))
10938 const_addend = XEXP (x, 1);
10942 if (GET_CODE (x) == UNSPEC
10943 && ((XINT (x, 1) == UNSPEC_GOT && MEM_P (orig_x))
10944 || (XINT (x, 1) == UNSPEC_GOTOFF && !MEM_P (orig_x))))
10945 result = XVECEXP (x, 0, 0);
10947 if (TARGET_MACHO && darwin_local_data_pic (x)
10948 && !MEM_P (orig_x))
10949 result = XVECEXP (x, 0, 0);
10955 result = gen_rtx_CONST (Pmode, gen_rtx_PLUS (Pmode, result, const_addend));
10957 result = gen_rtx_PLUS (Pmode, reg_addend, result);
10961 /* If X is a machine specific address (i.e. a symbol or label being
10962 referenced as a displacement from the GOT implemented using an
10963 UNSPEC), then return the base term. Otherwise return X. */
10966 ix86_find_base_term (rtx x)
10972 if (GET_CODE (x) != CONST)
10974 term = XEXP (x, 0);
10975 if (GET_CODE (term) == PLUS
10976 && (CONST_INT_P (XEXP (term, 1))
10977 || GET_CODE (XEXP (term, 1)) == CONST_DOUBLE))
10978 term = XEXP (term, 0);
10979 if (GET_CODE (term) != UNSPEC
10980 || XINT (term, 1) != UNSPEC_GOTPCREL)
10983 return XVECEXP (term, 0, 0);
10986 return ix86_delegitimize_address (x);
10990 put_condition_code (enum rtx_code code, enum machine_mode mode, int reverse,
10991 int fp, FILE *file)
10993 const char *suffix;
10995 if (mode == CCFPmode || mode == CCFPUmode)
10997 code = ix86_fp_compare_code_to_integer (code);
11001 code = reverse_condition (code);
11052 gcc_assert (mode == CCmode || mode == CCNOmode || mode == CCGCmode);
11056 /* ??? Use "nbe" instead of "a" for fcmov lossage on some assemblers.
11057 Those same assemblers have the same but opposite lossage on cmov. */
11058 if (mode == CCmode)
11059 suffix = fp ? "nbe" : "a";
11060 else if (mode == CCCmode)
11063 gcc_unreachable ();
11079 gcc_unreachable ();
11083 gcc_assert (mode == CCmode || mode == CCCmode);
11100 gcc_unreachable ();
11104 /* ??? As above. */
11105 gcc_assert (mode == CCmode || mode == CCCmode);
11106 suffix = fp ? "nb" : "ae";
11109 gcc_assert (mode == CCmode || mode == CCGCmode || mode == CCNOmode);
11113 /* ??? As above. */
11114 if (mode == CCmode)
11116 else if (mode == CCCmode)
11117 suffix = fp ? "nb" : "ae";
11119 gcc_unreachable ();
11122 suffix = fp ? "u" : "p";
11125 suffix = fp ? "nu" : "np";
11128 gcc_unreachable ();
11130 fputs (suffix, file);
11133 /* Print the name of register X to FILE based on its machine mode and number.
11134 If CODE is 'w', pretend the mode is HImode.
11135 If CODE is 'b', pretend the mode is QImode.
11136 If CODE is 'k', pretend the mode is SImode.
11137 If CODE is 'q', pretend the mode is DImode.
11138 If CODE is 'x', pretend the mode is V4SFmode.
11139 If CODE is 't', pretend the mode is V8SFmode.
11140 If CODE is 'h', pretend the reg is the 'high' byte register.
11141 If CODE is 'y', print "st(0)" instead of "st", if the reg is stack op.
11142 If CODE is 'd', duplicate the operand for AVX instruction.
11146 print_reg (rtx x, int code, FILE *file)
11149 bool duplicated = code == 'd' && TARGET_AVX;
11151 gcc_assert (x == pc_rtx
11152 || (REGNO (x) != ARG_POINTER_REGNUM
11153 && REGNO (x) != FRAME_POINTER_REGNUM
11154 && REGNO (x) != FLAGS_REG
11155 && REGNO (x) != FPSR_REG
11156 && REGNO (x) != FPCR_REG));
11158 if (ASSEMBLER_DIALECT == ASM_ATT)
11163 gcc_assert (TARGET_64BIT);
11164 fputs ("rip", file);
11168 if (code == 'w' || MMX_REG_P (x))
11170 else if (code == 'b')
11172 else if (code == 'k')
11174 else if (code == 'q')
11176 else if (code == 'y')
11178 else if (code == 'h')
11180 else if (code == 'x')
11182 else if (code == 't')
11185 code = GET_MODE_SIZE (GET_MODE (x));
11187 /* Irritatingly, AMD extended registers use different naming convention
11188 from the normal registers. */
11189 if (REX_INT_REG_P (x))
11191 gcc_assert (TARGET_64BIT);
11195 error ("extended registers have no high halves");
11198 fprintf (file, "r%ib", REGNO (x) - FIRST_REX_INT_REG + 8);
11201 fprintf (file, "r%iw", REGNO (x) - FIRST_REX_INT_REG + 8);
11204 fprintf (file, "r%id", REGNO (x) - FIRST_REX_INT_REG + 8);
11207 fprintf (file, "r%i", REGNO (x) - FIRST_REX_INT_REG + 8);
11210 error ("unsupported operand size for extended register");
11220 if (STACK_TOP_P (x))
11229 if (! ANY_FP_REG_P (x))
11230 putc (code == 8 && TARGET_64BIT ? 'r' : 'e', file);
11235 reg = hi_reg_name[REGNO (x)];
11238 if (REGNO (x) >= ARRAY_SIZE (qi_reg_name))
11240 reg = qi_reg_name[REGNO (x)];
11243 if (REGNO (x) >= ARRAY_SIZE (qi_high_reg_name))
11245 reg = qi_high_reg_name[REGNO (x)];
11250 gcc_assert (!duplicated);
11252 fputs (hi_reg_name[REGNO (x)] + 1, file);
11257 gcc_unreachable ();
11263 if (ASSEMBLER_DIALECT == ASM_ATT)
11264 fprintf (file, ", %%%s", reg);
11266 fprintf (file, ", %s", reg);
11270 /* Locate some local-dynamic symbol still in use by this function
11271 so that we can print its name in some tls_local_dynamic_base
11275 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
11279 if (GET_CODE (x) == SYMBOL_REF
11280 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
11282 cfun->machine->some_ld_name = XSTR (x, 0);
11289 static const char *
11290 get_some_local_dynamic_name (void)
11294 if (cfun->machine->some_ld_name)
11295 return cfun->machine->some_ld_name;
11297 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
11299 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
11300 return cfun->machine->some_ld_name;
11305 /* Meaning of CODE:
11306 L,W,B,Q,S,T -- print the opcode suffix for specified size of operand.
11307 C -- print opcode suffix for set/cmov insn.
11308 c -- like C, but print reversed condition
11309 E,e -- likewise, but for compare-and-branch fused insn.
11310 F,f -- likewise, but for floating-point.
11311 O -- if HAVE_AS_IX86_CMOV_SUN_SYNTAX, expand to "w.", "l." or "q.",
11313 R -- print the prefix for register names.
11314 z -- print the opcode suffix for the size of the current operand.
11315 Z -- likewise, with special suffixes for x87 instructions.
11316 * -- print a star (in certain assembler syntax)
11317 A -- print an absolute memory reference.
11318 w -- print the operand as if it's a "word" (HImode) even if it isn't.
11319 s -- print a shift double count, followed by the assemblers argument
11321 b -- print the QImode name of the register for the indicated operand.
11322 %b0 would print %al if operands[0] is reg 0.
11323 w -- likewise, print the HImode name of the register.
11324 k -- likewise, print the SImode name of the register.
11325 q -- likewise, print the DImode name of the register.
11326 x -- likewise, print the V4SFmode name of the register.
11327 t -- likewise, print the V8SFmode name of the register.
11328 h -- print the QImode name for a "high" register, either ah, bh, ch or dh.
11329 y -- print "st(0)" instead of "st" as a register.
11330 d -- print duplicated register operand for AVX instruction.
11331 D -- print condition for SSE cmp instruction.
11332 P -- if PIC, print an @PLT suffix.
11333 X -- don't print any sort of PIC '@' suffix for a symbol.
11334 & -- print some in-use local-dynamic symbol name.
11335 H -- print a memory address offset by 8; used for sse high-parts
11336 Y -- print condition for XOP pcom* instruction.
11337 + -- print a branch hint as 'cs' or 'ds' prefix
11338 ; -- print a semicolon (after prefixes due to bug in older gas).
11342 print_operand (FILE *file, rtx x, int code)
11349 if (ASSEMBLER_DIALECT == ASM_ATT)
11355 const char *name = get_some_local_dynamic_name ();
11357 output_operand_lossage ("'%%&' used without any "
11358 "local dynamic TLS references");
11360 assemble_name (file, name);
11365 switch (ASSEMBLER_DIALECT)
11372 /* Intel syntax. For absolute addresses, registers should not
11373 be surrounded by braces. */
11377 PRINT_OPERAND (file, x, 0);
11384 gcc_unreachable ();
11387 PRINT_OPERAND (file, x, 0);
11392 if (ASSEMBLER_DIALECT == ASM_ATT)
11397 if (ASSEMBLER_DIALECT == ASM_ATT)
11402 if (ASSEMBLER_DIALECT == ASM_ATT)
11407 if (ASSEMBLER_DIALECT == ASM_ATT)
11412 if (ASSEMBLER_DIALECT == ASM_ATT)
11417 if (ASSEMBLER_DIALECT == ASM_ATT)
11422 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT)
11424 /* Opcodes don't get size suffixes if using Intel opcodes. */
11425 if (ASSEMBLER_DIALECT == ASM_INTEL)
11428 switch (GET_MODE_SIZE (GET_MODE (x)))
11447 output_operand_lossage
11448 ("invalid operand size for operand code '%c'", code);
11453 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
11455 (0, "non-integer operand used with operand code '%c'", code);
11459 /* 387 opcodes don't get size suffixes if using Intel opcodes. */
11460 if (ASSEMBLER_DIALECT == ASM_INTEL)
11463 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT)
11465 switch (GET_MODE_SIZE (GET_MODE (x)))
11468 #ifdef HAVE_AS_IX86_FILDS
11478 #ifdef HAVE_AS_IX86_FILDQ
11481 fputs ("ll", file);
11489 else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
11491 /* 387 opcodes don't get size suffixes
11492 if the operands are registers. */
11493 if (STACK_REG_P (x))
11496 switch (GET_MODE_SIZE (GET_MODE (x)))
11517 output_operand_lossage
11518 ("invalid operand type used with operand code '%c'", code);
11522 output_operand_lossage
11523 ("invalid operand size for operand code '%c'", code);
11540 if (CONST_INT_P (x) || ! SHIFT_DOUBLE_OMITS_COUNT)
11542 PRINT_OPERAND (file, x, 0);
11543 fputs (", ", file);
11548 /* Little bit of braindamage here. The SSE compare instructions
11549 does use completely different names for the comparisons that the
11550 fp conditional moves. */
11553 switch (GET_CODE (x))
11556 fputs ("eq", file);
11559 fputs ("eq_us", file);
11562 fputs ("lt", file);
11565 fputs ("nge", file);
11568 fputs ("le", file);
11571 fputs ("ngt", file);
11574 fputs ("unord", file);
11577 fputs ("neq", file);
11580 fputs ("neq_oq", file);
11583 fputs ("ge", file);
11586 fputs ("nlt", file);
11589 fputs ("gt", file);
11592 fputs ("nle", file);
11595 fputs ("ord", file);
11598 output_operand_lossage ("operand is not a condition code, "
11599 "invalid operand code 'D'");
11605 switch (GET_CODE (x))
11609 fputs ("eq", file);
11613 fputs ("lt", file);
11617 fputs ("le", file);
11620 fputs ("unord", file);
11624 fputs ("neq", file);
11628 fputs ("nlt", file);
11632 fputs ("nle", file);
11635 fputs ("ord", file);
11638 output_operand_lossage ("operand is not a condition code, "
11639 "invalid operand code 'D'");
11645 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
11646 if (ASSEMBLER_DIALECT == ASM_ATT)
11648 switch (GET_MODE (x))
11650 case HImode: putc ('w', file); break;
11652 case SFmode: putc ('l', file); break;
11654 case DFmode: putc ('q', file); break;
11655 default: gcc_unreachable ();
11662 if (!COMPARISON_P (x))
11664 output_operand_lossage ("operand is neither a constant nor a "
11665 "condition code, invalid operand code "
11669 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 0, file);
11672 if (!COMPARISON_P (x))
11674 output_operand_lossage ("operand is neither a constant nor a "
11675 "condition code, invalid operand code "
11679 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
11680 if (ASSEMBLER_DIALECT == ASM_ATT)
11683 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 1, file);
11686 /* Like above, but reverse condition */
11688 /* Check to see if argument to %c is really a constant
11689 and not a condition code which needs to be reversed. */
11690 if (!COMPARISON_P (x))
11692 output_operand_lossage ("operand is neither a constant nor a "
11693 "condition code, invalid operand "
11697 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 0, file);
11700 if (!COMPARISON_P (x))
11702 output_operand_lossage ("operand is neither a constant nor a "
11703 "condition code, invalid operand "
11707 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
11708 if (ASSEMBLER_DIALECT == ASM_ATT)
11711 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 1, file);
11715 put_condition_code (GET_CODE (x), CCmode, 0, 0, file);
11719 put_condition_code (GET_CODE (x), CCmode, 1, 0, file);
11723 /* It doesn't actually matter what mode we use here, as we're
11724 only going to use this for printing. */
11725 x = adjust_address_nv (x, DImode, 8);
11733 || optimize_function_for_size_p (cfun) || !TARGET_BRANCH_PREDICTION_HINTS)
11736 x = find_reg_note (current_output_insn, REG_BR_PROB, 0);
11739 int pred_val = INTVAL (XEXP (x, 0));
11741 if (pred_val < REG_BR_PROB_BASE * 45 / 100
11742 || pred_val > REG_BR_PROB_BASE * 55 / 100)
11744 int taken = pred_val > REG_BR_PROB_BASE / 2;
11745 int cputaken = final_forward_branch_p (current_output_insn) == 0;
11747 /* Emit hints only in the case default branch prediction
11748 heuristics would fail. */
11749 if (taken != cputaken)
11751 /* We use 3e (DS) prefix for taken branches and
11752 2e (CS) prefix for not taken branches. */
11754 fputs ("ds ; ", file);
11756 fputs ("cs ; ", file);
11764 switch (GET_CODE (x))
11767 fputs ("neq", file);
11770 fputs ("eq", file);
11774 fputs (INTEGRAL_MODE_P (GET_MODE (x)) ? "ge" : "unlt", file);
11778 fputs (INTEGRAL_MODE_P (GET_MODE (x)) ? "gt" : "unle", file);
11782 fputs ("le", file);
11786 fputs ("lt", file);
11789 fputs ("unord", file);
11792 fputs ("ord", file);
11795 fputs ("ueq", file);
11798 fputs ("nlt", file);
11801 fputs ("nle", file);
11804 fputs ("ule", file);
11807 fputs ("ult", file);
11810 fputs ("une", file);
11813 output_operand_lossage ("operand is not a condition code, "
11814 "invalid operand code 'Y'");
11821 fputs (" ; ", file);
11828 output_operand_lossage ("invalid operand code '%c'", code);
11833 print_reg (x, code, file);
11835 else if (MEM_P (x))
11837 /* No `byte ptr' prefix for call instructions or BLKmode operands. */
11838 if (ASSEMBLER_DIALECT == ASM_INTEL && code != 'X' && code != 'P'
11839 && GET_MODE (x) != BLKmode)
11842 switch (GET_MODE_SIZE (GET_MODE (x)))
11844 case 1: size = "BYTE"; break;
11845 case 2: size = "WORD"; break;
11846 case 4: size = "DWORD"; break;
11847 case 8: size = "QWORD"; break;
11848 case 12: size = "TBYTE"; break;
11850 if (GET_MODE (x) == XFmode)
11855 case 32: size = "YMMWORD"; break;
11857 gcc_unreachable ();
11860 /* Check for explicit size override (codes 'b', 'w' and 'k') */
11863 else if (code == 'w')
11865 else if (code == 'k')
11868 fputs (size, file);
11869 fputs (" PTR ", file);
11873 /* Avoid (%rip) for call operands. */
11874 if (CONSTANT_ADDRESS_P (x) && code == 'P'
11875 && !CONST_INT_P (x))
11876 output_addr_const (file, x);
11877 else if (this_is_asm_operands && ! address_operand (x, VOIDmode))
11878 output_operand_lossage ("invalid constraints for operand");
11880 output_address (x);
11883 else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == SFmode)
11888 REAL_VALUE_FROM_CONST_DOUBLE (r, x);
11889 REAL_VALUE_TO_TARGET_SINGLE (r, l);
11891 if (ASSEMBLER_DIALECT == ASM_ATT)
11893 fprintf (file, "0x%08lx", (long unsigned int) l);
11896 /* These float cases don't actually occur as immediate operands. */
11897 else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == DFmode)
11901 real_to_decimal (dstr, CONST_DOUBLE_REAL_VALUE (x), sizeof (dstr), 0, 1);
11902 fputs (dstr, file);
11905 else if (GET_CODE (x) == CONST_DOUBLE
11906 && GET_MODE (x) == XFmode)
11910 real_to_decimal (dstr, CONST_DOUBLE_REAL_VALUE (x), sizeof (dstr), 0, 1);
11911 fputs (dstr, file);
11916 /* We have patterns that allow zero sets of memory, for instance.
11917 In 64-bit mode, we should probably support all 8-byte vectors,
11918 since we can in fact encode that into an immediate. */
11919 if (GET_CODE (x) == CONST_VECTOR)
11921 gcc_assert (x == CONST0_RTX (GET_MODE (x)));
11927 if (CONST_INT_P (x) || GET_CODE (x) == CONST_DOUBLE)
11929 if (ASSEMBLER_DIALECT == ASM_ATT)
11932 else if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF
11933 || GET_CODE (x) == LABEL_REF)
11935 if (ASSEMBLER_DIALECT == ASM_ATT)
11938 fputs ("OFFSET FLAT:", file);
11941 if (CONST_INT_P (x))
11942 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x));
11944 output_pic_addr_const (file, x, code);
11946 output_addr_const (file, x);
11950 /* Print a memory operand whose address is ADDR. */
11953 print_operand_address (FILE *file, rtx addr)
11955 struct ix86_address parts;
11956 rtx base, index, disp;
11958 int ok = ix86_decompose_address (addr, &parts);
11963 index = parts.index;
11965 scale = parts.scale;
11973 if (ASSEMBLER_DIALECT == ASM_ATT)
11975 fputs ((parts.seg == SEG_FS ? "fs:" : "gs:"), file);
11978 gcc_unreachable ();
11981 /* Use one byte shorter RIP relative addressing for 64bit mode. */
11982 if (TARGET_64BIT && !base && !index)
11986 if (GET_CODE (disp) == CONST
11987 && GET_CODE (XEXP (disp, 0)) == PLUS
11988 && CONST_INT_P (XEXP (XEXP (disp, 0), 1)))
11989 symbol = XEXP (XEXP (disp, 0), 0);
11991 if (GET_CODE (symbol) == LABEL_REF
11992 || (GET_CODE (symbol) == SYMBOL_REF
11993 && SYMBOL_REF_TLS_MODEL (symbol) == 0))
11996 if (!base && !index)
11998 /* Displacement only requires special attention. */
12000 if (CONST_INT_P (disp))
12002 if (ASSEMBLER_DIALECT == ASM_INTEL && parts.seg == SEG_DEFAULT)
12003 fputs ("ds:", file);
12004 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (disp));
12007 output_pic_addr_const (file, disp, 0);
12009 output_addr_const (file, disp);
12013 if (ASSEMBLER_DIALECT == ASM_ATT)
12018 output_pic_addr_const (file, disp, 0);
12019 else if (GET_CODE (disp) == LABEL_REF)
12020 output_asm_label (disp);
12022 output_addr_const (file, disp);
12027 print_reg (base, 0, file);
12031 print_reg (index, 0, file);
12033 fprintf (file, ",%d", scale);
12039 rtx offset = NULL_RTX;
12043 /* Pull out the offset of a symbol; print any symbol itself. */
12044 if (GET_CODE (disp) == CONST
12045 && GET_CODE (XEXP (disp, 0)) == PLUS
12046 && CONST_INT_P (XEXP (XEXP (disp, 0), 1)))
12048 offset = XEXP (XEXP (disp, 0), 1);
12049 disp = gen_rtx_CONST (VOIDmode,
12050 XEXP (XEXP (disp, 0), 0));
12054 output_pic_addr_const (file, disp, 0);
12055 else if (GET_CODE (disp) == LABEL_REF)
12056 output_asm_label (disp);
12057 else if (CONST_INT_P (disp))
12060 output_addr_const (file, disp);
12066 print_reg (base, 0, file);
12069 if (INTVAL (offset) >= 0)
12071 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset));
12075 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset));
12082 print_reg (index, 0, file);
12084 fprintf (file, "*%d", scale);
12092 output_addr_const_extra (FILE *file, rtx x)
12096 if (GET_CODE (x) != UNSPEC)
12099 op = XVECEXP (x, 0, 0);
12100 switch (XINT (x, 1))
12102 case UNSPEC_GOTTPOFF:
12103 output_addr_const (file, op);
12104 /* FIXME: This might be @TPOFF in Sun ld. */
12105 fputs ("@GOTTPOFF", file);
12108 output_addr_const (file, op);
12109 fputs ("@TPOFF", file);
12111 case UNSPEC_NTPOFF:
12112 output_addr_const (file, op);
12114 fputs ("@TPOFF", file);
12116 fputs ("@NTPOFF", file);
12118 case UNSPEC_DTPOFF:
12119 output_addr_const (file, op);
12120 fputs ("@DTPOFF", file);
12122 case UNSPEC_GOTNTPOFF:
12123 output_addr_const (file, op);
12125 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
12126 "@GOTTPOFF(%rip)" : "@GOTTPOFF[rip]", file);
12128 fputs ("@GOTNTPOFF", file);
12130 case UNSPEC_INDNTPOFF:
12131 output_addr_const (file, op);
12132 fputs ("@INDNTPOFF", file);
12135 case UNSPEC_MACHOPIC_OFFSET:
12136 output_addr_const (file, op);
12138 machopic_output_function_base_name (file);
12149 /* Split one or more DImode RTL references into pairs of SImode
12150 references. The RTL can be REG, offsettable MEM, integer constant, or
12151 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
12152 split and "num" is its length. lo_half and hi_half are output arrays
12153 that parallel "operands". */
12156 split_di (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
12160 rtx op = operands[num];
12162 /* simplify_subreg refuse to split volatile memory addresses,
12163 but we still have to handle it. */
12166 lo_half[num] = adjust_address (op, SImode, 0);
12167 hi_half[num] = adjust_address (op, SImode, 4);
12171 lo_half[num] = simplify_gen_subreg (SImode, op,
12172 GET_MODE (op) == VOIDmode
12173 ? DImode : GET_MODE (op), 0);
12174 hi_half[num] = simplify_gen_subreg (SImode, op,
12175 GET_MODE (op) == VOIDmode
12176 ? DImode : GET_MODE (op), 4);
12180 /* Split one or more TImode RTL references into pairs of DImode
12181 references. The RTL can be REG, offsettable MEM, integer constant, or
12182 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
12183 split and "num" is its length. lo_half and hi_half are output arrays
12184 that parallel "operands". */
12187 split_ti (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
12191 rtx op = operands[num];
12193 /* simplify_subreg refuse to split volatile memory addresses, but we
12194 still have to handle it. */
12197 lo_half[num] = adjust_address (op, DImode, 0);
12198 hi_half[num] = adjust_address (op, DImode, 8);
12202 lo_half[num] = simplify_gen_subreg (DImode, op, TImode, 0);
12203 hi_half[num] = simplify_gen_subreg (DImode, op, TImode, 8);
12208 /* Output code to perform a 387 binary operation in INSN, one of PLUS,
12209 MINUS, MULT or DIV. OPERANDS are the insn operands, where operands[3]
12210 is the expression of the binary operation. The output may either be
12211 emitted here, or returned to the caller, like all output_* functions.
12213 There is no guarantee that the operands are the same mode, as they
12214 might be within FLOAT or FLOAT_EXTEND expressions. */
12216 #ifndef SYSV386_COMPAT
12217 /* Set to 1 for compatibility with brain-damaged assemblers. No-one
12218 wants to fix the assemblers because that causes incompatibility
12219 with gcc. No-one wants to fix gcc because that causes
12220 incompatibility with assemblers... You can use the option of
12221 -DSYSV386_COMPAT=0 if you recompile both gcc and gas this way. */
12222 #define SYSV386_COMPAT 1
12226 output_387_binary_op (rtx insn, rtx *operands)
12228 static char buf[40];
12231 int is_sse = SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]) || SSE_REG_P (operands[2]);
12233 #ifdef ENABLE_CHECKING
12234 /* Even if we do not want to check the inputs, this documents input
12235 constraints. Which helps in understanding the following code. */
12236 if (STACK_REG_P (operands[0])
12237 && ((REG_P (operands[1])
12238 && REGNO (operands[0]) == REGNO (operands[1])
12239 && (STACK_REG_P (operands[2]) || MEM_P (operands[2])))
12240 || (REG_P (operands[2])
12241 && REGNO (operands[0]) == REGNO (operands[2])
12242 && (STACK_REG_P (operands[1]) || MEM_P (operands[1]))))
12243 && (STACK_TOP_P (operands[1]) || STACK_TOP_P (operands[2])))
12246 gcc_assert (is_sse);
12249 switch (GET_CODE (operands[3]))
12252 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
12253 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
12261 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
12262 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
12270 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
12271 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
12279 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
12280 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
12288 gcc_unreachable ();
12295 strcpy (buf, ssep);
12296 if (GET_MODE (operands[0]) == SFmode)
12297 strcat (buf, "ss\t{%2, %1, %0|%0, %1, %2}");
12299 strcat (buf, "sd\t{%2, %1, %0|%0, %1, %2}");
12303 strcpy (buf, ssep + 1);
12304 if (GET_MODE (operands[0]) == SFmode)
12305 strcat (buf, "ss\t{%2, %0|%0, %2}");
12307 strcat (buf, "sd\t{%2, %0|%0, %2}");
12313 switch (GET_CODE (operands[3]))
12317 if (REG_P (operands[2]) && REGNO (operands[0]) == REGNO (operands[2]))
12319 rtx temp = operands[2];
12320 operands[2] = operands[1];
12321 operands[1] = temp;
12324 /* know operands[0] == operands[1]. */
12326 if (MEM_P (operands[2]))
12332 if (find_regno_note (insn, REG_DEAD, REGNO (operands[2])))
12334 if (STACK_TOP_P (operands[0]))
12335 /* How is it that we are storing to a dead operand[2]?
12336 Well, presumably operands[1] is dead too. We can't
12337 store the result to st(0) as st(0) gets popped on this
12338 instruction. Instead store to operands[2] (which I
12339 think has to be st(1)). st(1) will be popped later.
12340 gcc <= 2.8.1 didn't have this check and generated
12341 assembly code that the Unixware assembler rejected. */
12342 p = "p\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
12344 p = "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
12348 if (STACK_TOP_P (operands[0]))
12349 p = "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
12351 p = "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
12356 if (MEM_P (operands[1]))
12362 if (MEM_P (operands[2]))
12368 if (find_regno_note (insn, REG_DEAD, REGNO (operands[2])))
12371 /* The SystemV/386 SVR3.2 assembler, and probably all AT&T
12372 derived assemblers, confusingly reverse the direction of
12373 the operation for fsub{r} and fdiv{r} when the
12374 destination register is not st(0). The Intel assembler
12375 doesn't have this brain damage. Read !SYSV386_COMPAT to
12376 figure out what the hardware really does. */
12377 if (STACK_TOP_P (operands[0]))
12378 p = "{p\t%0, %2|rp\t%2, %0}";
12380 p = "{rp\t%2, %0|p\t%0, %2}";
12382 if (STACK_TOP_P (operands[0]))
12383 /* As above for fmul/fadd, we can't store to st(0). */
12384 p = "rp\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
12386 p = "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
12391 if (find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
12394 if (STACK_TOP_P (operands[0]))
12395 p = "{rp\t%0, %1|p\t%1, %0}";
12397 p = "{p\t%1, %0|rp\t%0, %1}";
12399 if (STACK_TOP_P (operands[0]))
12400 p = "p\t{%0, %1|%1, %0}"; /* st(1) = st(1) op st(0); pop */
12402 p = "rp\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2); pop */
12407 if (STACK_TOP_P (operands[0]))
12409 if (STACK_TOP_P (operands[1]))
12410 p = "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
12412 p = "r\t{%y1, %0|%0, %y1}"; /* st(0) = st(r1) op st(0) */
12415 else if (STACK_TOP_P (operands[1]))
12418 p = "{\t%1, %0|r\t%0, %1}";
12420 p = "r\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2) */
12426 p = "{r\t%2, %0|\t%0, %2}";
12428 p = "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
12434 gcc_unreachable ();
12441 /* Return needed mode for entity in optimize_mode_switching pass. */
12444 ix86_mode_needed (int entity, rtx insn)
12446 enum attr_i387_cw mode;
12448 /* The mode UNINITIALIZED is used to store control word after a
12449 function call or ASM pattern. The mode ANY specify that function
12450 has no requirements on the control word and make no changes in the
12451 bits we are interested in. */
12454 || (NONJUMP_INSN_P (insn)
12455 && (asm_noperands (PATTERN (insn)) >= 0
12456 || GET_CODE (PATTERN (insn)) == ASM_INPUT)))
12457 return I387_CW_UNINITIALIZED;
12459 if (recog_memoized (insn) < 0)
12460 return I387_CW_ANY;
12462 mode = get_attr_i387_cw (insn);
12467 if (mode == I387_CW_TRUNC)
12472 if (mode == I387_CW_FLOOR)
12477 if (mode == I387_CW_CEIL)
12482 if (mode == I387_CW_MASK_PM)
12487 gcc_unreachable ();
12490 return I387_CW_ANY;
12493 /* Output code to initialize control word copies used by trunc?f?i and
12494 rounding patterns. CURRENT_MODE is set to current control word,
12495 while NEW_MODE is set to new control word. */
12498 emit_i387_cw_initialization (int mode)
12500 rtx stored_mode = assign_386_stack_local (HImode, SLOT_CW_STORED);
12503 enum ix86_stack_slot slot;
12505 rtx reg = gen_reg_rtx (HImode);
12507 emit_insn (gen_x86_fnstcw_1 (stored_mode));
12508 emit_move_insn (reg, copy_rtx (stored_mode));
12510 if (TARGET_64BIT || TARGET_PARTIAL_REG_STALL
12511 || optimize_function_for_size_p (cfun))
12515 case I387_CW_TRUNC:
12516 /* round toward zero (truncate) */
12517 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0c00)));
12518 slot = SLOT_CW_TRUNC;
12521 case I387_CW_FLOOR:
12522 /* round down toward -oo */
12523 emit_insn (gen_andhi3 (reg, reg, GEN_INT (~0x0c00)));
12524 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0400)));
12525 slot = SLOT_CW_FLOOR;
12529 /* round up toward +oo */
12530 emit_insn (gen_andhi3 (reg, reg, GEN_INT (~0x0c00)));
12531 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0800)));
12532 slot = SLOT_CW_CEIL;
12535 case I387_CW_MASK_PM:
12536 /* mask precision exception for nearbyint() */
12537 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0020)));
12538 slot = SLOT_CW_MASK_PM;
12542 gcc_unreachable ();
12549 case I387_CW_TRUNC:
12550 /* round toward zero (truncate) */
12551 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0xc)));
12552 slot = SLOT_CW_TRUNC;
12555 case I387_CW_FLOOR:
12556 /* round down toward -oo */
12557 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0x4)));
12558 slot = SLOT_CW_FLOOR;
12562 /* round up toward +oo */
12563 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0x8)));
12564 slot = SLOT_CW_CEIL;
12567 case I387_CW_MASK_PM:
12568 /* mask precision exception for nearbyint() */
12569 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0020)));
12570 slot = SLOT_CW_MASK_PM;
12574 gcc_unreachable ();
12578 gcc_assert (slot < MAX_386_STACK_LOCALS);
12580 new_mode = assign_386_stack_local (HImode, slot);
12581 emit_move_insn (new_mode, reg);
12584 /* Output code for INSN to convert a float to a signed int. OPERANDS
12585 are the insn operands. The output may be [HSD]Imode and the input
12586 operand may be [SDX]Fmode. */
12589 output_fix_trunc (rtx insn, rtx *operands, int fisttp)
12591 int stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0;
12592 int dimode_p = GET_MODE (operands[0]) == DImode;
12593 int round_mode = get_attr_i387_cw (insn);
12595 /* Jump through a hoop or two for DImode, since the hardware has no
12596 non-popping instruction. We used to do this a different way, but
12597 that was somewhat fragile and broke with post-reload splitters. */
12598 if ((dimode_p || fisttp) && !stack_top_dies)
12599 output_asm_insn ("fld\t%y1", operands);
12601 gcc_assert (STACK_TOP_P (operands[1]));
12602 gcc_assert (MEM_P (operands[0]));
12603 gcc_assert (GET_MODE (operands[1]) != TFmode);
12606 output_asm_insn ("fisttp%Z0\t%0", operands);
12609 if (round_mode != I387_CW_ANY)
12610 output_asm_insn ("fldcw\t%3", operands);
12611 if (stack_top_dies || dimode_p)
12612 output_asm_insn ("fistp%Z0\t%0", operands);
12614 output_asm_insn ("fist%Z0\t%0", operands);
12615 if (round_mode != I387_CW_ANY)
12616 output_asm_insn ("fldcw\t%2", operands);
12622 /* Output code for x87 ffreep insn. The OPNO argument, which may only
12623 have the values zero or one, indicates the ffreep insn's operand
12624 from the OPERANDS array. */
12626 static const char *
12627 output_387_ffreep (rtx *operands ATTRIBUTE_UNUSED, int opno)
12629 if (TARGET_USE_FFREEP)
12630 #ifdef HAVE_AS_IX86_FFREEP
12631 return opno ? "ffreep\t%y1" : "ffreep\t%y0";
12634 static char retval[32];
12635 int regno = REGNO (operands[opno]);
12637 gcc_assert (FP_REGNO_P (regno));
12639 regno -= FIRST_STACK_REG;
12641 snprintf (retval, sizeof (retval), ASM_SHORT "0xc%ddf", regno);
12646 return opno ? "fstp\t%y1" : "fstp\t%y0";
12650 /* Output code for INSN to compare OPERANDS. EFLAGS_P is 1 when fcomi
12651 should be used. UNORDERED_P is true when fucom should be used. */
12654 output_fp_compare (rtx insn, rtx *operands, int eflags_p, int unordered_p)
12656 int stack_top_dies;
12657 rtx cmp_op0, cmp_op1;
12658 int is_sse = SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]);
12662 cmp_op0 = operands[0];
12663 cmp_op1 = operands[1];
12667 cmp_op0 = operands[1];
12668 cmp_op1 = operands[2];
12673 static const char ucomiss[] = "vucomiss\t{%1, %0|%0, %1}";
12674 static const char ucomisd[] = "vucomisd\t{%1, %0|%0, %1}";
12675 static const char comiss[] = "vcomiss\t{%1, %0|%0, %1}";
12676 static const char comisd[] = "vcomisd\t{%1, %0|%0, %1}";
12678 if (GET_MODE (operands[0]) == SFmode)
12680 return &ucomiss[TARGET_AVX ? 0 : 1];
12682 return &comiss[TARGET_AVX ? 0 : 1];
12685 return &ucomisd[TARGET_AVX ? 0 : 1];
12687 return &comisd[TARGET_AVX ? 0 : 1];
12690 gcc_assert (STACK_TOP_P (cmp_op0));
12692 stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0;
12694 if (cmp_op1 == CONST0_RTX (GET_MODE (cmp_op1)))
12696 if (stack_top_dies)
12698 output_asm_insn ("ftst\n\tfnstsw\t%0", operands);
12699 return output_387_ffreep (operands, 1);
12702 return "ftst\n\tfnstsw\t%0";
12705 if (STACK_REG_P (cmp_op1)
12707 && find_regno_note (insn, REG_DEAD, REGNO (cmp_op1))
12708 && REGNO (cmp_op1) != FIRST_STACK_REG)
12710 /* If both the top of the 387 stack dies, and the other operand
12711 is also a stack register that dies, then this must be a
12712 `fcompp' float compare */
12716 /* There is no double popping fcomi variant. Fortunately,
12717 eflags is immune from the fstp's cc clobbering. */
12719 output_asm_insn ("fucomip\t{%y1, %0|%0, %y1}", operands);
12721 output_asm_insn ("fcomip\t{%y1, %0|%0, %y1}", operands);
12722 return output_387_ffreep (operands, 0);
12727 return "fucompp\n\tfnstsw\t%0";
12729 return "fcompp\n\tfnstsw\t%0";
12734 /* Encoded here as eflags_p | intmode | unordered_p | stack_top_dies. */
12736 static const char * const alt[16] =
12738 "fcom%Z2\t%y2\n\tfnstsw\t%0",
12739 "fcomp%Z2\t%y2\n\tfnstsw\t%0",
12740 "fucom%Z2\t%y2\n\tfnstsw\t%0",
12741 "fucomp%Z2\t%y2\n\tfnstsw\t%0",
12743 "ficom%Z2\t%y2\n\tfnstsw\t%0",
12744 "ficomp%Z2\t%y2\n\tfnstsw\t%0",
12748 "fcomi\t{%y1, %0|%0, %y1}",
12749 "fcomip\t{%y1, %0|%0, %y1}",
12750 "fucomi\t{%y1, %0|%0, %y1}",
12751 "fucomip\t{%y1, %0|%0, %y1}",
12762 mask = eflags_p << 3;
12763 mask |= (GET_MODE_CLASS (GET_MODE (cmp_op1)) == MODE_INT) << 2;
12764 mask |= unordered_p << 1;
12765 mask |= stack_top_dies;
12767 gcc_assert (mask < 16);
12776 ix86_output_addr_vec_elt (FILE *file, int value)
12778 const char *directive = ASM_LONG;
12782 directive = ASM_QUAD;
12784 gcc_assert (!TARGET_64BIT);
12787 fprintf (file, "%s" LPREFIX "%d\n", directive, value);
12791 ix86_output_addr_diff_elt (FILE *file, int value, int rel)
12793 const char *directive = ASM_LONG;
12796 if (TARGET_64BIT && CASE_VECTOR_MODE == DImode)
12797 directive = ASM_QUAD;
12799 gcc_assert (!TARGET_64BIT);
12801 /* We can't use @GOTOFF for text labels on VxWorks; see gotoff_operand. */
12802 if (TARGET_64BIT || TARGET_VXWORKS_RTP)
12803 fprintf (file, "%s" LPREFIX "%d-" LPREFIX "%d\n",
12804 directive, value, rel);
12805 else if (HAVE_AS_GOTOFF_IN_DATA)
12806 fprintf (file, ASM_LONG LPREFIX "%d@GOTOFF\n", value);
12808 else if (TARGET_MACHO)
12810 fprintf (file, ASM_LONG LPREFIX "%d-", value);
12811 machopic_output_function_base_name (file);
12816 asm_fprintf (file, ASM_LONG "%U%s+[.-" LPREFIX "%d]\n",
12817 GOT_SYMBOL_NAME, value);
12820 /* Generate either "mov $0, reg" or "xor reg, reg", as appropriate
12824 ix86_expand_clear (rtx dest)
12828 /* We play register width games, which are only valid after reload. */
12829 gcc_assert (reload_completed);
12831 /* Avoid HImode and its attendant prefix byte. */
12832 if (GET_MODE_SIZE (GET_MODE (dest)) < 4)
12833 dest = gen_rtx_REG (SImode, REGNO (dest));
12834 tmp = gen_rtx_SET (VOIDmode, dest, const0_rtx);
12836 /* This predicate should match that for movsi_xor and movdi_xor_rex64. */
12837 if (!TARGET_USE_MOV0 || optimize_insn_for_speed_p ())
12839 rtx clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
12840 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob));
12846 /* X is an unchanging MEM. If it is a constant pool reference, return
12847 the constant pool rtx, else NULL. */
12850 maybe_get_pool_constant (rtx x)
12852 x = ix86_delegitimize_address (XEXP (x, 0));
12854 if (GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x))
12855 return get_pool_constant (x);
12861 ix86_expand_move (enum machine_mode mode, rtx operands[])
12864 enum tls_model model;
12869 if (GET_CODE (op1) == SYMBOL_REF)
12871 model = SYMBOL_REF_TLS_MODEL (op1);
12874 op1 = legitimize_tls_address (op1, model, true);
12875 op1 = force_operand (op1, op0);
12879 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
12880 && SYMBOL_REF_DLLIMPORT_P (op1))
12881 op1 = legitimize_dllimport_symbol (op1, false);
12883 else if (GET_CODE (op1) == CONST
12884 && GET_CODE (XEXP (op1, 0)) == PLUS
12885 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SYMBOL_REF)
12887 rtx addend = XEXP (XEXP (op1, 0), 1);
12888 rtx symbol = XEXP (XEXP (op1, 0), 0);
12891 model = SYMBOL_REF_TLS_MODEL (symbol);
12893 tmp = legitimize_tls_address (symbol, model, true);
12894 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
12895 && SYMBOL_REF_DLLIMPORT_P (symbol))
12896 tmp = legitimize_dllimport_symbol (symbol, true);
12900 tmp = force_operand (tmp, NULL);
12901 tmp = expand_simple_binop (Pmode, PLUS, tmp, addend,
12902 op0, 1, OPTAB_DIRECT);
12908 if (flag_pic && mode == Pmode && symbolic_operand (op1, Pmode))
12910 if (TARGET_MACHO && !TARGET_64BIT)
12915 rtx temp = ((reload_in_progress
12916 || ((op0 && REG_P (op0))
12918 ? op0 : gen_reg_rtx (Pmode));
12919 op1 = machopic_indirect_data_reference (op1, temp);
12920 op1 = machopic_legitimize_pic_address (op1, mode,
12921 temp == op1 ? 0 : temp);
12923 else if (MACHOPIC_INDIRECT)
12924 op1 = machopic_indirect_data_reference (op1, 0);
12932 op1 = force_reg (Pmode, op1);
12933 else if (!TARGET_64BIT || !x86_64_movabs_operand (op1, Pmode))
12935 rtx reg = can_create_pseudo_p () ? NULL_RTX : op0;
12936 op1 = legitimize_pic_address (op1, reg);
12945 && (PUSH_ROUNDING (GET_MODE_SIZE (mode)) != GET_MODE_SIZE (mode)
12946 || !push_operand (op0, mode))
12948 op1 = force_reg (mode, op1);
12950 if (push_operand (op0, mode)
12951 && ! general_no_elim_operand (op1, mode))
12952 op1 = copy_to_mode_reg (mode, op1);
12954 /* Force large constants in 64bit compilation into register
12955 to get them CSEed. */
12956 if (can_create_pseudo_p ()
12957 && (mode == DImode) && TARGET_64BIT
12958 && immediate_operand (op1, mode)
12959 && !x86_64_zext_immediate_operand (op1, VOIDmode)
12960 && !register_operand (op0, mode)
12962 op1 = copy_to_mode_reg (mode, op1);
12964 if (can_create_pseudo_p ()
12965 && FLOAT_MODE_P (mode)
12966 && GET_CODE (op1) == CONST_DOUBLE)
12968 /* If we are loading a floating point constant to a register,
12969 force the value to memory now, since we'll get better code
12970 out the back end. */
12972 op1 = validize_mem (force_const_mem (mode, op1));
12973 if (!register_operand (op0, mode))
12975 rtx temp = gen_reg_rtx (mode);
12976 emit_insn (gen_rtx_SET (VOIDmode, temp, op1));
12977 emit_move_insn (op0, temp);
12983 emit_insn (gen_rtx_SET (VOIDmode, op0, op1));
12987 ix86_expand_vector_move (enum machine_mode mode, rtx operands[])
12989 rtx op0 = operands[0], op1 = operands[1];
12990 unsigned int align = GET_MODE_ALIGNMENT (mode);
12992 /* Force constants other than zero into memory. We do not know how
12993 the instructions used to build constants modify the upper 64 bits
12994 of the register, once we have that information we may be able
12995 to handle some of them more efficiently. */
12996 if (can_create_pseudo_p ()
12997 && register_operand (op0, mode)
12998 && (CONSTANT_P (op1)
12999 || (GET_CODE (op1) == SUBREG
13000 && CONSTANT_P (SUBREG_REG (op1))))
13001 && !standard_sse_constant_p (op1))
13002 op1 = validize_mem (force_const_mem (mode, op1));
13004 /* We need to check memory alignment for SSE mode since attribute
13005 can make operands unaligned. */
13006 if (can_create_pseudo_p ()
13007 && SSE_REG_MODE_P (mode)
13008 && ((MEM_P (op0) && (MEM_ALIGN (op0) < align))
13009 || (MEM_P (op1) && (MEM_ALIGN (op1) < align))))
13013 /* ix86_expand_vector_move_misalign() does not like constants ... */
13014 if (CONSTANT_P (op1)
13015 || (GET_CODE (op1) == SUBREG
13016 && CONSTANT_P (SUBREG_REG (op1))))
13017 op1 = validize_mem (force_const_mem (mode, op1));
13019 /* ... nor both arguments in memory. */
13020 if (!register_operand (op0, mode)
13021 && !register_operand (op1, mode))
13022 op1 = force_reg (mode, op1);
13024 tmp[0] = op0; tmp[1] = op1;
13025 ix86_expand_vector_move_misalign (mode, tmp);
13029 /* Make operand1 a register if it isn't already. */
13030 if (can_create_pseudo_p ()
13031 && !register_operand (op0, mode)
13032 && !register_operand (op1, mode))
13034 emit_move_insn (op0, force_reg (GET_MODE (op0), op1));
13038 emit_insn (gen_rtx_SET (VOIDmode, op0, op1));
13041 /* Implement the movmisalign patterns for SSE. Non-SSE modes go
13042 straight to ix86_expand_vector_move. */
13043 /* Code generation for scalar reg-reg moves of single and double precision data:
13044 if (x86_sse_partial_reg_dependency == true | x86_sse_split_regs == true)
13048 if (x86_sse_partial_reg_dependency == true)
13053 Code generation for scalar loads of double precision data:
13054 if (x86_sse_split_regs == true)
13055 movlpd mem, reg (gas syntax)
13059 Code generation for unaligned packed loads of single precision data
13060 (x86_sse_unaligned_move_optimal overrides x86_sse_partial_reg_dependency):
13061 if (x86_sse_unaligned_move_optimal)
13064 if (x86_sse_partial_reg_dependency == true)
13076 Code generation for unaligned packed loads of double precision data
13077 (x86_sse_unaligned_move_optimal overrides x86_sse_split_regs):
13078 if (x86_sse_unaligned_move_optimal)
13081 if (x86_sse_split_regs == true)
13094 ix86_expand_vector_move_misalign (enum machine_mode mode, rtx operands[])
13103 switch (GET_MODE_CLASS (mode))
13105 case MODE_VECTOR_INT:
13107 switch (GET_MODE_SIZE (mode))
13110 op0 = gen_lowpart (V16QImode, op0);
13111 op1 = gen_lowpart (V16QImode, op1);
13112 emit_insn (gen_avx_movdqu (op0, op1));
13115 op0 = gen_lowpart (V32QImode, op0);
13116 op1 = gen_lowpart (V32QImode, op1);
13117 emit_insn (gen_avx_movdqu256 (op0, op1));
13120 gcc_unreachable ();
13123 case MODE_VECTOR_FLOAT:
13124 op0 = gen_lowpart (mode, op0);
13125 op1 = gen_lowpart (mode, op1);
13130 emit_insn (gen_avx_movups (op0, op1));
13133 emit_insn (gen_avx_movups256 (op0, op1));
13136 emit_insn (gen_avx_movupd (op0, op1));
13139 emit_insn (gen_avx_movupd256 (op0, op1));
13142 gcc_unreachable ();
13147 gcc_unreachable ();
13155 /* If we're optimizing for size, movups is the smallest. */
13156 if (optimize_insn_for_size_p ())
13158 op0 = gen_lowpart (V4SFmode, op0);
13159 op1 = gen_lowpart (V4SFmode, op1);
13160 emit_insn (gen_sse_movups (op0, op1));
13164 /* ??? If we have typed data, then it would appear that using
13165 movdqu is the only way to get unaligned data loaded with
13167 if (TARGET_SSE2 && GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
13169 op0 = gen_lowpart (V16QImode, op0);
13170 op1 = gen_lowpart (V16QImode, op1);
13171 emit_insn (gen_sse2_movdqu (op0, op1));
13175 if (TARGET_SSE2 && mode == V2DFmode)
13179 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL)
13181 op0 = gen_lowpart (V2DFmode, op0);
13182 op1 = gen_lowpart (V2DFmode, op1);
13183 emit_insn (gen_sse2_movupd (op0, op1));
13187 /* When SSE registers are split into halves, we can avoid
13188 writing to the top half twice. */
13189 if (TARGET_SSE_SPLIT_REGS)
13191 emit_clobber (op0);
13196 /* ??? Not sure about the best option for the Intel chips.
13197 The following would seem to satisfy; the register is
13198 entirely cleared, breaking the dependency chain. We
13199 then store to the upper half, with a dependency depth
13200 of one. A rumor has it that Intel recommends two movsd
13201 followed by an unpacklpd, but this is unconfirmed. And
13202 given that the dependency depth of the unpacklpd would
13203 still be one, I'm not sure why this would be better. */
13204 zero = CONST0_RTX (V2DFmode);
13207 m = adjust_address (op1, DFmode, 0);
13208 emit_insn (gen_sse2_loadlpd (op0, zero, m));
13209 m = adjust_address (op1, DFmode, 8);
13210 emit_insn (gen_sse2_loadhpd (op0, op0, m));
13214 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL)
13216 op0 = gen_lowpart (V4SFmode, op0);
13217 op1 = gen_lowpart (V4SFmode, op1);
13218 emit_insn (gen_sse_movups (op0, op1));
13222 if (TARGET_SSE_PARTIAL_REG_DEPENDENCY)
13223 emit_move_insn (op0, CONST0_RTX (mode));
13225 emit_clobber (op0);
13227 if (mode != V4SFmode)
13228 op0 = gen_lowpart (V4SFmode, op0);
13229 m = adjust_address (op1, V2SFmode, 0);
13230 emit_insn (gen_sse_loadlps (op0, op0, m));
13231 m = adjust_address (op1, V2SFmode, 8);
13232 emit_insn (gen_sse_loadhps (op0, op0, m));
13235 else if (MEM_P (op0))
13237 /* If we're optimizing for size, movups is the smallest. */
13238 if (optimize_insn_for_size_p ())
13240 op0 = gen_lowpart (V4SFmode, op0);
13241 op1 = gen_lowpart (V4SFmode, op1);
13242 emit_insn (gen_sse_movups (op0, op1));
13246 /* ??? Similar to above, only less clear because of quote
13247 typeless stores unquote. */
13248 if (TARGET_SSE2 && !TARGET_SSE_TYPELESS_STORES
13249 && GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
13251 op0 = gen_lowpart (V16QImode, op0);
13252 op1 = gen_lowpart (V16QImode, op1);
13253 emit_insn (gen_sse2_movdqu (op0, op1));
13257 if (TARGET_SSE2 && mode == V2DFmode)
13259 m = adjust_address (op0, DFmode, 0);
13260 emit_insn (gen_sse2_storelpd (m, op1));
13261 m = adjust_address (op0, DFmode, 8);
13262 emit_insn (gen_sse2_storehpd (m, op1));
13266 if (mode != V4SFmode)
13267 op1 = gen_lowpart (V4SFmode, op1);
13268 m = adjust_address (op0, V2SFmode, 0);
13269 emit_insn (gen_sse_storelps (m, op1));
13270 m = adjust_address (op0, V2SFmode, 8);
13271 emit_insn (gen_sse_storehps (m, op1));
13275 gcc_unreachable ();
13278 /* Expand a push in MODE. This is some mode for which we do not support
13279 proper push instructions, at least from the registers that we expect
13280 the value to live in. */
13283 ix86_expand_push (enum machine_mode mode, rtx x)
13287 tmp = expand_simple_binop (Pmode, PLUS, stack_pointer_rtx,
13288 GEN_INT (-GET_MODE_SIZE (mode)),
13289 stack_pointer_rtx, 1, OPTAB_DIRECT);
13290 if (tmp != stack_pointer_rtx)
13291 emit_move_insn (stack_pointer_rtx, tmp);
13293 tmp = gen_rtx_MEM (mode, stack_pointer_rtx);
13295 /* When we push an operand onto stack, it has to be aligned at least
13296 at the function argument boundary. However since we don't have
13297 the argument type, we can't determine the actual argument
13299 emit_move_insn (tmp, x);
13302 /* Helper function of ix86_fixup_binary_operands to canonicalize
13303 operand order. Returns true if the operands should be swapped. */
13306 ix86_swap_binary_operands_p (enum rtx_code code, enum machine_mode mode,
13309 rtx dst = operands[0];
13310 rtx src1 = operands[1];
13311 rtx src2 = operands[2];
13313 /* If the operation is not commutative, we can't do anything. */
13314 if (GET_RTX_CLASS (code) != RTX_COMM_ARITH)
13317 /* Highest priority is that src1 should match dst. */
13318 if (rtx_equal_p (dst, src1))
13320 if (rtx_equal_p (dst, src2))
13323 /* Next highest priority is that immediate constants come second. */
13324 if (immediate_operand (src2, mode))
13326 if (immediate_operand (src1, mode))
13329 /* Lowest priority is that memory references should come second. */
13339 /* Fix up OPERANDS to satisfy ix86_binary_operator_ok. Return the
13340 destination to use for the operation. If different from the true
13341 destination in operands[0], a copy operation will be required. */
13344 ix86_fixup_binary_operands (enum rtx_code code, enum machine_mode mode,
13347 rtx dst = operands[0];
13348 rtx src1 = operands[1];
13349 rtx src2 = operands[2];
13351 /* Canonicalize operand order. */
13352 if (ix86_swap_binary_operands_p (code, mode, operands))
13356 /* It is invalid to swap operands of different modes. */
13357 gcc_assert (GET_MODE (src1) == GET_MODE (src2));
13364 /* Both source operands cannot be in memory. */
13365 if (MEM_P (src1) && MEM_P (src2))
13367 /* Optimization: Only read from memory once. */
13368 if (rtx_equal_p (src1, src2))
13370 src2 = force_reg (mode, src2);
13374 src2 = force_reg (mode, src2);
13377 /* If the destination is memory, and we do not have matching source
13378 operands, do things in registers. */
13379 if (MEM_P (dst) && !rtx_equal_p (dst, src1))
13380 dst = gen_reg_rtx (mode);
13382 /* Source 1 cannot be a constant. */
13383 if (CONSTANT_P (src1))
13384 src1 = force_reg (mode, src1);
13386 /* Source 1 cannot be a non-matching memory. */
13387 if (MEM_P (src1) && !rtx_equal_p (dst, src1))
13388 src1 = force_reg (mode, src1);
13390 /* In order for the multiply-add patterns to get matched, we need
13391 to aid combine by forcing all operands into registers to start. */
13392 if (optimize && TARGET_FMA4)
13395 src2 = force_reg (GET_MODE (src2), src2);
13396 else if (MEM_P (src1))
13397 src1 = force_reg (GET_MODE (src1), src1);
13400 operands[1] = src1;
13401 operands[2] = src2;
13405 /* Similarly, but assume that the destination has already been
13406 set up properly. */
13409 ix86_fixup_binary_operands_no_copy (enum rtx_code code,
13410 enum machine_mode mode, rtx operands[])
13412 rtx dst = ix86_fixup_binary_operands (code, mode, operands);
13413 gcc_assert (dst == operands[0]);
13416 /* Attempt to expand a binary operator. Make the expansion closer to the
13417 actual machine, then just general_operand, which will allow 3 separate
13418 memory references (one output, two input) in a single insn. */
13421 ix86_expand_binary_operator (enum rtx_code code, enum machine_mode mode,
13424 rtx src1, src2, dst, op, clob;
13426 dst = ix86_fixup_binary_operands (code, mode, operands);
13427 src1 = operands[1];
13428 src2 = operands[2];
13430 /* Emit the instruction. */
13432 op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_ee (code, mode, src1, src2));
13433 if (reload_in_progress)
13435 /* Reload doesn't know about the flags register, and doesn't know that
13436 it doesn't want to clobber it. We can only do this with PLUS. */
13437 gcc_assert (code == PLUS);
13442 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
13443 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob)));
13446 /* Fix up the destination if needed. */
13447 if (dst != operands[0])
13448 emit_move_insn (operands[0], dst);
13451 /* Return TRUE or FALSE depending on whether the binary operator meets the
13452 appropriate constraints. */
13455 ix86_binary_operator_ok (enum rtx_code code, enum machine_mode mode,
13458 rtx dst = operands[0];
13459 rtx src1 = operands[1];
13460 rtx src2 = operands[2];
13462 /* Both source operands cannot be in memory. */
13463 if (MEM_P (src1) && MEM_P (src2))
13466 /* Canonicalize operand order for commutative operators. */
13467 if (ix86_swap_binary_operands_p (code, mode, operands))
13474 /* If the destination is memory, we must have a matching source operand. */
13475 if (MEM_P (dst) && !rtx_equal_p (dst, src1))
13478 /* Source 1 cannot be a constant. */
13479 if (CONSTANT_P (src1))
13482 /* Source 1 cannot be a non-matching memory. */
13483 if (MEM_P (src1) && !rtx_equal_p (dst, src1))
13489 /* Attempt to expand a unary operator. Make the expansion closer to the
13490 actual machine, then just general_operand, which will allow 2 separate
13491 memory references (one output, one input) in a single insn. */
13494 ix86_expand_unary_operator (enum rtx_code code, enum machine_mode mode,
13497 int matching_memory;
13498 rtx src, dst, op, clob;
13503 /* If the destination is memory, and we do not have matching source
13504 operands, do things in registers. */
13505 matching_memory = 0;
13508 if (rtx_equal_p (dst, src))
13509 matching_memory = 1;
13511 dst = gen_reg_rtx (mode);
13514 /* When source operand is memory, destination must match. */
13515 if (MEM_P (src) && !matching_memory)
13516 src = force_reg (mode, src);
13518 /* Emit the instruction. */
13520 op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_e (code, mode, src));
13521 if (reload_in_progress || code == NOT)
13523 /* Reload doesn't know about the flags register, and doesn't know that
13524 it doesn't want to clobber it. */
13525 gcc_assert (code == NOT);
13530 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
13531 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob)));
13534 /* Fix up the destination if needed. */
13535 if (dst != operands[0])
13536 emit_move_insn (operands[0], dst);
13539 #define LEA_SEARCH_THRESHOLD 12
13541 /* Search backward for non-agu definition of register number REGNO1
13542 or register number REGNO2 in INSN's basic block until
13543 1. Pass LEA_SEARCH_THRESHOLD instructions, or
13544 2. Reach BB boundary, or
13545 3. Reach agu definition.
13546 Returns the distance between the non-agu definition point and INSN.
13547 If no definition point, returns -1. */
13550 distance_non_agu_define (unsigned int regno1, unsigned int regno2,
13553 basic_block bb = BLOCK_FOR_INSN (insn);
13556 enum attr_type insn_type;
13558 if (insn != BB_HEAD (bb))
13560 rtx prev = PREV_INSN (insn);
13561 while (prev && distance < LEA_SEARCH_THRESHOLD)
13566 for (def_rec = DF_INSN_DEFS (prev); *def_rec; def_rec++)
13567 if (DF_REF_TYPE (*def_rec) == DF_REF_REG_DEF
13568 && !DF_REF_IS_ARTIFICIAL (*def_rec)
13569 && (regno1 == DF_REF_REGNO (*def_rec)
13570 || regno2 == DF_REF_REGNO (*def_rec)))
13572 insn_type = get_attr_type (prev);
13573 if (insn_type != TYPE_LEA)
13577 if (prev == BB_HEAD (bb))
13579 prev = PREV_INSN (prev);
13583 if (distance < LEA_SEARCH_THRESHOLD)
13587 bool simple_loop = false;
13589 FOR_EACH_EDGE (e, ei, bb->preds)
13592 simple_loop = true;
13598 rtx prev = BB_END (bb);
13601 && distance < LEA_SEARCH_THRESHOLD)
13606 for (def_rec = DF_INSN_DEFS (prev); *def_rec; def_rec++)
13607 if (DF_REF_TYPE (*def_rec) == DF_REF_REG_DEF
13608 && !DF_REF_IS_ARTIFICIAL (*def_rec)
13609 && (regno1 == DF_REF_REGNO (*def_rec)
13610 || regno2 == DF_REF_REGNO (*def_rec)))
13612 insn_type = get_attr_type (prev);
13613 if (insn_type != TYPE_LEA)
13617 prev = PREV_INSN (prev);
13625 /* get_attr_type may modify recog data. We want to make sure
13626 that recog data is valid for instruction INSN, on which
13627 distance_non_agu_define is called. INSN is unchanged here. */
13628 extract_insn_cached (insn);
13632 /* Return the distance between INSN and the next insn that uses
13633 register number REGNO0 in memory address. Return -1 if no such
13634 a use is found within LEA_SEARCH_THRESHOLD or REGNO0 is set. */
13637 distance_agu_use (unsigned int regno0, rtx insn)
13639 basic_block bb = BLOCK_FOR_INSN (insn);
13644 if (insn != BB_END (bb))
13646 rtx next = NEXT_INSN (insn);
13647 while (next && distance < LEA_SEARCH_THRESHOLD)
13653 for (use_rec = DF_INSN_USES (next); *use_rec; use_rec++)
13654 if ((DF_REF_TYPE (*use_rec) == DF_REF_REG_MEM_LOAD
13655 || DF_REF_TYPE (*use_rec) == DF_REF_REG_MEM_STORE)
13656 && regno0 == DF_REF_REGNO (*use_rec))
13658 /* Return DISTANCE if OP0 is used in memory
13659 address in NEXT. */
13663 for (def_rec = DF_INSN_DEFS (next); *def_rec; def_rec++)
13664 if (DF_REF_TYPE (*def_rec) == DF_REF_REG_DEF
13665 && !DF_REF_IS_ARTIFICIAL (*def_rec)
13666 && regno0 == DF_REF_REGNO (*def_rec))
13668 /* Return -1 if OP0 is set in NEXT. */
13672 if (next == BB_END (bb))
13674 next = NEXT_INSN (next);
13678 if (distance < LEA_SEARCH_THRESHOLD)
13682 bool simple_loop = false;
13684 FOR_EACH_EDGE (e, ei, bb->succs)
13687 simple_loop = true;
13693 rtx next = BB_HEAD (bb);
13696 && distance < LEA_SEARCH_THRESHOLD)
13702 for (use_rec = DF_INSN_USES (next); *use_rec; use_rec++)
13703 if ((DF_REF_TYPE (*use_rec) == DF_REF_REG_MEM_LOAD
13704 || DF_REF_TYPE (*use_rec) == DF_REF_REG_MEM_STORE)
13705 && regno0 == DF_REF_REGNO (*use_rec))
13707 /* Return DISTANCE if OP0 is used in memory
13708 address in NEXT. */
13712 for (def_rec = DF_INSN_DEFS (next); *def_rec; def_rec++)
13713 if (DF_REF_TYPE (*def_rec) == DF_REF_REG_DEF
13714 && !DF_REF_IS_ARTIFICIAL (*def_rec)
13715 && regno0 == DF_REF_REGNO (*def_rec))
13717 /* Return -1 if OP0 is set in NEXT. */
13722 next = NEXT_INSN (next);
13730 /* Define this macro to tune LEA priority vs ADD, it take effect when
13731 there is a dilemma of choicing LEA or ADD
13732 Negative value: ADD is more preferred than LEA
13734 Positive value: LEA is more preferred than ADD*/
13735 #define IX86_LEA_PRIORITY 2
13737 /* Return true if it is ok to optimize an ADD operation to LEA
13738 operation to avoid flag register consumation. For the processors
13739 like ATOM, if the destination register of LEA holds an actual
13740 address which will be used soon, LEA is better and otherwise ADD
13744 ix86_lea_for_add_ok (enum rtx_code code ATTRIBUTE_UNUSED,
13745 rtx insn, rtx operands[])
13747 unsigned int regno0 = true_regnum (operands[0]);
13748 unsigned int regno1 = true_regnum (operands[1]);
13749 unsigned int regno2;
13751 if (!TARGET_OPT_AGU || optimize_function_for_size_p (cfun))
13752 return regno0 != regno1;
13754 regno2 = true_regnum (operands[2]);
13756 /* If a = b + c, (a!=b && a!=c), must use lea form. */
13757 if (regno0 != regno1 && regno0 != regno2)
13761 int dist_define, dist_use;
13762 dist_define = distance_non_agu_define (regno1, regno2, insn);
13763 if (dist_define <= 0)
13766 /* If this insn has both backward non-agu dependence and forward
13767 agu dependence, the one with short distance take effect. */
13768 dist_use = distance_agu_use (regno0, insn);
13770 || (dist_define + IX86_LEA_PRIORITY) < dist_use)
13777 /* Return true if destination reg of SET_BODY is shift count of
13781 ix86_dep_by_shift_count_body (const_rtx set_body, const_rtx use_body)
13787 /* Retrieve destination of SET_BODY. */
13788 switch (GET_CODE (set_body))
13791 set_dest = SET_DEST (set_body);
13792 if (!set_dest || !REG_P (set_dest))
13796 for (i = XVECLEN (set_body, 0) - 1; i >= 0; i--)
13797 if (ix86_dep_by_shift_count_body (XVECEXP (set_body, 0, i),
13805 /* Retrieve shift count of USE_BODY. */
13806 switch (GET_CODE (use_body))
13809 shift_rtx = XEXP (use_body, 1);
13812 for (i = XVECLEN (use_body, 0) - 1; i >= 0; i--)
13813 if (ix86_dep_by_shift_count_body (set_body,
13814 XVECEXP (use_body, 0, i)))
13822 && (GET_CODE (shift_rtx) == ASHIFT
13823 || GET_CODE (shift_rtx) == LSHIFTRT
13824 || GET_CODE (shift_rtx) == ASHIFTRT
13825 || GET_CODE (shift_rtx) == ROTATE
13826 || GET_CODE (shift_rtx) == ROTATERT))
13828 rtx shift_count = XEXP (shift_rtx, 1);
13830 /* Return true if shift count is dest of SET_BODY. */
13831 if (REG_P (shift_count)
13832 && true_regnum (set_dest) == true_regnum (shift_count))
13839 /* Return true if destination reg of SET_INSN is shift count of
13843 ix86_dep_by_shift_count (const_rtx set_insn, const_rtx use_insn)
13845 return ix86_dep_by_shift_count_body (PATTERN (set_insn),
13846 PATTERN (use_insn));
13849 /* Return TRUE or FALSE depending on whether the unary operator meets the
13850 appropriate constraints. */
13853 ix86_unary_operator_ok (enum rtx_code code ATTRIBUTE_UNUSED,
13854 enum machine_mode mode ATTRIBUTE_UNUSED,
13855 rtx operands[2] ATTRIBUTE_UNUSED)
13857 /* If one of operands is memory, source and destination must match. */
13858 if ((MEM_P (operands[0])
13859 || MEM_P (operands[1]))
13860 && ! rtx_equal_p (operands[0], operands[1]))
13865 /* Return TRUE if the operands to a vec_interleave_{high,low}v2df
13866 are ok, keeping in mind the possible movddup alternative. */
13869 ix86_vec_interleave_v2df_operator_ok (rtx operands[3], bool high)
13871 if (MEM_P (operands[0]))
13872 return rtx_equal_p (operands[0], operands[1 + high]);
13873 if (MEM_P (operands[1]) && MEM_P (operands[2]))
13874 return TARGET_SSE3 && rtx_equal_p (operands[1], operands[2]);
13878 /* Post-reload splitter for converting an SF or DFmode value in an
13879 SSE register into an unsigned SImode. */
13882 ix86_split_convert_uns_si_sse (rtx operands[])
13884 enum machine_mode vecmode;
13885 rtx value, large, zero_or_two31, input, two31, x;
13887 large = operands[1];
13888 zero_or_two31 = operands[2];
13889 input = operands[3];
13890 two31 = operands[4];
13891 vecmode = GET_MODE (large);
13892 value = gen_rtx_REG (vecmode, REGNO (operands[0]));
13894 /* Load up the value into the low element. We must ensure that the other
13895 elements are valid floats -- zero is the easiest such value. */
13898 if (vecmode == V4SFmode)
13899 emit_insn (gen_vec_setv4sf_0 (value, CONST0_RTX (V4SFmode), input));
13901 emit_insn (gen_sse2_loadlpd (value, CONST0_RTX (V2DFmode), input));
13905 input = gen_rtx_REG (vecmode, REGNO (input));
13906 emit_move_insn (value, CONST0_RTX (vecmode));
13907 if (vecmode == V4SFmode)
13908 emit_insn (gen_sse_movss (value, value, input));
13910 emit_insn (gen_sse2_movsd (value, value, input));
13913 emit_move_insn (large, two31);
13914 emit_move_insn (zero_or_two31, MEM_P (two31) ? large : two31);
13916 x = gen_rtx_fmt_ee (LE, vecmode, large, value);
13917 emit_insn (gen_rtx_SET (VOIDmode, large, x));
13919 x = gen_rtx_AND (vecmode, zero_or_two31, large);
13920 emit_insn (gen_rtx_SET (VOIDmode, zero_or_two31, x));
13922 x = gen_rtx_MINUS (vecmode, value, zero_or_two31);
13923 emit_insn (gen_rtx_SET (VOIDmode, value, x));
13925 large = gen_rtx_REG (V4SImode, REGNO (large));
13926 emit_insn (gen_ashlv4si3 (large, large, GEN_INT (31)));
13928 x = gen_rtx_REG (V4SImode, REGNO (value));
13929 if (vecmode == V4SFmode)
13930 emit_insn (gen_sse2_cvttps2dq (x, value));
13932 emit_insn (gen_sse2_cvttpd2dq (x, value));
13935 emit_insn (gen_xorv4si3 (value, value, large));
13938 /* Convert an unsigned DImode value into a DFmode, using only SSE.
13939 Expects the 64-bit DImode to be supplied in a pair of integral
13940 registers. Requires SSE2; will use SSE3 if available. For x86_32,
13941 -mfpmath=sse, !optimize_size only. */
13944 ix86_expand_convert_uns_didf_sse (rtx target, rtx input)
13946 REAL_VALUE_TYPE bias_lo_rvt, bias_hi_rvt;
13947 rtx int_xmm, fp_xmm;
13948 rtx biases, exponents;
13951 int_xmm = gen_reg_rtx (V4SImode);
13952 if (TARGET_INTER_UNIT_MOVES)
13953 emit_insn (gen_movdi_to_sse (int_xmm, input));
13954 else if (TARGET_SSE_SPLIT_REGS)
13956 emit_clobber (int_xmm);
13957 emit_move_insn (gen_lowpart (DImode, int_xmm), input);
13961 x = gen_reg_rtx (V2DImode);
13962 ix86_expand_vector_init_one_nonzero (false, V2DImode, x, input, 0);
13963 emit_move_insn (int_xmm, gen_lowpart (V4SImode, x));
13966 x = gen_rtx_CONST_VECTOR (V4SImode,
13967 gen_rtvec (4, GEN_INT (0x43300000UL),
13968 GEN_INT (0x45300000UL),
13969 const0_rtx, const0_rtx));
13970 exponents = validize_mem (force_const_mem (V4SImode, x));
13972 /* int_xmm = {0x45300000UL, fp_xmm/hi, 0x43300000, fp_xmm/lo } */
13973 emit_insn (gen_vec_interleave_lowv4si (int_xmm, int_xmm, exponents));
13975 /* Concatenating (juxtaposing) (0x43300000UL ## fp_value_low_xmm)
13976 yields a valid DF value equal to (0x1.0p52 + double(fp_value_lo_xmm)).
13977 Similarly (0x45300000UL ## fp_value_hi_xmm) yields
13978 (0x1.0p84 + double(fp_value_hi_xmm)).
13979 Note these exponents differ by 32. */
13981 fp_xmm = copy_to_mode_reg (V2DFmode, gen_lowpart (V2DFmode, int_xmm));
13983 /* Subtract off those 0x1.0p52 and 0x1.0p84 biases, to produce values
13984 in [0,2**32-1] and [0]+[2**32,2**64-1] respectively. */
13985 real_ldexp (&bias_lo_rvt, &dconst1, 52);
13986 real_ldexp (&bias_hi_rvt, &dconst1, 84);
13987 biases = const_double_from_real_value (bias_lo_rvt, DFmode);
13988 x = const_double_from_real_value (bias_hi_rvt, DFmode);
13989 biases = gen_rtx_CONST_VECTOR (V2DFmode, gen_rtvec (2, biases, x));
13990 biases = validize_mem (force_const_mem (V2DFmode, biases));
13991 emit_insn (gen_subv2df3 (fp_xmm, fp_xmm, biases));
13993 /* Add the upper and lower DFmode values together. */
13995 emit_insn (gen_sse3_haddv2df3 (fp_xmm, fp_xmm, fp_xmm));
13998 x = copy_to_mode_reg (V2DFmode, fp_xmm);
13999 emit_insn (gen_vec_interleave_highv2df (fp_xmm, fp_xmm, fp_xmm));
14000 emit_insn (gen_addv2df3 (fp_xmm, fp_xmm, x));
14003 ix86_expand_vector_extract (false, target, fp_xmm, 0);
14006 /* Not used, but eases macroization of patterns. */
14008 ix86_expand_convert_uns_sixf_sse (rtx target ATTRIBUTE_UNUSED,
14009 rtx input ATTRIBUTE_UNUSED)
14011 gcc_unreachable ();
14014 /* Convert an unsigned SImode value into a DFmode. Only currently used
14015 for SSE, but applicable anywhere. */
14018 ix86_expand_convert_uns_sidf_sse (rtx target, rtx input)
14020 REAL_VALUE_TYPE TWO31r;
14023 x = expand_simple_binop (SImode, PLUS, input, GEN_INT (-2147483647 - 1),
14024 NULL, 1, OPTAB_DIRECT);
14026 fp = gen_reg_rtx (DFmode);
14027 emit_insn (gen_floatsidf2 (fp, x));
14029 real_ldexp (&TWO31r, &dconst1, 31);
14030 x = const_double_from_real_value (TWO31r, DFmode);
14032 x = expand_simple_binop (DFmode, PLUS, fp, x, target, 0, OPTAB_DIRECT);
14034 emit_move_insn (target, x);
14037 /* Convert a signed DImode value into a DFmode. Only used for SSE in
14038 32-bit mode; otherwise we have a direct convert instruction. */
14041 ix86_expand_convert_sign_didf_sse (rtx target, rtx input)
14043 REAL_VALUE_TYPE TWO32r;
14044 rtx fp_lo, fp_hi, x;
14046 fp_lo = gen_reg_rtx (DFmode);
14047 fp_hi = gen_reg_rtx (DFmode);
14049 emit_insn (gen_floatsidf2 (fp_hi, gen_highpart (SImode, input)));
14051 real_ldexp (&TWO32r, &dconst1, 32);
14052 x = const_double_from_real_value (TWO32r, DFmode);
14053 fp_hi = expand_simple_binop (DFmode, MULT, fp_hi, x, fp_hi, 0, OPTAB_DIRECT);
14055 ix86_expand_convert_uns_sidf_sse (fp_lo, gen_lowpart (SImode, input));
14057 x = expand_simple_binop (DFmode, PLUS, fp_hi, fp_lo, target,
14060 emit_move_insn (target, x);
14063 /* Convert an unsigned SImode value into a SFmode, using only SSE.
14064 For x86_32, -mfpmath=sse, !optimize_size only. */
14066 ix86_expand_convert_uns_sisf_sse (rtx target, rtx input)
14068 REAL_VALUE_TYPE ONE16r;
14069 rtx fp_hi, fp_lo, int_hi, int_lo, x;
14071 real_ldexp (&ONE16r, &dconst1, 16);
14072 x = const_double_from_real_value (ONE16r, SFmode);
14073 int_lo = expand_simple_binop (SImode, AND, input, GEN_INT(0xffff),
14074 NULL, 0, OPTAB_DIRECT);
14075 int_hi = expand_simple_binop (SImode, LSHIFTRT, input, GEN_INT(16),
14076 NULL, 0, OPTAB_DIRECT);
14077 fp_hi = gen_reg_rtx (SFmode);
14078 fp_lo = gen_reg_rtx (SFmode);
14079 emit_insn (gen_floatsisf2 (fp_hi, int_hi));
14080 emit_insn (gen_floatsisf2 (fp_lo, int_lo));
14081 fp_hi = expand_simple_binop (SFmode, MULT, fp_hi, x, fp_hi,
14083 fp_hi = expand_simple_binop (SFmode, PLUS, fp_hi, fp_lo, target,
14085 if (!rtx_equal_p (target, fp_hi))
14086 emit_move_insn (target, fp_hi);
14089 /* A subroutine of ix86_build_signbit_mask. If VECT is true,
14090 then replicate the value for all elements of the vector
14094 ix86_build_const_vector (enum machine_mode mode, bool vect, rtx value)
14101 v = gen_rtvec (4, value, value, value, value);
14102 return gen_rtx_CONST_VECTOR (V4SImode, v);
14106 v = gen_rtvec (2, value, value);
14107 return gen_rtx_CONST_VECTOR (V2DImode, v);
14111 v = gen_rtvec (4, value, value, value, value);
14113 v = gen_rtvec (4, value, CONST0_RTX (SFmode),
14114 CONST0_RTX (SFmode), CONST0_RTX (SFmode));
14115 return gen_rtx_CONST_VECTOR (V4SFmode, v);
14119 v = gen_rtvec (2, value, value);
14121 v = gen_rtvec (2, value, CONST0_RTX (DFmode));
14122 return gen_rtx_CONST_VECTOR (V2DFmode, v);
14125 gcc_unreachable ();
14129 /* A subroutine of ix86_expand_fp_absneg_operator, copysign expanders
14130 and ix86_expand_int_vcond. Create a mask for the sign bit in MODE
14131 for an SSE register. If VECT is true, then replicate the mask for
14132 all elements of the vector register. If INVERT is true, then create
14133 a mask excluding the sign bit. */
14136 ix86_build_signbit_mask (enum machine_mode mode, bool vect, bool invert)
14138 enum machine_mode vec_mode, imode;
14139 HOST_WIDE_INT hi, lo;
14144 /* Find the sign bit, sign extended to 2*HWI. */
14150 vec_mode = (mode == SImode) ? V4SImode : V4SFmode;
14151 lo = 0x80000000, hi = lo < 0;
14157 vec_mode = (mode == DImode) ? V2DImode : V2DFmode;
14158 if (HOST_BITS_PER_WIDE_INT >= 64)
14159 lo = (HOST_WIDE_INT)1 << shift, hi = -1;
14161 lo = 0, hi = (HOST_WIDE_INT)1 << (shift - HOST_BITS_PER_WIDE_INT);
14166 vec_mode = VOIDmode;
14167 if (HOST_BITS_PER_WIDE_INT >= 64)
14170 lo = 0, hi = (HOST_WIDE_INT)1 << shift;
14177 lo = 0, hi = (HOST_WIDE_INT)1 << (shift - HOST_BITS_PER_WIDE_INT);
14181 lo = ~lo, hi = ~hi;
14187 mask = immed_double_const (lo, hi, imode);
14189 vec = gen_rtvec (2, v, mask);
14190 v = gen_rtx_CONST_VECTOR (V2DImode, vec);
14191 v = copy_to_mode_reg (mode, gen_lowpart (mode, v));
14198 gcc_unreachable ();
14202 lo = ~lo, hi = ~hi;
14204 /* Force this value into the low part of a fp vector constant. */
14205 mask = immed_double_const (lo, hi, imode);
14206 mask = gen_lowpart (mode, mask);
14208 if (vec_mode == VOIDmode)
14209 return force_reg (mode, mask);
14211 v = ix86_build_const_vector (mode, vect, mask);
14212 return force_reg (vec_mode, v);
14215 /* Generate code for floating point ABS or NEG. */
14218 ix86_expand_fp_absneg_operator (enum rtx_code code, enum machine_mode mode,
14221 rtx mask, set, use, clob, dst, src;
14222 bool use_sse = false;
14223 bool vector_mode = VECTOR_MODE_P (mode);
14224 enum machine_mode elt_mode = mode;
14228 elt_mode = GET_MODE_INNER (mode);
14231 else if (mode == TFmode)
14233 else if (TARGET_SSE_MATH)
14234 use_sse = SSE_FLOAT_MODE_P (mode);
14236 /* NEG and ABS performed with SSE use bitwise mask operations.
14237 Create the appropriate mask now. */
14239 mask = ix86_build_signbit_mask (elt_mode, vector_mode, code == ABS);
14248 set = gen_rtx_fmt_ee (code == NEG ? XOR : AND, mode, src, mask);
14249 set = gen_rtx_SET (VOIDmode, dst, set);
14254 set = gen_rtx_fmt_e (code, mode, src);
14255 set = gen_rtx_SET (VOIDmode, dst, set);
14258 use = gen_rtx_USE (VOIDmode, mask);
14259 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
14260 emit_insn (gen_rtx_PARALLEL (VOIDmode,
14261 gen_rtvec (3, set, use, clob)));
14268 /* Expand a copysign operation. Special case operand 0 being a constant. */
14271 ix86_expand_copysign (rtx operands[])
14273 enum machine_mode mode;
14274 rtx dest, op0, op1, mask, nmask;
14276 dest = operands[0];
14280 mode = GET_MODE (dest);
14282 if (GET_CODE (op0) == CONST_DOUBLE)
14284 rtx (*copysign_insn)(rtx, rtx, rtx, rtx);
14286 if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0)))
14287 op0 = simplify_unary_operation (ABS, mode, op0, mode);
14289 if (mode == SFmode || mode == DFmode)
14291 enum machine_mode vmode;
14293 vmode = mode == SFmode ? V4SFmode : V2DFmode;
14295 if (op0 == CONST0_RTX (mode))
14296 op0 = CONST0_RTX (vmode);
14299 rtx v = ix86_build_const_vector (mode, false, op0);
14301 op0 = force_reg (vmode, v);
14304 else if (op0 != CONST0_RTX (mode))
14305 op0 = force_reg (mode, op0);
14307 mask = ix86_build_signbit_mask (mode, 0, 0);
14309 if (mode == SFmode)
14310 copysign_insn = gen_copysignsf3_const;
14311 else if (mode == DFmode)
14312 copysign_insn = gen_copysigndf3_const;
14314 copysign_insn = gen_copysigntf3_const;
14316 emit_insn (copysign_insn (dest, op0, op1, mask));
14320 rtx (*copysign_insn)(rtx, rtx, rtx, rtx, rtx, rtx);
14322 nmask = ix86_build_signbit_mask (mode, 0, 1);
14323 mask = ix86_build_signbit_mask (mode, 0, 0);
14325 if (mode == SFmode)
14326 copysign_insn = gen_copysignsf3_var;
14327 else if (mode == DFmode)
14328 copysign_insn = gen_copysigndf3_var;
14330 copysign_insn = gen_copysigntf3_var;
14332 emit_insn (copysign_insn (dest, NULL_RTX, op0, op1, nmask, mask));
14336 /* Deconstruct a copysign operation into bit masks. Operand 0 is known to
14337 be a constant, and so has already been expanded into a vector constant. */
14340 ix86_split_copysign_const (rtx operands[])
14342 enum machine_mode mode, vmode;
14343 rtx dest, op0, mask, x;
14345 dest = operands[0];
14347 mask = operands[3];
14349 mode = GET_MODE (dest);
14350 vmode = GET_MODE (mask);
14352 dest = simplify_gen_subreg (vmode, dest, mode, 0);
14353 x = gen_rtx_AND (vmode, dest, mask);
14354 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
14356 if (op0 != CONST0_RTX (vmode))
14358 x = gen_rtx_IOR (vmode, dest, op0);
14359 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
14363 /* Deconstruct a copysign operation into bit masks. Operand 0 is variable,
14364 so we have to do two masks. */
14367 ix86_split_copysign_var (rtx operands[])
14369 enum machine_mode mode, vmode;
14370 rtx dest, scratch, op0, op1, mask, nmask, x;
14372 dest = operands[0];
14373 scratch = operands[1];
14376 nmask = operands[4];
14377 mask = operands[5];
14379 mode = GET_MODE (dest);
14380 vmode = GET_MODE (mask);
14382 if (rtx_equal_p (op0, op1))
14384 /* Shouldn't happen often (it's useless, obviously), but when it does
14385 we'd generate incorrect code if we continue below. */
14386 emit_move_insn (dest, op0);
14390 if (REG_P (mask) && REGNO (dest) == REGNO (mask)) /* alternative 0 */
14392 gcc_assert (REGNO (op1) == REGNO (scratch));
14394 x = gen_rtx_AND (vmode, scratch, mask);
14395 emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
14398 op0 = simplify_gen_subreg (vmode, op0, mode, 0);
14399 x = gen_rtx_NOT (vmode, dest);
14400 x = gen_rtx_AND (vmode, x, op0);
14401 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
14405 if (REGNO (op1) == REGNO (scratch)) /* alternative 1,3 */
14407 x = gen_rtx_AND (vmode, scratch, mask);
14409 else /* alternative 2,4 */
14411 gcc_assert (REGNO (mask) == REGNO (scratch));
14412 op1 = simplify_gen_subreg (vmode, op1, mode, 0);
14413 x = gen_rtx_AND (vmode, scratch, op1);
14415 emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
14417 if (REGNO (op0) == REGNO (dest)) /* alternative 1,2 */
14419 dest = simplify_gen_subreg (vmode, op0, mode, 0);
14420 x = gen_rtx_AND (vmode, dest, nmask);
14422 else /* alternative 3,4 */
14424 gcc_assert (REGNO (nmask) == REGNO (dest));
14426 op0 = simplify_gen_subreg (vmode, op0, mode, 0);
14427 x = gen_rtx_AND (vmode, dest, op0);
14429 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
14432 x = gen_rtx_IOR (vmode, dest, scratch);
14433 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
14436 /* Return TRUE or FALSE depending on whether the first SET in INSN
14437 has source and destination with matching CC modes, and that the
14438 CC mode is at least as constrained as REQ_MODE. */
14441 ix86_match_ccmode (rtx insn, enum machine_mode req_mode)
14444 enum machine_mode set_mode;
14446 set = PATTERN (insn);
14447 if (GET_CODE (set) == PARALLEL)
14448 set = XVECEXP (set, 0, 0);
14449 gcc_assert (GET_CODE (set) == SET);
14450 gcc_assert (GET_CODE (SET_SRC (set)) == COMPARE);
14452 set_mode = GET_MODE (SET_DEST (set));
14456 if (req_mode != CCNOmode
14457 && (req_mode != CCmode
14458 || XEXP (SET_SRC (set), 1) != const0_rtx))
14462 if (req_mode == CCGCmode)
14466 if (req_mode == CCGOCmode || req_mode == CCNOmode)
14470 if (req_mode == CCZmode)
14481 gcc_unreachable ();
14484 return (GET_MODE (SET_SRC (set)) == set_mode);
14487 /* Generate insn patterns to do an integer compare of OPERANDS. */
14490 ix86_expand_int_compare (enum rtx_code code, rtx op0, rtx op1)
14492 enum machine_mode cmpmode;
14495 cmpmode = SELECT_CC_MODE (code, op0, op1);
14496 flags = gen_rtx_REG (cmpmode, FLAGS_REG);
14498 /* This is very simple, but making the interface the same as in the
14499 FP case makes the rest of the code easier. */
14500 tmp = gen_rtx_COMPARE (cmpmode, op0, op1);
14501 emit_insn (gen_rtx_SET (VOIDmode, flags, tmp));
14503 /* Return the test that should be put into the flags user, i.e.
14504 the bcc, scc, or cmov instruction. */
14505 return gen_rtx_fmt_ee (code, VOIDmode, flags, const0_rtx);
14508 /* Figure out whether to use ordered or unordered fp comparisons.
14509 Return the appropriate mode to use. */
14512 ix86_fp_compare_mode (enum rtx_code code ATTRIBUTE_UNUSED)
14514 /* ??? In order to make all comparisons reversible, we do all comparisons
14515 non-trapping when compiling for IEEE. Once gcc is able to distinguish
14516 all forms trapping and nontrapping comparisons, we can make inequality
14517 comparisons trapping again, since it results in better code when using
14518 FCOM based compares. */
14519 return TARGET_IEEE_FP ? CCFPUmode : CCFPmode;
14523 ix86_cc_mode (enum rtx_code code, rtx op0, rtx op1)
14525 enum machine_mode mode = GET_MODE (op0);
14527 if (SCALAR_FLOAT_MODE_P (mode))
14529 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode));
14530 return ix86_fp_compare_mode (code);
14535 /* Only zero flag is needed. */
14536 case EQ: /* ZF=0 */
14537 case NE: /* ZF!=0 */
14539 /* Codes needing carry flag. */
14540 case GEU: /* CF=0 */
14541 case LTU: /* CF=1 */
14542 /* Detect overflow checks. They need just the carry flag. */
14543 if (GET_CODE (op0) == PLUS
14544 && rtx_equal_p (op1, XEXP (op0, 0)))
14548 case GTU: /* CF=0 & ZF=0 */
14549 case LEU: /* CF=1 | ZF=1 */
14550 /* Detect overflow checks. They need just the carry flag. */
14551 if (GET_CODE (op0) == MINUS
14552 && rtx_equal_p (op1, XEXP (op0, 0)))
14556 /* Codes possibly doable only with sign flag when
14557 comparing against zero. */
14558 case GE: /* SF=OF or SF=0 */
14559 case LT: /* SF<>OF or SF=1 */
14560 if (op1 == const0_rtx)
14563 /* For other cases Carry flag is not required. */
14565 /* Codes doable only with sign flag when comparing
14566 against zero, but we miss jump instruction for it
14567 so we need to use relational tests against overflow
14568 that thus needs to be zero. */
14569 case GT: /* ZF=0 & SF=OF */
14570 case LE: /* ZF=1 | SF<>OF */
14571 if (op1 == const0_rtx)
14575 /* strcmp pattern do (use flags) and combine may ask us for proper
14580 gcc_unreachable ();
14584 /* Return the fixed registers used for condition codes. */
14587 ix86_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2)
14594 /* If two condition code modes are compatible, return a condition code
14595 mode which is compatible with both. Otherwise, return
14598 static enum machine_mode
14599 ix86_cc_modes_compatible (enum machine_mode m1, enum machine_mode m2)
14604 if (GET_MODE_CLASS (m1) != MODE_CC || GET_MODE_CLASS (m2) != MODE_CC)
14607 if ((m1 == CCGCmode && m2 == CCGOCmode)
14608 || (m1 == CCGOCmode && m2 == CCGCmode))
14614 gcc_unreachable ();
14644 /* These are only compatible with themselves, which we already
14651 /* Return a comparison we can do and that it is equivalent to
14652 swap_condition (code) apart possibly from orderedness.
14653 But, never change orderedness if TARGET_IEEE_FP, returning
14654 UNKNOWN in that case if necessary. */
14656 static enum rtx_code
14657 ix86_fp_swap_condition (enum rtx_code code)
14661 case GT: /* GTU - CF=0 & ZF=0 */
14662 return TARGET_IEEE_FP ? UNKNOWN : UNLT;
14663 case GE: /* GEU - CF=0 */
14664 return TARGET_IEEE_FP ? UNKNOWN : UNLE;
14665 case UNLT: /* LTU - CF=1 */
14666 return TARGET_IEEE_FP ? UNKNOWN : GT;
14667 case UNLE: /* LEU - CF=1 | ZF=1 */
14668 return TARGET_IEEE_FP ? UNKNOWN : GE;
14670 return swap_condition (code);
14674 /* Return cost of comparison CODE using the best strategy for performance.
14675 All following functions do use number of instructions as a cost metrics.
14676 In future this should be tweaked to compute bytes for optimize_size and
14677 take into account performance of various instructions on various CPUs. */
14680 ix86_fp_comparison_cost (enum rtx_code code)
14684 /* The cost of code using bit-twiddling on %ah. */
14701 arith_cost = TARGET_IEEE_FP ? 5 : 4;
14705 arith_cost = TARGET_IEEE_FP ? 6 : 4;
14708 gcc_unreachable ();
14711 switch (ix86_fp_comparison_strategy (code))
14713 case IX86_FPCMP_COMI:
14714 return arith_cost > 4 ? 3 : 2;
14715 case IX86_FPCMP_SAHF:
14716 return arith_cost > 4 ? 4 : 3;
14722 /* Return strategy to use for floating-point. We assume that fcomi is always
14723 preferrable where available, since that is also true when looking at size
14724 (2 bytes, vs. 3 for fnstsw+sahf and at least 5 for fnstsw+test). */
14726 enum ix86_fpcmp_strategy
14727 ix86_fp_comparison_strategy (enum rtx_code code ATTRIBUTE_UNUSED)
14729 /* Do fcomi/sahf based test when profitable. */
14732 return IX86_FPCMP_COMI;
14734 if (TARGET_SAHF && (TARGET_USE_SAHF || optimize_function_for_size_p (cfun)))
14735 return IX86_FPCMP_SAHF;
14737 return IX86_FPCMP_ARITH;
14740 /* Swap, force into registers, or otherwise massage the two operands
14741 to a fp comparison. The operands are updated in place; the new
14742 comparison code is returned. */
14744 static enum rtx_code
14745 ix86_prepare_fp_compare_args (enum rtx_code code, rtx *pop0, rtx *pop1)
14747 enum machine_mode fpcmp_mode = ix86_fp_compare_mode (code);
14748 rtx op0 = *pop0, op1 = *pop1;
14749 enum machine_mode op_mode = GET_MODE (op0);
14750 int is_sse = TARGET_SSE_MATH && SSE_FLOAT_MODE_P (op_mode);
14752 /* All of the unordered compare instructions only work on registers.
14753 The same is true of the fcomi compare instructions. The XFmode
14754 compare instructions require registers except when comparing
14755 against zero or when converting operand 1 from fixed point to
14759 && (fpcmp_mode == CCFPUmode
14760 || (op_mode == XFmode
14761 && ! (standard_80387_constant_p (op0) == 1
14762 || standard_80387_constant_p (op1) == 1)
14763 && GET_CODE (op1) != FLOAT)
14764 || ix86_fp_comparison_strategy (code) == IX86_FPCMP_COMI))
14766 op0 = force_reg (op_mode, op0);
14767 op1 = force_reg (op_mode, op1);
14771 /* %%% We only allow op1 in memory; op0 must be st(0). So swap
14772 things around if they appear profitable, otherwise force op0
14773 into a register. */
14775 if (standard_80387_constant_p (op0) == 0
14777 && ! (standard_80387_constant_p (op1) == 0
14780 enum rtx_code new_code = ix86_fp_swap_condition (code);
14781 if (new_code != UNKNOWN)
14784 tmp = op0, op0 = op1, op1 = tmp;
14790 op0 = force_reg (op_mode, op0);
14792 if (CONSTANT_P (op1))
14794 int tmp = standard_80387_constant_p (op1);
14796 op1 = validize_mem (force_const_mem (op_mode, op1));
14800 op1 = force_reg (op_mode, op1);
14803 op1 = force_reg (op_mode, op1);
14807 /* Try to rearrange the comparison to make it cheaper. */
14808 if (ix86_fp_comparison_cost (code)
14809 > ix86_fp_comparison_cost (swap_condition (code))
14810 && (REG_P (op1) || can_create_pseudo_p ()))
14813 tmp = op0, op0 = op1, op1 = tmp;
14814 code = swap_condition (code);
14816 op0 = force_reg (op_mode, op0);
14824 /* Convert comparison codes we use to represent FP comparison to integer
14825 code that will result in proper branch. Return UNKNOWN if no such code
14829 ix86_fp_compare_code_to_integer (enum rtx_code code)
14858 /* Generate insn patterns to do a floating point compare of OPERANDS. */
14861 ix86_expand_fp_compare (enum rtx_code code, rtx op0, rtx op1, rtx scratch)
14863 enum machine_mode fpcmp_mode, intcmp_mode;
14866 fpcmp_mode = ix86_fp_compare_mode (code);
14867 code = ix86_prepare_fp_compare_args (code, &op0, &op1);
14869 /* Do fcomi/sahf based test when profitable. */
14870 switch (ix86_fp_comparison_strategy (code))
14872 case IX86_FPCMP_COMI:
14873 intcmp_mode = fpcmp_mode;
14874 tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1);
14875 tmp = gen_rtx_SET (VOIDmode, gen_rtx_REG (fpcmp_mode, FLAGS_REG),
14880 case IX86_FPCMP_SAHF:
14881 intcmp_mode = fpcmp_mode;
14882 tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1);
14883 tmp = gen_rtx_SET (VOIDmode, gen_rtx_REG (fpcmp_mode, FLAGS_REG),
14887 scratch = gen_reg_rtx (HImode);
14888 tmp2 = gen_rtx_CLOBBER (VOIDmode, scratch);
14889 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, tmp2)));
14892 case IX86_FPCMP_ARITH:
14893 /* Sadness wrt reg-stack pops killing fpsr -- gotta get fnstsw first. */
14894 tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1);
14895 tmp2 = gen_rtx_UNSPEC (HImode, gen_rtvec (1, tmp), UNSPEC_FNSTSW);
14897 scratch = gen_reg_rtx (HImode);
14898 emit_insn (gen_rtx_SET (VOIDmode, scratch, tmp2));
14900 /* In the unordered case, we have to check C2 for NaN's, which
14901 doesn't happen to work out to anything nice combination-wise.
14902 So do some bit twiddling on the value we've got in AH to come
14903 up with an appropriate set of condition codes. */
14905 intcmp_mode = CCNOmode;
14910 if (code == GT || !TARGET_IEEE_FP)
14912 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x45)));
14917 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14918 emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx));
14919 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x44)));
14920 intcmp_mode = CCmode;
14926 if (code == LT && TARGET_IEEE_FP)
14928 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14929 emit_insn (gen_cmpqi_ext_3 (scratch, const1_rtx));
14930 intcmp_mode = CCmode;
14935 emit_insn (gen_testqi_ext_ccno_0 (scratch, const1_rtx));
14941 if (code == GE || !TARGET_IEEE_FP)
14943 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x05)));
14948 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14949 emit_insn (gen_xorqi_cc_ext_1 (scratch, scratch, const1_rtx));
14955 if (code == LE && TARGET_IEEE_FP)
14957 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14958 emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx));
14959 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40)));
14960 intcmp_mode = CCmode;
14965 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x45)));
14971 if (code == EQ && TARGET_IEEE_FP)
14973 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14974 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40)));
14975 intcmp_mode = CCmode;
14980 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x40)));
14986 if (code == NE && TARGET_IEEE_FP)
14988 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
14989 emit_insn (gen_xorqi_cc_ext_1 (scratch, scratch,
14995 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x40)));
15001 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x04)));
15005 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x04)));
15010 gcc_unreachable ();
15018 /* Return the test that should be put into the flags user, i.e.
15019 the bcc, scc, or cmov instruction. */
15020 return gen_rtx_fmt_ee (code, VOIDmode,
15021 gen_rtx_REG (intcmp_mode, FLAGS_REG),
15026 ix86_expand_compare (enum rtx_code code)
15029 op0 = ix86_compare_op0;
15030 op1 = ix86_compare_op1;
15032 if (GET_MODE_CLASS (GET_MODE (ix86_compare_op0)) == MODE_CC)
15033 ret = gen_rtx_fmt_ee (code, VOIDmode, ix86_compare_op0, ix86_compare_op1);
15035 else if (SCALAR_FLOAT_MODE_P (GET_MODE (op0)))
15037 gcc_assert (!DECIMAL_FLOAT_MODE_P (GET_MODE (op0)));
15038 ret = ix86_expand_fp_compare (code, op0, op1, NULL_RTX);
15041 ret = ix86_expand_int_compare (code, op0, op1);
15047 ix86_expand_branch (enum rtx_code code, rtx label)
15051 switch (GET_MODE (ix86_compare_op0))
15060 tmp = ix86_expand_compare (code);
15061 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
15062 gen_rtx_LABEL_REF (VOIDmode, label),
15064 emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
15071 /* Expand DImode branch into multiple compare+branch. */
15073 rtx lo[2], hi[2], label2;
15074 enum rtx_code code1, code2, code3;
15075 enum machine_mode submode;
15077 if (CONSTANT_P (ix86_compare_op0) && ! CONSTANT_P (ix86_compare_op1))
15079 tmp = ix86_compare_op0;
15080 ix86_compare_op0 = ix86_compare_op1;
15081 ix86_compare_op1 = tmp;
15082 code = swap_condition (code);
15084 if (GET_MODE (ix86_compare_op0) == DImode)
15086 split_di (&ix86_compare_op0, 1, lo+0, hi+0);
15087 split_di (&ix86_compare_op1, 1, lo+1, hi+1);
15092 split_ti (&ix86_compare_op0, 1, lo+0, hi+0);
15093 split_ti (&ix86_compare_op1, 1, lo+1, hi+1);
15097 /* When comparing for equality, we can use (hi0^hi1)|(lo0^lo1) to
15098 avoid two branches. This costs one extra insn, so disable when
15099 optimizing for size. */
15101 if ((code == EQ || code == NE)
15102 && (!optimize_insn_for_size_p ()
15103 || hi[1] == const0_rtx || lo[1] == const0_rtx))
15108 if (hi[1] != const0_rtx)
15109 xor1 = expand_binop (submode, xor_optab, xor1, hi[1],
15110 NULL_RTX, 0, OPTAB_WIDEN);
15113 if (lo[1] != const0_rtx)
15114 xor0 = expand_binop (submode, xor_optab, xor0, lo[1],
15115 NULL_RTX, 0, OPTAB_WIDEN);
15117 tmp = expand_binop (submode, ior_optab, xor1, xor0,
15118 NULL_RTX, 0, OPTAB_WIDEN);
15120 ix86_compare_op0 = tmp;
15121 ix86_compare_op1 = const0_rtx;
15122 ix86_expand_branch (code, label);
15126 /* Otherwise, if we are doing less-than or greater-or-equal-than,
15127 op1 is a constant and the low word is zero, then we can just
15128 examine the high word. Similarly for low word -1 and
15129 less-or-equal-than or greater-than. */
15131 if (CONST_INT_P (hi[1]))
15134 case LT: case LTU: case GE: case GEU:
15135 if (lo[1] == const0_rtx)
15137 ix86_compare_op0 = hi[0];
15138 ix86_compare_op1 = hi[1];
15139 ix86_expand_branch (code, label);
15143 case LE: case LEU: case GT: case GTU:
15144 if (lo[1] == constm1_rtx)
15146 ix86_compare_op0 = hi[0];
15147 ix86_compare_op1 = hi[1];
15148 ix86_expand_branch (code, label);
15156 /* Otherwise, we need two or three jumps. */
15158 label2 = gen_label_rtx ();
15161 code2 = swap_condition (code);
15162 code3 = unsigned_condition (code);
15166 case LT: case GT: case LTU: case GTU:
15169 case LE: code1 = LT; code2 = GT; break;
15170 case GE: code1 = GT; code2 = LT; break;
15171 case LEU: code1 = LTU; code2 = GTU; break;
15172 case GEU: code1 = GTU; code2 = LTU; break;
15174 case EQ: code1 = UNKNOWN; code2 = NE; break;
15175 case NE: code2 = UNKNOWN; break;
15178 gcc_unreachable ();
15183 * if (hi(a) < hi(b)) goto true;
15184 * if (hi(a) > hi(b)) goto false;
15185 * if (lo(a) < lo(b)) goto true;
15189 ix86_compare_op0 = hi[0];
15190 ix86_compare_op1 = hi[1];
15192 if (code1 != UNKNOWN)
15193 ix86_expand_branch (code1, label);
15194 if (code2 != UNKNOWN)
15195 ix86_expand_branch (code2, label2);
15197 ix86_compare_op0 = lo[0];
15198 ix86_compare_op1 = lo[1];
15199 ix86_expand_branch (code3, label);
15201 if (code2 != UNKNOWN)
15202 emit_label (label2);
15207 /* If we have already emitted a compare insn, go straight to simple.
15208 ix86_expand_compare won't emit anything if ix86_compare_emitted
15210 gcc_assert (GET_MODE_CLASS (GET_MODE (ix86_compare_op0)) == MODE_CC);
15215 /* Split branch based on floating point condition. */
15217 ix86_split_fp_branch (enum rtx_code code, rtx op1, rtx op2,
15218 rtx target1, rtx target2, rtx tmp, rtx pushed)
15223 if (target2 != pc_rtx)
15226 code = reverse_condition_maybe_unordered (code);
15231 condition = ix86_expand_fp_compare (code, op1, op2,
15234 /* Remove pushed operand from stack. */
15236 ix86_free_from_memory (GET_MODE (pushed));
15238 i = emit_jump_insn (gen_rtx_SET
15240 gen_rtx_IF_THEN_ELSE (VOIDmode,
15241 condition, target1, target2)));
15242 if (split_branch_probability >= 0)
15243 add_reg_note (i, REG_BR_PROB, GEN_INT (split_branch_probability));
15247 ix86_expand_setcc (enum rtx_code code, rtx dest)
15251 gcc_assert (GET_MODE (dest) == QImode);
15253 ret = ix86_expand_compare (code);
15254 PUT_MODE (ret, QImode);
15255 emit_insn (gen_rtx_SET (VOIDmode, dest, ret));
15258 /* Expand comparison setting or clearing carry flag. Return true when
15259 successful and set pop for the operation. */
15261 ix86_expand_carry_flag_compare (enum rtx_code code, rtx op0, rtx op1, rtx *pop)
15263 enum machine_mode mode =
15264 GET_MODE (op0) != VOIDmode ? GET_MODE (op0) : GET_MODE (op1);
15266 /* Do not handle DImode compares that go through special path. */
15267 if (mode == (TARGET_64BIT ? TImode : DImode))
15270 if (SCALAR_FLOAT_MODE_P (mode))
15272 rtx compare_op, compare_seq;
15274 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode));
15276 /* Shortcut: following common codes never translate
15277 into carry flag compares. */
15278 if (code == EQ || code == NE || code == UNEQ || code == LTGT
15279 || code == ORDERED || code == UNORDERED)
15282 /* These comparisons require zero flag; swap operands so they won't. */
15283 if ((code == GT || code == UNLE || code == LE || code == UNGT)
15284 && !TARGET_IEEE_FP)
15289 code = swap_condition (code);
15292 /* Try to expand the comparison and verify that we end up with
15293 carry flag based comparison. This fails to be true only when
15294 we decide to expand comparison using arithmetic that is not
15295 too common scenario. */
15297 compare_op = ix86_expand_fp_compare (code, op0, op1, NULL_RTX);
15298 compare_seq = get_insns ();
15301 if (GET_MODE (XEXP (compare_op, 0)) == CCFPmode
15302 || GET_MODE (XEXP (compare_op, 0)) == CCFPUmode)
15303 code = ix86_fp_compare_code_to_integer (GET_CODE (compare_op));
15305 code = GET_CODE (compare_op);
15307 if (code != LTU && code != GEU)
15310 emit_insn (compare_seq);
15315 if (!INTEGRAL_MODE_P (mode))
15324 /* Convert a==0 into (unsigned)a<1. */
15327 if (op1 != const0_rtx)
15330 code = (code == EQ ? LTU : GEU);
15333 /* Convert a>b into b<a or a>=b-1. */
15336 if (CONST_INT_P (op1))
15338 op1 = gen_int_mode (INTVAL (op1) + 1, GET_MODE (op0));
15339 /* Bail out on overflow. We still can swap operands but that
15340 would force loading of the constant into register. */
15341 if (op1 == const0_rtx
15342 || !x86_64_immediate_operand (op1, GET_MODE (op1)))
15344 code = (code == GTU ? GEU : LTU);
15351 code = (code == GTU ? LTU : GEU);
15355 /* Convert a>=0 into (unsigned)a<0x80000000. */
15358 if (mode == DImode || op1 != const0_rtx)
15360 op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode);
15361 code = (code == LT ? GEU : LTU);
15365 if (mode == DImode || op1 != constm1_rtx)
15367 op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode);
15368 code = (code == LE ? GEU : LTU);
15374 /* Swapping operands may cause constant to appear as first operand. */
15375 if (!nonimmediate_operand (op0, VOIDmode))
15377 if (!can_create_pseudo_p ())
15379 op0 = force_reg (mode, op0);
15381 ix86_compare_op0 = op0;
15382 ix86_compare_op1 = op1;
15383 *pop = ix86_expand_compare (code);
15384 gcc_assert (GET_CODE (*pop) == LTU || GET_CODE (*pop) == GEU);
15389 ix86_expand_int_movcc (rtx operands[])
15391 enum rtx_code code = GET_CODE (operands[1]), compare_code;
15392 rtx compare_seq, compare_op;
15393 enum machine_mode mode = GET_MODE (operands[0]);
15394 bool sign_bit_compare_p = false;;
15397 ix86_compare_op0 = XEXP (operands[1], 0);
15398 ix86_compare_op1 = XEXP (operands[1], 1);
15399 compare_op = ix86_expand_compare (code);
15400 compare_seq = get_insns ();
15403 compare_code = GET_CODE (compare_op);
15405 if ((ix86_compare_op1 == const0_rtx && (code == GE || code == LT))
15406 || (ix86_compare_op1 == constm1_rtx && (code == GT || code == LE)))
15407 sign_bit_compare_p = true;
15409 /* Don't attempt mode expansion here -- if we had to expand 5 or 6
15410 HImode insns, we'd be swallowed in word prefix ops. */
15412 if ((mode != HImode || TARGET_FAST_PREFIX)
15413 && (mode != (TARGET_64BIT ? TImode : DImode))
15414 && CONST_INT_P (operands[2])
15415 && CONST_INT_P (operands[3]))
15417 rtx out = operands[0];
15418 HOST_WIDE_INT ct = INTVAL (operands[2]);
15419 HOST_WIDE_INT cf = INTVAL (operands[3]);
15420 HOST_WIDE_INT diff;
15423 /* Sign bit compares are better done using shifts than we do by using
15425 if (sign_bit_compare_p
15426 || ix86_expand_carry_flag_compare (code, ix86_compare_op0,
15427 ix86_compare_op1, &compare_op))
15429 /* Detect overlap between destination and compare sources. */
15432 if (!sign_bit_compare_p)
15435 rtx (*insn)(rtx, rtx, rtx);
15436 bool fpcmp = false;
15438 compare_code = GET_CODE (compare_op);
15440 flags = XEXP (compare_op, 0);
15442 if (GET_MODE (flags) == CCFPmode
15443 || GET_MODE (flags) == CCFPUmode)
15447 = ix86_fp_compare_code_to_integer (compare_code);
15450 /* To simplify rest of code, restrict to the GEU case. */
15451 if (compare_code == LTU)
15453 HOST_WIDE_INT tmp = ct;
15456 compare_code = reverse_condition (compare_code);
15457 code = reverse_condition (code);
15462 PUT_CODE (compare_op,
15463 reverse_condition_maybe_unordered
15464 (GET_CODE (compare_op)));
15466 PUT_CODE (compare_op,
15467 reverse_condition (GET_CODE (compare_op)));
15471 if (reg_overlap_mentioned_p (out, ix86_compare_op0)
15472 || reg_overlap_mentioned_p (out, ix86_compare_op1))
15473 tmp = gen_reg_rtx (mode);
15475 if (mode == DImode)
15476 insn = gen_x86_movdicc_0_m1;
15478 insn = gen_x86_movsicc_0_m1;
15480 emit_insn (insn (tmp, flags, compare_op));
15484 if (code == GT || code == GE)
15485 code = reverse_condition (code);
15488 HOST_WIDE_INT tmp = ct;
15493 tmp = emit_store_flag (tmp, code, ix86_compare_op0,
15494 ix86_compare_op1, VOIDmode, 0, -1);
15507 tmp = expand_simple_binop (mode, PLUS,
15509 copy_rtx (tmp), 1, OPTAB_DIRECT);
15520 tmp = expand_simple_binop (mode, IOR,
15522 copy_rtx (tmp), 1, OPTAB_DIRECT);
15524 else if (diff == -1 && ct)
15534 tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1);
15536 tmp = expand_simple_binop (mode, PLUS,
15537 copy_rtx (tmp), GEN_INT (cf),
15538 copy_rtx (tmp), 1, OPTAB_DIRECT);
15546 * andl cf - ct, dest
15556 tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1);
15559 tmp = expand_simple_binop (mode, AND,
15561 gen_int_mode (cf - ct, mode),
15562 copy_rtx (tmp), 1, OPTAB_DIRECT);
15564 tmp = expand_simple_binop (mode, PLUS,
15565 copy_rtx (tmp), GEN_INT (ct),
15566 copy_rtx (tmp), 1, OPTAB_DIRECT);
15569 if (!rtx_equal_p (tmp, out))
15570 emit_move_insn (copy_rtx (out), copy_rtx (tmp));
15572 return 1; /* DONE */
15577 enum machine_mode cmp_mode = GET_MODE (ix86_compare_op0);
15580 tmp = ct, ct = cf, cf = tmp;
15583 if (SCALAR_FLOAT_MODE_P (cmp_mode))
15585 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode));
15587 /* We may be reversing unordered compare to normal compare, that
15588 is not valid in general (we may convert non-trapping condition
15589 to trapping one), however on i386 we currently emit all
15590 comparisons unordered. */
15591 compare_code = reverse_condition_maybe_unordered (compare_code);
15592 code = reverse_condition_maybe_unordered (code);
15596 compare_code = reverse_condition (compare_code);
15597 code = reverse_condition (code);
15601 compare_code = UNKNOWN;
15602 if (GET_MODE_CLASS (GET_MODE (ix86_compare_op0)) == MODE_INT
15603 && CONST_INT_P (ix86_compare_op1))
15605 if (ix86_compare_op1 == const0_rtx
15606 && (code == LT || code == GE))
15607 compare_code = code;
15608 else if (ix86_compare_op1 == constm1_rtx)
15612 else if (code == GT)
15617 /* Optimize dest = (op0 < 0) ? -1 : cf. */
15618 if (compare_code != UNKNOWN
15619 && GET_MODE (ix86_compare_op0) == GET_MODE (out)
15620 && (cf == -1 || ct == -1))
15622 /* If lea code below could be used, only optimize
15623 if it results in a 2 insn sequence. */
15625 if (! (diff == 1 || diff == 2 || diff == 4 || diff == 8
15626 || diff == 3 || diff == 5 || diff == 9)
15627 || (compare_code == LT && ct == -1)
15628 || (compare_code == GE && cf == -1))
15631 * notl op1 (if necessary)
15639 code = reverse_condition (code);
15642 out = emit_store_flag (out, code, ix86_compare_op0,
15643 ix86_compare_op1, VOIDmode, 0, -1);
15645 out = expand_simple_binop (mode, IOR,
15647 out, 1, OPTAB_DIRECT);
15648 if (out != operands[0])
15649 emit_move_insn (operands[0], out);
15651 return 1; /* DONE */
15656 if ((diff == 1 || diff == 2 || diff == 4 || diff == 8
15657 || diff == 3 || diff == 5 || diff == 9)
15658 && ((mode != QImode && mode != HImode) || !TARGET_PARTIAL_REG_STALL)
15660 || x86_64_immediate_operand (GEN_INT (cf), VOIDmode)))
15666 * lea cf(dest*(ct-cf)),dest
15670 * This also catches the degenerate setcc-only case.
15676 out = emit_store_flag (out, code, ix86_compare_op0,
15677 ix86_compare_op1, VOIDmode, 0, 1);
15680 /* On x86_64 the lea instruction operates on Pmode, so we need
15681 to get arithmetics done in proper mode to match. */
15683 tmp = copy_rtx (out);
15687 out1 = copy_rtx (out);
15688 tmp = gen_rtx_MULT (mode, out1, GEN_INT (diff & ~1));
15692 tmp = gen_rtx_PLUS (mode, tmp, out1);
15698 tmp = gen_rtx_PLUS (mode, tmp, GEN_INT (cf));
15701 if (!rtx_equal_p (tmp, out))
15704 out = force_operand (tmp, copy_rtx (out));
15706 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (out), copy_rtx (tmp)));
15708 if (!rtx_equal_p (out, operands[0]))
15709 emit_move_insn (operands[0], copy_rtx (out));
15711 return 1; /* DONE */
15715 * General case: Jumpful:
15716 * xorl dest,dest cmpl op1, op2
15717 * cmpl op1, op2 movl ct, dest
15718 * setcc dest jcc 1f
15719 * decl dest movl cf, dest
15720 * andl (cf-ct),dest 1:
15723 * Size 20. Size 14.
15725 * This is reasonably steep, but branch mispredict costs are
15726 * high on modern cpus, so consider failing only if optimizing
15730 if ((!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL))
15731 && BRANCH_COST (optimize_insn_for_speed_p (),
15736 enum machine_mode cmp_mode = GET_MODE (ix86_compare_op0);
15741 if (SCALAR_FLOAT_MODE_P (cmp_mode))
15743 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode));
15745 /* We may be reversing unordered compare to normal compare,
15746 that is not valid in general (we may convert non-trapping
15747 condition to trapping one), however on i386 we currently
15748 emit all comparisons unordered. */
15749 code = reverse_condition_maybe_unordered (code);
15753 code = reverse_condition (code);
15754 if (compare_code != UNKNOWN)
15755 compare_code = reverse_condition (compare_code);
15759 if (compare_code != UNKNOWN)
15761 /* notl op1 (if needed)
15766 For x < 0 (resp. x <= -1) there will be no notl,
15767 so if possible swap the constants to get rid of the
15769 True/false will be -1/0 while code below (store flag
15770 followed by decrement) is 0/-1, so the constants need
15771 to be exchanged once more. */
15773 if (compare_code == GE || !cf)
15775 code = reverse_condition (code);
15780 HOST_WIDE_INT tmp = cf;
15785 out = emit_store_flag (out, code, ix86_compare_op0,
15786 ix86_compare_op1, VOIDmode, 0, -1);
15790 out = emit_store_flag (out, code, ix86_compare_op0,
15791 ix86_compare_op1, VOIDmode, 0, 1);
15793 out = expand_simple_binop (mode, PLUS, copy_rtx (out), constm1_rtx,
15794 copy_rtx (out), 1, OPTAB_DIRECT);
15797 out = expand_simple_binop (mode, AND, copy_rtx (out),
15798 gen_int_mode (cf - ct, mode),
15799 copy_rtx (out), 1, OPTAB_DIRECT);
15801 out = expand_simple_binop (mode, PLUS, copy_rtx (out), GEN_INT (ct),
15802 copy_rtx (out), 1, OPTAB_DIRECT);
15803 if (!rtx_equal_p (out, operands[0]))
15804 emit_move_insn (operands[0], copy_rtx (out));
15806 return 1; /* DONE */
15810 if (!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL))
15812 /* Try a few things more with specific constants and a variable. */
15815 rtx var, orig_out, out, tmp;
15817 if (BRANCH_COST (optimize_insn_for_speed_p (), false) <= 2)
15818 return 0; /* FAIL */
15820 /* If one of the two operands is an interesting constant, load a
15821 constant with the above and mask it in with a logical operation. */
15823 if (CONST_INT_P (operands[2]))
15826 if (INTVAL (operands[2]) == 0 && operands[3] != constm1_rtx)
15827 operands[3] = constm1_rtx, op = and_optab;
15828 else if (INTVAL (operands[2]) == -1 && operands[3] != const0_rtx)
15829 operands[3] = const0_rtx, op = ior_optab;
15831 return 0; /* FAIL */
15833 else if (CONST_INT_P (operands[3]))
15836 if (INTVAL (operands[3]) == 0 && operands[2] != constm1_rtx)
15837 operands[2] = constm1_rtx, op = and_optab;
15838 else if (INTVAL (operands[3]) == -1 && operands[3] != const0_rtx)
15839 operands[2] = const0_rtx, op = ior_optab;
15841 return 0; /* FAIL */
15844 return 0; /* FAIL */
15846 orig_out = operands[0];
15847 tmp = gen_reg_rtx (mode);
15850 /* Recurse to get the constant loaded. */
15851 if (ix86_expand_int_movcc (operands) == 0)
15852 return 0; /* FAIL */
15854 /* Mask in the interesting variable. */
15855 out = expand_binop (mode, op, var, tmp, orig_out, 0,
15857 if (!rtx_equal_p (out, orig_out))
15858 emit_move_insn (copy_rtx (orig_out), copy_rtx (out));
15860 return 1; /* DONE */
15864 * For comparison with above,
15874 if (! nonimmediate_operand (operands[2], mode))
15875 operands[2] = force_reg (mode, operands[2]);
15876 if (! nonimmediate_operand (operands[3], mode))
15877 operands[3] = force_reg (mode, operands[3]);
15879 if (! register_operand (operands[2], VOIDmode)
15881 || ! register_operand (operands[3], VOIDmode)))
15882 operands[2] = force_reg (mode, operands[2]);
15885 && ! register_operand (operands[3], VOIDmode))
15886 operands[3] = force_reg (mode, operands[3]);
15888 emit_insn (compare_seq);
15889 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
15890 gen_rtx_IF_THEN_ELSE (mode,
15891 compare_op, operands[2],
15894 return 1; /* DONE */
15897 /* Swap, force into registers, or otherwise massage the two operands
15898 to an sse comparison with a mask result. Thus we differ a bit from
15899 ix86_prepare_fp_compare_args which expects to produce a flags result.
15901 The DEST operand exists to help determine whether to commute commutative
15902 operators. The POP0/POP1 operands are updated in place. The new
15903 comparison code is returned, or UNKNOWN if not implementable. */
15905 static enum rtx_code
15906 ix86_prepare_sse_fp_compare_args (rtx dest, enum rtx_code code,
15907 rtx *pop0, rtx *pop1)
15915 /* We have no LTGT as an operator. We could implement it with
15916 NE & ORDERED, but this requires an extra temporary. It's
15917 not clear that it's worth it. */
15924 /* These are supported directly. */
15931 /* For commutative operators, try to canonicalize the destination
15932 operand to be first in the comparison - this helps reload to
15933 avoid extra moves. */
15934 if (!dest || !rtx_equal_p (dest, *pop1))
15942 /* These are not supported directly. Swap the comparison operands
15943 to transform into something that is supported. */
15947 code = swap_condition (code);
15951 gcc_unreachable ();
15957 /* Detect conditional moves that exactly match min/max operational
15958 semantics. Note that this is IEEE safe, as long as we don't
15959 interchange the operands.
15961 Returns FALSE if this conditional move doesn't match a MIN/MAX,
15962 and TRUE if the operation is successful and instructions are emitted. */
15965 ix86_expand_sse_fp_minmax (rtx dest, enum rtx_code code, rtx cmp_op0,
15966 rtx cmp_op1, rtx if_true, rtx if_false)
15968 enum machine_mode mode;
15974 else if (code == UNGE)
15977 if_true = if_false;
15983 if (rtx_equal_p (cmp_op0, if_true) && rtx_equal_p (cmp_op1, if_false))
15985 else if (rtx_equal_p (cmp_op1, if_true) && rtx_equal_p (cmp_op0, if_false))
15990 mode = GET_MODE (dest);
15992 /* We want to check HONOR_NANS and HONOR_SIGNED_ZEROS here,
15993 but MODE may be a vector mode and thus not appropriate. */
15994 if (!flag_finite_math_only || !flag_unsafe_math_optimizations)
15996 int u = is_min ? UNSPEC_IEEE_MIN : UNSPEC_IEEE_MAX;
15999 if_true = force_reg (mode, if_true);
16000 v = gen_rtvec (2, if_true, if_false);
16001 tmp = gen_rtx_UNSPEC (mode, v, u);
16005 code = is_min ? SMIN : SMAX;
16006 tmp = gen_rtx_fmt_ee (code, mode, if_true, if_false);
16009 emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
16013 /* Expand an sse vector comparison. Return the register with the result. */
16016 ix86_expand_sse_cmp (rtx dest, enum rtx_code code, rtx cmp_op0, rtx cmp_op1,
16017 rtx op_true, rtx op_false)
16019 enum machine_mode mode = GET_MODE (dest);
16022 cmp_op0 = force_reg (mode, cmp_op0);
16023 if (!nonimmediate_operand (cmp_op1, mode))
16024 cmp_op1 = force_reg (mode, cmp_op1);
16027 || reg_overlap_mentioned_p (dest, op_true)
16028 || reg_overlap_mentioned_p (dest, op_false))
16029 dest = gen_reg_rtx (mode);
16031 x = gen_rtx_fmt_ee (code, mode, cmp_op0, cmp_op1);
16032 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
16037 /* Expand DEST = CMP ? OP_TRUE : OP_FALSE into a sequence of logical
16038 operations. This is used for both scalar and vector conditional moves. */
16041 ix86_expand_sse_movcc (rtx dest, rtx cmp, rtx op_true, rtx op_false)
16043 enum machine_mode mode = GET_MODE (dest);
16046 if (op_false == CONST0_RTX (mode))
16048 op_true = force_reg (mode, op_true);
16049 x = gen_rtx_AND (mode, cmp, op_true);
16050 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
16052 else if (op_true == CONST0_RTX (mode))
16054 op_false = force_reg (mode, op_false);
16055 x = gen_rtx_NOT (mode, cmp);
16056 x = gen_rtx_AND (mode, x, op_false);
16057 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
16059 else if (TARGET_XOP)
16061 rtx pcmov = gen_rtx_SET (mode, dest,
16062 gen_rtx_IF_THEN_ELSE (mode, cmp,
16069 op_true = force_reg (mode, op_true);
16070 op_false = force_reg (mode, op_false);
16072 t2 = gen_reg_rtx (mode);
16074 t3 = gen_reg_rtx (mode);
16078 x = gen_rtx_AND (mode, op_true, cmp);
16079 emit_insn (gen_rtx_SET (VOIDmode, t2, x));
16081 x = gen_rtx_NOT (mode, cmp);
16082 x = gen_rtx_AND (mode, x, op_false);
16083 emit_insn (gen_rtx_SET (VOIDmode, t3, x));
16085 x = gen_rtx_IOR (mode, t3, t2);
16086 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
16090 /* Expand a floating-point conditional move. Return true if successful. */
16093 ix86_expand_fp_movcc (rtx operands[])
16095 enum machine_mode mode = GET_MODE (operands[0]);
16096 enum rtx_code code = GET_CODE (operands[1]);
16097 rtx tmp, compare_op;
16099 ix86_compare_op0 = XEXP (operands[1], 0);
16100 ix86_compare_op1 = XEXP (operands[1], 1);
16101 if (TARGET_SSE_MATH && SSE_FLOAT_MODE_P (mode))
16103 enum machine_mode cmode;
16105 /* Since we've no cmove for sse registers, don't force bad register
16106 allocation just to gain access to it. Deny movcc when the
16107 comparison mode doesn't match the move mode. */
16108 cmode = GET_MODE (ix86_compare_op0);
16109 if (cmode == VOIDmode)
16110 cmode = GET_MODE (ix86_compare_op1);
16114 code = ix86_prepare_sse_fp_compare_args (operands[0], code,
16116 &ix86_compare_op1);
16117 if (code == UNKNOWN)
16120 if (ix86_expand_sse_fp_minmax (operands[0], code, ix86_compare_op0,
16121 ix86_compare_op1, operands[2],
16125 tmp = ix86_expand_sse_cmp (operands[0], code, ix86_compare_op0,
16126 ix86_compare_op1, operands[2], operands[3]);
16127 ix86_expand_sse_movcc (operands[0], tmp, operands[2], operands[3]);
16131 /* The floating point conditional move instructions don't directly
16132 support conditions resulting from a signed integer comparison. */
16134 compare_op = ix86_expand_compare (code);
16135 if (!fcmov_comparison_operator (compare_op, VOIDmode))
16137 tmp = gen_reg_rtx (QImode);
16138 ix86_expand_setcc (code, tmp);
16140 ix86_compare_op0 = tmp;
16141 ix86_compare_op1 = const0_rtx;
16142 compare_op = ix86_expand_compare (code);
16145 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
16146 gen_rtx_IF_THEN_ELSE (mode, compare_op,
16147 operands[2], operands[3])));
16152 /* Expand a floating-point vector conditional move; a vcond operation
16153 rather than a movcc operation. */
16156 ix86_expand_fp_vcond (rtx operands[])
16158 enum rtx_code code = GET_CODE (operands[3]);
16161 code = ix86_prepare_sse_fp_compare_args (operands[0], code,
16162 &operands[4], &operands[5]);
16163 if (code == UNKNOWN)
16166 if (ix86_expand_sse_fp_minmax (operands[0], code, operands[4],
16167 operands[5], operands[1], operands[2]))
16170 cmp = ix86_expand_sse_cmp (operands[0], code, operands[4], operands[5],
16171 operands[1], operands[2]);
16172 ix86_expand_sse_movcc (operands[0], cmp, operands[1], operands[2]);
16176 /* Expand a signed/unsigned integral vector conditional move. */
16179 ix86_expand_int_vcond (rtx operands[])
16181 enum machine_mode mode = GET_MODE (operands[0]);
16182 enum rtx_code code = GET_CODE (operands[3]);
16183 bool negate = false;
16186 cop0 = operands[4];
16187 cop1 = operands[5];
16189 /* XOP supports all of the comparisons on all vector int types. */
16192 /* Canonicalize the comparison to EQ, GT, GTU. */
16203 code = reverse_condition (code);
16209 code = reverse_condition (code);
16215 code = swap_condition (code);
16216 x = cop0, cop0 = cop1, cop1 = x;
16220 gcc_unreachable ();
16223 /* Only SSE4.1/SSE4.2 supports V2DImode. */
16224 if (mode == V2DImode)
16229 /* SSE4.1 supports EQ. */
16230 if (!TARGET_SSE4_1)
16236 /* SSE4.2 supports GT/GTU. */
16237 if (!TARGET_SSE4_2)
16242 gcc_unreachable ();
16246 /* Unsigned parallel compare is not supported by the hardware. Play some
16247 tricks to turn this into a signed comparison against 0. */
16250 cop0 = force_reg (mode, cop0);
16259 /* Perform a parallel modulo subtraction. */
16260 t1 = gen_reg_rtx (mode);
16261 emit_insn ((mode == V4SImode
16263 : gen_subv2di3) (t1, cop0, cop1));
16265 /* Extract the original sign bit of op0. */
16266 mask = ix86_build_signbit_mask (GET_MODE_INNER (mode),
16268 t2 = gen_reg_rtx (mode);
16269 emit_insn ((mode == V4SImode
16271 : gen_andv2di3) (t2, cop0, mask));
16273 /* XOR it back into the result of the subtraction. This results
16274 in the sign bit set iff we saw unsigned underflow. */
16275 x = gen_reg_rtx (mode);
16276 emit_insn ((mode == V4SImode
16278 : gen_xorv2di3) (x, t1, t2));
16286 /* Perform a parallel unsigned saturating subtraction. */
16287 x = gen_reg_rtx (mode);
16288 emit_insn (gen_rtx_SET (VOIDmode, x,
16289 gen_rtx_US_MINUS (mode, cop0, cop1)));
16296 gcc_unreachable ();
16300 cop1 = CONST0_RTX (mode);
16304 x = ix86_expand_sse_cmp (operands[0], code, cop0, cop1,
16305 operands[1+negate], operands[2-negate]);
16307 ix86_expand_sse_movcc (operands[0], x, operands[1+negate],
16308 operands[2-negate]);
16312 /* Unpack OP[1] into the next wider integer vector type. UNSIGNED_P is
16313 true if we should do zero extension, else sign extension. HIGH_P is
16314 true if we want the N/2 high elements, else the low elements. */
16317 ix86_expand_sse_unpack (rtx operands[2], bool unsigned_p, bool high_p)
16319 enum machine_mode imode = GET_MODE (operands[1]);
16320 rtx (*unpack)(rtx, rtx, rtx);
16327 unpack = gen_vec_interleave_highv16qi;
16329 unpack = gen_vec_interleave_lowv16qi;
16333 unpack = gen_vec_interleave_highv8hi;
16335 unpack = gen_vec_interleave_lowv8hi;
16339 unpack = gen_vec_interleave_highv4si;
16341 unpack = gen_vec_interleave_lowv4si;
16344 gcc_unreachable ();
16347 dest = gen_lowpart (imode, operands[0]);
16350 se = force_reg (imode, CONST0_RTX (imode));
16352 se = ix86_expand_sse_cmp (gen_reg_rtx (imode), GT, CONST0_RTX (imode),
16353 operands[1], pc_rtx, pc_rtx);
16355 emit_insn (unpack (dest, operands[1], se));
16358 /* This function performs the same task as ix86_expand_sse_unpack,
16359 but with SSE4.1 instructions. */
16362 ix86_expand_sse4_unpack (rtx operands[2], bool unsigned_p, bool high_p)
16364 enum machine_mode imode = GET_MODE (operands[1]);
16365 rtx (*unpack)(rtx, rtx);
16372 unpack = gen_sse4_1_zero_extendv8qiv8hi2;
16374 unpack = gen_sse4_1_extendv8qiv8hi2;
16378 unpack = gen_sse4_1_zero_extendv4hiv4si2;
16380 unpack = gen_sse4_1_extendv4hiv4si2;
16384 unpack = gen_sse4_1_zero_extendv2siv2di2;
16386 unpack = gen_sse4_1_extendv2siv2di2;
16389 gcc_unreachable ();
16392 dest = operands[0];
16395 /* Shift higher 8 bytes to lower 8 bytes. */
16396 src = gen_reg_rtx (imode);
16397 emit_insn (gen_sse2_lshrv1ti3 (gen_lowpart (V1TImode, src),
16398 gen_lowpart (V1TImode, operands[1]),
16404 emit_insn (unpack (dest, src));
16407 /* Expand conditional increment or decrement using adb/sbb instructions.
16408 The default case using setcc followed by the conditional move can be
16409 done by generic code. */
16411 ix86_expand_int_addcc (rtx operands[])
16413 enum rtx_code code = GET_CODE (operands[1]);
16415 rtx (*insn)(rtx, rtx, rtx, rtx, rtx);
16417 rtx val = const0_rtx;
16418 bool fpcmp = false;
16419 enum machine_mode mode;
16421 ix86_compare_op0 = XEXP (operands[1], 0);
16422 ix86_compare_op1 = XEXP (operands[1], 1);
16423 if (operands[3] != const1_rtx
16424 && operands[3] != constm1_rtx)
16426 if (!ix86_expand_carry_flag_compare (code, ix86_compare_op0,
16427 ix86_compare_op1, &compare_op))
16429 code = GET_CODE (compare_op);
16431 flags = XEXP (compare_op, 0);
16433 if (GET_MODE (flags) == CCFPmode
16434 || GET_MODE (flags) == CCFPUmode)
16437 code = ix86_fp_compare_code_to_integer (code);
16444 PUT_CODE (compare_op,
16445 reverse_condition_maybe_unordered
16446 (GET_CODE (compare_op)));
16448 PUT_CODE (compare_op, reverse_condition (GET_CODE (compare_op)));
16451 mode = GET_MODE (operands[0]);
16453 /* Construct either adc or sbb insn. */
16454 if ((code == LTU) == (operands[3] == constm1_rtx))
16459 insn = gen_subqi3_carry;
16462 insn = gen_subhi3_carry;
16465 insn = gen_subsi3_carry;
16468 insn = gen_subdi3_carry;
16471 gcc_unreachable ();
16479 insn = gen_addqi3_carry;
16482 insn = gen_addhi3_carry;
16485 insn = gen_addsi3_carry;
16488 insn = gen_adddi3_carry;
16491 gcc_unreachable ();
16494 emit_insn (insn (operands[0], operands[2], val, flags, compare_op));
16496 return 1; /* DONE */
16500 /* Split operands 0 and 1 into SImode parts. Similar to split_di, but
16501 works for floating pointer parameters and nonoffsetable memories.
16502 For pushes, it returns just stack offsets; the values will be saved
16503 in the right order. Maximally three parts are generated. */
16506 ix86_split_to_parts (rtx operand, rtx *parts, enum machine_mode mode)
16511 size = mode==XFmode ? 3 : GET_MODE_SIZE (mode) / 4;
16513 size = (GET_MODE_SIZE (mode) + 4) / 8;
16515 gcc_assert (!REG_P (operand) || !MMX_REGNO_P (REGNO (operand)));
16516 gcc_assert (size >= 2 && size <= 4);
16518 /* Optimize constant pool reference to immediates. This is used by fp
16519 moves, that force all constants to memory to allow combining. */
16520 if (MEM_P (operand) && MEM_READONLY_P (operand))
16522 rtx tmp = maybe_get_pool_constant (operand);
16527 if (MEM_P (operand) && !offsettable_memref_p (operand))
16529 /* The only non-offsetable memories we handle are pushes. */
16530 int ok = push_operand (operand, VOIDmode);
16534 operand = copy_rtx (operand);
16535 PUT_MODE (operand, Pmode);
16536 parts[0] = parts[1] = parts[2] = parts[3] = operand;
16540 if (GET_CODE (operand) == CONST_VECTOR)
16542 enum machine_mode imode = int_mode_for_mode (mode);
16543 /* Caution: if we looked through a constant pool memory above,
16544 the operand may actually have a different mode now. That's
16545 ok, since we want to pun this all the way back to an integer. */
16546 operand = simplify_subreg (imode, operand, GET_MODE (operand), 0);
16547 gcc_assert (operand != NULL);
16553 if (mode == DImode)
16554 split_di (&operand, 1, &parts[0], &parts[1]);
16559 if (REG_P (operand))
16561 gcc_assert (reload_completed);
16562 for (i = 0; i < size; i++)
16563 parts[i] = gen_rtx_REG (SImode, REGNO (operand) + i);
16565 else if (offsettable_memref_p (operand))
16567 operand = adjust_address (operand, SImode, 0);
16568 parts[0] = operand;
16569 for (i = 1; i < size; i++)
16570 parts[i] = adjust_address (operand, SImode, 4 * i);
16572 else if (GET_CODE (operand) == CONST_DOUBLE)
16577 REAL_VALUE_FROM_CONST_DOUBLE (r, operand);
16581 real_to_target (l, &r, mode);
16582 parts[3] = gen_int_mode (l[3], SImode);
16583 parts[2] = gen_int_mode (l[2], SImode);
16586 REAL_VALUE_TO_TARGET_LONG_DOUBLE (r, l);
16587 parts[2] = gen_int_mode (l[2], SImode);
16590 REAL_VALUE_TO_TARGET_DOUBLE (r, l);
16593 gcc_unreachable ();
16595 parts[1] = gen_int_mode (l[1], SImode);
16596 parts[0] = gen_int_mode (l[0], SImode);
16599 gcc_unreachable ();
16604 if (mode == TImode)
16605 split_ti (&operand, 1, &parts[0], &parts[1]);
16606 if (mode == XFmode || mode == TFmode)
16608 enum machine_mode upper_mode = mode==XFmode ? SImode : DImode;
16609 if (REG_P (operand))
16611 gcc_assert (reload_completed);
16612 parts[0] = gen_rtx_REG (DImode, REGNO (operand) + 0);
16613 parts[1] = gen_rtx_REG (upper_mode, REGNO (operand) + 1);
16615 else if (offsettable_memref_p (operand))
16617 operand = adjust_address (operand, DImode, 0);
16618 parts[0] = operand;
16619 parts[1] = adjust_address (operand, upper_mode, 8);
16621 else if (GET_CODE (operand) == CONST_DOUBLE)
16626 REAL_VALUE_FROM_CONST_DOUBLE (r, operand);
16627 real_to_target (l, &r, mode);
16629 /* Do not use shift by 32 to avoid warning on 32bit systems. */
16630 if (HOST_BITS_PER_WIDE_INT >= 64)
16633 ((l[0] & (((HOST_WIDE_INT) 2 << 31) - 1))
16634 + ((((HOST_WIDE_INT) l[1]) << 31) << 1),
16637 parts[0] = immed_double_const (l[0], l[1], DImode);
16639 if (upper_mode == SImode)
16640 parts[1] = gen_int_mode (l[2], SImode);
16641 else if (HOST_BITS_PER_WIDE_INT >= 64)
16644 ((l[2] & (((HOST_WIDE_INT) 2 << 31) - 1))
16645 + ((((HOST_WIDE_INT) l[3]) << 31) << 1),
16648 parts[1] = immed_double_const (l[2], l[3], DImode);
16651 gcc_unreachable ();
16658 /* Emit insns to perform a move or push of DI, DF, XF, and TF values.
16659 Return false when normal moves are needed; true when all required
16660 insns have been emitted. Operands 2-4 contain the input values
16661 int the correct order; operands 5-7 contain the output values. */
16664 ix86_split_long_move (rtx operands[])
16669 int collisions = 0;
16670 enum machine_mode mode = GET_MODE (operands[0]);
16671 bool collisionparts[4];
16673 /* The DFmode expanders may ask us to move double.
16674 For 64bit target this is single move. By hiding the fact
16675 here we simplify i386.md splitters. */
16676 if (GET_MODE_SIZE (GET_MODE (operands[0])) == 8 && TARGET_64BIT)
16678 /* Optimize constant pool reference to immediates. This is used by
16679 fp moves, that force all constants to memory to allow combining. */
16681 if (MEM_P (operands[1])
16682 && GET_CODE (XEXP (operands[1], 0)) == SYMBOL_REF
16683 && CONSTANT_POOL_ADDRESS_P (XEXP (operands[1], 0)))
16684 operands[1] = get_pool_constant (XEXP (operands[1], 0));
16685 if (push_operand (operands[0], VOIDmode))
16687 operands[0] = copy_rtx (operands[0]);
16688 PUT_MODE (operands[0], Pmode);
16691 operands[0] = gen_lowpart (DImode, operands[0]);
16692 operands[1] = gen_lowpart (DImode, operands[1]);
16693 emit_move_insn (operands[0], operands[1]);
16697 /* The only non-offsettable memory we handle is push. */
16698 if (push_operand (operands[0], VOIDmode))
16701 gcc_assert (!MEM_P (operands[0])
16702 || offsettable_memref_p (operands[0]));
16704 nparts = ix86_split_to_parts (operands[1], part[1], GET_MODE (operands[0]));
16705 ix86_split_to_parts (operands[0], part[0], GET_MODE (operands[0]));
16707 /* When emitting push, take care for source operands on the stack. */
16708 if (push && MEM_P (operands[1])
16709 && reg_overlap_mentioned_p (stack_pointer_rtx, operands[1]))
16711 rtx src_base = XEXP (part[1][nparts - 1], 0);
16713 /* Compensate for the stack decrement by 4. */
16714 if (!TARGET_64BIT && nparts == 3
16715 && mode == XFmode && TARGET_128BIT_LONG_DOUBLE)
16716 src_base = plus_constant (src_base, 4);
16718 /* src_base refers to the stack pointer and is
16719 automatically decreased by emitted push. */
16720 for (i = 0; i < nparts; i++)
16721 part[1][i] = change_address (part[1][i],
16722 GET_MODE (part[1][i]), src_base);
16725 /* We need to do copy in the right order in case an address register
16726 of the source overlaps the destination. */
16727 if (REG_P (part[0][0]) && MEM_P (part[1][0]))
16731 for (i = 0; i < nparts; i++)
16734 = reg_overlap_mentioned_p (part[0][i], XEXP (part[1][0], 0));
16735 if (collisionparts[i])
16739 /* Collision in the middle part can be handled by reordering. */
16740 if (collisions == 1 && nparts == 3 && collisionparts [1])
16742 tmp = part[0][1]; part[0][1] = part[0][2]; part[0][2] = tmp;
16743 tmp = part[1][1]; part[1][1] = part[1][2]; part[1][2] = tmp;
16745 else if (collisions == 1
16747 && (collisionparts [1] || collisionparts [2]))
16749 if (collisionparts [1])
16751 tmp = part[0][1]; part[0][1] = part[0][2]; part[0][2] = tmp;
16752 tmp = part[1][1]; part[1][1] = part[1][2]; part[1][2] = tmp;
16756 tmp = part[0][2]; part[0][2] = part[0][3]; part[0][3] = tmp;
16757 tmp = part[1][2]; part[1][2] = part[1][3]; part[1][3] = tmp;
16761 /* If there are more collisions, we can't handle it by reordering.
16762 Do an lea to the last part and use only one colliding move. */
16763 else if (collisions > 1)
16769 base = part[0][nparts - 1];
16771 /* Handle the case when the last part isn't valid for lea.
16772 Happens in 64-bit mode storing the 12-byte XFmode. */
16773 if (GET_MODE (base) != Pmode)
16774 base = gen_rtx_REG (Pmode, REGNO (base));
16776 emit_insn (gen_rtx_SET (VOIDmode, base, XEXP (part[1][0], 0)));
16777 part[1][0] = replace_equiv_address (part[1][0], base);
16778 for (i = 1; i < nparts; i++)
16780 tmp = plus_constant (base, UNITS_PER_WORD * i);
16781 part[1][i] = replace_equiv_address (part[1][i], tmp);
16792 if (TARGET_128BIT_LONG_DOUBLE && mode == XFmode)
16793 emit_insn (gen_addsi3 (stack_pointer_rtx,
16794 stack_pointer_rtx, GEN_INT (-4)));
16795 emit_move_insn (part[0][2], part[1][2]);
16797 else if (nparts == 4)
16799 emit_move_insn (part[0][3], part[1][3]);
16800 emit_move_insn (part[0][2], part[1][2]);
16805 /* In 64bit mode we don't have 32bit push available. In case this is
16806 register, it is OK - we will just use larger counterpart. We also
16807 retype memory - these comes from attempt to avoid REX prefix on
16808 moving of second half of TFmode value. */
16809 if (GET_MODE (part[1][1]) == SImode)
16811 switch (GET_CODE (part[1][1]))
16814 part[1][1] = adjust_address (part[1][1], DImode, 0);
16818 part[1][1] = gen_rtx_REG (DImode, REGNO (part[1][1]));
16822 gcc_unreachable ();
16825 if (GET_MODE (part[1][0]) == SImode)
16826 part[1][0] = part[1][1];
16829 emit_move_insn (part[0][1], part[1][1]);
16830 emit_move_insn (part[0][0], part[1][0]);
16834 /* Choose correct order to not overwrite the source before it is copied. */
16835 if ((REG_P (part[0][0])
16836 && REG_P (part[1][1])
16837 && (REGNO (part[0][0]) == REGNO (part[1][1])
16839 && REGNO (part[0][0]) == REGNO (part[1][2]))
16841 && REGNO (part[0][0]) == REGNO (part[1][3]))))
16843 && reg_overlap_mentioned_p (part[0][0], XEXP (part[1][0], 0))))
16845 for (i = 0, j = nparts - 1; i < nparts; i++, j--)
16847 operands[2 + i] = part[0][j];
16848 operands[6 + i] = part[1][j];
16853 for (i = 0; i < nparts; i++)
16855 operands[2 + i] = part[0][i];
16856 operands[6 + i] = part[1][i];
16860 /* If optimizing for size, attempt to locally unCSE nonzero constants. */
16861 if (optimize_insn_for_size_p ())
16863 for (j = 0; j < nparts - 1; j++)
16864 if (CONST_INT_P (operands[6 + j])
16865 && operands[6 + j] != const0_rtx
16866 && REG_P (operands[2 + j]))
16867 for (i = j; i < nparts - 1; i++)
16868 if (CONST_INT_P (operands[7 + i])
16869 && INTVAL (operands[7 + i]) == INTVAL (operands[6 + j]))
16870 operands[7 + i] = operands[2 + j];
16873 for (i = 0; i < nparts; i++)
16874 emit_move_insn (operands[2 + i], operands[6 + i]);
16879 /* Helper function of ix86_split_ashl used to generate an SImode/DImode
16880 left shift by a constant, either using a single shift or
16881 a sequence of add instructions. */
16884 ix86_expand_ashl_const (rtx operand, int count, enum machine_mode mode)
16888 emit_insn ((mode == DImode
16890 : gen_adddi3) (operand, operand, operand));
16892 else if (!optimize_insn_for_size_p ()
16893 && count * ix86_cost->add <= ix86_cost->shift_const)
16896 for (i=0; i<count; i++)
16898 emit_insn ((mode == DImode
16900 : gen_adddi3) (operand, operand, operand));
16904 emit_insn ((mode == DImode
16906 : gen_ashldi3) (operand, operand, GEN_INT (count)));
16910 ix86_split_ashl (rtx *operands, rtx scratch, enum machine_mode mode)
16912 rtx low[2], high[2];
16914 const int single_width = mode == DImode ? 32 : 64;
16916 if (CONST_INT_P (operands[2]))
16918 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
16919 count = INTVAL (operands[2]) & (single_width * 2 - 1);
16921 if (count >= single_width)
16923 emit_move_insn (high[0], low[1]);
16924 emit_move_insn (low[0], const0_rtx);
16926 if (count > single_width)
16927 ix86_expand_ashl_const (high[0], count - single_width, mode);
16931 if (!rtx_equal_p (operands[0], operands[1]))
16932 emit_move_insn (operands[0], operands[1]);
16933 emit_insn ((mode == DImode
16935 : gen_x86_64_shld) (high[0], low[0], GEN_INT (count)));
16936 ix86_expand_ashl_const (low[0], count, mode);
16941 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
16943 if (operands[1] == const1_rtx)
16945 /* Assuming we've chosen a QImode capable registers, then 1 << N
16946 can be done with two 32/64-bit shifts, no branches, no cmoves. */
16947 if (ANY_QI_REG_P (low[0]) && ANY_QI_REG_P (high[0]))
16949 rtx s, d, flags = gen_rtx_REG (CCZmode, FLAGS_REG);
16951 ix86_expand_clear (low[0]);
16952 ix86_expand_clear (high[0]);
16953 emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (single_width)));
16955 d = gen_lowpart (QImode, low[0]);
16956 d = gen_rtx_STRICT_LOW_PART (VOIDmode, d);
16957 s = gen_rtx_EQ (QImode, flags, const0_rtx);
16958 emit_insn (gen_rtx_SET (VOIDmode, d, s));
16960 d = gen_lowpart (QImode, high[0]);
16961 d = gen_rtx_STRICT_LOW_PART (VOIDmode, d);
16962 s = gen_rtx_NE (QImode, flags, const0_rtx);
16963 emit_insn (gen_rtx_SET (VOIDmode, d, s));
16966 /* Otherwise, we can get the same results by manually performing
16967 a bit extract operation on bit 5/6, and then performing the two
16968 shifts. The two methods of getting 0/1 into low/high are exactly
16969 the same size. Avoiding the shift in the bit extract case helps
16970 pentium4 a bit; no one else seems to care much either way. */
16975 if (TARGET_PARTIAL_REG_STALL && !optimize_insn_for_size_p ())
16976 x = gen_rtx_ZERO_EXTEND (mode == DImode ? SImode : DImode, operands[2]);
16978 x = gen_lowpart (mode == DImode ? SImode : DImode, operands[2]);
16979 emit_insn (gen_rtx_SET (VOIDmode, high[0], x));
16981 emit_insn ((mode == DImode
16983 : gen_lshrdi3) (high[0], high[0],
16984 GEN_INT (mode == DImode ? 5 : 6)));
16985 emit_insn ((mode == DImode
16987 : gen_anddi3) (high[0], high[0], const1_rtx));
16988 emit_move_insn (low[0], high[0]);
16989 emit_insn ((mode == DImode
16991 : gen_xordi3) (low[0], low[0], const1_rtx));
16994 emit_insn ((mode == DImode
16996 : gen_ashldi3) (low[0], low[0], operands[2]));
16997 emit_insn ((mode == DImode
16999 : gen_ashldi3) (high[0], high[0], operands[2]));
17003 if (operands[1] == constm1_rtx)
17005 /* For -1 << N, we can avoid the shld instruction, because we
17006 know that we're shifting 0...31/63 ones into a -1. */
17007 emit_move_insn (low[0], constm1_rtx);
17008 if (optimize_insn_for_size_p ())
17009 emit_move_insn (high[0], low[0]);
17011 emit_move_insn (high[0], constm1_rtx);
17015 if (!rtx_equal_p (operands[0], operands[1]))
17016 emit_move_insn (operands[0], operands[1]);
17018 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
17019 emit_insn ((mode == DImode
17021 : gen_x86_64_shld) (high[0], low[0], operands[2]));
17024 emit_insn ((mode == DImode ? gen_ashlsi3 : gen_ashldi3) (low[0], low[0], operands[2]));
17026 if (TARGET_CMOVE && scratch)
17028 ix86_expand_clear (scratch);
17029 emit_insn ((mode == DImode
17030 ? gen_x86_shift_adj_1
17031 : gen_x86_64_shift_adj_1) (high[0], low[0], operands[2],
17035 emit_insn ((mode == DImode
17036 ? gen_x86_shift_adj_2
17037 : gen_x86_64_shift_adj_2) (high[0], low[0], operands[2]));
17041 ix86_split_ashr (rtx *operands, rtx scratch, enum machine_mode mode)
17043 rtx low[2], high[2];
17045 const int single_width = mode == DImode ? 32 : 64;
17047 if (CONST_INT_P (operands[2]))
17049 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
17050 count = INTVAL (operands[2]) & (single_width * 2 - 1);
17052 if (count == single_width * 2 - 1)
17054 emit_move_insn (high[0], high[1]);
17055 emit_insn ((mode == DImode
17057 : gen_ashrdi3) (high[0], high[0],
17058 GEN_INT (single_width - 1)));
17059 emit_move_insn (low[0], high[0]);
17062 else if (count >= single_width)
17064 emit_move_insn (low[0], high[1]);
17065 emit_move_insn (high[0], low[0]);
17066 emit_insn ((mode == DImode
17068 : gen_ashrdi3) (high[0], high[0],
17069 GEN_INT (single_width - 1)));
17070 if (count > single_width)
17071 emit_insn ((mode == DImode
17073 : gen_ashrdi3) (low[0], low[0],
17074 GEN_INT (count - single_width)));
17078 if (!rtx_equal_p (operands[0], operands[1]))
17079 emit_move_insn (operands[0], operands[1]);
17080 emit_insn ((mode == DImode
17082 : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
17083 emit_insn ((mode == DImode
17085 : gen_ashrdi3) (high[0], high[0], GEN_INT (count)));
17090 if (!rtx_equal_p (operands[0], operands[1]))
17091 emit_move_insn (operands[0], operands[1]);
17093 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
17095 emit_insn ((mode == DImode
17097 : gen_x86_64_shrd) (low[0], high[0], operands[2]));
17098 emit_insn ((mode == DImode
17100 : gen_ashrdi3) (high[0], high[0], operands[2]));
17102 if (TARGET_CMOVE && scratch)
17104 emit_move_insn (scratch, high[0]);
17105 emit_insn ((mode == DImode
17107 : gen_ashrdi3) (scratch, scratch,
17108 GEN_INT (single_width - 1)));
17109 emit_insn ((mode == DImode
17110 ? gen_x86_shift_adj_1
17111 : gen_x86_64_shift_adj_1) (low[0], high[0], operands[2],
17115 emit_insn ((mode == DImode
17116 ? gen_x86_shift_adj_3
17117 : gen_x86_64_shift_adj_3) (low[0], high[0], operands[2]));
17122 ix86_split_lshr (rtx *operands, rtx scratch, enum machine_mode mode)
17124 rtx low[2], high[2];
17126 const int single_width = mode == DImode ? 32 : 64;
17128 if (CONST_INT_P (operands[2]))
17130 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
17131 count = INTVAL (operands[2]) & (single_width * 2 - 1);
17133 if (count >= single_width)
17135 emit_move_insn (low[0], high[1]);
17136 ix86_expand_clear (high[0]);
17138 if (count > single_width)
17139 emit_insn ((mode == DImode
17141 : gen_lshrdi3) (low[0], low[0],
17142 GEN_INT (count - single_width)));
17146 if (!rtx_equal_p (operands[0], operands[1]))
17147 emit_move_insn (operands[0], operands[1]);
17148 emit_insn ((mode == DImode
17150 : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
17151 emit_insn ((mode == DImode
17153 : gen_lshrdi3) (high[0], high[0], GEN_INT (count)));
17158 if (!rtx_equal_p (operands[0], operands[1]))
17159 emit_move_insn (operands[0], operands[1]);
17161 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
17163 emit_insn ((mode == DImode
17165 : gen_x86_64_shrd) (low[0], high[0], operands[2]));
17166 emit_insn ((mode == DImode
17168 : gen_lshrdi3) (high[0], high[0], operands[2]));
17170 /* Heh. By reversing the arguments, we can reuse this pattern. */
17171 if (TARGET_CMOVE && scratch)
17173 ix86_expand_clear (scratch);
17174 emit_insn ((mode == DImode
17175 ? gen_x86_shift_adj_1
17176 : gen_x86_64_shift_adj_1) (low[0], high[0], operands[2],
17180 emit_insn ((mode == DImode
17181 ? gen_x86_shift_adj_2
17182 : gen_x86_64_shift_adj_2) (low[0], high[0], operands[2]));
17186 /* Predict just emitted jump instruction to be taken with probability PROB. */
17188 predict_jump (int prob)
17190 rtx insn = get_last_insn ();
17191 gcc_assert (JUMP_P (insn));
17192 add_reg_note (insn, REG_BR_PROB, GEN_INT (prob));
17195 /* Helper function for the string operations below. Dest VARIABLE whether
17196 it is aligned to VALUE bytes. If true, jump to the label. */
17198 ix86_expand_aligntest (rtx variable, int value, bool epilogue)
17200 rtx label = gen_label_rtx ();
17201 rtx tmpcount = gen_reg_rtx (GET_MODE (variable));
17202 if (GET_MODE (variable) == DImode)
17203 emit_insn (gen_anddi3 (tmpcount, variable, GEN_INT (value)));
17205 emit_insn (gen_andsi3 (tmpcount, variable, GEN_INT (value)));
17206 emit_cmp_and_jump_insns (tmpcount, const0_rtx, EQ, 0, GET_MODE (variable),
17209 predict_jump (REG_BR_PROB_BASE * 50 / 100);
17211 predict_jump (REG_BR_PROB_BASE * 90 / 100);
17215 /* Adjust COUNTER by the VALUE. */
17217 ix86_adjust_counter (rtx countreg, HOST_WIDE_INT value)
17219 if (GET_MODE (countreg) == DImode)
17220 emit_insn (gen_adddi3 (countreg, countreg, GEN_INT (-value)));
17222 emit_insn (gen_addsi3 (countreg, countreg, GEN_INT (-value)));
17225 /* Zero extend possibly SImode EXP to Pmode register. */
17227 ix86_zero_extend_to_Pmode (rtx exp)
17230 if (GET_MODE (exp) == VOIDmode)
17231 return force_reg (Pmode, exp);
17232 if (GET_MODE (exp) == Pmode)
17233 return copy_to_mode_reg (Pmode, exp);
17234 r = gen_reg_rtx (Pmode);
17235 emit_insn (gen_zero_extendsidi2 (r, exp));
17239 /* Divide COUNTREG by SCALE. */
17241 scale_counter (rtx countreg, int scale)
17247 if (CONST_INT_P (countreg))
17248 return GEN_INT (INTVAL (countreg) / scale);
17249 gcc_assert (REG_P (countreg));
17251 sc = expand_simple_binop (GET_MODE (countreg), LSHIFTRT, countreg,
17252 GEN_INT (exact_log2 (scale)),
17253 NULL, 1, OPTAB_DIRECT);
17257 /* Return mode for the memcpy/memset loop counter. Prefer SImode over
17258 DImode for constant loop counts. */
17260 static enum machine_mode
17261 counter_mode (rtx count_exp)
17263 if (GET_MODE (count_exp) != VOIDmode)
17264 return GET_MODE (count_exp);
17265 if (!CONST_INT_P (count_exp))
17267 if (TARGET_64BIT && (INTVAL (count_exp) & ~0xffffffff))
17272 /* When SRCPTR is non-NULL, output simple loop to move memory
17273 pointer to SRCPTR to DESTPTR via chunks of MODE unrolled UNROLL times,
17274 overall size is COUNT specified in bytes. When SRCPTR is NULL, output the
17275 equivalent loop to set memory by VALUE (supposed to be in MODE).
17277 The size is rounded down to whole number of chunk size moved at once.
17278 SRCMEM and DESTMEM provide MEMrtx to feed proper aliasing info. */
17282 expand_set_or_movmem_via_loop (rtx destmem, rtx srcmem,
17283 rtx destptr, rtx srcptr, rtx value,
17284 rtx count, enum machine_mode mode, int unroll,
17287 rtx out_label, top_label, iter, tmp;
17288 enum machine_mode iter_mode = counter_mode (count);
17289 rtx piece_size = GEN_INT (GET_MODE_SIZE (mode) * unroll);
17290 rtx piece_size_mask = GEN_INT (~((GET_MODE_SIZE (mode) * unroll) - 1));
17296 top_label = gen_label_rtx ();
17297 out_label = gen_label_rtx ();
17298 iter = gen_reg_rtx (iter_mode);
17300 size = expand_simple_binop (iter_mode, AND, count, piece_size_mask,
17301 NULL, 1, OPTAB_DIRECT);
17302 /* Those two should combine. */
17303 if (piece_size == const1_rtx)
17305 emit_cmp_and_jump_insns (size, const0_rtx, EQ, NULL_RTX, iter_mode,
17307 predict_jump (REG_BR_PROB_BASE * 10 / 100);
17309 emit_move_insn (iter, const0_rtx);
17311 emit_label (top_label);
17313 tmp = convert_modes (Pmode, iter_mode, iter, true);
17314 x_addr = gen_rtx_PLUS (Pmode, destptr, tmp);
17315 destmem = change_address (destmem, mode, x_addr);
17319 y_addr = gen_rtx_PLUS (Pmode, srcptr, copy_rtx (tmp));
17320 srcmem = change_address (srcmem, mode, y_addr);
17322 /* When unrolling for chips that reorder memory reads and writes,
17323 we can save registers by using single temporary.
17324 Also using 4 temporaries is overkill in 32bit mode. */
17325 if (!TARGET_64BIT && 0)
17327 for (i = 0; i < unroll; i++)
17332 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
17334 adjust_address (copy_rtx (srcmem), mode, GET_MODE_SIZE (mode));
17336 emit_move_insn (destmem, srcmem);
17342 gcc_assert (unroll <= 4);
17343 for (i = 0; i < unroll; i++)
17345 tmpreg[i] = gen_reg_rtx (mode);
17349 adjust_address (copy_rtx (srcmem), mode, GET_MODE_SIZE (mode));
17351 emit_move_insn (tmpreg[i], srcmem);
17353 for (i = 0; i < unroll; i++)
17358 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
17360 emit_move_insn (destmem, tmpreg[i]);
17365 for (i = 0; i < unroll; i++)
17369 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
17370 emit_move_insn (destmem, value);
17373 tmp = expand_simple_binop (iter_mode, PLUS, iter, piece_size, iter,
17374 true, OPTAB_LIB_WIDEN);
17376 emit_move_insn (iter, tmp);
17378 emit_cmp_and_jump_insns (iter, size, LT, NULL_RTX, iter_mode,
17380 if (expected_size != -1)
17382 expected_size /= GET_MODE_SIZE (mode) * unroll;
17383 if (expected_size == 0)
17385 else if (expected_size > REG_BR_PROB_BASE)
17386 predict_jump (REG_BR_PROB_BASE - 1);
17388 predict_jump (REG_BR_PROB_BASE - (REG_BR_PROB_BASE + expected_size / 2) / expected_size);
17391 predict_jump (REG_BR_PROB_BASE * 80 / 100);
17392 iter = ix86_zero_extend_to_Pmode (iter);
17393 tmp = expand_simple_binop (Pmode, PLUS, destptr, iter, destptr,
17394 true, OPTAB_LIB_WIDEN);
17395 if (tmp != destptr)
17396 emit_move_insn (destptr, tmp);
17399 tmp = expand_simple_binop (Pmode, PLUS, srcptr, iter, srcptr,
17400 true, OPTAB_LIB_WIDEN);
17402 emit_move_insn (srcptr, tmp);
17404 emit_label (out_label);
17407 /* Output "rep; mov" instruction.
17408 Arguments have same meaning as for previous function */
17410 expand_movmem_via_rep_mov (rtx destmem, rtx srcmem,
17411 rtx destptr, rtx srcptr,
17413 enum machine_mode mode)
17419 /* If the size is known, it is shorter to use rep movs. */
17420 if (mode == QImode && CONST_INT_P (count)
17421 && !(INTVAL (count) & 3))
17424 if (destptr != XEXP (destmem, 0) || GET_MODE (destmem) != BLKmode)
17425 destmem = adjust_automodify_address_nv (destmem, BLKmode, destptr, 0);
17426 if (srcptr != XEXP (srcmem, 0) || GET_MODE (srcmem) != BLKmode)
17427 srcmem = adjust_automodify_address_nv (srcmem, BLKmode, srcptr, 0);
17428 countreg = ix86_zero_extend_to_Pmode (scale_counter (count, GET_MODE_SIZE (mode)));
17429 if (mode != QImode)
17431 destexp = gen_rtx_ASHIFT (Pmode, countreg,
17432 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
17433 destexp = gen_rtx_PLUS (Pmode, destexp, destptr);
17434 srcexp = gen_rtx_ASHIFT (Pmode, countreg,
17435 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
17436 srcexp = gen_rtx_PLUS (Pmode, srcexp, srcptr);
17440 destexp = gen_rtx_PLUS (Pmode, destptr, countreg);
17441 srcexp = gen_rtx_PLUS (Pmode, srcptr, countreg);
17443 if (CONST_INT_P (count))
17445 count = GEN_INT (INTVAL (count)
17446 & ~((HOST_WIDE_INT) GET_MODE_SIZE (mode) - 1));
17447 destmem = shallow_copy_rtx (destmem);
17448 srcmem = shallow_copy_rtx (srcmem);
17449 set_mem_size (destmem, count);
17450 set_mem_size (srcmem, count);
17454 if (MEM_SIZE (destmem))
17455 set_mem_size (destmem, NULL_RTX);
17456 if (MEM_SIZE (srcmem))
17457 set_mem_size (srcmem, NULL_RTX);
17459 emit_insn (gen_rep_mov (destptr, destmem, srcptr, srcmem, countreg,
17463 /* Output "rep; stos" instruction.
17464 Arguments have same meaning as for previous function */
17466 expand_setmem_via_rep_stos (rtx destmem, rtx destptr, rtx value,
17467 rtx count, enum machine_mode mode,
17473 if (destptr != XEXP (destmem, 0) || GET_MODE (destmem) != BLKmode)
17474 destmem = adjust_automodify_address_nv (destmem, BLKmode, destptr, 0);
17475 value = force_reg (mode, gen_lowpart (mode, value));
17476 countreg = ix86_zero_extend_to_Pmode (scale_counter (count, GET_MODE_SIZE (mode)));
17477 if (mode != QImode)
17479 destexp = gen_rtx_ASHIFT (Pmode, countreg,
17480 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
17481 destexp = gen_rtx_PLUS (Pmode, destexp, destptr);
17484 destexp = gen_rtx_PLUS (Pmode, destptr, countreg);
17485 if (orig_value == const0_rtx && CONST_INT_P (count))
17487 count = GEN_INT (INTVAL (count)
17488 & ~((HOST_WIDE_INT) GET_MODE_SIZE (mode) - 1));
17489 destmem = shallow_copy_rtx (destmem);
17490 set_mem_size (destmem, count);
17492 else if (MEM_SIZE (destmem))
17493 set_mem_size (destmem, NULL_RTX);
17494 emit_insn (gen_rep_stos (destptr, countreg, destmem, value, destexp));
17498 emit_strmov (rtx destmem, rtx srcmem,
17499 rtx destptr, rtx srcptr, enum machine_mode mode, int offset)
17501 rtx src = adjust_automodify_address_nv (srcmem, mode, srcptr, offset);
17502 rtx dest = adjust_automodify_address_nv (destmem, mode, destptr, offset);
17503 emit_insn (gen_strmov (destptr, dest, srcptr, src));
17506 /* Output code to copy at most count & (max_size - 1) bytes from SRC to DEST. */
17508 expand_movmem_epilogue (rtx destmem, rtx srcmem,
17509 rtx destptr, rtx srcptr, rtx count, int max_size)
17512 if (CONST_INT_P (count))
17514 HOST_WIDE_INT countval = INTVAL (count);
17517 if ((countval & 0x10) && max_size > 16)
17521 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset);
17522 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset + 8);
17525 gcc_unreachable ();
17528 if ((countval & 0x08) && max_size > 8)
17531 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset);
17534 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset);
17535 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset + 4);
17539 if ((countval & 0x04) && max_size > 4)
17541 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset);
17544 if ((countval & 0x02) && max_size > 2)
17546 emit_strmov (destmem, srcmem, destptr, srcptr, HImode, offset);
17549 if ((countval & 0x01) && max_size > 1)
17551 emit_strmov (destmem, srcmem, destptr, srcptr, QImode, offset);
17558 count = expand_simple_binop (GET_MODE (count), AND, count, GEN_INT (max_size - 1),
17559 count, 1, OPTAB_DIRECT);
17560 expand_set_or_movmem_via_loop (destmem, srcmem, destptr, srcptr, NULL,
17561 count, QImode, 1, 4);
17565 /* When there are stringops, we can cheaply increase dest and src pointers.
17566 Otherwise we save code size by maintaining offset (zero is readily
17567 available from preceding rep operation) and using x86 addressing modes.
17569 if (TARGET_SINGLE_STRINGOP)
17573 rtx label = ix86_expand_aligntest (count, 4, true);
17574 src = change_address (srcmem, SImode, srcptr);
17575 dest = change_address (destmem, SImode, destptr);
17576 emit_insn (gen_strmov (destptr, dest, srcptr, src));
17577 emit_label (label);
17578 LABEL_NUSES (label) = 1;
17582 rtx label = ix86_expand_aligntest (count, 2, true);
17583 src = change_address (srcmem, HImode, srcptr);
17584 dest = change_address (destmem, HImode, destptr);
17585 emit_insn (gen_strmov (destptr, dest, srcptr, src));
17586 emit_label (label);
17587 LABEL_NUSES (label) = 1;
17591 rtx label = ix86_expand_aligntest (count, 1, true);
17592 src = change_address (srcmem, QImode, srcptr);
17593 dest = change_address (destmem, QImode, destptr);
17594 emit_insn (gen_strmov (destptr, dest, srcptr, src));
17595 emit_label (label);
17596 LABEL_NUSES (label) = 1;
17601 rtx offset = force_reg (Pmode, const0_rtx);
17606 rtx label = ix86_expand_aligntest (count, 4, true);
17607 src = change_address (srcmem, SImode, srcptr);
17608 dest = change_address (destmem, SImode, destptr);
17609 emit_move_insn (dest, src);
17610 tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (4), NULL,
17611 true, OPTAB_LIB_WIDEN);
17613 emit_move_insn (offset, tmp);
17614 emit_label (label);
17615 LABEL_NUSES (label) = 1;
17619 rtx label = ix86_expand_aligntest (count, 2, true);
17620 tmp = gen_rtx_PLUS (Pmode, srcptr, offset);
17621 src = change_address (srcmem, HImode, tmp);
17622 tmp = gen_rtx_PLUS (Pmode, destptr, offset);
17623 dest = change_address (destmem, HImode, tmp);
17624 emit_move_insn (dest, src);
17625 tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (2), tmp,
17626 true, OPTAB_LIB_WIDEN);
17628 emit_move_insn (offset, tmp);
17629 emit_label (label);
17630 LABEL_NUSES (label) = 1;
17634 rtx label = ix86_expand_aligntest (count, 1, true);
17635 tmp = gen_rtx_PLUS (Pmode, srcptr, offset);
17636 src = change_address (srcmem, QImode, tmp);
17637 tmp = gen_rtx_PLUS (Pmode, destptr, offset);
17638 dest = change_address (destmem, QImode, tmp);
17639 emit_move_insn (dest, src);
17640 emit_label (label);
17641 LABEL_NUSES (label) = 1;
17646 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
17648 expand_setmem_epilogue_via_loop (rtx destmem, rtx destptr, rtx value,
17649 rtx count, int max_size)
17652 expand_simple_binop (counter_mode (count), AND, count,
17653 GEN_INT (max_size - 1), count, 1, OPTAB_DIRECT);
17654 expand_set_or_movmem_via_loop (destmem, NULL, destptr, NULL,
17655 gen_lowpart (QImode, value), count, QImode,
17659 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
17661 expand_setmem_epilogue (rtx destmem, rtx destptr, rtx value, rtx count, int max_size)
17665 if (CONST_INT_P (count))
17667 HOST_WIDE_INT countval = INTVAL (count);
17670 if ((countval & 0x10) && max_size > 16)
17674 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset);
17675 emit_insn (gen_strset (destptr, dest, value));
17676 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset + 8);
17677 emit_insn (gen_strset (destptr, dest, value));
17680 gcc_unreachable ();
17683 if ((countval & 0x08) && max_size > 8)
17687 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset);
17688 emit_insn (gen_strset (destptr, dest, value));
17692 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset);
17693 emit_insn (gen_strset (destptr, dest, value));
17694 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset + 4);
17695 emit_insn (gen_strset (destptr, dest, value));
17699 if ((countval & 0x04) && max_size > 4)
17701 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset);
17702 emit_insn (gen_strset (destptr, dest, gen_lowpart (SImode, value)));
17705 if ((countval & 0x02) && max_size > 2)
17707 dest = adjust_automodify_address_nv (destmem, HImode, destptr, offset);
17708 emit_insn (gen_strset (destptr, dest, gen_lowpart (HImode, value)));
17711 if ((countval & 0x01) && max_size > 1)
17713 dest = adjust_automodify_address_nv (destmem, QImode, destptr, offset);
17714 emit_insn (gen_strset (destptr, dest, gen_lowpart (QImode, value)));
17721 expand_setmem_epilogue_via_loop (destmem, destptr, value, count, max_size);
17726 rtx label = ix86_expand_aligntest (count, 16, true);
17729 dest = change_address (destmem, DImode, destptr);
17730 emit_insn (gen_strset (destptr, dest, value));
17731 emit_insn (gen_strset (destptr, dest, value));
17735 dest = change_address (destmem, SImode, destptr);
17736 emit_insn (gen_strset (destptr, dest, value));
17737 emit_insn (gen_strset (destptr, dest, value));
17738 emit_insn (gen_strset (destptr, dest, value));
17739 emit_insn (gen_strset (destptr, dest, value));
17741 emit_label (label);
17742 LABEL_NUSES (label) = 1;
17746 rtx label = ix86_expand_aligntest (count, 8, true);
17749 dest = change_address (destmem, DImode, destptr);
17750 emit_insn (gen_strset (destptr, dest, value));
17754 dest = change_address (destmem, SImode, destptr);
17755 emit_insn (gen_strset (destptr, dest, value));
17756 emit_insn (gen_strset (destptr, dest, value));
17758 emit_label (label);
17759 LABEL_NUSES (label) = 1;
17763 rtx label = ix86_expand_aligntest (count, 4, true);
17764 dest = change_address (destmem, SImode, destptr);
17765 emit_insn (gen_strset (destptr, dest, gen_lowpart (SImode, value)));
17766 emit_label (label);
17767 LABEL_NUSES (label) = 1;
17771 rtx label = ix86_expand_aligntest (count, 2, true);
17772 dest = change_address (destmem, HImode, destptr);
17773 emit_insn (gen_strset (destptr, dest, gen_lowpart (HImode, value)));
17774 emit_label (label);
17775 LABEL_NUSES (label) = 1;
17779 rtx label = ix86_expand_aligntest (count, 1, true);
17780 dest = change_address (destmem, QImode, destptr);
17781 emit_insn (gen_strset (destptr, dest, gen_lowpart (QImode, value)));
17782 emit_label (label);
17783 LABEL_NUSES (label) = 1;
17787 /* Copy enough from DEST to SRC to align DEST known to by aligned by ALIGN to
17788 DESIRED_ALIGNMENT. */
17790 expand_movmem_prologue (rtx destmem, rtx srcmem,
17791 rtx destptr, rtx srcptr, rtx count,
17792 int align, int desired_alignment)
17794 if (align <= 1 && desired_alignment > 1)
17796 rtx label = ix86_expand_aligntest (destptr, 1, false);
17797 srcmem = change_address (srcmem, QImode, srcptr);
17798 destmem = change_address (destmem, QImode, destptr);
17799 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17800 ix86_adjust_counter (count, 1);
17801 emit_label (label);
17802 LABEL_NUSES (label) = 1;
17804 if (align <= 2 && desired_alignment > 2)
17806 rtx label = ix86_expand_aligntest (destptr, 2, false);
17807 srcmem = change_address (srcmem, HImode, srcptr);
17808 destmem = change_address (destmem, HImode, destptr);
17809 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17810 ix86_adjust_counter (count, 2);
17811 emit_label (label);
17812 LABEL_NUSES (label) = 1;
17814 if (align <= 4 && desired_alignment > 4)
17816 rtx label = ix86_expand_aligntest (destptr, 4, false);
17817 srcmem = change_address (srcmem, SImode, srcptr);
17818 destmem = change_address (destmem, SImode, destptr);
17819 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17820 ix86_adjust_counter (count, 4);
17821 emit_label (label);
17822 LABEL_NUSES (label) = 1;
17824 gcc_assert (desired_alignment <= 8);
17827 /* Copy enough from DST to SRC to align DST known to DESIRED_ALIGN.
17828 ALIGN_BYTES is how many bytes need to be copied. */
17830 expand_constant_movmem_prologue (rtx dst, rtx *srcp, rtx destreg, rtx srcreg,
17831 int desired_align, int align_bytes)
17834 rtx src_size, dst_size;
17836 int src_align_bytes = get_mem_align_offset (src, desired_align * BITS_PER_UNIT);
17837 if (src_align_bytes >= 0)
17838 src_align_bytes = desired_align - src_align_bytes;
17839 src_size = MEM_SIZE (src);
17840 dst_size = MEM_SIZE (dst);
17841 if (align_bytes & 1)
17843 dst = adjust_automodify_address_nv (dst, QImode, destreg, 0);
17844 src = adjust_automodify_address_nv (src, QImode, srcreg, 0);
17846 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17848 if (align_bytes & 2)
17850 dst = adjust_automodify_address_nv (dst, HImode, destreg, off);
17851 src = adjust_automodify_address_nv (src, HImode, srcreg, off);
17852 if (MEM_ALIGN (dst) < 2 * BITS_PER_UNIT)
17853 set_mem_align (dst, 2 * BITS_PER_UNIT);
17854 if (src_align_bytes >= 0
17855 && (src_align_bytes & 1) == (align_bytes & 1)
17856 && MEM_ALIGN (src) < 2 * BITS_PER_UNIT)
17857 set_mem_align (src, 2 * BITS_PER_UNIT);
17859 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17861 if (align_bytes & 4)
17863 dst = adjust_automodify_address_nv (dst, SImode, destreg, off);
17864 src = adjust_automodify_address_nv (src, SImode, srcreg, off);
17865 if (MEM_ALIGN (dst) < 4 * BITS_PER_UNIT)
17866 set_mem_align (dst, 4 * BITS_PER_UNIT);
17867 if (src_align_bytes >= 0)
17869 unsigned int src_align = 0;
17870 if ((src_align_bytes & 3) == (align_bytes & 3))
17872 else if ((src_align_bytes & 1) == (align_bytes & 1))
17874 if (MEM_ALIGN (src) < src_align * BITS_PER_UNIT)
17875 set_mem_align (src, src_align * BITS_PER_UNIT);
17878 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17880 dst = adjust_automodify_address_nv (dst, BLKmode, destreg, off);
17881 src = adjust_automodify_address_nv (src, BLKmode, srcreg, off);
17882 if (MEM_ALIGN (dst) < (unsigned int) desired_align * BITS_PER_UNIT)
17883 set_mem_align (dst, desired_align * BITS_PER_UNIT);
17884 if (src_align_bytes >= 0)
17886 unsigned int src_align = 0;
17887 if ((src_align_bytes & 7) == (align_bytes & 7))
17889 else if ((src_align_bytes & 3) == (align_bytes & 3))
17891 else if ((src_align_bytes & 1) == (align_bytes & 1))
17893 if (src_align > (unsigned int) desired_align)
17894 src_align = desired_align;
17895 if (MEM_ALIGN (src) < src_align * BITS_PER_UNIT)
17896 set_mem_align (src, src_align * BITS_PER_UNIT);
17899 set_mem_size (dst, GEN_INT (INTVAL (dst_size) - align_bytes));
17901 set_mem_size (dst, GEN_INT (INTVAL (src_size) - align_bytes));
17906 /* Set enough from DEST to align DEST known to by aligned by ALIGN to
17907 DESIRED_ALIGNMENT. */
17909 expand_setmem_prologue (rtx destmem, rtx destptr, rtx value, rtx count,
17910 int align, int desired_alignment)
17912 if (align <= 1 && desired_alignment > 1)
17914 rtx label = ix86_expand_aligntest (destptr, 1, false);
17915 destmem = change_address (destmem, QImode, destptr);
17916 emit_insn (gen_strset (destptr, destmem, gen_lowpart (QImode, value)));
17917 ix86_adjust_counter (count, 1);
17918 emit_label (label);
17919 LABEL_NUSES (label) = 1;
17921 if (align <= 2 && desired_alignment > 2)
17923 rtx label = ix86_expand_aligntest (destptr, 2, false);
17924 destmem = change_address (destmem, HImode, destptr);
17925 emit_insn (gen_strset (destptr, destmem, gen_lowpart (HImode, value)));
17926 ix86_adjust_counter (count, 2);
17927 emit_label (label);
17928 LABEL_NUSES (label) = 1;
17930 if (align <= 4 && desired_alignment > 4)
17932 rtx label = ix86_expand_aligntest (destptr, 4, false);
17933 destmem = change_address (destmem, SImode, destptr);
17934 emit_insn (gen_strset (destptr, destmem, gen_lowpart (SImode, value)));
17935 ix86_adjust_counter (count, 4);
17936 emit_label (label);
17937 LABEL_NUSES (label) = 1;
17939 gcc_assert (desired_alignment <= 8);
17942 /* Set enough from DST to align DST known to by aligned by ALIGN to
17943 DESIRED_ALIGN. ALIGN_BYTES is how many bytes need to be stored. */
17945 expand_constant_setmem_prologue (rtx dst, rtx destreg, rtx value,
17946 int desired_align, int align_bytes)
17949 rtx dst_size = MEM_SIZE (dst);
17950 if (align_bytes & 1)
17952 dst = adjust_automodify_address_nv (dst, QImode, destreg, 0);
17954 emit_insn (gen_strset (destreg, dst,
17955 gen_lowpart (QImode, value)));
17957 if (align_bytes & 2)
17959 dst = adjust_automodify_address_nv (dst, HImode, destreg, off);
17960 if (MEM_ALIGN (dst) < 2 * BITS_PER_UNIT)
17961 set_mem_align (dst, 2 * BITS_PER_UNIT);
17963 emit_insn (gen_strset (destreg, dst,
17964 gen_lowpart (HImode, value)));
17966 if (align_bytes & 4)
17968 dst = adjust_automodify_address_nv (dst, SImode, destreg, off);
17969 if (MEM_ALIGN (dst) < 4 * BITS_PER_UNIT)
17970 set_mem_align (dst, 4 * BITS_PER_UNIT);
17972 emit_insn (gen_strset (destreg, dst,
17973 gen_lowpart (SImode, value)));
17975 dst = adjust_automodify_address_nv (dst, BLKmode, destreg, off);
17976 if (MEM_ALIGN (dst) < (unsigned int) desired_align * BITS_PER_UNIT)
17977 set_mem_align (dst, desired_align * BITS_PER_UNIT);
17979 set_mem_size (dst, GEN_INT (INTVAL (dst_size) - align_bytes));
17983 /* Given COUNT and EXPECTED_SIZE, decide on codegen of string operation. */
17984 static enum stringop_alg
17985 decide_alg (HOST_WIDE_INT count, HOST_WIDE_INT expected_size, bool memset,
17986 int *dynamic_check)
17988 const struct stringop_algs * algs;
17989 bool optimize_for_speed;
17990 /* Algorithms using the rep prefix want at least edi and ecx;
17991 additionally, memset wants eax and memcpy wants esi. Don't
17992 consider such algorithms if the user has appropriated those
17993 registers for their own purposes. */
17994 bool rep_prefix_usable = !(fixed_regs[CX_REG] || fixed_regs[DI_REG]
17996 ? fixed_regs[AX_REG] : fixed_regs[SI_REG]));
17998 #define ALG_USABLE_P(alg) (rep_prefix_usable \
17999 || (alg != rep_prefix_1_byte \
18000 && alg != rep_prefix_4_byte \
18001 && alg != rep_prefix_8_byte))
18002 const struct processor_costs *cost;
18004 /* Even if the string operation call is cold, we still might spend a lot
18005 of time processing large blocks. */
18006 if (optimize_function_for_size_p (cfun)
18007 || (optimize_insn_for_size_p ()
18008 && expected_size != -1 && expected_size < 256))
18009 optimize_for_speed = false;
18011 optimize_for_speed = true;
18013 cost = optimize_for_speed ? ix86_cost : &ix86_size_cost;
18015 *dynamic_check = -1;
18017 algs = &cost->memset[TARGET_64BIT != 0];
18019 algs = &cost->memcpy[TARGET_64BIT != 0];
18020 if (stringop_alg != no_stringop && ALG_USABLE_P (stringop_alg))
18021 return stringop_alg;
18022 /* rep; movq or rep; movl is the smallest variant. */
18023 else if (!optimize_for_speed)
18025 if (!count || (count & 3))
18026 return rep_prefix_usable ? rep_prefix_1_byte : loop_1_byte;
18028 return rep_prefix_usable ? rep_prefix_4_byte : loop;
18030 /* Very tiny blocks are best handled via the loop, REP is expensive to setup.
18032 else if (expected_size != -1 && expected_size < 4)
18033 return loop_1_byte;
18034 else if (expected_size != -1)
18037 enum stringop_alg alg = libcall;
18038 for (i = 0; i < NAX_STRINGOP_ALGS; i++)
18040 /* We get here if the algorithms that were not libcall-based
18041 were rep-prefix based and we are unable to use rep prefixes
18042 based on global register usage. Break out of the loop and
18043 use the heuristic below. */
18044 if (algs->size[i].max == 0)
18046 if (algs->size[i].max >= expected_size || algs->size[i].max == -1)
18048 enum stringop_alg candidate = algs->size[i].alg;
18050 if (candidate != libcall && ALG_USABLE_P (candidate))
18052 /* Honor TARGET_INLINE_ALL_STRINGOPS by picking
18053 last non-libcall inline algorithm. */
18054 if (TARGET_INLINE_ALL_STRINGOPS)
18056 /* When the current size is best to be copied by a libcall,
18057 but we are still forced to inline, run the heuristic below
18058 that will pick code for medium sized blocks. */
18059 if (alg != libcall)
18063 else if (ALG_USABLE_P (candidate))
18067 gcc_assert (TARGET_INLINE_ALL_STRINGOPS || !rep_prefix_usable);
18069 /* When asked to inline the call anyway, try to pick meaningful choice.
18070 We look for maximal size of block that is faster to copy by hand and
18071 take blocks of at most of that size guessing that average size will
18072 be roughly half of the block.
18074 If this turns out to be bad, we might simply specify the preferred
18075 choice in ix86_costs. */
18076 if ((TARGET_INLINE_ALL_STRINGOPS || TARGET_INLINE_STRINGOPS_DYNAMICALLY)
18077 && (algs->unknown_size == libcall || !ALG_USABLE_P (algs->unknown_size)))
18080 enum stringop_alg alg;
18082 bool any_alg_usable_p = true;
18084 for (i = 0; i < NAX_STRINGOP_ALGS; i++)
18086 enum stringop_alg candidate = algs->size[i].alg;
18087 any_alg_usable_p = any_alg_usable_p && ALG_USABLE_P (candidate);
18089 if (candidate != libcall && candidate
18090 && ALG_USABLE_P (candidate))
18091 max = algs->size[i].max;
18093 /* If there aren't any usable algorithms, then recursing on
18094 smaller sizes isn't going to find anything. Just return the
18095 simple byte-at-a-time copy loop. */
18096 if (!any_alg_usable_p)
18098 /* Pick something reasonable. */
18099 if (TARGET_INLINE_STRINGOPS_DYNAMICALLY)
18100 *dynamic_check = 128;
18101 return loop_1_byte;
18105 alg = decide_alg (count, max / 2, memset, dynamic_check);
18106 gcc_assert (*dynamic_check == -1);
18107 gcc_assert (alg != libcall);
18108 if (TARGET_INLINE_STRINGOPS_DYNAMICALLY)
18109 *dynamic_check = max;
18112 return ALG_USABLE_P (algs->unknown_size) ? algs->unknown_size : libcall;
18113 #undef ALG_USABLE_P
18116 /* Decide on alignment. We know that the operand is already aligned to ALIGN
18117 (ALIGN can be based on profile feedback and thus it is not 100% guaranteed). */
18119 decide_alignment (int align,
18120 enum stringop_alg alg,
18123 int desired_align = 0;
18127 gcc_unreachable ();
18129 case unrolled_loop:
18130 desired_align = GET_MODE_SIZE (Pmode);
18132 case rep_prefix_8_byte:
18135 case rep_prefix_4_byte:
18136 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
18137 copying whole cacheline at once. */
18138 if (TARGET_PENTIUMPRO)
18143 case rep_prefix_1_byte:
18144 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
18145 copying whole cacheline at once. */
18146 if (TARGET_PENTIUMPRO)
18160 if (desired_align < align)
18161 desired_align = align;
18162 if (expected_size != -1 && expected_size < 4)
18163 desired_align = align;
18164 return desired_align;
18167 /* Return the smallest power of 2 greater than VAL. */
18169 smallest_pow2_greater_than (int val)
18177 /* Expand string move (memcpy) operation. Use i386 string operations when
18178 profitable. expand_setmem contains similar code. The code depends upon
18179 architecture, block size and alignment, but always has the same
18182 1) Prologue guard: Conditional that jumps up to epilogues for small
18183 blocks that can be handled by epilogue alone. This is faster but
18184 also needed for correctness, since prologue assume the block is larger
18185 than the desired alignment.
18187 Optional dynamic check for size and libcall for large
18188 blocks is emitted here too, with -minline-stringops-dynamically.
18190 2) Prologue: copy first few bytes in order to get destination aligned
18191 to DESIRED_ALIGN. It is emitted only when ALIGN is less than
18192 DESIRED_ALIGN and and up to DESIRED_ALIGN - ALIGN bytes can be copied.
18193 We emit either a jump tree on power of two sized blocks, or a byte loop.
18195 3) Main body: the copying loop itself, copying in SIZE_NEEDED chunks
18196 with specified algorithm.
18198 4) Epilogue: code copying tail of the block that is too small to be
18199 handled by main body (or up to size guarded by prologue guard). */
18202 ix86_expand_movmem (rtx dst, rtx src, rtx count_exp, rtx align_exp,
18203 rtx expected_align_exp, rtx expected_size_exp)
18209 rtx jump_around_label = NULL;
18210 HOST_WIDE_INT align = 1;
18211 unsigned HOST_WIDE_INT count = 0;
18212 HOST_WIDE_INT expected_size = -1;
18213 int size_needed = 0, epilogue_size_needed;
18214 int desired_align = 0, align_bytes = 0;
18215 enum stringop_alg alg;
18217 bool need_zero_guard = false;
18219 if (CONST_INT_P (align_exp))
18220 align = INTVAL (align_exp);
18221 /* i386 can do misaligned access on reasonably increased cost. */
18222 if (CONST_INT_P (expected_align_exp)
18223 && INTVAL (expected_align_exp) > align)
18224 align = INTVAL (expected_align_exp);
18225 /* ALIGN is the minimum of destination and source alignment, but we care here
18226 just about destination alignment. */
18227 else if (MEM_ALIGN (dst) > (unsigned HOST_WIDE_INT) align * BITS_PER_UNIT)
18228 align = MEM_ALIGN (dst) / BITS_PER_UNIT;
18230 if (CONST_INT_P (count_exp))
18231 count = expected_size = INTVAL (count_exp);
18232 if (CONST_INT_P (expected_size_exp) && count == 0)
18233 expected_size = INTVAL (expected_size_exp);
18235 /* Make sure we don't need to care about overflow later on. */
18236 if (count > ((unsigned HOST_WIDE_INT) 1 << 30))
18239 /* Step 0: Decide on preferred algorithm, desired alignment and
18240 size of chunks to be copied by main loop. */
18242 alg = decide_alg (count, expected_size, false, &dynamic_check);
18243 desired_align = decide_alignment (align, alg, expected_size);
18245 if (!TARGET_ALIGN_STRINGOPS)
18246 align = desired_align;
18248 if (alg == libcall)
18250 gcc_assert (alg != no_stringop);
18252 count_exp = copy_to_mode_reg (GET_MODE (count_exp), count_exp);
18253 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
18254 srcreg = copy_to_mode_reg (Pmode, XEXP (src, 0));
18259 gcc_unreachable ();
18261 need_zero_guard = true;
18262 size_needed = GET_MODE_SIZE (Pmode);
18264 case unrolled_loop:
18265 need_zero_guard = true;
18266 size_needed = GET_MODE_SIZE (Pmode) * (TARGET_64BIT ? 4 : 2);
18268 case rep_prefix_8_byte:
18271 case rep_prefix_4_byte:
18274 case rep_prefix_1_byte:
18278 need_zero_guard = true;
18283 epilogue_size_needed = size_needed;
18285 /* Step 1: Prologue guard. */
18287 /* Alignment code needs count to be in register. */
18288 if (CONST_INT_P (count_exp) && desired_align > align)
18290 if (INTVAL (count_exp) > desired_align
18291 && INTVAL (count_exp) > size_needed)
18294 = get_mem_align_offset (dst, desired_align * BITS_PER_UNIT);
18295 if (align_bytes <= 0)
18298 align_bytes = desired_align - align_bytes;
18300 if (align_bytes == 0)
18301 count_exp = force_reg (counter_mode (count_exp), count_exp);
18303 gcc_assert (desired_align >= 1 && align >= 1);
18305 /* Ensure that alignment prologue won't copy past end of block. */
18306 if (size_needed > 1 || (desired_align > 1 && desired_align > align))
18308 epilogue_size_needed = MAX (size_needed - 1, desired_align - align);
18309 /* Epilogue always copies COUNT_EXP & EPILOGUE_SIZE_NEEDED bytes.
18310 Make sure it is power of 2. */
18311 epilogue_size_needed = smallest_pow2_greater_than (epilogue_size_needed);
18315 if (count < (unsigned HOST_WIDE_INT)epilogue_size_needed)
18317 /* If main algorithm works on QImode, no epilogue is needed.
18318 For small sizes just don't align anything. */
18319 if (size_needed == 1)
18320 desired_align = align;
18327 label = gen_label_rtx ();
18328 emit_cmp_and_jump_insns (count_exp,
18329 GEN_INT (epilogue_size_needed),
18330 LTU, 0, counter_mode (count_exp), 1, label);
18331 if (expected_size == -1 || expected_size < epilogue_size_needed)
18332 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18334 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18338 /* Emit code to decide on runtime whether library call or inline should be
18340 if (dynamic_check != -1)
18342 if (CONST_INT_P (count_exp))
18344 if (UINTVAL (count_exp) >= (unsigned HOST_WIDE_INT)dynamic_check)
18346 emit_block_move_via_libcall (dst, src, count_exp, false);
18347 count_exp = const0_rtx;
18353 rtx hot_label = gen_label_rtx ();
18354 jump_around_label = gen_label_rtx ();
18355 emit_cmp_and_jump_insns (count_exp, GEN_INT (dynamic_check - 1),
18356 LEU, 0, GET_MODE (count_exp), 1, hot_label);
18357 predict_jump (REG_BR_PROB_BASE * 90 / 100);
18358 emit_block_move_via_libcall (dst, src, count_exp, false);
18359 emit_jump (jump_around_label);
18360 emit_label (hot_label);
18364 /* Step 2: Alignment prologue. */
18366 if (desired_align > align)
18368 if (align_bytes == 0)
18370 /* Except for the first move in epilogue, we no longer know
18371 constant offset in aliasing info. It don't seems to worth
18372 the pain to maintain it for the first move, so throw away
18374 src = change_address (src, BLKmode, srcreg);
18375 dst = change_address (dst, BLKmode, destreg);
18376 expand_movmem_prologue (dst, src, destreg, srcreg, count_exp, align,
18381 /* If we know how many bytes need to be stored before dst is
18382 sufficiently aligned, maintain aliasing info accurately. */
18383 dst = expand_constant_movmem_prologue (dst, &src, destreg, srcreg,
18384 desired_align, align_bytes);
18385 count_exp = plus_constant (count_exp, -align_bytes);
18386 count -= align_bytes;
18388 if (need_zero_guard
18389 && (count < (unsigned HOST_WIDE_INT) size_needed
18390 || (align_bytes == 0
18391 && count < ((unsigned HOST_WIDE_INT) size_needed
18392 + desired_align - align))))
18394 /* It is possible that we copied enough so the main loop will not
18396 gcc_assert (size_needed > 1);
18397 if (label == NULL_RTX)
18398 label = gen_label_rtx ();
18399 emit_cmp_and_jump_insns (count_exp,
18400 GEN_INT (size_needed),
18401 LTU, 0, counter_mode (count_exp), 1, label);
18402 if (expected_size == -1
18403 || expected_size < (desired_align - align) / 2 + size_needed)
18404 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18406 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18409 if (label && size_needed == 1)
18411 emit_label (label);
18412 LABEL_NUSES (label) = 1;
18414 epilogue_size_needed = 1;
18416 else if (label == NULL_RTX)
18417 epilogue_size_needed = size_needed;
18419 /* Step 3: Main loop. */
18425 gcc_unreachable ();
18427 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
18428 count_exp, QImode, 1, expected_size);
18431 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
18432 count_exp, Pmode, 1, expected_size);
18434 case unrolled_loop:
18435 /* Unroll only by factor of 2 in 32bit mode, since we don't have enough
18436 registers for 4 temporaries anyway. */
18437 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
18438 count_exp, Pmode, TARGET_64BIT ? 4 : 2,
18441 case rep_prefix_8_byte:
18442 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
18445 case rep_prefix_4_byte:
18446 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
18449 case rep_prefix_1_byte:
18450 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
18454 /* Adjust properly the offset of src and dest memory for aliasing. */
18455 if (CONST_INT_P (count_exp))
18457 src = adjust_automodify_address_nv (src, BLKmode, srcreg,
18458 (count / size_needed) * size_needed);
18459 dst = adjust_automodify_address_nv (dst, BLKmode, destreg,
18460 (count / size_needed) * size_needed);
18464 src = change_address (src, BLKmode, srcreg);
18465 dst = change_address (dst, BLKmode, destreg);
18468 /* Step 4: Epilogue to copy the remaining bytes. */
18472 /* When the main loop is done, COUNT_EXP might hold original count,
18473 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
18474 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
18475 bytes. Compensate if needed. */
18477 if (size_needed < epilogue_size_needed)
18480 expand_simple_binop (counter_mode (count_exp), AND, count_exp,
18481 GEN_INT (size_needed - 1), count_exp, 1,
18483 if (tmp != count_exp)
18484 emit_move_insn (count_exp, tmp);
18486 emit_label (label);
18487 LABEL_NUSES (label) = 1;
18490 if (count_exp != const0_rtx && epilogue_size_needed > 1)
18491 expand_movmem_epilogue (dst, src, destreg, srcreg, count_exp,
18492 epilogue_size_needed);
18493 if (jump_around_label)
18494 emit_label (jump_around_label);
18498 /* Helper function for memcpy. For QImode value 0xXY produce
18499 0xXYXYXYXY of wide specified by MODE. This is essentially
18500 a * 0x10101010, but we can do slightly better than
18501 synth_mult by unwinding the sequence by hand on CPUs with
18504 promote_duplicated_reg (enum machine_mode mode, rtx val)
18506 enum machine_mode valmode = GET_MODE (val);
18508 int nops = mode == DImode ? 3 : 2;
18510 gcc_assert (mode == SImode || mode == DImode);
18511 if (val == const0_rtx)
18512 return copy_to_mode_reg (mode, const0_rtx);
18513 if (CONST_INT_P (val))
18515 HOST_WIDE_INT v = INTVAL (val) & 255;
18519 if (mode == DImode)
18520 v |= (v << 16) << 16;
18521 return copy_to_mode_reg (mode, gen_int_mode (v, mode));
18524 if (valmode == VOIDmode)
18526 if (valmode != QImode)
18527 val = gen_lowpart (QImode, val);
18528 if (mode == QImode)
18530 if (!TARGET_PARTIAL_REG_STALL)
18532 if (ix86_cost->mult_init[mode == DImode ? 3 : 2]
18533 + ix86_cost->mult_bit * (mode == DImode ? 8 : 4)
18534 <= (ix86_cost->shift_const + ix86_cost->add) * nops
18535 + (COSTS_N_INSNS (TARGET_PARTIAL_REG_STALL == 0)))
18537 rtx reg = convert_modes (mode, QImode, val, true);
18538 tmp = promote_duplicated_reg (mode, const1_rtx);
18539 return expand_simple_binop (mode, MULT, reg, tmp, NULL, 1,
18544 rtx reg = convert_modes (mode, QImode, val, true);
18546 if (!TARGET_PARTIAL_REG_STALL)
18547 if (mode == SImode)
18548 emit_insn (gen_movsi_insv_1 (reg, reg));
18550 emit_insn (gen_movdi_insv_1_rex64 (reg, reg));
18553 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (8),
18554 NULL, 1, OPTAB_DIRECT);
18556 expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
18558 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (16),
18559 NULL, 1, OPTAB_DIRECT);
18560 reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
18561 if (mode == SImode)
18563 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (32),
18564 NULL, 1, OPTAB_DIRECT);
18565 reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
18570 /* Duplicate value VAL using promote_duplicated_reg into maximal size that will
18571 be needed by main loop copying SIZE_NEEDED chunks and prologue getting
18572 alignment from ALIGN to DESIRED_ALIGN. */
18574 promote_duplicated_reg_to_size (rtx val, int size_needed, int desired_align, int align)
18579 && (size_needed > 4 || (desired_align > align && desired_align > 4)))
18580 promoted_val = promote_duplicated_reg (DImode, val);
18581 else if (size_needed > 2 || (desired_align > align && desired_align > 2))
18582 promoted_val = promote_duplicated_reg (SImode, val);
18583 else if (size_needed > 1 || (desired_align > align && desired_align > 1))
18584 promoted_val = promote_duplicated_reg (HImode, val);
18586 promoted_val = val;
18588 return promoted_val;
18591 /* Expand string clear operation (bzero). Use i386 string operations when
18592 profitable. See expand_movmem comment for explanation of individual
18593 steps performed. */
18595 ix86_expand_setmem (rtx dst, rtx count_exp, rtx val_exp, rtx align_exp,
18596 rtx expected_align_exp, rtx expected_size_exp)
18601 rtx jump_around_label = NULL;
18602 HOST_WIDE_INT align = 1;
18603 unsigned HOST_WIDE_INT count = 0;
18604 HOST_WIDE_INT expected_size = -1;
18605 int size_needed = 0, epilogue_size_needed;
18606 int desired_align = 0, align_bytes = 0;
18607 enum stringop_alg alg;
18608 rtx promoted_val = NULL;
18609 bool force_loopy_epilogue = false;
18611 bool need_zero_guard = false;
18613 if (CONST_INT_P (align_exp))
18614 align = INTVAL (align_exp);
18615 /* i386 can do misaligned access on reasonably increased cost. */
18616 if (CONST_INT_P (expected_align_exp)
18617 && INTVAL (expected_align_exp) > align)
18618 align = INTVAL (expected_align_exp);
18619 if (CONST_INT_P (count_exp))
18620 count = expected_size = INTVAL (count_exp);
18621 if (CONST_INT_P (expected_size_exp) && count == 0)
18622 expected_size = INTVAL (expected_size_exp);
18624 /* Make sure we don't need to care about overflow later on. */
18625 if (count > ((unsigned HOST_WIDE_INT) 1 << 30))
18628 /* Step 0: Decide on preferred algorithm, desired alignment and
18629 size of chunks to be copied by main loop. */
18631 alg = decide_alg (count, expected_size, true, &dynamic_check);
18632 desired_align = decide_alignment (align, alg, expected_size);
18634 if (!TARGET_ALIGN_STRINGOPS)
18635 align = desired_align;
18637 if (alg == libcall)
18639 gcc_assert (alg != no_stringop);
18641 count_exp = copy_to_mode_reg (counter_mode (count_exp), count_exp);
18642 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
18647 gcc_unreachable ();
18649 need_zero_guard = true;
18650 size_needed = GET_MODE_SIZE (Pmode);
18652 case unrolled_loop:
18653 need_zero_guard = true;
18654 size_needed = GET_MODE_SIZE (Pmode) * 4;
18656 case rep_prefix_8_byte:
18659 case rep_prefix_4_byte:
18662 case rep_prefix_1_byte:
18666 need_zero_guard = true;
18670 epilogue_size_needed = size_needed;
18672 /* Step 1: Prologue guard. */
18674 /* Alignment code needs count to be in register. */
18675 if (CONST_INT_P (count_exp) && desired_align > align)
18677 if (INTVAL (count_exp) > desired_align
18678 && INTVAL (count_exp) > size_needed)
18681 = get_mem_align_offset (dst, desired_align * BITS_PER_UNIT);
18682 if (align_bytes <= 0)
18685 align_bytes = desired_align - align_bytes;
18687 if (align_bytes == 0)
18689 enum machine_mode mode = SImode;
18690 if (TARGET_64BIT && (count & ~0xffffffff))
18692 count_exp = force_reg (mode, count_exp);
18695 /* Do the cheap promotion to allow better CSE across the
18696 main loop and epilogue (ie one load of the big constant in the
18697 front of all code. */
18698 if (CONST_INT_P (val_exp))
18699 promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed,
18700 desired_align, align);
18701 /* Ensure that alignment prologue won't copy past end of block. */
18702 if (size_needed > 1 || (desired_align > 1 && desired_align > align))
18704 epilogue_size_needed = MAX (size_needed - 1, desired_align - align);
18705 /* Epilogue always copies COUNT_EXP & (EPILOGUE_SIZE_NEEDED - 1) bytes.
18706 Make sure it is power of 2. */
18707 epilogue_size_needed = smallest_pow2_greater_than (epilogue_size_needed);
18709 /* To improve performance of small blocks, we jump around the VAL
18710 promoting mode. This mean that if the promoted VAL is not constant,
18711 we might not use it in the epilogue and have to use byte
18713 if (epilogue_size_needed > 2 && !promoted_val)
18714 force_loopy_epilogue = true;
18717 if (count < (unsigned HOST_WIDE_INT)epilogue_size_needed)
18719 /* If main algorithm works on QImode, no epilogue is needed.
18720 For small sizes just don't align anything. */
18721 if (size_needed == 1)
18722 desired_align = align;
18729 label = gen_label_rtx ();
18730 emit_cmp_and_jump_insns (count_exp,
18731 GEN_INT (epilogue_size_needed),
18732 LTU, 0, counter_mode (count_exp), 1, label);
18733 if (expected_size == -1 || expected_size <= epilogue_size_needed)
18734 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18736 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18739 if (dynamic_check != -1)
18741 rtx hot_label = gen_label_rtx ();
18742 jump_around_label = gen_label_rtx ();
18743 emit_cmp_and_jump_insns (count_exp, GEN_INT (dynamic_check - 1),
18744 LEU, 0, counter_mode (count_exp), 1, hot_label);
18745 predict_jump (REG_BR_PROB_BASE * 90 / 100);
18746 set_storage_via_libcall (dst, count_exp, val_exp, false);
18747 emit_jump (jump_around_label);
18748 emit_label (hot_label);
18751 /* Step 2: Alignment prologue. */
18753 /* Do the expensive promotion once we branched off the small blocks. */
18755 promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed,
18756 desired_align, align);
18757 gcc_assert (desired_align >= 1 && align >= 1);
18759 if (desired_align > align)
18761 if (align_bytes == 0)
18763 /* Except for the first move in epilogue, we no longer know
18764 constant offset in aliasing info. It don't seems to worth
18765 the pain to maintain it for the first move, so throw away
18767 dst = change_address (dst, BLKmode, destreg);
18768 expand_setmem_prologue (dst, destreg, promoted_val, count_exp, align,
18773 /* If we know how many bytes need to be stored before dst is
18774 sufficiently aligned, maintain aliasing info accurately. */
18775 dst = expand_constant_setmem_prologue (dst, destreg, promoted_val,
18776 desired_align, align_bytes);
18777 count_exp = plus_constant (count_exp, -align_bytes);
18778 count -= align_bytes;
18780 if (need_zero_guard
18781 && (count < (unsigned HOST_WIDE_INT) size_needed
18782 || (align_bytes == 0
18783 && count < ((unsigned HOST_WIDE_INT) size_needed
18784 + desired_align - align))))
18786 /* It is possible that we copied enough so the main loop will not
18788 gcc_assert (size_needed > 1);
18789 if (label == NULL_RTX)
18790 label = gen_label_rtx ();
18791 emit_cmp_and_jump_insns (count_exp,
18792 GEN_INT (size_needed),
18793 LTU, 0, counter_mode (count_exp), 1, label);
18794 if (expected_size == -1
18795 || expected_size < (desired_align - align) / 2 + size_needed)
18796 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18798 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18801 if (label && size_needed == 1)
18803 emit_label (label);
18804 LABEL_NUSES (label) = 1;
18806 promoted_val = val_exp;
18807 epilogue_size_needed = 1;
18809 else if (label == NULL_RTX)
18810 epilogue_size_needed = size_needed;
18812 /* Step 3: Main loop. */
18818 gcc_unreachable ();
18820 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18821 count_exp, QImode, 1, expected_size);
18824 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18825 count_exp, Pmode, 1, expected_size);
18827 case unrolled_loop:
18828 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18829 count_exp, Pmode, 4, expected_size);
18831 case rep_prefix_8_byte:
18832 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18835 case rep_prefix_4_byte:
18836 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18839 case rep_prefix_1_byte:
18840 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18844 /* Adjust properly the offset of src and dest memory for aliasing. */
18845 if (CONST_INT_P (count_exp))
18846 dst = adjust_automodify_address_nv (dst, BLKmode, destreg,
18847 (count / size_needed) * size_needed);
18849 dst = change_address (dst, BLKmode, destreg);
18851 /* Step 4: Epilogue to copy the remaining bytes. */
18855 /* When the main loop is done, COUNT_EXP might hold original count,
18856 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
18857 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
18858 bytes. Compensate if needed. */
18860 if (size_needed < epilogue_size_needed)
18863 expand_simple_binop (counter_mode (count_exp), AND, count_exp,
18864 GEN_INT (size_needed - 1), count_exp, 1,
18866 if (tmp != count_exp)
18867 emit_move_insn (count_exp, tmp);
18869 emit_label (label);
18870 LABEL_NUSES (label) = 1;
18873 if (count_exp != const0_rtx && epilogue_size_needed > 1)
18875 if (force_loopy_epilogue)
18876 expand_setmem_epilogue_via_loop (dst, destreg, val_exp, count_exp,
18877 epilogue_size_needed);
18879 expand_setmem_epilogue (dst, destreg, promoted_val, count_exp,
18880 epilogue_size_needed);
18882 if (jump_around_label)
18883 emit_label (jump_around_label);
18887 /* Expand the appropriate insns for doing strlen if not just doing
18890 out = result, initialized with the start address
18891 align_rtx = alignment of the address.
18892 scratch = scratch register, initialized with the startaddress when
18893 not aligned, otherwise undefined
18895 This is just the body. It needs the initializations mentioned above and
18896 some address computing at the end. These things are done in i386.md. */
18899 ix86_expand_strlensi_unroll_1 (rtx out, rtx src, rtx align_rtx)
18903 rtx align_2_label = NULL_RTX;
18904 rtx align_3_label = NULL_RTX;
18905 rtx align_4_label = gen_label_rtx ();
18906 rtx end_0_label = gen_label_rtx ();
18908 rtx tmpreg = gen_reg_rtx (SImode);
18909 rtx scratch = gen_reg_rtx (SImode);
18913 if (CONST_INT_P (align_rtx))
18914 align = INTVAL (align_rtx);
18916 /* Loop to check 1..3 bytes for null to get an aligned pointer. */
18918 /* Is there a known alignment and is it less than 4? */
18921 rtx scratch1 = gen_reg_rtx (Pmode);
18922 emit_move_insn (scratch1, out);
18923 /* Is there a known alignment and is it not 2? */
18926 align_3_label = gen_label_rtx (); /* Label when aligned to 3-byte */
18927 align_2_label = gen_label_rtx (); /* Label when aligned to 2-byte */
18929 /* Leave just the 3 lower bits. */
18930 align_rtx = expand_binop (Pmode, and_optab, scratch1, GEN_INT (3),
18931 NULL_RTX, 0, OPTAB_WIDEN);
18933 emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL,
18934 Pmode, 1, align_4_label);
18935 emit_cmp_and_jump_insns (align_rtx, const2_rtx, EQ, NULL,
18936 Pmode, 1, align_2_label);
18937 emit_cmp_and_jump_insns (align_rtx, const2_rtx, GTU, NULL,
18938 Pmode, 1, align_3_label);
18942 /* Since the alignment is 2, we have to check 2 or 0 bytes;
18943 check if is aligned to 4 - byte. */
18945 align_rtx = expand_binop (Pmode, and_optab, scratch1, const2_rtx,
18946 NULL_RTX, 0, OPTAB_WIDEN);
18948 emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL,
18949 Pmode, 1, align_4_label);
18952 mem = change_address (src, QImode, out);
18954 /* Now compare the bytes. */
18956 /* Compare the first n unaligned byte on a byte per byte basis. */
18957 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL,
18958 QImode, 1, end_0_label);
18960 /* Increment the address. */
18961 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18963 /* Not needed with an alignment of 2 */
18966 emit_label (align_2_label);
18968 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1,
18971 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18973 emit_label (align_3_label);
18976 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1,
18979 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18982 /* Generate loop to check 4 bytes at a time. It is not a good idea to
18983 align this loop. It gives only huge programs, but does not help to
18985 emit_label (align_4_label);
18987 mem = change_address (src, SImode, out);
18988 emit_move_insn (scratch, mem);
18989 emit_insn ((*ix86_gen_add3) (out, out, GEN_INT (4)));
18991 /* This formula yields a nonzero result iff one of the bytes is zero.
18992 This saves three branches inside loop and many cycles. */
18994 emit_insn (gen_addsi3 (tmpreg, scratch, GEN_INT (-0x01010101)));
18995 emit_insn (gen_one_cmplsi2 (scratch, scratch));
18996 emit_insn (gen_andsi3 (tmpreg, tmpreg, scratch));
18997 emit_insn (gen_andsi3 (tmpreg, tmpreg,
18998 gen_int_mode (0x80808080, SImode)));
18999 emit_cmp_and_jump_insns (tmpreg, const0_rtx, EQ, 0, SImode, 1,
19004 rtx reg = gen_reg_rtx (SImode);
19005 rtx reg2 = gen_reg_rtx (Pmode);
19006 emit_move_insn (reg, tmpreg);
19007 emit_insn (gen_lshrsi3 (reg, reg, GEN_INT (16)));
19009 /* If zero is not in the first two bytes, move two bytes forward. */
19010 emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080)));
19011 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
19012 tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
19013 emit_insn (gen_rtx_SET (VOIDmode, tmpreg,
19014 gen_rtx_IF_THEN_ELSE (SImode, tmp,
19017 /* Emit lea manually to avoid clobbering of flags. */
19018 emit_insn (gen_rtx_SET (SImode, reg2,
19019 gen_rtx_PLUS (Pmode, out, const2_rtx)));
19021 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
19022 tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
19023 emit_insn (gen_rtx_SET (VOIDmode, out,
19024 gen_rtx_IF_THEN_ELSE (Pmode, tmp,
19030 rtx end_2_label = gen_label_rtx ();
19031 /* Is zero in the first two bytes? */
19033 emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080)));
19034 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
19035 tmp = gen_rtx_NE (VOIDmode, tmp, const0_rtx);
19036 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
19037 gen_rtx_LABEL_REF (VOIDmode, end_2_label),
19039 tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
19040 JUMP_LABEL (tmp) = end_2_label;
19042 /* Not in the first two. Move two bytes forward. */
19043 emit_insn (gen_lshrsi3 (tmpreg, tmpreg, GEN_INT (16)));
19044 emit_insn ((*ix86_gen_add3) (out, out, const2_rtx));
19046 emit_label (end_2_label);
19050 /* Avoid branch in fixing the byte. */
19051 tmpreg = gen_lowpart (QImode, tmpreg);
19052 emit_insn (gen_addqi3_cc (tmpreg, tmpreg, tmpreg));
19053 tmp = gen_rtx_REG (CCmode, FLAGS_REG);
19054 cmp = gen_rtx_LTU (VOIDmode, tmp, const0_rtx);
19055 emit_insn ((*ix86_gen_sub3_carry) (out, out, GEN_INT (3), tmp, cmp));
19057 emit_label (end_0_label);
19060 /* Expand strlen. */
19063 ix86_expand_strlen (rtx out, rtx src, rtx eoschar, rtx align)
19065 rtx addr, scratch1, scratch2, scratch3, scratch4;
19067 /* The generic case of strlen expander is long. Avoid it's
19068 expanding unless TARGET_INLINE_ALL_STRINGOPS. */
19070 if (TARGET_UNROLL_STRLEN && eoschar == const0_rtx && optimize > 1
19071 && !TARGET_INLINE_ALL_STRINGOPS
19072 && !optimize_insn_for_size_p ()
19073 && (!CONST_INT_P (align) || INTVAL (align) < 4))
19076 addr = force_reg (Pmode, XEXP (src, 0));
19077 scratch1 = gen_reg_rtx (Pmode);
19079 if (TARGET_UNROLL_STRLEN && eoschar == const0_rtx && optimize > 1
19080 && !optimize_insn_for_size_p ())
19082 /* Well it seems that some optimizer does not combine a call like
19083 foo(strlen(bar), strlen(bar));
19084 when the move and the subtraction is done here. It does calculate
19085 the length just once when these instructions are done inside of
19086 output_strlen_unroll(). But I think since &bar[strlen(bar)] is
19087 often used and I use one fewer register for the lifetime of
19088 output_strlen_unroll() this is better. */
19090 emit_move_insn (out, addr);
19092 ix86_expand_strlensi_unroll_1 (out, src, align);
19094 /* strlensi_unroll_1 returns the address of the zero at the end of
19095 the string, like memchr(), so compute the length by subtracting
19096 the start address. */
19097 emit_insn ((*ix86_gen_sub3) (out, out, addr));
19103 /* Can't use this if the user has appropriated eax, ecx, or edi. */
19104 if (fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])
19107 scratch2 = gen_reg_rtx (Pmode);
19108 scratch3 = gen_reg_rtx (Pmode);
19109 scratch4 = force_reg (Pmode, constm1_rtx);
19111 emit_move_insn (scratch3, addr);
19112 eoschar = force_reg (QImode, eoschar);
19114 src = replace_equiv_address_nv (src, scratch3);
19116 /* If .md starts supporting :P, this can be done in .md. */
19117 unspec = gen_rtx_UNSPEC (Pmode, gen_rtvec (4, src, eoschar, align,
19118 scratch4), UNSPEC_SCAS);
19119 emit_insn (gen_strlenqi_1 (scratch1, scratch3, unspec));
19120 emit_insn ((*ix86_gen_one_cmpl2) (scratch2, scratch1));
19121 emit_insn ((*ix86_gen_add3) (out, scratch2, constm1_rtx));
19126 /* For given symbol (function) construct code to compute address of it's PLT
19127 entry in large x86-64 PIC model. */
19129 construct_plt_address (rtx symbol)
19131 rtx tmp = gen_reg_rtx (Pmode);
19132 rtx unspec = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, symbol), UNSPEC_PLTOFF);
19134 gcc_assert (GET_CODE (symbol) == SYMBOL_REF);
19135 gcc_assert (ix86_cmodel == CM_LARGE_PIC);
19137 emit_move_insn (tmp, gen_rtx_CONST (Pmode, unspec));
19138 emit_insn (gen_adddi3 (tmp, tmp, pic_offset_table_rtx));
19143 ix86_expand_call (rtx retval, rtx fnaddr, rtx callarg1,
19145 rtx pop, int sibcall)
19147 rtx use = NULL, call;
19149 if (pop == const0_rtx)
19151 gcc_assert (!TARGET_64BIT || !pop);
19153 if (TARGET_MACHO && !TARGET_64BIT)
19156 if (flag_pic && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF)
19157 fnaddr = machopic_indirect_call_target (fnaddr);
19162 /* Static functions and indirect calls don't need the pic register. */
19163 if (flag_pic && (!TARGET_64BIT || ix86_cmodel == CM_LARGE_PIC)
19164 && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF
19165 && ! SYMBOL_REF_LOCAL_P (XEXP (fnaddr, 0)))
19166 use_reg (&use, pic_offset_table_rtx);
19169 if (TARGET_64BIT && INTVAL (callarg2) >= 0)
19171 rtx al = gen_rtx_REG (QImode, AX_REG);
19172 emit_move_insn (al, callarg2);
19173 use_reg (&use, al);
19176 if (ix86_cmodel == CM_LARGE_PIC
19178 && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF
19179 && !local_symbolic_operand (XEXP (fnaddr, 0), VOIDmode))
19180 fnaddr = gen_rtx_MEM (QImode, construct_plt_address (XEXP (fnaddr, 0)));
19182 ? !sibcall_insn_operand (XEXP (fnaddr, 0), Pmode)
19183 : !call_insn_operand (XEXP (fnaddr, 0), Pmode))
19185 fnaddr = copy_to_mode_reg (Pmode, XEXP (fnaddr, 0));
19186 fnaddr = gen_rtx_MEM (QImode, fnaddr);
19189 call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1);
19191 call = gen_rtx_SET (VOIDmode, retval, call);
19194 pop = gen_rtx_PLUS (Pmode, stack_pointer_rtx, pop);
19195 pop = gen_rtx_SET (VOIDmode, stack_pointer_rtx, pop);
19196 call = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, call, pop));
19199 && ix86_cfun_abi () == MS_ABI
19200 && (!callarg2 || INTVAL (callarg2) != -2))
19202 /* We need to represent that SI and DI registers are clobbered
19204 static int clobbered_registers[] = {
19205 XMM6_REG, XMM7_REG, XMM8_REG,
19206 XMM9_REG, XMM10_REG, XMM11_REG,
19207 XMM12_REG, XMM13_REG, XMM14_REG,
19208 XMM15_REG, SI_REG, DI_REG
19211 rtx vec[ARRAY_SIZE (clobbered_registers) + 2];
19212 rtx unspec = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, const0_rtx),
19213 UNSPEC_MS_TO_SYSV_CALL);
19217 for (i = 0; i < ARRAY_SIZE (clobbered_registers); i++)
19218 vec[i + 2] = gen_rtx_CLOBBER (SSE_REGNO_P (clobbered_registers[i])
19221 (SSE_REGNO_P (clobbered_registers[i])
19223 clobbered_registers[i]));
19225 call = gen_rtx_PARALLEL (VOIDmode,
19226 gen_rtvec_v (ARRAY_SIZE (clobbered_registers)
19230 call = emit_call_insn (call);
19232 CALL_INSN_FUNCTION_USAGE (call) = use;
19236 /* Clear stack slot assignments remembered from previous functions.
19237 This is called from INIT_EXPANDERS once before RTL is emitted for each
19240 static struct machine_function *
19241 ix86_init_machine_status (void)
19243 struct machine_function *f;
19245 f = GGC_CNEW (struct machine_function);
19246 f->use_fast_prologue_epilogue_nregs = -1;
19247 f->tls_descriptor_call_expanded_p = 0;
19248 f->call_abi = ix86_abi;
19253 /* Return a MEM corresponding to a stack slot with mode MODE.
19254 Allocate a new slot if necessary.
19256 The RTL for a function can have several slots available: N is
19257 which slot to use. */
19260 assign_386_stack_local (enum machine_mode mode, enum ix86_stack_slot n)
19262 struct stack_local_entry *s;
19264 gcc_assert (n < MAX_386_STACK_LOCALS);
19266 /* Virtual slot is valid only before vregs are instantiated. */
19267 gcc_assert ((n == SLOT_VIRTUAL) == !virtuals_instantiated);
19269 for (s = ix86_stack_locals; s; s = s->next)
19270 if (s->mode == mode && s->n == n)
19271 return copy_rtx (s->rtl);
19273 s = (struct stack_local_entry *)
19274 ggc_alloc (sizeof (struct stack_local_entry));
19277 s->rtl = assign_stack_local (mode, GET_MODE_SIZE (mode), 0);
19279 s->next = ix86_stack_locals;
19280 ix86_stack_locals = s;
19284 /* Construct the SYMBOL_REF for the tls_get_addr function. */
19286 static GTY(()) rtx ix86_tls_symbol;
19288 ix86_tls_get_addr (void)
19291 if (!ix86_tls_symbol)
19293 ix86_tls_symbol = gen_rtx_SYMBOL_REF (Pmode,
19294 (TARGET_ANY_GNU_TLS
19296 ? "___tls_get_addr"
19297 : "__tls_get_addr");
19300 return ix86_tls_symbol;
19303 /* Construct the SYMBOL_REF for the _TLS_MODULE_BASE_ symbol. */
19305 static GTY(()) rtx ix86_tls_module_base_symbol;
19307 ix86_tls_module_base (void)
19310 if (!ix86_tls_module_base_symbol)
19312 ix86_tls_module_base_symbol = gen_rtx_SYMBOL_REF (Pmode,
19313 "_TLS_MODULE_BASE_");
19314 SYMBOL_REF_FLAGS (ix86_tls_module_base_symbol)
19315 |= TLS_MODEL_GLOBAL_DYNAMIC << SYMBOL_FLAG_TLS_SHIFT;
19318 return ix86_tls_module_base_symbol;
19321 /* Calculate the length of the memory address in the instruction
19322 encoding. Does not include the one-byte modrm, opcode, or prefix. */
19325 memory_address_length (rtx addr)
19327 struct ix86_address parts;
19328 rtx base, index, disp;
19332 if (GET_CODE (addr) == PRE_DEC
19333 || GET_CODE (addr) == POST_INC
19334 || GET_CODE (addr) == PRE_MODIFY
19335 || GET_CODE (addr) == POST_MODIFY)
19338 ok = ix86_decompose_address (addr, &parts);
19341 if (parts.base && GET_CODE (parts.base) == SUBREG)
19342 parts.base = SUBREG_REG (parts.base);
19343 if (parts.index && GET_CODE (parts.index) == SUBREG)
19344 parts.index = SUBREG_REG (parts.index);
19347 index = parts.index;
19352 - esp as the base always wants an index,
19353 - ebp as the base always wants a displacement,
19354 - r12 as the base always wants an index,
19355 - r13 as the base always wants a displacement. */
19357 /* Register Indirect. */
19358 if (base && !index && !disp)
19360 /* esp (for its index) and ebp (for its displacement) need
19361 the two-byte modrm form. Similarly for r12 and r13 in 64-bit
19364 && (addr == arg_pointer_rtx
19365 || addr == frame_pointer_rtx
19366 || REGNO (addr) == SP_REG
19367 || REGNO (addr) == BP_REG
19368 || REGNO (addr) == R12_REG
19369 || REGNO (addr) == R13_REG))
19373 /* Direct Addressing. In 64-bit mode mod 00 r/m 5
19374 is not disp32, but disp32(%rip), so for disp32
19375 SIB byte is needed, unless print_operand_address
19376 optimizes it into disp32(%rip) or (%rip) is implied
19378 else if (disp && !base && !index)
19385 if (GET_CODE (disp) == CONST)
19386 symbol = XEXP (disp, 0);
19387 if (GET_CODE (symbol) == PLUS
19388 && CONST_INT_P (XEXP (symbol, 1)))
19389 symbol = XEXP (symbol, 0);
19391 if (GET_CODE (symbol) != LABEL_REF
19392 && (GET_CODE (symbol) != SYMBOL_REF
19393 || SYMBOL_REF_TLS_MODEL (symbol) != 0)
19394 && (GET_CODE (symbol) != UNSPEC
19395 || (XINT (symbol, 1) != UNSPEC_GOTPCREL
19396 && XINT (symbol, 1) != UNSPEC_GOTNTPOFF)))
19403 /* Find the length of the displacement constant. */
19406 if (base && satisfies_constraint_K (disp))
19411 /* ebp always wants a displacement. Similarly r13. */
19412 else if (base && REG_P (base)
19413 && (REGNO (base) == BP_REG || REGNO (base) == R13_REG))
19416 /* An index requires the two-byte modrm form.... */
19418 /* ...like esp (or r12), which always wants an index. */
19419 || base == arg_pointer_rtx
19420 || base == frame_pointer_rtx
19421 || (base && REG_P (base)
19422 && (REGNO (base) == SP_REG || REGNO (base) == R12_REG)))
19439 /* Compute default value for "length_immediate" attribute. When SHORTFORM
19440 is set, expect that insn have 8bit immediate alternative. */
19442 ix86_attr_length_immediate_default (rtx insn, int shortform)
19446 extract_insn_cached (insn);
19447 for (i = recog_data.n_operands - 1; i >= 0; --i)
19448 if (CONSTANT_P (recog_data.operand[i]))
19450 enum attr_mode mode = get_attr_mode (insn);
19453 if (shortform && CONST_INT_P (recog_data.operand[i]))
19455 HOST_WIDE_INT ival = INTVAL (recog_data.operand[i]);
19462 ival = trunc_int_for_mode (ival, HImode);
19465 ival = trunc_int_for_mode (ival, SImode);
19470 if (IN_RANGE (ival, -128, 127))
19487 /* Immediates for DImode instructions are encoded as 32bit sign extended values. */
19492 fatal_insn ("unknown insn mode", insn);
19497 /* Compute default value for "length_address" attribute. */
19499 ix86_attr_length_address_default (rtx insn)
19503 if (get_attr_type (insn) == TYPE_LEA)
19505 rtx set = PATTERN (insn), addr;
19507 if (GET_CODE (set) == PARALLEL)
19508 set = XVECEXP (set, 0, 0);
19510 gcc_assert (GET_CODE (set) == SET);
19512 addr = SET_SRC (set);
19513 if (TARGET_64BIT && get_attr_mode (insn) == MODE_SI)
19515 if (GET_CODE (addr) == ZERO_EXTEND)
19516 addr = XEXP (addr, 0);
19517 if (GET_CODE (addr) == SUBREG)
19518 addr = SUBREG_REG (addr);
19521 return memory_address_length (addr);
19524 extract_insn_cached (insn);
19525 for (i = recog_data.n_operands - 1; i >= 0; --i)
19526 if (MEM_P (recog_data.operand[i]))
19528 constrain_operands_cached (reload_completed);
19529 if (which_alternative != -1)
19531 const char *constraints = recog_data.constraints[i];
19532 int alt = which_alternative;
19534 while (*constraints == '=' || *constraints == '+')
19537 while (*constraints++ != ',')
19539 /* Skip ignored operands. */
19540 if (*constraints == 'X')
19543 return memory_address_length (XEXP (recog_data.operand[i], 0));
19548 /* Compute default value for "length_vex" attribute. It includes
19549 2 or 3 byte VEX prefix and 1 opcode byte. */
19552 ix86_attr_length_vex_default (rtx insn, int has_0f_opcode,
19557 /* Only 0f opcode can use 2 byte VEX prefix and VEX W bit uses 3
19558 byte VEX prefix. */
19559 if (!has_0f_opcode || has_vex_w)
19562 /* We can always use 2 byte VEX prefix in 32bit. */
19566 extract_insn_cached (insn);
19568 for (i = recog_data.n_operands - 1; i >= 0; --i)
19569 if (REG_P (recog_data.operand[i]))
19571 /* REX.W bit uses 3 byte VEX prefix. */
19572 if (GET_MODE (recog_data.operand[i]) == DImode
19573 && GENERAL_REG_P (recog_data.operand[i]))
19578 /* REX.X or REX.B bits use 3 byte VEX prefix. */
19579 if (MEM_P (recog_data.operand[i])
19580 && x86_extended_reg_mentioned_p (recog_data.operand[i]))
19587 /* Return the maximum number of instructions a cpu can issue. */
19590 ix86_issue_rate (void)
19594 case PROCESSOR_PENTIUM:
19595 case PROCESSOR_ATOM:
19599 case PROCESSOR_PENTIUMPRO:
19600 case PROCESSOR_PENTIUM4:
19601 case PROCESSOR_ATHLON:
19603 case PROCESSOR_AMDFAM10:
19604 case PROCESSOR_NOCONA:
19605 case PROCESSOR_GENERIC32:
19606 case PROCESSOR_GENERIC64:
19609 case PROCESSOR_CORE2:
19617 /* A subroutine of ix86_adjust_cost -- return true iff INSN reads flags set
19618 by DEP_INSN and nothing set by DEP_INSN. */
19621 ix86_flags_dependent (rtx insn, rtx dep_insn, enum attr_type insn_type)
19625 /* Simplify the test for uninteresting insns. */
19626 if (insn_type != TYPE_SETCC
19627 && insn_type != TYPE_ICMOV
19628 && insn_type != TYPE_FCMOV
19629 && insn_type != TYPE_IBR)
19632 if ((set = single_set (dep_insn)) != 0)
19634 set = SET_DEST (set);
19637 else if (GET_CODE (PATTERN (dep_insn)) == PARALLEL
19638 && XVECLEN (PATTERN (dep_insn), 0) == 2
19639 && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 0)) == SET
19640 && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 1)) == SET)
19642 set = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0));
19643 set2 = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0));
19648 if (!REG_P (set) || REGNO (set) != FLAGS_REG)
19651 /* This test is true if the dependent insn reads the flags but
19652 not any other potentially set register. */
19653 if (!reg_overlap_mentioned_p (set, PATTERN (insn)))
19656 if (set2 && reg_overlap_mentioned_p (set2, PATTERN (insn)))
19662 /* Return true iff USE_INSN has a memory address with operands set by
19666 ix86_agi_dependent (rtx set_insn, rtx use_insn)
19669 extract_insn_cached (use_insn);
19670 for (i = recog_data.n_operands - 1; i >= 0; --i)
19671 if (MEM_P (recog_data.operand[i]))
19673 rtx addr = XEXP (recog_data.operand[i], 0);
19674 return modified_in_p (addr, set_insn) != 0;
19680 ix86_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
19682 enum attr_type insn_type, dep_insn_type;
19683 enum attr_memory memory;
19685 int dep_insn_code_number;
19687 /* Anti and output dependencies have zero cost on all CPUs. */
19688 if (REG_NOTE_KIND (link) != 0)
19691 dep_insn_code_number = recog_memoized (dep_insn);
19693 /* If we can't recognize the insns, we can't really do anything. */
19694 if (dep_insn_code_number < 0 || recog_memoized (insn) < 0)
19697 insn_type = get_attr_type (insn);
19698 dep_insn_type = get_attr_type (dep_insn);
19702 case PROCESSOR_PENTIUM:
19703 /* Address Generation Interlock adds a cycle of latency. */
19704 if (insn_type == TYPE_LEA)
19706 rtx addr = PATTERN (insn);
19708 if (GET_CODE (addr) == PARALLEL)
19709 addr = XVECEXP (addr, 0, 0);
19711 gcc_assert (GET_CODE (addr) == SET);
19713 addr = SET_SRC (addr);
19714 if (modified_in_p (addr, dep_insn))
19717 else if (ix86_agi_dependent (dep_insn, insn))
19720 /* ??? Compares pair with jump/setcc. */
19721 if (ix86_flags_dependent (insn, dep_insn, insn_type))
19724 /* Floating point stores require value to be ready one cycle earlier. */
19725 if (insn_type == TYPE_FMOV
19726 && get_attr_memory (insn) == MEMORY_STORE
19727 && !ix86_agi_dependent (dep_insn, insn))
19731 case PROCESSOR_PENTIUMPRO:
19732 memory = get_attr_memory (insn);
19734 /* INT->FP conversion is expensive. */
19735 if (get_attr_fp_int_src (dep_insn))
19738 /* There is one cycle extra latency between an FP op and a store. */
19739 if (insn_type == TYPE_FMOV
19740 && (set = single_set (dep_insn)) != NULL_RTX
19741 && (set2 = single_set (insn)) != NULL_RTX
19742 && rtx_equal_p (SET_DEST (set), SET_SRC (set2))
19743 && MEM_P (SET_DEST (set2)))
19746 /* Show ability of reorder buffer to hide latency of load by executing
19747 in parallel with previous instruction in case
19748 previous instruction is not needed to compute the address. */
19749 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
19750 && !ix86_agi_dependent (dep_insn, insn))
19752 /* Claim moves to take one cycle, as core can issue one load
19753 at time and the next load can start cycle later. */
19754 if (dep_insn_type == TYPE_IMOV
19755 || dep_insn_type == TYPE_FMOV)
19763 memory = get_attr_memory (insn);
19765 /* The esp dependency is resolved before the instruction is really
19767 if ((insn_type == TYPE_PUSH || insn_type == TYPE_POP)
19768 && (dep_insn_type == TYPE_PUSH || dep_insn_type == TYPE_POP))
19771 /* INT->FP conversion is expensive. */
19772 if (get_attr_fp_int_src (dep_insn))
19775 /* Show ability of reorder buffer to hide latency of load by executing
19776 in parallel with previous instruction in case
19777 previous instruction is not needed to compute the address. */
19778 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
19779 && !ix86_agi_dependent (dep_insn, insn))
19781 /* Claim moves to take one cycle, as core can issue one load
19782 at time and the next load can start cycle later. */
19783 if (dep_insn_type == TYPE_IMOV
19784 || dep_insn_type == TYPE_FMOV)
19793 case PROCESSOR_ATHLON:
19795 case PROCESSOR_AMDFAM10:
19796 case PROCESSOR_ATOM:
19797 case PROCESSOR_GENERIC32:
19798 case PROCESSOR_GENERIC64:
19799 memory = get_attr_memory (insn);
19801 /* Show ability of reorder buffer to hide latency of load by executing
19802 in parallel with previous instruction in case
19803 previous instruction is not needed to compute the address. */
19804 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
19805 && !ix86_agi_dependent (dep_insn, insn))
19807 enum attr_unit unit = get_attr_unit (insn);
19810 /* Because of the difference between the length of integer and
19811 floating unit pipeline preparation stages, the memory operands
19812 for floating point are cheaper.
19814 ??? For Athlon it the difference is most probably 2. */
19815 if (unit == UNIT_INTEGER || unit == UNIT_UNKNOWN)
19818 loadcost = TARGET_ATHLON ? 2 : 0;
19820 if (cost >= loadcost)
19833 /* How many alternative schedules to try. This should be as wide as the
19834 scheduling freedom in the DFA, but no wider. Making this value too
19835 large results extra work for the scheduler. */
19838 ia32_multipass_dfa_lookahead (void)
19842 case PROCESSOR_PENTIUM:
19845 case PROCESSOR_PENTIUMPRO:
19855 /* Compute the alignment given to a constant that is being placed in memory.
19856 EXP is the constant and ALIGN is the alignment that the object would
19858 The value of this function is used instead of that alignment to align
19862 ix86_constant_alignment (tree exp, int align)
19864 if (TREE_CODE (exp) == REAL_CST || TREE_CODE (exp) == VECTOR_CST
19865 || TREE_CODE (exp) == INTEGER_CST)
19867 if (TYPE_MODE (TREE_TYPE (exp)) == DFmode && align < 64)
19869 else if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (exp))) && align < 128)
19872 else if (!optimize_size && TREE_CODE (exp) == STRING_CST
19873 && TREE_STRING_LENGTH (exp) >= 31 && align < BITS_PER_WORD)
19874 return BITS_PER_WORD;
19879 /* Compute the alignment for a static variable.
19880 TYPE is the data type, and ALIGN is the alignment that
19881 the object would ordinarily have. The value of this function is used
19882 instead of that alignment to align the object. */
19885 ix86_data_alignment (tree type, int align)
19887 int max_align = optimize_size ? BITS_PER_WORD : MIN (256, MAX_OFILE_ALIGNMENT);
19889 if (AGGREGATE_TYPE_P (type)
19890 && TYPE_SIZE (type)
19891 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19892 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= (unsigned) max_align
19893 || TREE_INT_CST_HIGH (TYPE_SIZE (type)))
19894 && align < max_align)
19897 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
19898 to 16byte boundary. */
19901 if (AGGREGATE_TYPE_P (type)
19902 && TYPE_SIZE (type)
19903 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19904 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= 128
19905 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 128)
19909 if (TREE_CODE (type) == ARRAY_TYPE)
19911 if (TYPE_MODE (TREE_TYPE (type)) == DFmode && align < 64)
19913 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type))) && align < 128)
19916 else if (TREE_CODE (type) == COMPLEX_TYPE)
19919 if (TYPE_MODE (type) == DCmode && align < 64)
19921 if ((TYPE_MODE (type) == XCmode
19922 || TYPE_MODE (type) == TCmode) && align < 128)
19925 else if ((TREE_CODE (type) == RECORD_TYPE
19926 || TREE_CODE (type) == UNION_TYPE
19927 || TREE_CODE (type) == QUAL_UNION_TYPE)
19928 && TYPE_FIELDS (type))
19930 if (DECL_MODE (TYPE_FIELDS (type)) == DFmode && align < 64)
19932 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type))) && align < 128)
19935 else if (TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == VECTOR_TYPE
19936 || TREE_CODE (type) == INTEGER_TYPE)
19938 if (TYPE_MODE (type) == DFmode && align < 64)
19940 if (ALIGN_MODE_128 (TYPE_MODE (type)) && align < 128)
19947 /* Compute the alignment for a local variable or a stack slot. EXP is
19948 the data type or decl itself, MODE is the widest mode available and
19949 ALIGN is the alignment that the object would ordinarily have. The
19950 value of this macro is used instead of that alignment to align the
19954 ix86_local_alignment (tree exp, enum machine_mode mode,
19955 unsigned int align)
19959 if (exp && DECL_P (exp))
19961 type = TREE_TYPE (exp);
19970 /* Don't do dynamic stack realignment for long long objects with
19971 -mpreferred-stack-boundary=2. */
19974 && ix86_preferred_stack_boundary < 64
19975 && (mode == DImode || (type && TYPE_MODE (type) == DImode))
19976 && (!type || !TYPE_USER_ALIGN (type))
19977 && (!decl || !DECL_USER_ALIGN (decl)))
19980 /* If TYPE is NULL, we are allocating a stack slot for caller-save
19981 register in MODE. We will return the largest alignment of XF
19985 if (mode == XFmode && align < GET_MODE_ALIGNMENT (DFmode))
19986 align = GET_MODE_ALIGNMENT (DFmode);
19990 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
19991 to 16byte boundary. */
19994 if (AGGREGATE_TYPE_P (type)
19995 && TYPE_SIZE (type)
19996 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19997 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= 16
19998 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 128)
20001 if (TREE_CODE (type) == ARRAY_TYPE)
20003 if (TYPE_MODE (TREE_TYPE (type)) == DFmode && align < 64)
20005 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type))) && align < 128)
20008 else if (TREE_CODE (type) == COMPLEX_TYPE)
20010 if (TYPE_MODE (type) == DCmode && align < 64)
20012 if ((TYPE_MODE (type) == XCmode
20013 || TYPE_MODE (type) == TCmode) && align < 128)
20016 else if ((TREE_CODE (type) == RECORD_TYPE
20017 || TREE_CODE (type) == UNION_TYPE
20018 || TREE_CODE (type) == QUAL_UNION_TYPE)
20019 && TYPE_FIELDS (type))
20021 if (DECL_MODE (TYPE_FIELDS (type)) == DFmode && align < 64)
20023 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type))) && align < 128)
20026 else if (TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == VECTOR_TYPE
20027 || TREE_CODE (type) == INTEGER_TYPE)
20030 if (TYPE_MODE (type) == DFmode && align < 64)
20032 if (ALIGN_MODE_128 (TYPE_MODE (type)) && align < 128)
20038 /* Compute the minimum required alignment for dynamic stack realignment
20039 purposes for a local variable, parameter or a stack slot. EXP is
20040 the data type or decl itself, MODE is its mode and ALIGN is the
20041 alignment that the object would ordinarily have. */
20044 ix86_minimum_alignment (tree exp, enum machine_mode mode,
20045 unsigned int align)
20049 if (TARGET_64BIT || align != 64 || ix86_preferred_stack_boundary >= 64)
20052 if (exp && DECL_P (exp))
20054 type = TREE_TYPE (exp);
20063 /* Don't do dynamic stack realignment for long long objects with
20064 -mpreferred-stack-boundary=2. */
20065 if ((mode == DImode || (type && TYPE_MODE (type) == DImode))
20066 && (!type || !TYPE_USER_ALIGN (type))
20067 && (!decl || !DECL_USER_ALIGN (decl)))
20073 /* Find a location for the static chain incoming to a nested function.
20074 This is a register, unless all free registers are used by arguments. */
20077 ix86_static_chain (const_tree fndecl, bool incoming_p)
20081 if (!DECL_STATIC_CHAIN (fndecl))
20086 /* We always use R10 in 64-bit mode. */
20092 /* By default in 32-bit mode we use ECX to pass the static chain. */
20095 fntype = TREE_TYPE (fndecl);
20096 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (fntype)))
20098 /* Fastcall functions use ecx/edx for arguments, which leaves
20099 us with EAX for the static chain. */
20102 else if (ix86_function_regparm (fntype, fndecl) == 3)
20104 /* For regparm 3, we have no free call-clobbered registers in
20105 which to store the static chain. In order to implement this,
20106 we have the trampoline push the static chain to the stack.
20107 However, we can't push a value below the return address when
20108 we call the nested function directly, so we have to use an
20109 alternate entry point. For this we use ESI, and have the
20110 alternate entry point push ESI, so that things appear the
20111 same once we're executing the nested function. */
20114 if (fndecl == current_function_decl)
20115 ix86_static_chain_on_stack = true;
20116 return gen_frame_mem (SImode,
20117 plus_constant (arg_pointer_rtx, -8));
20123 return gen_rtx_REG (Pmode, regno);
20126 /* Emit RTL insns to initialize the variable parts of a trampoline.
20127 FNDECL is the decl of the target address; M_TRAMP is a MEM for
20128 the trampoline, and CHAIN_VALUE is an RTX for the static chain
20129 to be passed to the target function. */
20132 ix86_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value)
20136 fnaddr = XEXP (DECL_RTL (fndecl), 0);
20143 /* Depending on the static chain location, either load a register
20144 with a constant, or push the constant to the stack. All of the
20145 instructions are the same size. */
20146 chain = ix86_static_chain (fndecl, true);
20149 if (REGNO (chain) == CX_REG)
20151 else if (REGNO (chain) == AX_REG)
20154 gcc_unreachable ();
20159 mem = adjust_address (m_tramp, QImode, 0);
20160 emit_move_insn (mem, gen_int_mode (opcode, QImode));
20162 mem = adjust_address (m_tramp, SImode, 1);
20163 emit_move_insn (mem, chain_value);
20165 /* Compute offset from the end of the jmp to the target function.
20166 In the case in which the trampoline stores the static chain on
20167 the stack, we need to skip the first insn which pushes the
20168 (call-saved) register static chain; this push is 1 byte. */
20169 disp = expand_binop (SImode, sub_optab, fnaddr,
20170 plus_constant (XEXP (m_tramp, 0),
20171 MEM_P (chain) ? 9 : 10),
20172 NULL_RTX, 1, OPTAB_DIRECT);
20174 mem = adjust_address (m_tramp, QImode, 5);
20175 emit_move_insn (mem, gen_int_mode (0xe9, QImode));
20177 mem = adjust_address (m_tramp, SImode, 6);
20178 emit_move_insn (mem, disp);
20184 /* Load the function address to r11. Try to load address using
20185 the shorter movl instead of movabs. We may want to support
20186 movq for kernel mode, but kernel does not use trampolines at
20188 if (x86_64_zext_immediate_operand (fnaddr, VOIDmode))
20190 fnaddr = copy_to_mode_reg (DImode, fnaddr);
20192 mem = adjust_address (m_tramp, HImode, offset);
20193 emit_move_insn (mem, gen_int_mode (0xbb41, HImode));
20195 mem = adjust_address (m_tramp, SImode, offset + 2);
20196 emit_move_insn (mem, gen_lowpart (SImode, fnaddr));
20201 mem = adjust_address (m_tramp, HImode, offset);
20202 emit_move_insn (mem, gen_int_mode (0xbb49, HImode));
20204 mem = adjust_address (m_tramp, DImode, offset + 2);
20205 emit_move_insn (mem, fnaddr);
20209 /* Load static chain using movabs to r10. */
20210 mem = adjust_address (m_tramp, HImode, offset);
20211 emit_move_insn (mem, gen_int_mode (0xba49, HImode));
20213 mem = adjust_address (m_tramp, DImode, offset + 2);
20214 emit_move_insn (mem, chain_value);
20217 /* Jump to r11; the last (unused) byte is a nop, only there to
20218 pad the write out to a single 32-bit store. */
20219 mem = adjust_address (m_tramp, SImode, offset);
20220 emit_move_insn (mem, gen_int_mode (0x90e3ff49, SImode));
20223 gcc_assert (offset <= TRAMPOLINE_SIZE);
20226 #ifdef ENABLE_EXECUTE_STACK
20227 #ifdef CHECK_EXECUTE_STACK_ENABLED
20228 if (CHECK_EXECUTE_STACK_ENABLED)
20230 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
20231 LCT_NORMAL, VOIDmode, 1, XEXP (m_tramp, 0), Pmode);
20235 /* The following file contains several enumerations and data structures
20236 built from the definitions in i386-builtin-types.def. */
20238 #include "i386-builtin-types.inc"
20240 /* Table for the ix86 builtin non-function types. */
20241 static GTY(()) tree ix86_builtin_type_tab[(int) IX86_BT_LAST_CPTR + 1];
20243 /* Retrieve an element from the above table, building some of
20244 the types lazily. */
20247 ix86_get_builtin_type (enum ix86_builtin_type tcode)
20249 unsigned int index;
20252 gcc_assert ((unsigned)tcode < ARRAY_SIZE(ix86_builtin_type_tab));
20254 type = ix86_builtin_type_tab[(int) tcode];
20258 gcc_assert (tcode > IX86_BT_LAST_PRIM);
20259 if (tcode <= IX86_BT_LAST_VECT)
20261 enum machine_mode mode;
20263 index = tcode - IX86_BT_LAST_PRIM - 1;
20264 itype = ix86_get_builtin_type (ix86_builtin_type_vect_base[index]);
20265 mode = ix86_builtin_type_vect_mode[index];
20267 type = build_vector_type_for_mode (itype, mode);
20273 index = tcode - IX86_BT_LAST_VECT - 1;
20274 if (tcode <= IX86_BT_LAST_PTR)
20275 quals = TYPE_UNQUALIFIED;
20277 quals = TYPE_QUAL_CONST;
20279 itype = ix86_get_builtin_type (ix86_builtin_type_ptr_base[index]);
20280 if (quals != TYPE_UNQUALIFIED)
20281 itype = build_qualified_type (itype, quals);
20283 type = build_pointer_type (itype);
20286 ix86_builtin_type_tab[(int) tcode] = type;
20290 /* Table for the ix86 builtin function types. */
20291 static GTY(()) tree ix86_builtin_func_type_tab[(int) IX86_BT_LAST_ALIAS + 1];
20293 /* Retrieve an element from the above table, building some of
20294 the types lazily. */
20297 ix86_get_builtin_func_type (enum ix86_builtin_func_type tcode)
20301 gcc_assert ((unsigned)tcode < ARRAY_SIZE (ix86_builtin_func_type_tab));
20303 type = ix86_builtin_func_type_tab[(int) tcode];
20307 if (tcode <= IX86_BT_LAST_FUNC)
20309 unsigned start = ix86_builtin_func_start[(int) tcode];
20310 unsigned after = ix86_builtin_func_start[(int) tcode + 1];
20311 tree rtype, atype, args = void_list_node;
20314 rtype = ix86_get_builtin_type (ix86_builtin_func_args[start]);
20315 for (i = after - 1; i > start; --i)
20317 atype = ix86_get_builtin_type (ix86_builtin_func_args[i]);
20318 args = tree_cons (NULL, atype, args);
20321 type = build_function_type (rtype, args);
20325 unsigned index = tcode - IX86_BT_LAST_FUNC - 1;
20326 enum ix86_builtin_func_type icode;
20328 icode = ix86_builtin_func_alias_base[index];
20329 type = ix86_get_builtin_func_type (icode);
20332 ix86_builtin_func_type_tab[(int) tcode] = type;
20337 /* Codes for all the SSE/MMX builtins. */
20340 IX86_BUILTIN_ADDPS,
20341 IX86_BUILTIN_ADDSS,
20342 IX86_BUILTIN_DIVPS,
20343 IX86_BUILTIN_DIVSS,
20344 IX86_BUILTIN_MULPS,
20345 IX86_BUILTIN_MULSS,
20346 IX86_BUILTIN_SUBPS,
20347 IX86_BUILTIN_SUBSS,
20349 IX86_BUILTIN_CMPEQPS,
20350 IX86_BUILTIN_CMPLTPS,
20351 IX86_BUILTIN_CMPLEPS,
20352 IX86_BUILTIN_CMPGTPS,
20353 IX86_BUILTIN_CMPGEPS,
20354 IX86_BUILTIN_CMPNEQPS,
20355 IX86_BUILTIN_CMPNLTPS,
20356 IX86_BUILTIN_CMPNLEPS,
20357 IX86_BUILTIN_CMPNGTPS,
20358 IX86_BUILTIN_CMPNGEPS,
20359 IX86_BUILTIN_CMPORDPS,
20360 IX86_BUILTIN_CMPUNORDPS,
20361 IX86_BUILTIN_CMPEQSS,
20362 IX86_BUILTIN_CMPLTSS,
20363 IX86_BUILTIN_CMPLESS,
20364 IX86_BUILTIN_CMPNEQSS,
20365 IX86_BUILTIN_CMPNLTSS,
20366 IX86_BUILTIN_CMPNLESS,
20367 IX86_BUILTIN_CMPNGTSS,
20368 IX86_BUILTIN_CMPNGESS,
20369 IX86_BUILTIN_CMPORDSS,
20370 IX86_BUILTIN_CMPUNORDSS,
20372 IX86_BUILTIN_COMIEQSS,
20373 IX86_BUILTIN_COMILTSS,
20374 IX86_BUILTIN_COMILESS,
20375 IX86_BUILTIN_COMIGTSS,
20376 IX86_BUILTIN_COMIGESS,
20377 IX86_BUILTIN_COMINEQSS,
20378 IX86_BUILTIN_UCOMIEQSS,
20379 IX86_BUILTIN_UCOMILTSS,
20380 IX86_BUILTIN_UCOMILESS,
20381 IX86_BUILTIN_UCOMIGTSS,
20382 IX86_BUILTIN_UCOMIGESS,
20383 IX86_BUILTIN_UCOMINEQSS,
20385 IX86_BUILTIN_CVTPI2PS,
20386 IX86_BUILTIN_CVTPS2PI,
20387 IX86_BUILTIN_CVTSI2SS,
20388 IX86_BUILTIN_CVTSI642SS,
20389 IX86_BUILTIN_CVTSS2SI,
20390 IX86_BUILTIN_CVTSS2SI64,
20391 IX86_BUILTIN_CVTTPS2PI,
20392 IX86_BUILTIN_CVTTSS2SI,
20393 IX86_BUILTIN_CVTTSS2SI64,
20395 IX86_BUILTIN_MAXPS,
20396 IX86_BUILTIN_MAXSS,
20397 IX86_BUILTIN_MINPS,
20398 IX86_BUILTIN_MINSS,
20400 IX86_BUILTIN_LOADUPS,
20401 IX86_BUILTIN_STOREUPS,
20402 IX86_BUILTIN_MOVSS,
20404 IX86_BUILTIN_MOVHLPS,
20405 IX86_BUILTIN_MOVLHPS,
20406 IX86_BUILTIN_LOADHPS,
20407 IX86_BUILTIN_LOADLPS,
20408 IX86_BUILTIN_STOREHPS,
20409 IX86_BUILTIN_STORELPS,
20411 IX86_BUILTIN_MASKMOVQ,
20412 IX86_BUILTIN_MOVMSKPS,
20413 IX86_BUILTIN_PMOVMSKB,
20415 IX86_BUILTIN_MOVNTPS,
20416 IX86_BUILTIN_MOVNTQ,
20418 IX86_BUILTIN_LOADDQU,
20419 IX86_BUILTIN_STOREDQU,
20421 IX86_BUILTIN_PACKSSWB,
20422 IX86_BUILTIN_PACKSSDW,
20423 IX86_BUILTIN_PACKUSWB,
20425 IX86_BUILTIN_PADDB,
20426 IX86_BUILTIN_PADDW,
20427 IX86_BUILTIN_PADDD,
20428 IX86_BUILTIN_PADDQ,
20429 IX86_BUILTIN_PADDSB,
20430 IX86_BUILTIN_PADDSW,
20431 IX86_BUILTIN_PADDUSB,
20432 IX86_BUILTIN_PADDUSW,
20433 IX86_BUILTIN_PSUBB,
20434 IX86_BUILTIN_PSUBW,
20435 IX86_BUILTIN_PSUBD,
20436 IX86_BUILTIN_PSUBQ,
20437 IX86_BUILTIN_PSUBSB,
20438 IX86_BUILTIN_PSUBSW,
20439 IX86_BUILTIN_PSUBUSB,
20440 IX86_BUILTIN_PSUBUSW,
20443 IX86_BUILTIN_PANDN,
20447 IX86_BUILTIN_PAVGB,
20448 IX86_BUILTIN_PAVGW,
20450 IX86_BUILTIN_PCMPEQB,
20451 IX86_BUILTIN_PCMPEQW,
20452 IX86_BUILTIN_PCMPEQD,
20453 IX86_BUILTIN_PCMPGTB,
20454 IX86_BUILTIN_PCMPGTW,
20455 IX86_BUILTIN_PCMPGTD,
20457 IX86_BUILTIN_PMADDWD,
20459 IX86_BUILTIN_PMAXSW,
20460 IX86_BUILTIN_PMAXUB,
20461 IX86_BUILTIN_PMINSW,
20462 IX86_BUILTIN_PMINUB,
20464 IX86_BUILTIN_PMULHUW,
20465 IX86_BUILTIN_PMULHW,
20466 IX86_BUILTIN_PMULLW,
20468 IX86_BUILTIN_PSADBW,
20469 IX86_BUILTIN_PSHUFW,
20471 IX86_BUILTIN_PSLLW,
20472 IX86_BUILTIN_PSLLD,
20473 IX86_BUILTIN_PSLLQ,
20474 IX86_BUILTIN_PSRAW,
20475 IX86_BUILTIN_PSRAD,
20476 IX86_BUILTIN_PSRLW,
20477 IX86_BUILTIN_PSRLD,
20478 IX86_BUILTIN_PSRLQ,
20479 IX86_BUILTIN_PSLLWI,
20480 IX86_BUILTIN_PSLLDI,
20481 IX86_BUILTIN_PSLLQI,
20482 IX86_BUILTIN_PSRAWI,
20483 IX86_BUILTIN_PSRADI,
20484 IX86_BUILTIN_PSRLWI,
20485 IX86_BUILTIN_PSRLDI,
20486 IX86_BUILTIN_PSRLQI,
20488 IX86_BUILTIN_PUNPCKHBW,
20489 IX86_BUILTIN_PUNPCKHWD,
20490 IX86_BUILTIN_PUNPCKHDQ,
20491 IX86_BUILTIN_PUNPCKLBW,
20492 IX86_BUILTIN_PUNPCKLWD,
20493 IX86_BUILTIN_PUNPCKLDQ,
20495 IX86_BUILTIN_SHUFPS,
20497 IX86_BUILTIN_RCPPS,
20498 IX86_BUILTIN_RCPSS,
20499 IX86_BUILTIN_RSQRTPS,
20500 IX86_BUILTIN_RSQRTPS_NR,
20501 IX86_BUILTIN_RSQRTSS,
20502 IX86_BUILTIN_RSQRTF,
20503 IX86_BUILTIN_SQRTPS,
20504 IX86_BUILTIN_SQRTPS_NR,
20505 IX86_BUILTIN_SQRTSS,
20507 IX86_BUILTIN_UNPCKHPS,
20508 IX86_BUILTIN_UNPCKLPS,
20510 IX86_BUILTIN_ANDPS,
20511 IX86_BUILTIN_ANDNPS,
20513 IX86_BUILTIN_XORPS,
20516 IX86_BUILTIN_LDMXCSR,
20517 IX86_BUILTIN_STMXCSR,
20518 IX86_BUILTIN_SFENCE,
20520 /* 3DNow! Original */
20521 IX86_BUILTIN_FEMMS,
20522 IX86_BUILTIN_PAVGUSB,
20523 IX86_BUILTIN_PF2ID,
20524 IX86_BUILTIN_PFACC,
20525 IX86_BUILTIN_PFADD,
20526 IX86_BUILTIN_PFCMPEQ,
20527 IX86_BUILTIN_PFCMPGE,
20528 IX86_BUILTIN_PFCMPGT,
20529 IX86_BUILTIN_PFMAX,
20530 IX86_BUILTIN_PFMIN,
20531 IX86_BUILTIN_PFMUL,
20532 IX86_BUILTIN_PFRCP,
20533 IX86_BUILTIN_PFRCPIT1,
20534 IX86_BUILTIN_PFRCPIT2,
20535 IX86_BUILTIN_PFRSQIT1,
20536 IX86_BUILTIN_PFRSQRT,
20537 IX86_BUILTIN_PFSUB,
20538 IX86_BUILTIN_PFSUBR,
20539 IX86_BUILTIN_PI2FD,
20540 IX86_BUILTIN_PMULHRW,
20542 /* 3DNow! Athlon Extensions */
20543 IX86_BUILTIN_PF2IW,
20544 IX86_BUILTIN_PFNACC,
20545 IX86_BUILTIN_PFPNACC,
20546 IX86_BUILTIN_PI2FW,
20547 IX86_BUILTIN_PSWAPDSI,
20548 IX86_BUILTIN_PSWAPDSF,
20551 IX86_BUILTIN_ADDPD,
20552 IX86_BUILTIN_ADDSD,
20553 IX86_BUILTIN_DIVPD,
20554 IX86_BUILTIN_DIVSD,
20555 IX86_BUILTIN_MULPD,
20556 IX86_BUILTIN_MULSD,
20557 IX86_BUILTIN_SUBPD,
20558 IX86_BUILTIN_SUBSD,
20560 IX86_BUILTIN_CMPEQPD,
20561 IX86_BUILTIN_CMPLTPD,
20562 IX86_BUILTIN_CMPLEPD,
20563 IX86_BUILTIN_CMPGTPD,
20564 IX86_BUILTIN_CMPGEPD,
20565 IX86_BUILTIN_CMPNEQPD,
20566 IX86_BUILTIN_CMPNLTPD,
20567 IX86_BUILTIN_CMPNLEPD,
20568 IX86_BUILTIN_CMPNGTPD,
20569 IX86_BUILTIN_CMPNGEPD,
20570 IX86_BUILTIN_CMPORDPD,
20571 IX86_BUILTIN_CMPUNORDPD,
20572 IX86_BUILTIN_CMPEQSD,
20573 IX86_BUILTIN_CMPLTSD,
20574 IX86_BUILTIN_CMPLESD,
20575 IX86_BUILTIN_CMPNEQSD,
20576 IX86_BUILTIN_CMPNLTSD,
20577 IX86_BUILTIN_CMPNLESD,
20578 IX86_BUILTIN_CMPORDSD,
20579 IX86_BUILTIN_CMPUNORDSD,
20581 IX86_BUILTIN_COMIEQSD,
20582 IX86_BUILTIN_COMILTSD,
20583 IX86_BUILTIN_COMILESD,
20584 IX86_BUILTIN_COMIGTSD,
20585 IX86_BUILTIN_COMIGESD,
20586 IX86_BUILTIN_COMINEQSD,
20587 IX86_BUILTIN_UCOMIEQSD,
20588 IX86_BUILTIN_UCOMILTSD,
20589 IX86_BUILTIN_UCOMILESD,
20590 IX86_BUILTIN_UCOMIGTSD,
20591 IX86_BUILTIN_UCOMIGESD,
20592 IX86_BUILTIN_UCOMINEQSD,
20594 IX86_BUILTIN_MAXPD,
20595 IX86_BUILTIN_MAXSD,
20596 IX86_BUILTIN_MINPD,
20597 IX86_BUILTIN_MINSD,
20599 IX86_BUILTIN_ANDPD,
20600 IX86_BUILTIN_ANDNPD,
20602 IX86_BUILTIN_XORPD,
20604 IX86_BUILTIN_SQRTPD,
20605 IX86_BUILTIN_SQRTSD,
20607 IX86_BUILTIN_UNPCKHPD,
20608 IX86_BUILTIN_UNPCKLPD,
20610 IX86_BUILTIN_SHUFPD,
20612 IX86_BUILTIN_LOADUPD,
20613 IX86_BUILTIN_STOREUPD,
20614 IX86_BUILTIN_MOVSD,
20616 IX86_BUILTIN_LOADHPD,
20617 IX86_BUILTIN_LOADLPD,
20619 IX86_BUILTIN_CVTDQ2PD,
20620 IX86_BUILTIN_CVTDQ2PS,
20622 IX86_BUILTIN_CVTPD2DQ,
20623 IX86_BUILTIN_CVTPD2PI,
20624 IX86_BUILTIN_CVTPD2PS,
20625 IX86_BUILTIN_CVTTPD2DQ,
20626 IX86_BUILTIN_CVTTPD2PI,
20628 IX86_BUILTIN_CVTPI2PD,
20629 IX86_BUILTIN_CVTSI2SD,
20630 IX86_BUILTIN_CVTSI642SD,
20632 IX86_BUILTIN_CVTSD2SI,
20633 IX86_BUILTIN_CVTSD2SI64,
20634 IX86_BUILTIN_CVTSD2SS,
20635 IX86_BUILTIN_CVTSS2SD,
20636 IX86_BUILTIN_CVTTSD2SI,
20637 IX86_BUILTIN_CVTTSD2SI64,
20639 IX86_BUILTIN_CVTPS2DQ,
20640 IX86_BUILTIN_CVTPS2PD,
20641 IX86_BUILTIN_CVTTPS2DQ,
20643 IX86_BUILTIN_MOVNTI,
20644 IX86_BUILTIN_MOVNTPD,
20645 IX86_BUILTIN_MOVNTDQ,
20647 IX86_BUILTIN_MOVQ128,
20650 IX86_BUILTIN_MASKMOVDQU,
20651 IX86_BUILTIN_MOVMSKPD,
20652 IX86_BUILTIN_PMOVMSKB128,
20654 IX86_BUILTIN_PACKSSWB128,
20655 IX86_BUILTIN_PACKSSDW128,
20656 IX86_BUILTIN_PACKUSWB128,
20658 IX86_BUILTIN_PADDB128,
20659 IX86_BUILTIN_PADDW128,
20660 IX86_BUILTIN_PADDD128,
20661 IX86_BUILTIN_PADDQ128,
20662 IX86_BUILTIN_PADDSB128,
20663 IX86_BUILTIN_PADDSW128,
20664 IX86_BUILTIN_PADDUSB128,
20665 IX86_BUILTIN_PADDUSW128,
20666 IX86_BUILTIN_PSUBB128,
20667 IX86_BUILTIN_PSUBW128,
20668 IX86_BUILTIN_PSUBD128,
20669 IX86_BUILTIN_PSUBQ128,
20670 IX86_BUILTIN_PSUBSB128,
20671 IX86_BUILTIN_PSUBSW128,
20672 IX86_BUILTIN_PSUBUSB128,
20673 IX86_BUILTIN_PSUBUSW128,
20675 IX86_BUILTIN_PAND128,
20676 IX86_BUILTIN_PANDN128,
20677 IX86_BUILTIN_POR128,
20678 IX86_BUILTIN_PXOR128,
20680 IX86_BUILTIN_PAVGB128,
20681 IX86_BUILTIN_PAVGW128,
20683 IX86_BUILTIN_PCMPEQB128,
20684 IX86_BUILTIN_PCMPEQW128,
20685 IX86_BUILTIN_PCMPEQD128,
20686 IX86_BUILTIN_PCMPGTB128,
20687 IX86_BUILTIN_PCMPGTW128,
20688 IX86_BUILTIN_PCMPGTD128,
20690 IX86_BUILTIN_PMADDWD128,
20692 IX86_BUILTIN_PMAXSW128,
20693 IX86_BUILTIN_PMAXUB128,
20694 IX86_BUILTIN_PMINSW128,
20695 IX86_BUILTIN_PMINUB128,
20697 IX86_BUILTIN_PMULUDQ,
20698 IX86_BUILTIN_PMULUDQ128,
20699 IX86_BUILTIN_PMULHUW128,
20700 IX86_BUILTIN_PMULHW128,
20701 IX86_BUILTIN_PMULLW128,
20703 IX86_BUILTIN_PSADBW128,
20704 IX86_BUILTIN_PSHUFHW,
20705 IX86_BUILTIN_PSHUFLW,
20706 IX86_BUILTIN_PSHUFD,
20708 IX86_BUILTIN_PSLLDQI128,
20709 IX86_BUILTIN_PSLLWI128,
20710 IX86_BUILTIN_PSLLDI128,
20711 IX86_BUILTIN_PSLLQI128,
20712 IX86_BUILTIN_PSRAWI128,
20713 IX86_BUILTIN_PSRADI128,
20714 IX86_BUILTIN_PSRLDQI128,
20715 IX86_BUILTIN_PSRLWI128,
20716 IX86_BUILTIN_PSRLDI128,
20717 IX86_BUILTIN_PSRLQI128,
20719 IX86_BUILTIN_PSLLDQ128,
20720 IX86_BUILTIN_PSLLW128,
20721 IX86_BUILTIN_PSLLD128,
20722 IX86_BUILTIN_PSLLQ128,
20723 IX86_BUILTIN_PSRAW128,
20724 IX86_BUILTIN_PSRAD128,
20725 IX86_BUILTIN_PSRLW128,
20726 IX86_BUILTIN_PSRLD128,
20727 IX86_BUILTIN_PSRLQ128,
20729 IX86_BUILTIN_PUNPCKHBW128,
20730 IX86_BUILTIN_PUNPCKHWD128,
20731 IX86_BUILTIN_PUNPCKHDQ128,
20732 IX86_BUILTIN_PUNPCKHQDQ128,
20733 IX86_BUILTIN_PUNPCKLBW128,
20734 IX86_BUILTIN_PUNPCKLWD128,
20735 IX86_BUILTIN_PUNPCKLDQ128,
20736 IX86_BUILTIN_PUNPCKLQDQ128,
20738 IX86_BUILTIN_CLFLUSH,
20739 IX86_BUILTIN_MFENCE,
20740 IX86_BUILTIN_LFENCE,
20742 IX86_BUILTIN_BSRSI,
20743 IX86_BUILTIN_BSRDI,
20744 IX86_BUILTIN_RDPMC,
20745 IX86_BUILTIN_RDTSC,
20746 IX86_BUILTIN_RDTSCP,
20747 IX86_BUILTIN_ROLQI,
20748 IX86_BUILTIN_ROLHI,
20749 IX86_BUILTIN_RORQI,
20750 IX86_BUILTIN_RORHI,
20753 IX86_BUILTIN_ADDSUBPS,
20754 IX86_BUILTIN_HADDPS,
20755 IX86_BUILTIN_HSUBPS,
20756 IX86_BUILTIN_MOVSHDUP,
20757 IX86_BUILTIN_MOVSLDUP,
20758 IX86_BUILTIN_ADDSUBPD,
20759 IX86_BUILTIN_HADDPD,
20760 IX86_BUILTIN_HSUBPD,
20761 IX86_BUILTIN_LDDQU,
20763 IX86_BUILTIN_MONITOR,
20764 IX86_BUILTIN_MWAIT,
20767 IX86_BUILTIN_PHADDW,
20768 IX86_BUILTIN_PHADDD,
20769 IX86_BUILTIN_PHADDSW,
20770 IX86_BUILTIN_PHSUBW,
20771 IX86_BUILTIN_PHSUBD,
20772 IX86_BUILTIN_PHSUBSW,
20773 IX86_BUILTIN_PMADDUBSW,
20774 IX86_BUILTIN_PMULHRSW,
20775 IX86_BUILTIN_PSHUFB,
20776 IX86_BUILTIN_PSIGNB,
20777 IX86_BUILTIN_PSIGNW,
20778 IX86_BUILTIN_PSIGND,
20779 IX86_BUILTIN_PALIGNR,
20780 IX86_BUILTIN_PABSB,
20781 IX86_BUILTIN_PABSW,
20782 IX86_BUILTIN_PABSD,
20784 IX86_BUILTIN_PHADDW128,
20785 IX86_BUILTIN_PHADDD128,
20786 IX86_BUILTIN_PHADDSW128,
20787 IX86_BUILTIN_PHSUBW128,
20788 IX86_BUILTIN_PHSUBD128,
20789 IX86_BUILTIN_PHSUBSW128,
20790 IX86_BUILTIN_PMADDUBSW128,
20791 IX86_BUILTIN_PMULHRSW128,
20792 IX86_BUILTIN_PSHUFB128,
20793 IX86_BUILTIN_PSIGNB128,
20794 IX86_BUILTIN_PSIGNW128,
20795 IX86_BUILTIN_PSIGND128,
20796 IX86_BUILTIN_PALIGNR128,
20797 IX86_BUILTIN_PABSB128,
20798 IX86_BUILTIN_PABSW128,
20799 IX86_BUILTIN_PABSD128,
20801 /* AMDFAM10 - SSE4A New Instructions. */
20802 IX86_BUILTIN_MOVNTSD,
20803 IX86_BUILTIN_MOVNTSS,
20804 IX86_BUILTIN_EXTRQI,
20805 IX86_BUILTIN_EXTRQ,
20806 IX86_BUILTIN_INSERTQI,
20807 IX86_BUILTIN_INSERTQ,
20810 IX86_BUILTIN_BLENDPD,
20811 IX86_BUILTIN_BLENDPS,
20812 IX86_BUILTIN_BLENDVPD,
20813 IX86_BUILTIN_BLENDVPS,
20814 IX86_BUILTIN_PBLENDVB128,
20815 IX86_BUILTIN_PBLENDW128,
20820 IX86_BUILTIN_INSERTPS128,
20822 IX86_BUILTIN_MOVNTDQA,
20823 IX86_BUILTIN_MPSADBW128,
20824 IX86_BUILTIN_PACKUSDW128,
20825 IX86_BUILTIN_PCMPEQQ,
20826 IX86_BUILTIN_PHMINPOSUW128,
20828 IX86_BUILTIN_PMAXSB128,
20829 IX86_BUILTIN_PMAXSD128,
20830 IX86_BUILTIN_PMAXUD128,
20831 IX86_BUILTIN_PMAXUW128,
20833 IX86_BUILTIN_PMINSB128,
20834 IX86_BUILTIN_PMINSD128,
20835 IX86_BUILTIN_PMINUD128,
20836 IX86_BUILTIN_PMINUW128,
20838 IX86_BUILTIN_PMOVSXBW128,
20839 IX86_BUILTIN_PMOVSXBD128,
20840 IX86_BUILTIN_PMOVSXBQ128,
20841 IX86_BUILTIN_PMOVSXWD128,
20842 IX86_BUILTIN_PMOVSXWQ128,
20843 IX86_BUILTIN_PMOVSXDQ128,
20845 IX86_BUILTIN_PMOVZXBW128,
20846 IX86_BUILTIN_PMOVZXBD128,
20847 IX86_BUILTIN_PMOVZXBQ128,
20848 IX86_BUILTIN_PMOVZXWD128,
20849 IX86_BUILTIN_PMOVZXWQ128,
20850 IX86_BUILTIN_PMOVZXDQ128,
20852 IX86_BUILTIN_PMULDQ128,
20853 IX86_BUILTIN_PMULLD128,
20855 IX86_BUILTIN_ROUNDPD,
20856 IX86_BUILTIN_ROUNDPS,
20857 IX86_BUILTIN_ROUNDSD,
20858 IX86_BUILTIN_ROUNDSS,
20860 IX86_BUILTIN_PTESTZ,
20861 IX86_BUILTIN_PTESTC,
20862 IX86_BUILTIN_PTESTNZC,
20864 IX86_BUILTIN_VEC_INIT_V2SI,
20865 IX86_BUILTIN_VEC_INIT_V4HI,
20866 IX86_BUILTIN_VEC_INIT_V8QI,
20867 IX86_BUILTIN_VEC_EXT_V2DF,
20868 IX86_BUILTIN_VEC_EXT_V2DI,
20869 IX86_BUILTIN_VEC_EXT_V4SF,
20870 IX86_BUILTIN_VEC_EXT_V4SI,
20871 IX86_BUILTIN_VEC_EXT_V8HI,
20872 IX86_BUILTIN_VEC_EXT_V2SI,
20873 IX86_BUILTIN_VEC_EXT_V4HI,
20874 IX86_BUILTIN_VEC_EXT_V16QI,
20875 IX86_BUILTIN_VEC_SET_V2DI,
20876 IX86_BUILTIN_VEC_SET_V4SF,
20877 IX86_BUILTIN_VEC_SET_V4SI,
20878 IX86_BUILTIN_VEC_SET_V8HI,
20879 IX86_BUILTIN_VEC_SET_V4HI,
20880 IX86_BUILTIN_VEC_SET_V16QI,
20882 IX86_BUILTIN_VEC_PACK_SFIX,
20885 IX86_BUILTIN_CRC32QI,
20886 IX86_BUILTIN_CRC32HI,
20887 IX86_BUILTIN_CRC32SI,
20888 IX86_BUILTIN_CRC32DI,
20890 IX86_BUILTIN_PCMPESTRI128,
20891 IX86_BUILTIN_PCMPESTRM128,
20892 IX86_BUILTIN_PCMPESTRA128,
20893 IX86_BUILTIN_PCMPESTRC128,
20894 IX86_BUILTIN_PCMPESTRO128,
20895 IX86_BUILTIN_PCMPESTRS128,
20896 IX86_BUILTIN_PCMPESTRZ128,
20897 IX86_BUILTIN_PCMPISTRI128,
20898 IX86_BUILTIN_PCMPISTRM128,
20899 IX86_BUILTIN_PCMPISTRA128,
20900 IX86_BUILTIN_PCMPISTRC128,
20901 IX86_BUILTIN_PCMPISTRO128,
20902 IX86_BUILTIN_PCMPISTRS128,
20903 IX86_BUILTIN_PCMPISTRZ128,
20905 IX86_BUILTIN_PCMPGTQ,
20907 /* AES instructions */
20908 IX86_BUILTIN_AESENC128,
20909 IX86_BUILTIN_AESENCLAST128,
20910 IX86_BUILTIN_AESDEC128,
20911 IX86_BUILTIN_AESDECLAST128,
20912 IX86_BUILTIN_AESIMC128,
20913 IX86_BUILTIN_AESKEYGENASSIST128,
20915 /* PCLMUL instruction */
20916 IX86_BUILTIN_PCLMULQDQ128,
20919 IX86_BUILTIN_ADDPD256,
20920 IX86_BUILTIN_ADDPS256,
20921 IX86_BUILTIN_ADDSUBPD256,
20922 IX86_BUILTIN_ADDSUBPS256,
20923 IX86_BUILTIN_ANDPD256,
20924 IX86_BUILTIN_ANDPS256,
20925 IX86_BUILTIN_ANDNPD256,
20926 IX86_BUILTIN_ANDNPS256,
20927 IX86_BUILTIN_BLENDPD256,
20928 IX86_BUILTIN_BLENDPS256,
20929 IX86_BUILTIN_BLENDVPD256,
20930 IX86_BUILTIN_BLENDVPS256,
20931 IX86_BUILTIN_DIVPD256,
20932 IX86_BUILTIN_DIVPS256,
20933 IX86_BUILTIN_DPPS256,
20934 IX86_BUILTIN_HADDPD256,
20935 IX86_BUILTIN_HADDPS256,
20936 IX86_BUILTIN_HSUBPD256,
20937 IX86_BUILTIN_HSUBPS256,
20938 IX86_BUILTIN_MAXPD256,
20939 IX86_BUILTIN_MAXPS256,
20940 IX86_BUILTIN_MINPD256,
20941 IX86_BUILTIN_MINPS256,
20942 IX86_BUILTIN_MULPD256,
20943 IX86_BUILTIN_MULPS256,
20944 IX86_BUILTIN_ORPD256,
20945 IX86_BUILTIN_ORPS256,
20946 IX86_BUILTIN_SHUFPD256,
20947 IX86_BUILTIN_SHUFPS256,
20948 IX86_BUILTIN_SUBPD256,
20949 IX86_BUILTIN_SUBPS256,
20950 IX86_BUILTIN_XORPD256,
20951 IX86_BUILTIN_XORPS256,
20952 IX86_BUILTIN_CMPSD,
20953 IX86_BUILTIN_CMPSS,
20954 IX86_BUILTIN_CMPPD,
20955 IX86_BUILTIN_CMPPS,
20956 IX86_BUILTIN_CMPPD256,
20957 IX86_BUILTIN_CMPPS256,
20958 IX86_BUILTIN_CVTDQ2PD256,
20959 IX86_BUILTIN_CVTDQ2PS256,
20960 IX86_BUILTIN_CVTPD2PS256,
20961 IX86_BUILTIN_CVTPS2DQ256,
20962 IX86_BUILTIN_CVTPS2PD256,
20963 IX86_BUILTIN_CVTTPD2DQ256,
20964 IX86_BUILTIN_CVTPD2DQ256,
20965 IX86_BUILTIN_CVTTPS2DQ256,
20966 IX86_BUILTIN_EXTRACTF128PD256,
20967 IX86_BUILTIN_EXTRACTF128PS256,
20968 IX86_BUILTIN_EXTRACTF128SI256,
20969 IX86_BUILTIN_VZEROALL,
20970 IX86_BUILTIN_VZEROUPPER,
20971 IX86_BUILTIN_VPERMILVARPD,
20972 IX86_BUILTIN_VPERMILVARPS,
20973 IX86_BUILTIN_VPERMILVARPD256,
20974 IX86_BUILTIN_VPERMILVARPS256,
20975 IX86_BUILTIN_VPERMILPD,
20976 IX86_BUILTIN_VPERMILPS,
20977 IX86_BUILTIN_VPERMILPD256,
20978 IX86_BUILTIN_VPERMILPS256,
20979 IX86_BUILTIN_VPERM2F128PD256,
20980 IX86_BUILTIN_VPERM2F128PS256,
20981 IX86_BUILTIN_VPERM2F128SI256,
20982 IX86_BUILTIN_VBROADCASTSS,
20983 IX86_BUILTIN_VBROADCASTSD256,
20984 IX86_BUILTIN_VBROADCASTSS256,
20985 IX86_BUILTIN_VBROADCASTPD256,
20986 IX86_BUILTIN_VBROADCASTPS256,
20987 IX86_BUILTIN_VINSERTF128PD256,
20988 IX86_BUILTIN_VINSERTF128PS256,
20989 IX86_BUILTIN_VINSERTF128SI256,
20990 IX86_BUILTIN_LOADUPD256,
20991 IX86_BUILTIN_LOADUPS256,
20992 IX86_BUILTIN_STOREUPD256,
20993 IX86_BUILTIN_STOREUPS256,
20994 IX86_BUILTIN_LDDQU256,
20995 IX86_BUILTIN_MOVNTDQ256,
20996 IX86_BUILTIN_MOVNTPD256,
20997 IX86_BUILTIN_MOVNTPS256,
20998 IX86_BUILTIN_LOADDQU256,
20999 IX86_BUILTIN_STOREDQU256,
21000 IX86_BUILTIN_MASKLOADPD,
21001 IX86_BUILTIN_MASKLOADPS,
21002 IX86_BUILTIN_MASKSTOREPD,
21003 IX86_BUILTIN_MASKSTOREPS,
21004 IX86_BUILTIN_MASKLOADPD256,
21005 IX86_BUILTIN_MASKLOADPS256,
21006 IX86_BUILTIN_MASKSTOREPD256,
21007 IX86_BUILTIN_MASKSTOREPS256,
21008 IX86_BUILTIN_MOVSHDUP256,
21009 IX86_BUILTIN_MOVSLDUP256,
21010 IX86_BUILTIN_MOVDDUP256,
21012 IX86_BUILTIN_SQRTPD256,
21013 IX86_BUILTIN_SQRTPS256,
21014 IX86_BUILTIN_SQRTPS_NR256,
21015 IX86_BUILTIN_RSQRTPS256,
21016 IX86_BUILTIN_RSQRTPS_NR256,
21018 IX86_BUILTIN_RCPPS256,
21020 IX86_BUILTIN_ROUNDPD256,
21021 IX86_BUILTIN_ROUNDPS256,
21023 IX86_BUILTIN_UNPCKHPD256,
21024 IX86_BUILTIN_UNPCKLPD256,
21025 IX86_BUILTIN_UNPCKHPS256,
21026 IX86_BUILTIN_UNPCKLPS256,
21028 IX86_BUILTIN_SI256_SI,
21029 IX86_BUILTIN_PS256_PS,
21030 IX86_BUILTIN_PD256_PD,
21031 IX86_BUILTIN_SI_SI256,
21032 IX86_BUILTIN_PS_PS256,
21033 IX86_BUILTIN_PD_PD256,
21035 IX86_BUILTIN_VTESTZPD,
21036 IX86_BUILTIN_VTESTCPD,
21037 IX86_BUILTIN_VTESTNZCPD,
21038 IX86_BUILTIN_VTESTZPS,
21039 IX86_BUILTIN_VTESTCPS,
21040 IX86_BUILTIN_VTESTNZCPS,
21041 IX86_BUILTIN_VTESTZPD256,
21042 IX86_BUILTIN_VTESTCPD256,
21043 IX86_BUILTIN_VTESTNZCPD256,
21044 IX86_BUILTIN_VTESTZPS256,
21045 IX86_BUILTIN_VTESTCPS256,
21046 IX86_BUILTIN_VTESTNZCPS256,
21047 IX86_BUILTIN_PTESTZ256,
21048 IX86_BUILTIN_PTESTC256,
21049 IX86_BUILTIN_PTESTNZC256,
21051 IX86_BUILTIN_MOVMSKPD256,
21052 IX86_BUILTIN_MOVMSKPS256,
21054 /* TFmode support builtins. */
21056 IX86_BUILTIN_HUGE_VALQ,
21057 IX86_BUILTIN_FABSQ,
21058 IX86_BUILTIN_COPYSIGNQ,
21060 /* Vectorizer support builtins. */
21061 IX86_BUILTIN_CPYSGNPS,
21062 IX86_BUILTIN_CPYSGNPD,
21064 IX86_BUILTIN_CVTUDQ2PS,
21066 IX86_BUILTIN_VEC_PERM_V2DF,
21067 IX86_BUILTIN_VEC_PERM_V4SF,
21068 IX86_BUILTIN_VEC_PERM_V2DI,
21069 IX86_BUILTIN_VEC_PERM_V4SI,
21070 IX86_BUILTIN_VEC_PERM_V8HI,
21071 IX86_BUILTIN_VEC_PERM_V16QI,
21072 IX86_BUILTIN_VEC_PERM_V2DI_U,
21073 IX86_BUILTIN_VEC_PERM_V4SI_U,
21074 IX86_BUILTIN_VEC_PERM_V8HI_U,
21075 IX86_BUILTIN_VEC_PERM_V16QI_U,
21076 IX86_BUILTIN_VEC_PERM_V4DF,
21077 IX86_BUILTIN_VEC_PERM_V8SF,
21079 /* FMA4 and XOP instructions. */
21080 IX86_BUILTIN_VFMADDSS,
21081 IX86_BUILTIN_VFMADDSD,
21082 IX86_BUILTIN_VFMADDPS,
21083 IX86_BUILTIN_VFMADDPD,
21084 IX86_BUILTIN_VFMSUBSS,
21085 IX86_BUILTIN_VFMSUBSD,
21086 IX86_BUILTIN_VFMSUBPS,
21087 IX86_BUILTIN_VFMSUBPD,
21088 IX86_BUILTIN_VFMADDSUBPS,
21089 IX86_BUILTIN_VFMADDSUBPD,
21090 IX86_BUILTIN_VFMSUBADDPS,
21091 IX86_BUILTIN_VFMSUBADDPD,
21092 IX86_BUILTIN_VFNMADDSS,
21093 IX86_BUILTIN_VFNMADDSD,
21094 IX86_BUILTIN_VFNMADDPS,
21095 IX86_BUILTIN_VFNMADDPD,
21096 IX86_BUILTIN_VFNMSUBSS,
21097 IX86_BUILTIN_VFNMSUBSD,
21098 IX86_BUILTIN_VFNMSUBPS,
21099 IX86_BUILTIN_VFNMSUBPD,
21100 IX86_BUILTIN_VFMADDPS256,
21101 IX86_BUILTIN_VFMADDPD256,
21102 IX86_BUILTIN_VFMSUBPS256,
21103 IX86_BUILTIN_VFMSUBPD256,
21104 IX86_BUILTIN_VFMADDSUBPS256,
21105 IX86_BUILTIN_VFMADDSUBPD256,
21106 IX86_BUILTIN_VFMSUBADDPS256,
21107 IX86_BUILTIN_VFMSUBADDPD256,
21108 IX86_BUILTIN_VFNMADDPS256,
21109 IX86_BUILTIN_VFNMADDPD256,
21110 IX86_BUILTIN_VFNMSUBPS256,
21111 IX86_BUILTIN_VFNMSUBPD256,
21113 IX86_BUILTIN_VPCMOV,
21114 IX86_BUILTIN_VPCMOV_V2DI,
21115 IX86_BUILTIN_VPCMOV_V4SI,
21116 IX86_BUILTIN_VPCMOV_V8HI,
21117 IX86_BUILTIN_VPCMOV_V16QI,
21118 IX86_BUILTIN_VPCMOV_V4SF,
21119 IX86_BUILTIN_VPCMOV_V2DF,
21120 IX86_BUILTIN_VPCMOV256,
21121 IX86_BUILTIN_VPCMOV_V4DI256,
21122 IX86_BUILTIN_VPCMOV_V8SI256,
21123 IX86_BUILTIN_VPCMOV_V16HI256,
21124 IX86_BUILTIN_VPCMOV_V32QI256,
21125 IX86_BUILTIN_VPCMOV_V8SF256,
21126 IX86_BUILTIN_VPCMOV_V4DF256,
21128 IX86_BUILTIN_VPPERM,
21130 IX86_BUILTIN_VPMACSSWW,
21131 IX86_BUILTIN_VPMACSWW,
21132 IX86_BUILTIN_VPMACSSWD,
21133 IX86_BUILTIN_VPMACSWD,
21134 IX86_BUILTIN_VPMACSSDD,
21135 IX86_BUILTIN_VPMACSDD,
21136 IX86_BUILTIN_VPMACSSDQL,
21137 IX86_BUILTIN_VPMACSSDQH,
21138 IX86_BUILTIN_VPMACSDQL,
21139 IX86_BUILTIN_VPMACSDQH,
21140 IX86_BUILTIN_VPMADCSSWD,
21141 IX86_BUILTIN_VPMADCSWD,
21143 IX86_BUILTIN_VPHADDBW,
21144 IX86_BUILTIN_VPHADDBD,
21145 IX86_BUILTIN_VPHADDBQ,
21146 IX86_BUILTIN_VPHADDWD,
21147 IX86_BUILTIN_VPHADDWQ,
21148 IX86_BUILTIN_VPHADDDQ,
21149 IX86_BUILTIN_VPHADDUBW,
21150 IX86_BUILTIN_VPHADDUBD,
21151 IX86_BUILTIN_VPHADDUBQ,
21152 IX86_BUILTIN_VPHADDUWD,
21153 IX86_BUILTIN_VPHADDUWQ,
21154 IX86_BUILTIN_VPHADDUDQ,
21155 IX86_BUILTIN_VPHSUBBW,
21156 IX86_BUILTIN_VPHSUBWD,
21157 IX86_BUILTIN_VPHSUBDQ,
21159 IX86_BUILTIN_VPROTB,
21160 IX86_BUILTIN_VPROTW,
21161 IX86_BUILTIN_VPROTD,
21162 IX86_BUILTIN_VPROTQ,
21163 IX86_BUILTIN_VPROTB_IMM,
21164 IX86_BUILTIN_VPROTW_IMM,
21165 IX86_BUILTIN_VPROTD_IMM,
21166 IX86_BUILTIN_VPROTQ_IMM,
21168 IX86_BUILTIN_VPSHLB,
21169 IX86_BUILTIN_VPSHLW,
21170 IX86_BUILTIN_VPSHLD,
21171 IX86_BUILTIN_VPSHLQ,
21172 IX86_BUILTIN_VPSHAB,
21173 IX86_BUILTIN_VPSHAW,
21174 IX86_BUILTIN_VPSHAD,
21175 IX86_BUILTIN_VPSHAQ,
21177 IX86_BUILTIN_VFRCZSS,
21178 IX86_BUILTIN_VFRCZSD,
21179 IX86_BUILTIN_VFRCZPS,
21180 IX86_BUILTIN_VFRCZPD,
21181 IX86_BUILTIN_VFRCZPS256,
21182 IX86_BUILTIN_VFRCZPD256,
21184 IX86_BUILTIN_VPCOMEQUB,
21185 IX86_BUILTIN_VPCOMNEUB,
21186 IX86_BUILTIN_VPCOMLTUB,
21187 IX86_BUILTIN_VPCOMLEUB,
21188 IX86_BUILTIN_VPCOMGTUB,
21189 IX86_BUILTIN_VPCOMGEUB,
21190 IX86_BUILTIN_VPCOMFALSEUB,
21191 IX86_BUILTIN_VPCOMTRUEUB,
21193 IX86_BUILTIN_VPCOMEQUW,
21194 IX86_BUILTIN_VPCOMNEUW,
21195 IX86_BUILTIN_VPCOMLTUW,
21196 IX86_BUILTIN_VPCOMLEUW,
21197 IX86_BUILTIN_VPCOMGTUW,
21198 IX86_BUILTIN_VPCOMGEUW,
21199 IX86_BUILTIN_VPCOMFALSEUW,
21200 IX86_BUILTIN_VPCOMTRUEUW,
21202 IX86_BUILTIN_VPCOMEQUD,
21203 IX86_BUILTIN_VPCOMNEUD,
21204 IX86_BUILTIN_VPCOMLTUD,
21205 IX86_BUILTIN_VPCOMLEUD,
21206 IX86_BUILTIN_VPCOMGTUD,
21207 IX86_BUILTIN_VPCOMGEUD,
21208 IX86_BUILTIN_VPCOMFALSEUD,
21209 IX86_BUILTIN_VPCOMTRUEUD,
21211 IX86_BUILTIN_VPCOMEQUQ,
21212 IX86_BUILTIN_VPCOMNEUQ,
21213 IX86_BUILTIN_VPCOMLTUQ,
21214 IX86_BUILTIN_VPCOMLEUQ,
21215 IX86_BUILTIN_VPCOMGTUQ,
21216 IX86_BUILTIN_VPCOMGEUQ,
21217 IX86_BUILTIN_VPCOMFALSEUQ,
21218 IX86_BUILTIN_VPCOMTRUEUQ,
21220 IX86_BUILTIN_VPCOMEQB,
21221 IX86_BUILTIN_VPCOMNEB,
21222 IX86_BUILTIN_VPCOMLTB,
21223 IX86_BUILTIN_VPCOMLEB,
21224 IX86_BUILTIN_VPCOMGTB,
21225 IX86_BUILTIN_VPCOMGEB,
21226 IX86_BUILTIN_VPCOMFALSEB,
21227 IX86_BUILTIN_VPCOMTRUEB,
21229 IX86_BUILTIN_VPCOMEQW,
21230 IX86_BUILTIN_VPCOMNEW,
21231 IX86_BUILTIN_VPCOMLTW,
21232 IX86_BUILTIN_VPCOMLEW,
21233 IX86_BUILTIN_VPCOMGTW,
21234 IX86_BUILTIN_VPCOMGEW,
21235 IX86_BUILTIN_VPCOMFALSEW,
21236 IX86_BUILTIN_VPCOMTRUEW,
21238 IX86_BUILTIN_VPCOMEQD,
21239 IX86_BUILTIN_VPCOMNED,
21240 IX86_BUILTIN_VPCOMLTD,
21241 IX86_BUILTIN_VPCOMLED,
21242 IX86_BUILTIN_VPCOMGTD,
21243 IX86_BUILTIN_VPCOMGED,
21244 IX86_BUILTIN_VPCOMFALSED,
21245 IX86_BUILTIN_VPCOMTRUED,
21247 IX86_BUILTIN_VPCOMEQQ,
21248 IX86_BUILTIN_VPCOMNEQ,
21249 IX86_BUILTIN_VPCOMLTQ,
21250 IX86_BUILTIN_VPCOMLEQ,
21251 IX86_BUILTIN_VPCOMGTQ,
21252 IX86_BUILTIN_VPCOMGEQ,
21253 IX86_BUILTIN_VPCOMFALSEQ,
21254 IX86_BUILTIN_VPCOMTRUEQ,
21256 /* LWP instructions. */
21257 IX86_BUILTIN_LLWPCB,
21258 IX86_BUILTIN_SLWPCB,
21259 IX86_BUILTIN_LWPVAL32,
21260 IX86_BUILTIN_LWPVAL64,
21261 IX86_BUILTIN_LWPINS32,
21262 IX86_BUILTIN_LWPINS64,
21269 /* Table for the ix86 builtin decls. */
21270 static GTY(()) tree ix86_builtins[(int) IX86_BUILTIN_MAX];
21272 /* Table of all of the builtin functions that are possible with different ISA's
21273 but are waiting to be built until a function is declared to use that
21275 struct builtin_isa {
21276 const char *name; /* function name */
21277 enum ix86_builtin_func_type tcode; /* type to use in the declaration */
21278 int isa; /* isa_flags this builtin is defined for */
21279 bool const_p; /* true if the declaration is constant */
21280 bool set_and_not_built_p;
21283 static struct builtin_isa ix86_builtins_isa[(int) IX86_BUILTIN_MAX];
21286 /* Add an ix86 target builtin function with CODE, NAME and TYPE. Save the MASK
21287 of which isa_flags to use in the ix86_builtins_isa array. Stores the
21288 function decl in the ix86_builtins array. Returns the function decl or
21289 NULL_TREE, if the builtin was not added.
21291 If the front end has a special hook for builtin functions, delay adding
21292 builtin functions that aren't in the current ISA until the ISA is changed
21293 with function specific optimization. Doing so, can save about 300K for the
21294 default compiler. When the builtin is expanded, check at that time whether
21297 If the front end doesn't have a special hook, record all builtins, even if
21298 it isn't an instruction set in the current ISA in case the user uses
21299 function specific options for a different ISA, so that we don't get scope
21300 errors if a builtin is added in the middle of a function scope. */
21303 def_builtin (int mask, const char *name, enum ix86_builtin_func_type tcode,
21304 enum ix86_builtins code)
21306 tree decl = NULL_TREE;
21308 if (!(mask & OPTION_MASK_ISA_64BIT) || TARGET_64BIT)
21310 ix86_builtins_isa[(int) code].isa = mask;
21313 || (mask & ix86_isa_flags) != 0
21314 || (lang_hooks.builtin_function
21315 == lang_hooks.builtin_function_ext_scope))
21318 tree type = ix86_get_builtin_func_type (tcode);
21319 decl = add_builtin_function (name, type, code, BUILT_IN_MD,
21321 ix86_builtins[(int) code] = decl;
21322 ix86_builtins_isa[(int) code].set_and_not_built_p = false;
21326 ix86_builtins[(int) code] = NULL_TREE;
21327 ix86_builtins_isa[(int) code].tcode = tcode;
21328 ix86_builtins_isa[(int) code].name = name;
21329 ix86_builtins_isa[(int) code].const_p = false;
21330 ix86_builtins_isa[(int) code].set_and_not_built_p = true;
21337 /* Like def_builtin, but also marks the function decl "const". */
21340 def_builtin_const (int mask, const char *name,
21341 enum ix86_builtin_func_type tcode, enum ix86_builtins code)
21343 tree decl = def_builtin (mask, name, tcode, code);
21345 TREE_READONLY (decl) = 1;
21347 ix86_builtins_isa[(int) code].const_p = true;
21352 /* Add any new builtin functions for a given ISA that may not have been
21353 declared. This saves a bit of space compared to adding all of the
21354 declarations to the tree, even if we didn't use them. */
21357 ix86_add_new_builtins (int isa)
21361 for (i = 0; i < (int)IX86_BUILTIN_MAX; i++)
21363 if ((ix86_builtins_isa[i].isa & isa) != 0
21364 && ix86_builtins_isa[i].set_and_not_built_p)
21368 /* Don't define the builtin again. */
21369 ix86_builtins_isa[i].set_and_not_built_p = false;
21371 type = ix86_get_builtin_func_type (ix86_builtins_isa[i].tcode);
21372 decl = add_builtin_function_ext_scope (ix86_builtins_isa[i].name,
21373 type, i, BUILT_IN_MD, NULL,
21376 ix86_builtins[i] = decl;
21377 if (ix86_builtins_isa[i].const_p)
21378 TREE_READONLY (decl) = 1;
21383 /* Bits for builtin_description.flag. */
21385 /* Set when we don't support the comparison natively, and should
21386 swap_comparison in order to support it. */
21387 #define BUILTIN_DESC_SWAP_OPERANDS 1
21389 struct builtin_description
21391 const unsigned int mask;
21392 const enum insn_code icode;
21393 const char *const name;
21394 const enum ix86_builtins code;
21395 const enum rtx_code comparison;
21399 static const struct builtin_description bdesc_comi[] =
21401 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comieq", IX86_BUILTIN_COMIEQSS, UNEQ, 0 },
21402 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comilt", IX86_BUILTIN_COMILTSS, UNLT, 0 },
21403 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comile", IX86_BUILTIN_COMILESS, UNLE, 0 },
21404 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comigt", IX86_BUILTIN_COMIGTSS, GT, 0 },
21405 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comige", IX86_BUILTIN_COMIGESS, GE, 0 },
21406 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comineq", IX86_BUILTIN_COMINEQSS, LTGT, 0 },
21407 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomieq", IX86_BUILTIN_UCOMIEQSS, UNEQ, 0 },
21408 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomilt", IX86_BUILTIN_UCOMILTSS, UNLT, 0 },
21409 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomile", IX86_BUILTIN_UCOMILESS, UNLE, 0 },
21410 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomigt", IX86_BUILTIN_UCOMIGTSS, GT, 0 },
21411 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomige", IX86_BUILTIN_UCOMIGESS, GE, 0 },
21412 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomineq", IX86_BUILTIN_UCOMINEQSS, LTGT, 0 },
21413 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdeq", IX86_BUILTIN_COMIEQSD, UNEQ, 0 },
21414 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdlt", IX86_BUILTIN_COMILTSD, UNLT, 0 },
21415 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdle", IX86_BUILTIN_COMILESD, UNLE, 0 },
21416 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdgt", IX86_BUILTIN_COMIGTSD, GT, 0 },
21417 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdge", IX86_BUILTIN_COMIGESD, GE, 0 },
21418 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdneq", IX86_BUILTIN_COMINEQSD, LTGT, 0 },
21419 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdeq", IX86_BUILTIN_UCOMIEQSD, UNEQ, 0 },
21420 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdlt", IX86_BUILTIN_UCOMILTSD, UNLT, 0 },
21421 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdle", IX86_BUILTIN_UCOMILESD, UNLE, 0 },
21422 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdgt", IX86_BUILTIN_UCOMIGTSD, GT, 0 },
21423 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdge", IX86_BUILTIN_UCOMIGESD, GE, 0 },
21424 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdneq", IX86_BUILTIN_UCOMINEQSD, LTGT, 0 },
21427 static const struct builtin_description bdesc_pcmpestr[] =
21430 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestri128", IX86_BUILTIN_PCMPESTRI128, UNKNOWN, 0 },
21431 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestrm128", IX86_BUILTIN_PCMPESTRM128, UNKNOWN, 0 },
21432 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestria128", IX86_BUILTIN_PCMPESTRA128, UNKNOWN, (int) CCAmode },
21433 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestric128", IX86_BUILTIN_PCMPESTRC128, UNKNOWN, (int) CCCmode },
21434 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestrio128", IX86_BUILTIN_PCMPESTRO128, UNKNOWN, (int) CCOmode },
21435 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestris128", IX86_BUILTIN_PCMPESTRS128, UNKNOWN, (int) CCSmode },
21436 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestriz128", IX86_BUILTIN_PCMPESTRZ128, UNKNOWN, (int) CCZmode },
21439 static const struct builtin_description bdesc_pcmpistr[] =
21442 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistri128", IX86_BUILTIN_PCMPISTRI128, UNKNOWN, 0 },
21443 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistrm128", IX86_BUILTIN_PCMPISTRM128, UNKNOWN, 0 },
21444 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistria128", IX86_BUILTIN_PCMPISTRA128, UNKNOWN, (int) CCAmode },
21445 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistric128", IX86_BUILTIN_PCMPISTRC128, UNKNOWN, (int) CCCmode },
21446 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistrio128", IX86_BUILTIN_PCMPISTRO128, UNKNOWN, (int) CCOmode },
21447 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistris128", IX86_BUILTIN_PCMPISTRS128, UNKNOWN, (int) CCSmode },
21448 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistriz128", IX86_BUILTIN_PCMPISTRZ128, UNKNOWN, (int) CCZmode },
21451 /* Special builtins with variable number of arguments. */
21452 static const struct builtin_description bdesc_special_args[] =
21454 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rdtsc, "__builtin_ia32_rdtsc", IX86_BUILTIN_RDTSC, UNKNOWN, (int) UINT64_FTYPE_VOID },
21455 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rdtscp, "__builtin_ia32_rdtscp", IX86_BUILTIN_RDTSCP, UNKNOWN, (int) UINT64_FTYPE_PUNSIGNED },
21458 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_emms, "__builtin_ia32_emms", IX86_BUILTIN_EMMS, UNKNOWN, (int) VOID_FTYPE_VOID },
21461 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_femms, "__builtin_ia32_femms", IX86_BUILTIN_FEMMS, UNKNOWN, (int) VOID_FTYPE_VOID },
21464 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movups, "__builtin_ia32_storeups", IX86_BUILTIN_STOREUPS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
21465 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movntv4sf, "__builtin_ia32_movntps", IX86_BUILTIN_MOVNTPS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
21466 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movups, "__builtin_ia32_loadups", IX86_BUILTIN_LOADUPS, UNKNOWN, (int) V4SF_FTYPE_PCFLOAT },
21468 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_loadhps_exp, "__builtin_ia32_loadhps", IX86_BUILTIN_LOADHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_PCV2SF },
21469 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_loadlps_exp, "__builtin_ia32_loadlps", IX86_BUILTIN_LOADLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_PCV2SF },
21470 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_storehps, "__builtin_ia32_storehps", IX86_BUILTIN_STOREHPS, UNKNOWN, (int) VOID_FTYPE_PV2SF_V4SF },
21471 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_storelps, "__builtin_ia32_storelps", IX86_BUILTIN_STORELPS, UNKNOWN, (int) VOID_FTYPE_PV2SF_V4SF },
21473 /* SSE or 3DNow!A */
21474 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_sse_sfence, "__builtin_ia32_sfence", IX86_BUILTIN_SFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
21475 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_sse_movntdi, "__builtin_ia32_movntq", IX86_BUILTIN_MOVNTQ, UNKNOWN, (int) VOID_FTYPE_PULONGLONG_ULONGLONG },
21478 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_lfence, "__builtin_ia32_lfence", IX86_BUILTIN_LFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
21479 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_mfence, 0, IX86_BUILTIN_MFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
21480 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movupd, "__builtin_ia32_storeupd", IX86_BUILTIN_STOREUPD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
21481 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movdqu, "__builtin_ia32_storedqu", IX86_BUILTIN_STOREDQU, UNKNOWN, (int) VOID_FTYPE_PCHAR_V16QI },
21482 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntv2df, "__builtin_ia32_movntpd", IX86_BUILTIN_MOVNTPD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
21483 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntv2di, "__builtin_ia32_movntdq", IX86_BUILTIN_MOVNTDQ, UNKNOWN, (int) VOID_FTYPE_PV2DI_V2DI },
21484 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntsi, "__builtin_ia32_movnti", IX86_BUILTIN_MOVNTI, UNKNOWN, (int) VOID_FTYPE_PINT_INT },
21485 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movupd, "__builtin_ia32_loadupd", IX86_BUILTIN_LOADUPD, UNKNOWN, (int) V2DF_FTYPE_PCDOUBLE },
21486 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movdqu, "__builtin_ia32_loaddqu", IX86_BUILTIN_LOADDQU, UNKNOWN, (int) V16QI_FTYPE_PCCHAR },
21488 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_loadhpd_exp, "__builtin_ia32_loadhpd", IX86_BUILTIN_LOADHPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_PCDOUBLE },
21489 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_loadlpd_exp, "__builtin_ia32_loadlpd", IX86_BUILTIN_LOADLPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_PCDOUBLE },
21492 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_lddqu, "__builtin_ia32_lddqu", IX86_BUILTIN_LDDQU, UNKNOWN, (int) V16QI_FTYPE_PCCHAR },
21495 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_movntdqa, "__builtin_ia32_movntdqa", IX86_BUILTIN_MOVNTDQA, UNKNOWN, (int) V2DI_FTYPE_PV2DI },
21498 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_vmmovntv2df, "__builtin_ia32_movntsd", IX86_BUILTIN_MOVNTSD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
21499 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_vmmovntv4sf, "__builtin_ia32_movntss", IX86_BUILTIN_MOVNTSS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
21502 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vzeroall, "__builtin_ia32_vzeroall", IX86_BUILTIN_VZEROALL, UNKNOWN, (int) VOID_FTYPE_VOID },
21503 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vzeroupper, "__builtin_ia32_vzeroupper", IX86_BUILTIN_VZEROUPPER, UNKNOWN, (int) VOID_FTYPE_VOID },
21505 { OPTION_MASK_ISA_AVX, CODE_FOR_vec_dupv4sf, "__builtin_ia32_vbroadcastss", IX86_BUILTIN_VBROADCASTSS, UNKNOWN, (int) V4SF_FTYPE_PCFLOAT },
21506 { OPTION_MASK_ISA_AVX, CODE_FOR_vec_dupv4df, "__builtin_ia32_vbroadcastsd256", IX86_BUILTIN_VBROADCASTSD256, UNKNOWN, (int) V4DF_FTYPE_PCDOUBLE },
21507 { OPTION_MASK_ISA_AVX, CODE_FOR_vec_dupv8sf, "__builtin_ia32_vbroadcastss256", IX86_BUILTIN_VBROADCASTSS256, UNKNOWN, (int) V8SF_FTYPE_PCFLOAT },
21508 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastf128_v4df, "__builtin_ia32_vbroadcastf128_pd256", IX86_BUILTIN_VBROADCASTPD256, UNKNOWN, (int) V4DF_FTYPE_PCV2DF },
21509 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastf128_v8sf, "__builtin_ia32_vbroadcastf128_ps256", IX86_BUILTIN_VBROADCASTPS256, UNKNOWN, (int) V8SF_FTYPE_PCV4SF },
21511 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movupd256, "__builtin_ia32_loadupd256", IX86_BUILTIN_LOADUPD256, UNKNOWN, (int) V4DF_FTYPE_PCDOUBLE },
21512 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movups256, "__builtin_ia32_loadups256", IX86_BUILTIN_LOADUPS256, UNKNOWN, (int) V8SF_FTYPE_PCFLOAT },
21513 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movupd256, "__builtin_ia32_storeupd256", IX86_BUILTIN_STOREUPD256, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V4DF },
21514 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movups256, "__builtin_ia32_storeups256", IX86_BUILTIN_STOREUPS256, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V8SF },
21515 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movdqu256, "__builtin_ia32_loaddqu256", IX86_BUILTIN_LOADDQU256, UNKNOWN, (int) V32QI_FTYPE_PCCHAR },
21516 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movdqu256, "__builtin_ia32_storedqu256", IX86_BUILTIN_STOREDQU256, UNKNOWN, (int) VOID_FTYPE_PCHAR_V32QI },
21517 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_lddqu256, "__builtin_ia32_lddqu256", IX86_BUILTIN_LDDQU256, UNKNOWN, (int) V32QI_FTYPE_PCCHAR },
21519 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv4di, "__builtin_ia32_movntdq256", IX86_BUILTIN_MOVNTDQ256, UNKNOWN, (int) VOID_FTYPE_PV4DI_V4DI },
21520 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv4df, "__builtin_ia32_movntpd256", IX86_BUILTIN_MOVNTPD256, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V4DF },
21521 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv8sf, "__builtin_ia32_movntps256", IX86_BUILTIN_MOVNTPS256, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V8SF },
21523 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadpd, "__builtin_ia32_maskloadpd", IX86_BUILTIN_MASKLOADPD, UNKNOWN, (int) V2DF_FTYPE_PCV2DF_V2DF },
21524 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadps, "__builtin_ia32_maskloadps", IX86_BUILTIN_MASKLOADPS, UNKNOWN, (int) V4SF_FTYPE_PCV4SF_V4SF },
21525 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadpd256, "__builtin_ia32_maskloadpd256", IX86_BUILTIN_MASKLOADPD256, UNKNOWN, (int) V4DF_FTYPE_PCV4DF_V4DF },
21526 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadps256, "__builtin_ia32_maskloadps256", IX86_BUILTIN_MASKLOADPS256, UNKNOWN, (int) V8SF_FTYPE_PCV8SF_V8SF },
21527 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstorepd, "__builtin_ia32_maskstorepd", IX86_BUILTIN_MASKSTOREPD, UNKNOWN, (int) VOID_FTYPE_PV2DF_V2DF_V2DF },
21528 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstoreps, "__builtin_ia32_maskstoreps", IX86_BUILTIN_MASKSTOREPS, UNKNOWN, (int) VOID_FTYPE_PV4SF_V4SF_V4SF },
21529 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstorepd256, "__builtin_ia32_maskstorepd256", IX86_BUILTIN_MASKSTOREPD256, UNKNOWN, (int) VOID_FTYPE_PV4DF_V4DF_V4DF },
21530 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstoreps256, "__builtin_ia32_maskstoreps256", IX86_BUILTIN_MASKSTOREPS256, UNKNOWN, (int) VOID_FTYPE_PV8SF_V8SF_V8SF },
21532 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_llwpcb, "__builtin_ia32_llwpcb", IX86_BUILTIN_LLWPCB, UNKNOWN, (int) VOID_FTYPE_PVOID },
21533 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_slwpcb, "__builtin_ia32_slwpcb", IX86_BUILTIN_SLWPCB, UNKNOWN, (int) PVOID_FTYPE_VOID },
21534 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_lwpvalsi3, "__builtin_ia32_lwpval32", IX86_BUILTIN_LWPVAL32, UNKNOWN, (int) VOID_FTYPE_UINT_UINT_UINT },
21535 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_lwpvaldi3, "__builtin_ia32_lwpval64", IX86_BUILTIN_LWPVAL64, UNKNOWN, (int) VOID_FTYPE_UINT64_UINT_UINT },
21536 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_lwpinssi3, "__builtin_ia32_lwpins32", IX86_BUILTIN_LWPINS32, UNKNOWN, (int) UCHAR_FTYPE_UINT_UINT_UINT },
21537 { OPTION_MASK_ISA_LWP, CODE_FOR_lwp_lwpinsdi3, "__builtin_ia32_lwpins64", IX86_BUILTIN_LWPINS64, UNKNOWN, (int) UCHAR_FTYPE_UINT64_UINT_UINT },
21541 /* Builtins with variable number of arguments. */
21542 static const struct builtin_description bdesc_args[] =
21544 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_bsr, "__builtin_ia32_bsrsi", IX86_BUILTIN_BSRSI, UNKNOWN, (int) INT_FTYPE_INT },
21545 { OPTION_MASK_ISA_64BIT, CODE_FOR_bsr_rex64, "__builtin_ia32_bsrdi", IX86_BUILTIN_BSRDI, UNKNOWN, (int) INT64_FTYPE_INT64 },
21546 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rdpmc, "__builtin_ia32_rdpmc", IX86_BUILTIN_RDPMC, UNKNOWN, (int) UINT64_FTYPE_INT },
21547 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rotlqi3, "__builtin_ia32_rolqi", IX86_BUILTIN_ROLQI, UNKNOWN, (int) UINT8_FTYPE_UINT8_INT },
21548 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rotlhi3, "__builtin_ia32_rolhi", IX86_BUILTIN_ROLHI, UNKNOWN, (int) UINT16_FTYPE_UINT16_INT },
21549 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rotrqi3, "__builtin_ia32_rorqi", IX86_BUILTIN_RORQI, UNKNOWN, (int) UINT8_FTYPE_UINT8_INT },
21550 { ~OPTION_MASK_ISA_64BIT, CODE_FOR_rotrhi3, "__builtin_ia32_rorhi", IX86_BUILTIN_RORHI, UNKNOWN, (int) UINT16_FTYPE_UINT16_INT },
21553 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv8qi3, "__builtin_ia32_paddb", IX86_BUILTIN_PADDB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21554 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv4hi3, "__builtin_ia32_paddw", IX86_BUILTIN_PADDW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21555 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv2si3, "__builtin_ia32_paddd", IX86_BUILTIN_PADDD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21556 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv8qi3, "__builtin_ia32_psubb", IX86_BUILTIN_PSUBB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21557 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv4hi3, "__builtin_ia32_psubw", IX86_BUILTIN_PSUBW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21558 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv2si3, "__builtin_ia32_psubd", IX86_BUILTIN_PSUBD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21560 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ssaddv8qi3, "__builtin_ia32_paddsb", IX86_BUILTIN_PADDSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21561 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ssaddv4hi3, "__builtin_ia32_paddsw", IX86_BUILTIN_PADDSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21562 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_sssubv8qi3, "__builtin_ia32_psubsb", IX86_BUILTIN_PSUBSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21563 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_sssubv4hi3, "__builtin_ia32_psubsw", IX86_BUILTIN_PSUBSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21564 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_usaddv8qi3, "__builtin_ia32_paddusb", IX86_BUILTIN_PADDUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21565 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_usaddv4hi3, "__builtin_ia32_paddusw", IX86_BUILTIN_PADDUSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21566 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ussubv8qi3, "__builtin_ia32_psubusb", IX86_BUILTIN_PSUBUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21567 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ussubv4hi3, "__builtin_ia32_psubusw", IX86_BUILTIN_PSUBUSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21569 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_mulv4hi3, "__builtin_ia32_pmullw", IX86_BUILTIN_PMULLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21570 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_smulv4hi3_highpart, "__builtin_ia32_pmulhw", IX86_BUILTIN_PMULHW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21572 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_andv2si3, "__builtin_ia32_pand", IX86_BUILTIN_PAND, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21573 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_andnotv2si3, "__builtin_ia32_pandn", IX86_BUILTIN_PANDN, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21574 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_iorv2si3, "__builtin_ia32_por", IX86_BUILTIN_POR, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21575 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_xorv2si3, "__builtin_ia32_pxor", IX86_BUILTIN_PXOR, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21577 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv8qi3, "__builtin_ia32_pcmpeqb", IX86_BUILTIN_PCMPEQB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21578 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv4hi3, "__builtin_ia32_pcmpeqw", IX86_BUILTIN_PCMPEQW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21579 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv2si3, "__builtin_ia32_pcmpeqd", IX86_BUILTIN_PCMPEQD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21580 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv8qi3, "__builtin_ia32_pcmpgtb", IX86_BUILTIN_PCMPGTB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21581 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv4hi3, "__builtin_ia32_pcmpgtw", IX86_BUILTIN_PCMPGTW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21582 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv2si3, "__builtin_ia32_pcmpgtd", IX86_BUILTIN_PCMPGTD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21584 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhbw, "__builtin_ia32_punpckhbw", IX86_BUILTIN_PUNPCKHBW, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21585 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhwd, "__builtin_ia32_punpckhwd", IX86_BUILTIN_PUNPCKHWD, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21586 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhdq, "__builtin_ia32_punpckhdq", IX86_BUILTIN_PUNPCKHDQ, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21587 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpcklbw, "__builtin_ia32_punpcklbw", IX86_BUILTIN_PUNPCKLBW, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21588 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpcklwd, "__builtin_ia32_punpcklwd", IX86_BUILTIN_PUNPCKLWD, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI},
21589 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckldq, "__builtin_ia32_punpckldq", IX86_BUILTIN_PUNPCKLDQ, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI},
21591 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packsswb, "__builtin_ia32_packsswb", IX86_BUILTIN_PACKSSWB, UNKNOWN, (int) V8QI_FTYPE_V4HI_V4HI },
21592 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packssdw, "__builtin_ia32_packssdw", IX86_BUILTIN_PACKSSDW, UNKNOWN, (int) V4HI_FTYPE_V2SI_V2SI },
21593 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packuswb, "__builtin_ia32_packuswb", IX86_BUILTIN_PACKUSWB, UNKNOWN, (int) V8QI_FTYPE_V4HI_V4HI },
21595 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_pmaddwd, "__builtin_ia32_pmaddwd", IX86_BUILTIN_PMADDWD, UNKNOWN, (int) V2SI_FTYPE_V4HI_V4HI },
21597 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv4hi3, "__builtin_ia32_psllwi", IX86_BUILTIN_PSLLWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
21598 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv2si3, "__builtin_ia32_pslldi", IX86_BUILTIN_PSLLDI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
21599 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv1di3, "__builtin_ia32_psllqi", IX86_BUILTIN_PSLLQI, UNKNOWN, (int) V1DI_FTYPE_V1DI_SI_COUNT },
21600 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv4hi3, "__builtin_ia32_psllw", IX86_BUILTIN_PSLLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
21601 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv2si3, "__builtin_ia32_pslld", IX86_BUILTIN_PSLLD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
21602 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv1di3, "__builtin_ia32_psllq", IX86_BUILTIN_PSLLQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI_COUNT },
21604 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv4hi3, "__builtin_ia32_psrlwi", IX86_BUILTIN_PSRLWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
21605 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv2si3, "__builtin_ia32_psrldi", IX86_BUILTIN_PSRLDI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
21606 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv1di3, "__builtin_ia32_psrlqi", IX86_BUILTIN_PSRLQI, UNKNOWN, (int) V1DI_FTYPE_V1DI_SI_COUNT },
21607 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv4hi3, "__builtin_ia32_psrlw", IX86_BUILTIN_PSRLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
21608 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv2si3, "__builtin_ia32_psrld", IX86_BUILTIN_PSRLD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
21609 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv1di3, "__builtin_ia32_psrlq", IX86_BUILTIN_PSRLQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI_COUNT },
21611 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv4hi3, "__builtin_ia32_psrawi", IX86_BUILTIN_PSRAWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
21612 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv2si3, "__builtin_ia32_psradi", IX86_BUILTIN_PSRADI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
21613 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv4hi3, "__builtin_ia32_psraw", IX86_BUILTIN_PSRAW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
21614 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv2si3, "__builtin_ia32_psrad", IX86_BUILTIN_PSRAD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
21617 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_pf2id, "__builtin_ia32_pf2id", IX86_BUILTIN_PF2ID, UNKNOWN, (int) V2SI_FTYPE_V2SF },
21618 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_floatv2si2, "__builtin_ia32_pi2fd", IX86_BUILTIN_PI2FD, UNKNOWN, (int) V2SF_FTYPE_V2SI },
21619 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpv2sf2, "__builtin_ia32_pfrcp", IX86_BUILTIN_PFRCP, UNKNOWN, (int) V2SF_FTYPE_V2SF },
21620 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rsqrtv2sf2, "__builtin_ia32_pfrsqrt", IX86_BUILTIN_PFRSQRT, UNKNOWN, (int) V2SF_FTYPE_V2SF },
21622 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_uavgv8qi3, "__builtin_ia32_pavgusb", IX86_BUILTIN_PAVGUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21623 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_haddv2sf3, "__builtin_ia32_pfacc", IX86_BUILTIN_PFACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21624 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_addv2sf3, "__builtin_ia32_pfadd", IX86_BUILTIN_PFADD, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21625 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_eqv2sf3, "__builtin_ia32_pfcmpeq", IX86_BUILTIN_PFCMPEQ, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
21626 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_gev2sf3, "__builtin_ia32_pfcmpge", IX86_BUILTIN_PFCMPGE, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
21627 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_gtv2sf3, "__builtin_ia32_pfcmpgt", IX86_BUILTIN_PFCMPGT, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
21628 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_smaxv2sf3, "__builtin_ia32_pfmax", IX86_BUILTIN_PFMAX, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21629 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_sminv2sf3, "__builtin_ia32_pfmin", IX86_BUILTIN_PFMIN, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21630 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_mulv2sf3, "__builtin_ia32_pfmul", IX86_BUILTIN_PFMUL, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21631 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpit1v2sf3, "__builtin_ia32_pfrcpit1", IX86_BUILTIN_PFRCPIT1, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21632 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpit2v2sf3, "__builtin_ia32_pfrcpit2", IX86_BUILTIN_PFRCPIT2, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21633 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rsqit1v2sf3, "__builtin_ia32_pfrsqit1", IX86_BUILTIN_PFRSQIT1, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21634 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_subv2sf3, "__builtin_ia32_pfsub", IX86_BUILTIN_PFSUB, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21635 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_subrv2sf3, "__builtin_ia32_pfsubr", IX86_BUILTIN_PFSUBR, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21636 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_pmulhrwv4hi3, "__builtin_ia32_pmulhrw", IX86_BUILTIN_PMULHRW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21639 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pf2iw, "__builtin_ia32_pf2iw", IX86_BUILTIN_PF2IW, UNKNOWN, (int) V2SI_FTYPE_V2SF },
21640 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pi2fw, "__builtin_ia32_pi2fw", IX86_BUILTIN_PI2FW, UNKNOWN, (int) V2SF_FTYPE_V2SI },
21641 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pswapdv2si2, "__builtin_ia32_pswapdsi", IX86_BUILTIN_PSWAPDSI, UNKNOWN, (int) V2SI_FTYPE_V2SI },
21642 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pswapdv2sf2, "__builtin_ia32_pswapdsf", IX86_BUILTIN_PSWAPDSF, UNKNOWN, (int) V2SF_FTYPE_V2SF },
21643 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_hsubv2sf3, "__builtin_ia32_pfnacc", IX86_BUILTIN_PFNACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21644 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_addsubv2sf3, "__builtin_ia32_pfpnacc", IX86_BUILTIN_PFPNACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
21647 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movmskps, "__builtin_ia32_movmskps", IX86_BUILTIN_MOVMSKPS, UNKNOWN, (int) INT_FTYPE_V4SF },
21648 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_sqrtv4sf2, "__builtin_ia32_sqrtps", IX86_BUILTIN_SQRTPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21649 { OPTION_MASK_ISA_SSE, CODE_FOR_sqrtv4sf2, "__builtin_ia32_sqrtps_nr", IX86_BUILTIN_SQRTPS_NR, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21650 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_rsqrtv4sf2, "__builtin_ia32_rsqrtps", IX86_BUILTIN_RSQRTPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21651 { OPTION_MASK_ISA_SSE, CODE_FOR_rsqrtv4sf2, "__builtin_ia32_rsqrtps_nr", IX86_BUILTIN_RSQRTPS_NR, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21652 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_rcpv4sf2, "__builtin_ia32_rcpps", IX86_BUILTIN_RCPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21653 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtps2pi, "__builtin_ia32_cvtps2pi", IX86_BUILTIN_CVTPS2PI, UNKNOWN, (int) V2SI_FTYPE_V4SF },
21654 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtss2si, "__builtin_ia32_cvtss2si", IX86_BUILTIN_CVTSS2SI, UNKNOWN, (int) INT_FTYPE_V4SF },
21655 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvtss2siq, "__builtin_ia32_cvtss2si64", IX86_BUILTIN_CVTSS2SI64, UNKNOWN, (int) INT64_FTYPE_V4SF },
21656 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvttps2pi, "__builtin_ia32_cvttps2pi", IX86_BUILTIN_CVTTPS2PI, UNKNOWN, (int) V2SI_FTYPE_V4SF },
21657 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvttss2si, "__builtin_ia32_cvttss2si", IX86_BUILTIN_CVTTSS2SI, UNKNOWN, (int) INT_FTYPE_V4SF },
21658 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvttss2siq, "__builtin_ia32_cvttss2si64", IX86_BUILTIN_CVTTSS2SI64, UNKNOWN, (int) INT64_FTYPE_V4SF },
21660 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_shufps, "__builtin_ia32_shufps", IX86_BUILTIN_SHUFPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21662 { OPTION_MASK_ISA_SSE, CODE_FOR_addv4sf3, "__builtin_ia32_addps", IX86_BUILTIN_ADDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21663 { OPTION_MASK_ISA_SSE, CODE_FOR_subv4sf3, "__builtin_ia32_subps", IX86_BUILTIN_SUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21664 { OPTION_MASK_ISA_SSE, CODE_FOR_mulv4sf3, "__builtin_ia32_mulps", IX86_BUILTIN_MULPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21665 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_divv4sf3, "__builtin_ia32_divps", IX86_BUILTIN_DIVPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21666 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmaddv4sf3, "__builtin_ia32_addss", IX86_BUILTIN_ADDSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21667 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsubv4sf3, "__builtin_ia32_subss", IX86_BUILTIN_SUBSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21668 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmulv4sf3, "__builtin_ia32_mulss", IX86_BUILTIN_MULSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21669 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmdivv4sf3, "__builtin_ia32_divss", IX86_BUILTIN_DIVSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21671 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpeqps", IX86_BUILTIN_CMPEQPS, EQ, (int) V4SF_FTYPE_V4SF_V4SF },
21672 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpltps", IX86_BUILTIN_CMPLTPS, LT, (int) V4SF_FTYPE_V4SF_V4SF },
21673 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpleps", IX86_BUILTIN_CMPLEPS, LE, (int) V4SF_FTYPE_V4SF_V4SF },
21674 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpgtps", IX86_BUILTIN_CMPGTPS, LT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
21675 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpgeps", IX86_BUILTIN_CMPGEPS, LE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
21676 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpunordps", IX86_BUILTIN_CMPUNORDPS, UNORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
21677 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpneqps", IX86_BUILTIN_CMPNEQPS, NE, (int) V4SF_FTYPE_V4SF_V4SF },
21678 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpnltps", IX86_BUILTIN_CMPNLTPS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF },
21679 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpnleps", IX86_BUILTIN_CMPNLEPS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF },
21680 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpngtps", IX86_BUILTIN_CMPNGTPS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
21681 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpngeps", IX86_BUILTIN_CMPNGEPS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP},
21682 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpordps", IX86_BUILTIN_CMPORDPS, ORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
21683 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpeqss", IX86_BUILTIN_CMPEQSS, EQ, (int) V4SF_FTYPE_V4SF_V4SF },
21684 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpltss", IX86_BUILTIN_CMPLTSS, LT, (int) V4SF_FTYPE_V4SF_V4SF },
21685 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpless", IX86_BUILTIN_CMPLESS, LE, (int) V4SF_FTYPE_V4SF_V4SF },
21686 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpunordss", IX86_BUILTIN_CMPUNORDSS, UNORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
21687 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpneqss", IX86_BUILTIN_CMPNEQSS, NE, (int) V4SF_FTYPE_V4SF_V4SF },
21688 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpnltss", IX86_BUILTIN_CMPNLTSS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF },
21689 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpnless", IX86_BUILTIN_CMPNLESS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF },
21690 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpngtss", IX86_BUILTIN_CMPNGTSS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
21691 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpngess", IX86_BUILTIN_CMPNGESS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
21692 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpordss", IX86_BUILTIN_CMPORDSS, ORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
21694 { OPTION_MASK_ISA_SSE, CODE_FOR_sminv4sf3, "__builtin_ia32_minps", IX86_BUILTIN_MINPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21695 { OPTION_MASK_ISA_SSE, CODE_FOR_smaxv4sf3, "__builtin_ia32_maxps", IX86_BUILTIN_MAXPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21696 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsminv4sf3, "__builtin_ia32_minss", IX86_BUILTIN_MINSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21697 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsmaxv4sf3, "__builtin_ia32_maxss", IX86_BUILTIN_MAXSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21699 { OPTION_MASK_ISA_SSE, CODE_FOR_andv4sf3, "__builtin_ia32_andps", IX86_BUILTIN_ANDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21700 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_andnotv4sf3, "__builtin_ia32_andnps", IX86_BUILTIN_ANDNPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21701 { OPTION_MASK_ISA_SSE, CODE_FOR_iorv4sf3, "__builtin_ia32_orps", IX86_BUILTIN_ORPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21702 { OPTION_MASK_ISA_SSE, CODE_FOR_xorv4sf3, "__builtin_ia32_xorps", IX86_BUILTIN_XORPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21704 { OPTION_MASK_ISA_SSE, CODE_FOR_copysignv4sf3, "__builtin_ia32_copysignps", IX86_BUILTIN_CPYSGNPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21706 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movss, "__builtin_ia32_movss", IX86_BUILTIN_MOVSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21707 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movhlps_exp, "__builtin_ia32_movhlps", IX86_BUILTIN_MOVHLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21708 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movlhps_exp, "__builtin_ia32_movlhps", IX86_BUILTIN_MOVLHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21709 { OPTION_MASK_ISA_SSE, CODE_FOR_vec_interleave_highv4sf, "__builtin_ia32_unpckhps", IX86_BUILTIN_UNPCKHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21710 { OPTION_MASK_ISA_SSE, CODE_FOR_vec_interleave_lowv4sf, "__builtin_ia32_unpcklps", IX86_BUILTIN_UNPCKLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21712 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtpi2ps, "__builtin_ia32_cvtpi2ps", IX86_BUILTIN_CVTPI2PS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V2SI },
21713 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtsi2ss, "__builtin_ia32_cvtsi2ss", IX86_BUILTIN_CVTSI2SS, UNKNOWN, (int) V4SF_FTYPE_V4SF_SI },
21714 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvtsi2ssq, "__builtin_ia32_cvtsi642ss", IX86_BUILTIN_CVTSI642SS, UNKNOWN, V4SF_FTYPE_V4SF_DI },
21716 { OPTION_MASK_ISA_SSE, CODE_FOR_rsqrtsf2, "__builtin_ia32_rsqrtf", IX86_BUILTIN_RSQRTF, UNKNOWN, (int) FLOAT_FTYPE_FLOAT },
21718 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsqrtv4sf2, "__builtin_ia32_sqrtss", IX86_BUILTIN_SQRTSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
21719 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmrsqrtv4sf2, "__builtin_ia32_rsqrtss", IX86_BUILTIN_RSQRTSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
21720 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmrcpv4sf2, "__builtin_ia32_rcpss", IX86_BUILTIN_RCPSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
21722 /* SSE MMX or 3Dnow!A */
21723 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_uavgv8qi3, "__builtin_ia32_pavgb", IX86_BUILTIN_PAVGB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21724 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_uavgv4hi3, "__builtin_ia32_pavgw", IX86_BUILTIN_PAVGW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21725 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_umulv4hi3_highpart, "__builtin_ia32_pmulhuw", IX86_BUILTIN_PMULHUW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21727 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_umaxv8qi3, "__builtin_ia32_pmaxub", IX86_BUILTIN_PMAXUB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21728 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_smaxv4hi3, "__builtin_ia32_pmaxsw", IX86_BUILTIN_PMAXSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21729 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_uminv8qi3, "__builtin_ia32_pminub", IX86_BUILTIN_PMINUB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21730 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_sminv4hi3, "__builtin_ia32_pminsw", IX86_BUILTIN_PMINSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21732 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_psadbw, "__builtin_ia32_psadbw", IX86_BUILTIN_PSADBW, UNKNOWN, (int) V1DI_FTYPE_V8QI_V8QI },
21733 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pmovmskb, "__builtin_ia32_pmovmskb", IX86_BUILTIN_PMOVMSKB, UNKNOWN, (int) INT_FTYPE_V8QI },
21735 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pshufw, "__builtin_ia32_pshufw", IX86_BUILTIN_PSHUFW, UNKNOWN, (int) V4HI_FTYPE_V4HI_INT },
21738 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_shufpd, "__builtin_ia32_shufpd", IX86_BUILTIN_SHUFPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21740 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v2df", IX86_BUILTIN_VEC_PERM_V2DF, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_V2DI },
21741 { OPTION_MASK_ISA_SSE, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v4sf", IX86_BUILTIN_VEC_PERM_V4SF, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_V4SI },
21742 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v2di", IX86_BUILTIN_VEC_PERM_V2DI, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_V2DI },
21743 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v4si", IX86_BUILTIN_VEC_PERM_V4SI, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_V4SI },
21744 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v8hi", IX86_BUILTIN_VEC_PERM_V8HI, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_V8HI },
21745 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v16qi", IX86_BUILTIN_VEC_PERM_V16QI, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI_V16QI },
21746 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v2di_u", IX86_BUILTIN_VEC_PERM_V2DI_U, UNKNOWN, (int) V2UDI_FTYPE_V2UDI_V2UDI_V2UDI },
21747 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v4si_u", IX86_BUILTIN_VEC_PERM_V4SI_U, UNKNOWN, (int) V4USI_FTYPE_V4USI_V4USI_V4USI },
21748 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v8hi_u", IX86_BUILTIN_VEC_PERM_V8HI_U, UNKNOWN, (int) V8UHI_FTYPE_V8UHI_V8UHI_V8UHI },
21749 { OPTION_MASK_ISA_SSE2, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v16qi_u", IX86_BUILTIN_VEC_PERM_V16QI_U, UNKNOWN, (int) V16UQI_FTYPE_V16UQI_V16UQI_V16UQI },
21750 { OPTION_MASK_ISA_AVX, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v4df", IX86_BUILTIN_VEC_PERM_V4DF, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_V4DI },
21751 { OPTION_MASK_ISA_AVX, CODE_FOR_nothing, "__builtin_ia32_vec_perm_v8sf", IX86_BUILTIN_VEC_PERM_V8SF, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_V8SI },
21753 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movmskpd, "__builtin_ia32_movmskpd", IX86_BUILTIN_MOVMSKPD, UNKNOWN, (int) INT_FTYPE_V2DF },
21754 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pmovmskb, "__builtin_ia32_pmovmskb128", IX86_BUILTIN_PMOVMSKB128, UNKNOWN, (int) INT_FTYPE_V16QI },
21755 { OPTION_MASK_ISA_SSE2, CODE_FOR_sqrtv2df2, "__builtin_ia32_sqrtpd", IX86_BUILTIN_SQRTPD, UNKNOWN, (int) V2DF_FTYPE_V2DF },
21756 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtdq2pd, "__builtin_ia32_cvtdq2pd", IX86_BUILTIN_CVTDQ2PD, UNKNOWN, (int) V2DF_FTYPE_V4SI },
21757 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtdq2ps, "__builtin_ia32_cvtdq2ps", IX86_BUILTIN_CVTDQ2PS, UNKNOWN, (int) V4SF_FTYPE_V4SI },
21758 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtudq2ps, "__builtin_ia32_cvtudq2ps", IX86_BUILTIN_CVTUDQ2PS, UNKNOWN, (int) V4SF_FTYPE_V4SI },
21760 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2dq, "__builtin_ia32_cvtpd2dq", IX86_BUILTIN_CVTPD2DQ, UNKNOWN, (int) V4SI_FTYPE_V2DF },
21761 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2pi, "__builtin_ia32_cvtpd2pi", IX86_BUILTIN_CVTPD2PI, UNKNOWN, (int) V2SI_FTYPE_V2DF },
21762 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2ps, "__builtin_ia32_cvtpd2ps", IX86_BUILTIN_CVTPD2PS, UNKNOWN, (int) V4SF_FTYPE_V2DF },
21763 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttpd2dq, "__builtin_ia32_cvttpd2dq", IX86_BUILTIN_CVTTPD2DQ, UNKNOWN, (int) V4SI_FTYPE_V2DF },
21764 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttpd2pi, "__builtin_ia32_cvttpd2pi", IX86_BUILTIN_CVTTPD2PI, UNKNOWN, (int) V2SI_FTYPE_V2DF },
21766 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpi2pd, "__builtin_ia32_cvtpi2pd", IX86_BUILTIN_CVTPI2PD, UNKNOWN, (int) V2DF_FTYPE_V2SI },
21768 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsd2si, "__builtin_ia32_cvtsd2si", IX86_BUILTIN_CVTSD2SI, UNKNOWN, (int) INT_FTYPE_V2DF },
21769 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttsd2si, "__builtin_ia32_cvttsd2si", IX86_BUILTIN_CVTTSD2SI, UNKNOWN, (int) INT_FTYPE_V2DF },
21770 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvtsd2siq, "__builtin_ia32_cvtsd2si64", IX86_BUILTIN_CVTSD2SI64, UNKNOWN, (int) INT64_FTYPE_V2DF },
21771 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvttsd2siq, "__builtin_ia32_cvttsd2si64", IX86_BUILTIN_CVTTSD2SI64, UNKNOWN, (int) INT64_FTYPE_V2DF },
21773 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtps2dq, "__builtin_ia32_cvtps2dq", IX86_BUILTIN_CVTPS2DQ, UNKNOWN, (int) V4SI_FTYPE_V4SF },
21774 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtps2pd, "__builtin_ia32_cvtps2pd", IX86_BUILTIN_CVTPS2PD, UNKNOWN, (int) V2DF_FTYPE_V4SF },
21775 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttps2dq, "__builtin_ia32_cvttps2dq", IX86_BUILTIN_CVTTPS2DQ, UNKNOWN, (int) V4SI_FTYPE_V4SF },
21777 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv2df3, "__builtin_ia32_addpd", IX86_BUILTIN_ADDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21778 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv2df3, "__builtin_ia32_subpd", IX86_BUILTIN_SUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21779 { OPTION_MASK_ISA_SSE2, CODE_FOR_mulv2df3, "__builtin_ia32_mulpd", IX86_BUILTIN_MULPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21780 { OPTION_MASK_ISA_SSE2, CODE_FOR_divv2df3, "__builtin_ia32_divpd", IX86_BUILTIN_DIVPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21781 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmaddv2df3, "__builtin_ia32_addsd", IX86_BUILTIN_ADDSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21782 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsubv2df3, "__builtin_ia32_subsd", IX86_BUILTIN_SUBSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21783 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmulv2df3, "__builtin_ia32_mulsd", IX86_BUILTIN_MULSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21784 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmdivv2df3, "__builtin_ia32_divsd", IX86_BUILTIN_DIVSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21786 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpeqpd", IX86_BUILTIN_CMPEQPD, EQ, (int) V2DF_FTYPE_V2DF_V2DF },
21787 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpltpd", IX86_BUILTIN_CMPLTPD, LT, (int) V2DF_FTYPE_V2DF_V2DF },
21788 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmplepd", IX86_BUILTIN_CMPLEPD, LE, (int) V2DF_FTYPE_V2DF_V2DF },
21789 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpgtpd", IX86_BUILTIN_CMPGTPD, LT, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
21790 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpgepd", IX86_BUILTIN_CMPGEPD, LE, (int) V2DF_FTYPE_V2DF_V2DF_SWAP},
21791 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpunordpd", IX86_BUILTIN_CMPUNORDPD, UNORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
21792 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpneqpd", IX86_BUILTIN_CMPNEQPD, NE, (int) V2DF_FTYPE_V2DF_V2DF },
21793 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpnltpd", IX86_BUILTIN_CMPNLTPD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF },
21794 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpnlepd", IX86_BUILTIN_CMPNLEPD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF },
21795 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpngtpd", IX86_BUILTIN_CMPNGTPD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
21796 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpngepd", IX86_BUILTIN_CMPNGEPD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
21797 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpordpd", IX86_BUILTIN_CMPORDPD, ORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
21798 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpeqsd", IX86_BUILTIN_CMPEQSD, EQ, (int) V2DF_FTYPE_V2DF_V2DF },
21799 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpltsd", IX86_BUILTIN_CMPLTSD, LT, (int) V2DF_FTYPE_V2DF_V2DF },
21800 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmplesd", IX86_BUILTIN_CMPLESD, LE, (int) V2DF_FTYPE_V2DF_V2DF },
21801 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpunordsd", IX86_BUILTIN_CMPUNORDSD, UNORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
21802 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpneqsd", IX86_BUILTIN_CMPNEQSD, NE, (int) V2DF_FTYPE_V2DF_V2DF },
21803 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpnltsd", IX86_BUILTIN_CMPNLTSD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF },
21804 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpnlesd", IX86_BUILTIN_CMPNLESD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF },
21805 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpordsd", IX86_BUILTIN_CMPORDSD, ORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
21807 { OPTION_MASK_ISA_SSE2, CODE_FOR_sminv2df3, "__builtin_ia32_minpd", IX86_BUILTIN_MINPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21808 { OPTION_MASK_ISA_SSE2, CODE_FOR_smaxv2df3, "__builtin_ia32_maxpd", IX86_BUILTIN_MAXPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21809 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsminv2df3, "__builtin_ia32_minsd", IX86_BUILTIN_MINSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21810 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsmaxv2df3, "__builtin_ia32_maxsd", IX86_BUILTIN_MAXSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21812 { OPTION_MASK_ISA_SSE2, CODE_FOR_andv2df3, "__builtin_ia32_andpd", IX86_BUILTIN_ANDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21813 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_andnotv2df3, "__builtin_ia32_andnpd", IX86_BUILTIN_ANDNPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21814 { OPTION_MASK_ISA_SSE2, CODE_FOR_iorv2df3, "__builtin_ia32_orpd", IX86_BUILTIN_ORPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21815 { OPTION_MASK_ISA_SSE2, CODE_FOR_xorv2df3, "__builtin_ia32_xorpd", IX86_BUILTIN_XORPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21817 { OPTION_MASK_ISA_SSE2, CODE_FOR_copysignv2df3, "__builtin_ia32_copysignpd", IX86_BUILTIN_CPYSGNPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21819 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movsd, "__builtin_ia32_movsd", IX86_BUILTIN_MOVSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21820 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_highv2df, "__builtin_ia32_unpckhpd", IX86_BUILTIN_UNPCKHPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21821 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_lowv2df, "__builtin_ia32_unpcklpd", IX86_BUILTIN_UNPCKLPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21823 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_pack_sfix_v2df, "__builtin_ia32_vec_pack_sfix", IX86_BUILTIN_VEC_PACK_SFIX, UNKNOWN, (int) V4SI_FTYPE_V2DF_V2DF },
21825 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv16qi3, "__builtin_ia32_paddb128", IX86_BUILTIN_PADDB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21826 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv8hi3, "__builtin_ia32_paddw128", IX86_BUILTIN_PADDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21827 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv4si3, "__builtin_ia32_paddd128", IX86_BUILTIN_PADDD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21828 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv2di3, "__builtin_ia32_paddq128", IX86_BUILTIN_PADDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21829 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv16qi3, "__builtin_ia32_psubb128", IX86_BUILTIN_PSUBB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21830 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv8hi3, "__builtin_ia32_psubw128", IX86_BUILTIN_PSUBW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21831 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv4si3, "__builtin_ia32_psubd128", IX86_BUILTIN_PSUBD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21832 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv2di3, "__builtin_ia32_psubq128", IX86_BUILTIN_PSUBQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21834 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ssaddv16qi3, "__builtin_ia32_paddsb128", IX86_BUILTIN_PADDSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21835 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ssaddv8hi3, "__builtin_ia32_paddsw128", IX86_BUILTIN_PADDSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21836 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_sssubv16qi3, "__builtin_ia32_psubsb128", IX86_BUILTIN_PSUBSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21837 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_sssubv8hi3, "__builtin_ia32_psubsw128", IX86_BUILTIN_PSUBSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21838 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_usaddv16qi3, "__builtin_ia32_paddusb128", IX86_BUILTIN_PADDUSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21839 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_usaddv8hi3, "__builtin_ia32_paddusw128", IX86_BUILTIN_PADDUSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21840 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ussubv16qi3, "__builtin_ia32_psubusb128", IX86_BUILTIN_PSUBUSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21841 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ussubv8hi3, "__builtin_ia32_psubusw128", IX86_BUILTIN_PSUBUSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21843 { OPTION_MASK_ISA_SSE2, CODE_FOR_mulv8hi3, "__builtin_ia32_pmullw128", IX86_BUILTIN_PMULLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21844 { OPTION_MASK_ISA_SSE2, CODE_FOR_smulv8hi3_highpart, "__builtin_ia32_pmulhw128", IX86_BUILTIN_PMULHW128, UNKNOWN,(int) V8HI_FTYPE_V8HI_V8HI },
21846 { OPTION_MASK_ISA_SSE2, CODE_FOR_andv2di3, "__builtin_ia32_pand128", IX86_BUILTIN_PAND128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21847 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_andnotv2di3, "__builtin_ia32_pandn128", IX86_BUILTIN_PANDN128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21848 { OPTION_MASK_ISA_SSE2, CODE_FOR_iorv2di3, "__builtin_ia32_por128", IX86_BUILTIN_POR128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21849 { OPTION_MASK_ISA_SSE2, CODE_FOR_xorv2di3, "__builtin_ia32_pxor128", IX86_BUILTIN_PXOR128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21851 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_uavgv16qi3, "__builtin_ia32_pavgb128", IX86_BUILTIN_PAVGB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21852 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_uavgv8hi3, "__builtin_ia32_pavgw128", IX86_BUILTIN_PAVGW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21854 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv16qi3, "__builtin_ia32_pcmpeqb128", IX86_BUILTIN_PCMPEQB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21855 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv8hi3, "__builtin_ia32_pcmpeqw128", IX86_BUILTIN_PCMPEQW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21856 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv4si3, "__builtin_ia32_pcmpeqd128", IX86_BUILTIN_PCMPEQD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21857 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv16qi3, "__builtin_ia32_pcmpgtb128", IX86_BUILTIN_PCMPGTB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21858 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv8hi3, "__builtin_ia32_pcmpgtw128", IX86_BUILTIN_PCMPGTW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21859 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv4si3, "__builtin_ia32_pcmpgtd128", IX86_BUILTIN_PCMPGTD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21861 { OPTION_MASK_ISA_SSE2, CODE_FOR_umaxv16qi3, "__builtin_ia32_pmaxub128", IX86_BUILTIN_PMAXUB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21862 { OPTION_MASK_ISA_SSE2, CODE_FOR_smaxv8hi3, "__builtin_ia32_pmaxsw128", IX86_BUILTIN_PMAXSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21863 { OPTION_MASK_ISA_SSE2, CODE_FOR_uminv16qi3, "__builtin_ia32_pminub128", IX86_BUILTIN_PMINUB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21864 { OPTION_MASK_ISA_SSE2, CODE_FOR_sminv8hi3, "__builtin_ia32_pminsw128", IX86_BUILTIN_PMINSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21866 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_highv16qi, "__builtin_ia32_punpckhbw128", IX86_BUILTIN_PUNPCKHBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21867 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_highv8hi, "__builtin_ia32_punpckhwd128", IX86_BUILTIN_PUNPCKHWD128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21868 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_highv4si, "__builtin_ia32_punpckhdq128", IX86_BUILTIN_PUNPCKHDQ128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21869 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_highv2di, "__builtin_ia32_punpckhqdq128", IX86_BUILTIN_PUNPCKHQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21870 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_lowv16qi, "__builtin_ia32_punpcklbw128", IX86_BUILTIN_PUNPCKLBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21871 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_lowv8hi, "__builtin_ia32_punpcklwd128", IX86_BUILTIN_PUNPCKLWD128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21872 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_lowv4si, "__builtin_ia32_punpckldq128", IX86_BUILTIN_PUNPCKLDQ128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21873 { OPTION_MASK_ISA_SSE2, CODE_FOR_vec_interleave_lowv2di, "__builtin_ia32_punpcklqdq128", IX86_BUILTIN_PUNPCKLQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21875 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packsswb, "__builtin_ia32_packsswb128", IX86_BUILTIN_PACKSSWB128, UNKNOWN, (int) V16QI_FTYPE_V8HI_V8HI },
21876 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packssdw, "__builtin_ia32_packssdw128", IX86_BUILTIN_PACKSSDW128, UNKNOWN, (int) V8HI_FTYPE_V4SI_V4SI },
21877 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packuswb, "__builtin_ia32_packuswb128", IX86_BUILTIN_PACKUSWB128, UNKNOWN, (int) V16QI_FTYPE_V8HI_V8HI },
21879 { OPTION_MASK_ISA_SSE2, CODE_FOR_umulv8hi3_highpart, "__builtin_ia32_pmulhuw128", IX86_BUILTIN_PMULHUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21880 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_psadbw, "__builtin_ia32_psadbw128", IX86_BUILTIN_PSADBW128, UNKNOWN, (int) V2DI_FTYPE_V16QI_V16QI },
21882 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_umulv1siv1di3, "__builtin_ia32_pmuludq", IX86_BUILTIN_PMULUDQ, UNKNOWN, (int) V1DI_FTYPE_V2SI_V2SI },
21883 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_umulv2siv2di3, "__builtin_ia32_pmuludq128", IX86_BUILTIN_PMULUDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI_V4SI },
21885 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pmaddwd, "__builtin_ia32_pmaddwd128", IX86_BUILTIN_PMADDWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI_V8HI },
21887 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsi2sd, "__builtin_ia32_cvtsi2sd", IX86_BUILTIN_CVTSI2SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_SI },
21888 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvtsi2sdq, "__builtin_ia32_cvtsi642sd", IX86_BUILTIN_CVTSI642SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_DI },
21889 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsd2ss, "__builtin_ia32_cvtsd2ss", IX86_BUILTIN_CVTSD2SS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V2DF },
21890 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtss2sd, "__builtin_ia32_cvtss2sd", IX86_BUILTIN_CVTSS2SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V4SF },
21892 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ashlv1ti3, "__builtin_ia32_pslldqi128", IX86_BUILTIN_PSLLDQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_INT_CONVERT },
21893 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv8hi3, "__builtin_ia32_psllwi128", IX86_BUILTIN_PSLLWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21894 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv4si3, "__builtin_ia32_pslldi128", IX86_BUILTIN_PSLLDI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21895 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv2di3, "__builtin_ia32_psllqi128", IX86_BUILTIN_PSLLQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_SI_COUNT },
21896 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv8hi3, "__builtin_ia32_psllw128", IX86_BUILTIN_PSLLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21897 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv4si3, "__builtin_ia32_pslld128", IX86_BUILTIN_PSLLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21898 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv2di3, "__builtin_ia32_psllq128", IX86_BUILTIN_PSLLQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_COUNT },
21900 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_lshrv1ti3, "__builtin_ia32_psrldqi128", IX86_BUILTIN_PSRLDQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_INT_CONVERT },
21901 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv8hi3, "__builtin_ia32_psrlwi128", IX86_BUILTIN_PSRLWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21902 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv4si3, "__builtin_ia32_psrldi128", IX86_BUILTIN_PSRLDI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21903 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv2di3, "__builtin_ia32_psrlqi128", IX86_BUILTIN_PSRLQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_SI_COUNT },
21904 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv8hi3, "__builtin_ia32_psrlw128", IX86_BUILTIN_PSRLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21905 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv4si3, "__builtin_ia32_psrld128", IX86_BUILTIN_PSRLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21906 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv2di3, "__builtin_ia32_psrlq128", IX86_BUILTIN_PSRLQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_COUNT },
21908 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv8hi3, "__builtin_ia32_psrawi128", IX86_BUILTIN_PSRAWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21909 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv4si3, "__builtin_ia32_psradi128", IX86_BUILTIN_PSRADI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21910 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv8hi3, "__builtin_ia32_psraw128", IX86_BUILTIN_PSRAW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21911 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv4si3, "__builtin_ia32_psrad128", IX86_BUILTIN_PSRAD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21913 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshufd, "__builtin_ia32_pshufd", IX86_BUILTIN_PSHUFD, UNKNOWN, (int) V4SI_FTYPE_V4SI_INT },
21914 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshuflw, "__builtin_ia32_pshuflw", IX86_BUILTIN_PSHUFLW, UNKNOWN, (int) V8HI_FTYPE_V8HI_INT },
21915 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshufhw, "__builtin_ia32_pshufhw", IX86_BUILTIN_PSHUFHW, UNKNOWN, (int) V8HI_FTYPE_V8HI_INT },
21917 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsqrtv2df2, "__builtin_ia32_sqrtsd", IX86_BUILTIN_SQRTSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_VEC_MERGE },
21919 { OPTION_MASK_ISA_SSE2, CODE_FOR_abstf2, 0, IX86_BUILTIN_FABSQ, UNKNOWN, (int) FLOAT128_FTYPE_FLOAT128 },
21920 { OPTION_MASK_ISA_SSE2, CODE_FOR_copysigntf3, 0, IX86_BUILTIN_COPYSIGNQ, UNKNOWN, (int) FLOAT128_FTYPE_FLOAT128_FLOAT128 },
21922 { OPTION_MASK_ISA_SSE, CODE_FOR_sse2_movq128, "__builtin_ia32_movq128", IX86_BUILTIN_MOVQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI },
21925 { OPTION_MASK_ISA_SSE2, CODE_FOR_mmx_addv1di3, "__builtin_ia32_paddq", IX86_BUILTIN_PADDQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI },
21926 { OPTION_MASK_ISA_SSE2, CODE_FOR_mmx_subv1di3, "__builtin_ia32_psubq", IX86_BUILTIN_PSUBQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI },
21929 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_movshdup, "__builtin_ia32_movshdup", IX86_BUILTIN_MOVSHDUP, UNKNOWN, (int) V4SF_FTYPE_V4SF},
21930 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_movsldup, "__builtin_ia32_movsldup", IX86_BUILTIN_MOVSLDUP, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21932 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_addsubv4sf3, "__builtin_ia32_addsubps", IX86_BUILTIN_ADDSUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21933 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_addsubv2df3, "__builtin_ia32_addsubpd", IX86_BUILTIN_ADDSUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21934 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_haddv4sf3, "__builtin_ia32_haddps", IX86_BUILTIN_HADDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21935 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_haddv2df3, "__builtin_ia32_haddpd", IX86_BUILTIN_HADDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21936 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_hsubv4sf3, "__builtin_ia32_hsubps", IX86_BUILTIN_HSUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21937 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_hsubv2df3, "__builtin_ia32_hsubpd", IX86_BUILTIN_HSUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21940 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv16qi2, "__builtin_ia32_pabsb128", IX86_BUILTIN_PABSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI },
21941 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv8qi2, "__builtin_ia32_pabsb", IX86_BUILTIN_PABSB, UNKNOWN, (int) V8QI_FTYPE_V8QI },
21942 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv8hi2, "__builtin_ia32_pabsw128", IX86_BUILTIN_PABSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI },
21943 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv4hi2, "__builtin_ia32_pabsw", IX86_BUILTIN_PABSW, UNKNOWN, (int) V4HI_FTYPE_V4HI },
21944 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv4si2, "__builtin_ia32_pabsd128", IX86_BUILTIN_PABSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI },
21945 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv2si2, "__builtin_ia32_pabsd", IX86_BUILTIN_PABSD, UNKNOWN, (int) V2SI_FTYPE_V2SI },
21947 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddwv8hi3, "__builtin_ia32_phaddw128", IX86_BUILTIN_PHADDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21948 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddwv4hi3, "__builtin_ia32_phaddw", IX86_BUILTIN_PHADDW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21949 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phadddv4si3, "__builtin_ia32_phaddd128", IX86_BUILTIN_PHADDD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21950 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phadddv2si3, "__builtin_ia32_phaddd", IX86_BUILTIN_PHADDD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21951 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddswv8hi3, "__builtin_ia32_phaddsw128", IX86_BUILTIN_PHADDSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21952 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddswv4hi3, "__builtin_ia32_phaddsw", IX86_BUILTIN_PHADDSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21953 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubwv8hi3, "__builtin_ia32_phsubw128", IX86_BUILTIN_PHSUBW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21954 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubwv4hi3, "__builtin_ia32_phsubw", IX86_BUILTIN_PHSUBW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21955 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubdv4si3, "__builtin_ia32_phsubd128", IX86_BUILTIN_PHSUBD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21956 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubdv2si3, "__builtin_ia32_phsubd", IX86_BUILTIN_PHSUBD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21957 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubswv8hi3, "__builtin_ia32_phsubsw128", IX86_BUILTIN_PHSUBSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21958 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubswv4hi3, "__builtin_ia32_phsubsw", IX86_BUILTIN_PHSUBSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21959 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmaddubsw128, "__builtin_ia32_pmaddubsw128", IX86_BUILTIN_PMADDUBSW128, UNKNOWN, (int) V8HI_FTYPE_V16QI_V16QI },
21960 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmaddubsw, "__builtin_ia32_pmaddubsw", IX86_BUILTIN_PMADDUBSW, UNKNOWN, (int) V4HI_FTYPE_V8QI_V8QI },
21961 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmulhrswv8hi3, "__builtin_ia32_pmulhrsw128", IX86_BUILTIN_PMULHRSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21962 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmulhrswv4hi3, "__builtin_ia32_pmulhrsw", IX86_BUILTIN_PMULHRSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21963 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pshufbv16qi3, "__builtin_ia32_pshufb128", IX86_BUILTIN_PSHUFB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21964 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pshufbv8qi3, "__builtin_ia32_pshufb", IX86_BUILTIN_PSHUFB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21965 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv16qi3, "__builtin_ia32_psignb128", IX86_BUILTIN_PSIGNB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21966 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv8qi3, "__builtin_ia32_psignb", IX86_BUILTIN_PSIGNB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21967 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv8hi3, "__builtin_ia32_psignw128", IX86_BUILTIN_PSIGNW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21968 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv4hi3, "__builtin_ia32_psignw", IX86_BUILTIN_PSIGNW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21969 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv4si3, "__builtin_ia32_psignd128", IX86_BUILTIN_PSIGND128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21970 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv2si3, "__builtin_ia32_psignd", IX86_BUILTIN_PSIGND, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21973 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_palignrti, "__builtin_ia32_palignr128", IX86_BUILTIN_PALIGNR128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_INT_CONVERT },
21974 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_palignrdi, "__builtin_ia32_palignr", IX86_BUILTIN_PALIGNR, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI_INT_CONVERT },
21977 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendpd, "__builtin_ia32_blendpd", IX86_BUILTIN_BLENDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21978 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendps, "__builtin_ia32_blendps", IX86_BUILTIN_BLENDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21979 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendvpd, "__builtin_ia32_blendvpd", IX86_BUILTIN_BLENDVPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_V2DF },
21980 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendvps, "__builtin_ia32_blendvps", IX86_BUILTIN_BLENDVPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_V4SF },
21981 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_dppd, "__builtin_ia32_dppd", IX86_BUILTIN_DPPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21982 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_dpps, "__builtin_ia32_dpps", IX86_BUILTIN_DPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21983 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_insertps, "__builtin_ia32_insertps128", IX86_BUILTIN_INSERTPS128, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21984 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_mpsadbw, "__builtin_ia32_mpsadbw128", IX86_BUILTIN_MPSADBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI_INT },
21985 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_pblendvb, "__builtin_ia32_pblendvb128", IX86_BUILTIN_PBLENDVB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI_V16QI },
21986 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_pblendw, "__builtin_ia32_pblendw128", IX86_BUILTIN_PBLENDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_INT },
21988 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv8qiv8hi2, "__builtin_ia32_pmovsxbw128", IX86_BUILTIN_PMOVSXBW128, UNKNOWN, (int) V8HI_FTYPE_V16QI },
21989 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv4qiv4si2, "__builtin_ia32_pmovsxbd128", IX86_BUILTIN_PMOVSXBD128, UNKNOWN, (int) V4SI_FTYPE_V16QI },
21990 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2qiv2di2, "__builtin_ia32_pmovsxbq128", IX86_BUILTIN_PMOVSXBQ128, UNKNOWN, (int) V2DI_FTYPE_V16QI },
21991 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv4hiv4si2, "__builtin_ia32_pmovsxwd128", IX86_BUILTIN_PMOVSXWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI },
21992 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2hiv2di2, "__builtin_ia32_pmovsxwq128", IX86_BUILTIN_PMOVSXWQ128, UNKNOWN, (int) V2DI_FTYPE_V8HI },
21993 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2siv2di2, "__builtin_ia32_pmovsxdq128", IX86_BUILTIN_PMOVSXDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI },
21994 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv8qiv8hi2, "__builtin_ia32_pmovzxbw128", IX86_BUILTIN_PMOVZXBW128, UNKNOWN, (int) V8HI_FTYPE_V16QI },
21995 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv4qiv4si2, "__builtin_ia32_pmovzxbd128", IX86_BUILTIN_PMOVZXBD128, UNKNOWN, (int) V4SI_FTYPE_V16QI },
21996 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2qiv2di2, "__builtin_ia32_pmovzxbq128", IX86_BUILTIN_PMOVZXBQ128, UNKNOWN, (int) V2DI_FTYPE_V16QI },
21997 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv4hiv4si2, "__builtin_ia32_pmovzxwd128", IX86_BUILTIN_PMOVZXWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI },
21998 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2hiv2di2, "__builtin_ia32_pmovzxwq128", IX86_BUILTIN_PMOVZXWQ128, UNKNOWN, (int) V2DI_FTYPE_V8HI },
21999 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2siv2di2, "__builtin_ia32_pmovzxdq128", IX86_BUILTIN_PMOVZXDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI },
22000 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_phminposuw, "__builtin_ia32_phminposuw128", IX86_BUILTIN_PHMINPOSUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI },
22002 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_packusdw, "__builtin_ia32_packusdw128", IX86_BUILTIN_PACKUSDW128, UNKNOWN, (int) V8HI_FTYPE_V4SI_V4SI },
22003 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_eqv2di3, "__builtin_ia32_pcmpeqq", IX86_BUILTIN_PCMPEQQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22004 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_smaxv16qi3, "__builtin_ia32_pmaxsb128", IX86_BUILTIN_PMAXSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
22005 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_smaxv4si3, "__builtin_ia32_pmaxsd128", IX86_BUILTIN_PMAXSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
22006 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_umaxv4si3, "__builtin_ia32_pmaxud128", IX86_BUILTIN_PMAXUD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
22007 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_umaxv8hi3, "__builtin_ia32_pmaxuw128", IX86_BUILTIN_PMAXUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
22008 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sminv16qi3, "__builtin_ia32_pminsb128", IX86_BUILTIN_PMINSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
22009 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sminv4si3, "__builtin_ia32_pminsd128", IX86_BUILTIN_PMINSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
22010 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_uminv4si3, "__builtin_ia32_pminud128", IX86_BUILTIN_PMINUD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
22011 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_uminv8hi3, "__builtin_ia32_pminuw128", IX86_BUILTIN_PMINUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
22012 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_mulv2siv2di3, "__builtin_ia32_pmuldq128", IX86_BUILTIN_PMULDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI_V4SI },
22013 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_mulv4si3, "__builtin_ia32_pmulld128", IX86_BUILTIN_PMULLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
22016 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd, "__builtin_ia32_roundpd", IX86_BUILTIN_ROUNDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_INT },
22017 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_roundps", IX86_BUILTIN_ROUNDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_INT },
22018 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundsd, "__builtin_ia32_roundsd", IX86_BUILTIN_ROUNDSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
22019 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundss, "__builtin_ia32_roundss", IX86_BUILTIN_ROUNDSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
22021 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestz128", IX86_BUILTIN_PTESTZ, EQ, (int) INT_FTYPE_V2DI_V2DI_PTEST },
22022 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestc128", IX86_BUILTIN_PTESTC, LTU, (int) INT_FTYPE_V2DI_V2DI_PTEST },
22023 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestnzc128", IX86_BUILTIN_PTESTNZC, GTU, (int) INT_FTYPE_V2DI_V2DI_PTEST },
22026 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_gtv2di3, "__builtin_ia32_pcmpgtq", IX86_BUILTIN_PCMPGTQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22027 { OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_CRC32, CODE_FOR_sse4_2_crc32qi, "__builtin_ia32_crc32qi", IX86_BUILTIN_CRC32QI, UNKNOWN, (int) UINT_FTYPE_UINT_UCHAR },
22028 { OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_CRC32, CODE_FOR_sse4_2_crc32hi, "__builtin_ia32_crc32hi", IX86_BUILTIN_CRC32HI, UNKNOWN, (int) UINT_FTYPE_UINT_USHORT },
22029 { OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_CRC32, CODE_FOR_sse4_2_crc32si, "__builtin_ia32_crc32si", IX86_BUILTIN_CRC32SI, UNKNOWN, (int) UINT_FTYPE_UINT_UINT },
22030 { OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_CRC32 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse4_2_crc32di, "__builtin_ia32_crc32di", IX86_BUILTIN_CRC32DI, UNKNOWN, (int) UINT64_FTYPE_UINT64_UINT64 },
22033 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_extrqi, "__builtin_ia32_extrqi", IX86_BUILTIN_EXTRQI, UNKNOWN, (int) V2DI_FTYPE_V2DI_UINT_UINT },
22034 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_extrq, "__builtin_ia32_extrq", IX86_BUILTIN_EXTRQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V16QI },
22035 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_insertqi, "__builtin_ia32_insertqi", IX86_BUILTIN_INSERTQI, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_UINT_UINT },
22036 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_insertq, "__builtin_ia32_insertq", IX86_BUILTIN_INSERTQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22039 { OPTION_MASK_ISA_SSE2, CODE_FOR_aeskeygenassist, 0, IX86_BUILTIN_AESKEYGENASSIST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_INT },
22040 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesimc, 0, IX86_BUILTIN_AESIMC128, UNKNOWN, (int) V2DI_FTYPE_V2DI },
22042 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesenc, 0, IX86_BUILTIN_AESENC128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22043 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesenclast, 0, IX86_BUILTIN_AESENCLAST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22044 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesdec, 0, IX86_BUILTIN_AESDEC128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22045 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesdeclast, 0, IX86_BUILTIN_AESDECLAST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
22048 { OPTION_MASK_ISA_SSE2, CODE_FOR_pclmulqdq, 0, IX86_BUILTIN_PCLMULQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_INT },
22051 { OPTION_MASK_ISA_AVX, CODE_FOR_addv4df3, "__builtin_ia32_addpd256", IX86_BUILTIN_ADDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22052 { OPTION_MASK_ISA_AVX, CODE_FOR_addv8sf3, "__builtin_ia32_addps256", IX86_BUILTIN_ADDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22053 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_addsubv4df3, "__builtin_ia32_addsubpd256", IX86_BUILTIN_ADDSUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22054 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_addsubv8sf3, "__builtin_ia32_addsubps256", IX86_BUILTIN_ADDSUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22055 { OPTION_MASK_ISA_AVX, CODE_FOR_andv4df3, "__builtin_ia32_andpd256", IX86_BUILTIN_ANDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22056 { OPTION_MASK_ISA_AVX, CODE_FOR_andv8sf3, "__builtin_ia32_andps256", IX86_BUILTIN_ANDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22057 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_andnotv4df3, "__builtin_ia32_andnpd256", IX86_BUILTIN_ANDNPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22058 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_andnotv8sf3, "__builtin_ia32_andnps256", IX86_BUILTIN_ANDNPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22059 { OPTION_MASK_ISA_AVX, CODE_FOR_divv4df3, "__builtin_ia32_divpd256", IX86_BUILTIN_DIVPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22060 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_divv8sf3, "__builtin_ia32_divps256", IX86_BUILTIN_DIVPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22061 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_haddv4df3, "__builtin_ia32_haddpd256", IX86_BUILTIN_HADDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22062 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_hsubv8sf3, "__builtin_ia32_hsubps256", IX86_BUILTIN_HSUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22063 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_hsubv4df3, "__builtin_ia32_hsubpd256", IX86_BUILTIN_HSUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22064 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_haddv8sf3, "__builtin_ia32_haddps256", IX86_BUILTIN_HADDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22065 { OPTION_MASK_ISA_AVX, CODE_FOR_smaxv4df3, "__builtin_ia32_maxpd256", IX86_BUILTIN_MAXPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22066 { OPTION_MASK_ISA_AVX, CODE_FOR_smaxv8sf3, "__builtin_ia32_maxps256", IX86_BUILTIN_MAXPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22067 { OPTION_MASK_ISA_AVX, CODE_FOR_sminv4df3, "__builtin_ia32_minpd256", IX86_BUILTIN_MINPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22068 { OPTION_MASK_ISA_AVX, CODE_FOR_sminv8sf3, "__builtin_ia32_minps256", IX86_BUILTIN_MINPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22069 { OPTION_MASK_ISA_AVX, CODE_FOR_mulv4df3, "__builtin_ia32_mulpd256", IX86_BUILTIN_MULPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22070 { OPTION_MASK_ISA_AVX, CODE_FOR_mulv8sf3, "__builtin_ia32_mulps256", IX86_BUILTIN_MULPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22071 { OPTION_MASK_ISA_AVX, CODE_FOR_iorv4df3, "__builtin_ia32_orpd256", IX86_BUILTIN_ORPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22072 { OPTION_MASK_ISA_AVX, CODE_FOR_iorv8sf3, "__builtin_ia32_orps256", IX86_BUILTIN_ORPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22073 { OPTION_MASK_ISA_AVX, CODE_FOR_subv4df3, "__builtin_ia32_subpd256", IX86_BUILTIN_SUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22074 { OPTION_MASK_ISA_AVX, CODE_FOR_subv8sf3, "__builtin_ia32_subps256", IX86_BUILTIN_SUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22075 { OPTION_MASK_ISA_AVX, CODE_FOR_xorv4df3, "__builtin_ia32_xorpd256", IX86_BUILTIN_XORPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22076 { OPTION_MASK_ISA_AVX, CODE_FOR_xorv8sf3, "__builtin_ia32_xorps256", IX86_BUILTIN_XORPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22078 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv2df3, "__builtin_ia32_vpermilvarpd", IX86_BUILTIN_VPERMILVARPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DI },
22079 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv4sf3, "__builtin_ia32_vpermilvarps", IX86_BUILTIN_VPERMILVARPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SI },
22080 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv4df3, "__builtin_ia32_vpermilvarpd256", IX86_BUILTIN_VPERMILVARPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DI },
22081 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv8sf3, "__builtin_ia32_vpermilvarps256", IX86_BUILTIN_VPERMILVARPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SI },
22083 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendpd256, "__builtin_ia32_blendpd256", IX86_BUILTIN_BLENDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
22084 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendps256, "__builtin_ia32_blendps256", IX86_BUILTIN_BLENDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
22085 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendvpd256, "__builtin_ia32_blendvpd256", IX86_BUILTIN_BLENDVPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_V4DF },
22086 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendvps256, "__builtin_ia32_blendvps256", IX86_BUILTIN_BLENDVPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_V8SF },
22087 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_dpps256, "__builtin_ia32_dpps256", IX86_BUILTIN_DPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
22088 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_shufpd256, "__builtin_ia32_shufpd256", IX86_BUILTIN_SHUFPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
22089 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_shufps256, "__builtin_ia32_shufps256", IX86_BUILTIN_SHUFPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
22090 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmpsdv2df3, "__builtin_ia32_cmpsd", IX86_BUILTIN_CMPSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
22091 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmpssv4sf3, "__builtin_ia32_cmpss", IX86_BUILTIN_CMPSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
22092 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppdv2df3, "__builtin_ia32_cmppd", IX86_BUILTIN_CMPPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
22093 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppsv4sf3, "__builtin_ia32_cmpps", IX86_BUILTIN_CMPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
22094 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppdv4df3, "__builtin_ia32_cmppd256", IX86_BUILTIN_CMPPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
22095 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppsv8sf3, "__builtin_ia32_cmpps256", IX86_BUILTIN_CMPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
22096 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v4df, "__builtin_ia32_vextractf128_pd256", IX86_BUILTIN_EXTRACTF128PD256, UNKNOWN, (int) V2DF_FTYPE_V4DF_INT },
22097 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v8sf, "__builtin_ia32_vextractf128_ps256", IX86_BUILTIN_EXTRACTF128PS256, UNKNOWN, (int) V4SF_FTYPE_V8SF_INT },
22098 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v8si, "__builtin_ia32_vextractf128_si256", IX86_BUILTIN_EXTRACTF128SI256, UNKNOWN, (int) V4SI_FTYPE_V8SI_INT },
22099 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtdq2pd256, "__builtin_ia32_cvtdq2pd256", IX86_BUILTIN_CVTDQ2PD256, UNKNOWN, (int) V4DF_FTYPE_V4SI },
22100 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtdq2ps256, "__builtin_ia32_cvtdq2ps256", IX86_BUILTIN_CVTDQ2PS256, UNKNOWN, (int) V8SF_FTYPE_V8SI },
22101 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtpd2ps256, "__builtin_ia32_cvtpd2ps256", IX86_BUILTIN_CVTPD2PS256, UNKNOWN, (int) V4SF_FTYPE_V4DF },
22102 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtps2dq256, "__builtin_ia32_cvtps2dq256", IX86_BUILTIN_CVTPS2DQ256, UNKNOWN, (int) V8SI_FTYPE_V8SF },
22103 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtps2pd256, "__builtin_ia32_cvtps2pd256", IX86_BUILTIN_CVTPS2PD256, UNKNOWN, (int) V4DF_FTYPE_V4SF },
22104 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvttpd2dq256, "__builtin_ia32_cvttpd2dq256", IX86_BUILTIN_CVTTPD2DQ256, UNKNOWN, (int) V4SI_FTYPE_V4DF },
22105 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtpd2dq256, "__builtin_ia32_cvtpd2dq256", IX86_BUILTIN_CVTPD2DQ256, UNKNOWN, (int) V4SI_FTYPE_V4DF },
22106 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvttps2dq256, "__builtin_ia32_cvttps2dq256", IX86_BUILTIN_CVTTPS2DQ256, UNKNOWN, (int) V8SI_FTYPE_V8SF },
22107 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v4df3, "__builtin_ia32_vperm2f128_pd256", IX86_BUILTIN_VPERM2F128PD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
22108 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v8sf3, "__builtin_ia32_vperm2f128_ps256", IX86_BUILTIN_VPERM2F128PS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
22109 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v8si3, "__builtin_ia32_vperm2f128_si256", IX86_BUILTIN_VPERM2F128SI256, UNKNOWN, (int) V8SI_FTYPE_V8SI_V8SI_INT },
22110 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv2df, "__builtin_ia32_vpermilpd", IX86_BUILTIN_VPERMILPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_INT },
22111 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv4sf, "__builtin_ia32_vpermilps", IX86_BUILTIN_VPERMILPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_INT },
22112 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv4df, "__builtin_ia32_vpermilpd256", IX86_BUILTIN_VPERMILPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_INT },
22113 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv8sf, "__builtin_ia32_vpermilps256", IX86_BUILTIN_VPERMILPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_INT },
22114 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v4df, "__builtin_ia32_vinsertf128_pd256", IX86_BUILTIN_VINSERTF128PD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V2DF_INT },
22115 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v8sf, "__builtin_ia32_vinsertf128_ps256", IX86_BUILTIN_VINSERTF128PS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V4SF_INT },
22116 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v8si, "__builtin_ia32_vinsertf128_si256", IX86_BUILTIN_VINSERTF128SI256, UNKNOWN, (int) V8SI_FTYPE_V8SI_V4SI_INT },
22118 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movshdup256, "__builtin_ia32_movshdup256", IX86_BUILTIN_MOVSHDUP256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22119 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movsldup256, "__builtin_ia32_movsldup256", IX86_BUILTIN_MOVSLDUP256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22120 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movddup256, "__builtin_ia32_movddup256", IX86_BUILTIN_MOVDDUP256, UNKNOWN, (int) V4DF_FTYPE_V4DF },
22122 { OPTION_MASK_ISA_AVX, CODE_FOR_sqrtv4df2, "__builtin_ia32_sqrtpd256", IX86_BUILTIN_SQRTPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF },
22123 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_sqrtv8sf2, "__builtin_ia32_sqrtps256", IX86_BUILTIN_SQRTPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22124 { OPTION_MASK_ISA_AVX, CODE_FOR_sqrtv8sf2, "__builtin_ia32_sqrtps_nr256", IX86_BUILTIN_SQRTPS_NR256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22125 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_rsqrtv8sf2, "__builtin_ia32_rsqrtps256", IX86_BUILTIN_RSQRTPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22126 { OPTION_MASK_ISA_AVX, CODE_FOR_rsqrtv8sf2, "__builtin_ia32_rsqrtps_nr256", IX86_BUILTIN_RSQRTPS_NR256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22128 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_rcpv8sf2, "__builtin_ia32_rcpps256", IX86_BUILTIN_RCPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
22130 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundpd256, "__builtin_ia32_roundpd256", IX86_BUILTIN_ROUNDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_INT },
22131 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_roundps256", IX86_BUILTIN_ROUNDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_INT },
22133 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpckhpd256, "__builtin_ia32_unpckhpd256", IX86_BUILTIN_UNPCKHPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22134 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpcklpd256, "__builtin_ia32_unpcklpd256", IX86_BUILTIN_UNPCKLPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
22135 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpckhps256, "__builtin_ia32_unpckhps256", IX86_BUILTIN_UNPCKHPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22136 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpcklps256, "__builtin_ia32_unpcklps256", IX86_BUILTIN_UNPCKLPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
22138 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_si256_si, "__builtin_ia32_si256_si", IX86_BUILTIN_SI256_SI, UNKNOWN, (int) V8SI_FTYPE_V4SI },
22139 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ps256_ps, "__builtin_ia32_ps256_ps", IX86_BUILTIN_PS256_PS, UNKNOWN, (int) V8SF_FTYPE_V4SF },
22140 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_pd256_pd, "__builtin_ia32_pd256_pd", IX86_BUILTIN_PD256_PD, UNKNOWN, (int) V4DF_FTYPE_V2DF },
22141 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_si_si256, "__builtin_ia32_si_si256", IX86_BUILTIN_SI_SI256, UNKNOWN, (int) V4SI_FTYPE_V8SI },
22142 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ps_ps256, "__builtin_ia32_ps_ps256", IX86_BUILTIN_PS_PS256, UNKNOWN, (int) V4SF_FTYPE_V8SF },
22143 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_pd_pd256, "__builtin_ia32_pd_pd256", IX86_BUILTIN_PD_PD256, UNKNOWN, (int) V2DF_FTYPE_V4DF },
22145 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestzpd", IX86_BUILTIN_VTESTZPD, EQ, (int) INT_FTYPE_V2DF_V2DF_PTEST },
22146 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestcpd", IX86_BUILTIN_VTESTCPD, LTU, (int) INT_FTYPE_V2DF_V2DF_PTEST },
22147 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestnzcpd", IX86_BUILTIN_VTESTNZCPD, GTU, (int) INT_FTYPE_V2DF_V2DF_PTEST },
22148 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestzps", IX86_BUILTIN_VTESTZPS, EQ, (int) INT_FTYPE_V4SF_V4SF_PTEST },
22149 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestcps", IX86_BUILTIN_VTESTCPS, LTU, (int) INT_FTYPE_V4SF_V4SF_PTEST },
22150 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestnzcps", IX86_BUILTIN_VTESTNZCPS, GTU, (int) INT_FTYPE_V4SF_V4SF_PTEST },
22151 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestzpd256", IX86_BUILTIN_VTESTZPD256, EQ, (int) INT_FTYPE_V4DF_V4DF_PTEST },
22152 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestcpd256", IX86_BUILTIN_VTESTCPD256, LTU, (int) INT_FTYPE_V4DF_V4DF_PTEST },
22153 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestnzcpd256", IX86_BUILTIN_VTESTNZCPD256, GTU, (int) INT_FTYPE_V4DF_V4DF_PTEST },
22154 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestzps256", IX86_BUILTIN_VTESTZPS256, EQ, (int) INT_FTYPE_V8SF_V8SF_PTEST },
22155 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestcps256", IX86_BUILTIN_VTESTCPS256, LTU, (int) INT_FTYPE_V8SF_V8SF_PTEST },
22156 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestnzcps256", IX86_BUILTIN_VTESTNZCPS256, GTU, (int) INT_FTYPE_V8SF_V8SF_PTEST },
22157 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestz256", IX86_BUILTIN_PTESTZ256, EQ, (int) INT_FTYPE_V4DI_V4DI_PTEST },
22158 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestc256", IX86_BUILTIN_PTESTC256, LTU, (int) INT_FTYPE_V4DI_V4DI_PTEST },
22159 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestnzc256", IX86_BUILTIN_PTESTNZC256, GTU, (int) INT_FTYPE_V4DI_V4DI_PTEST },
22161 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movmskpd256, "__builtin_ia32_movmskpd256", IX86_BUILTIN_MOVMSKPD256, UNKNOWN, (int) INT_FTYPE_V4DF },
22162 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movmskps256, "__builtin_ia32_movmskps256", IX86_BUILTIN_MOVMSKPS256, UNKNOWN, (int) INT_FTYPE_V8SF },
22164 { OPTION_MASK_ISA_ABM, CODE_FOR_clzhi2_abm, "__builtin_clzs", IX86_BUILTIN_CLZS, UNKNOWN, (int) UINT16_FTYPE_UINT16 },
22167 /* FMA4 and XOP. */
22168 #define MULTI_ARG_3_SF V4SF_FTYPE_V4SF_V4SF_V4SF
22169 #define MULTI_ARG_3_DF V2DF_FTYPE_V2DF_V2DF_V2DF
22170 #define MULTI_ARG_3_SF2 V8SF_FTYPE_V8SF_V8SF_V8SF
22171 #define MULTI_ARG_3_DF2 V4DF_FTYPE_V4DF_V4DF_V4DF
22172 #define MULTI_ARG_3_DI V2DI_FTYPE_V2DI_V2DI_V2DI
22173 #define MULTI_ARG_3_SI V4SI_FTYPE_V4SI_V4SI_V4SI
22174 #define MULTI_ARG_3_SI_DI V4SI_FTYPE_V4SI_V4SI_V2DI
22175 #define MULTI_ARG_3_HI V8HI_FTYPE_V8HI_V8HI_V8HI
22176 #define MULTI_ARG_3_HI_SI V8HI_FTYPE_V8HI_V8HI_V4SI
22177 #define MULTI_ARG_3_QI V16QI_FTYPE_V16QI_V16QI_V16QI
22178 #define MULTI_ARG_3_DI2 V4DI_FTYPE_V4DI_V4DI_V4DI
22179 #define MULTI_ARG_3_SI2 V8SI_FTYPE_V8SI_V8SI_V8SI
22180 #define MULTI_ARG_3_HI2 V16HI_FTYPE_V16HI_V16HI_V16HI
22181 #define MULTI_ARG_3_QI2 V32QI_FTYPE_V32QI_V32QI_V32QI
22182 #define MULTI_ARG_2_SF V4SF_FTYPE_V4SF_V4SF
22183 #define MULTI_ARG_2_DF V2DF_FTYPE_V2DF_V2DF
22184 #define MULTI_ARG_2_DI V2DI_FTYPE_V2DI_V2DI
22185 #define MULTI_ARG_2_SI V4SI_FTYPE_V4SI_V4SI
22186 #define MULTI_ARG_2_HI V8HI_FTYPE_V8HI_V8HI
22187 #define MULTI_ARG_2_QI V16QI_FTYPE_V16QI_V16QI
22188 #define MULTI_ARG_2_DI_IMM V2DI_FTYPE_V2DI_SI
22189 #define MULTI_ARG_2_SI_IMM V4SI_FTYPE_V4SI_SI
22190 #define MULTI_ARG_2_HI_IMM V8HI_FTYPE_V8HI_SI
22191 #define MULTI_ARG_2_QI_IMM V16QI_FTYPE_V16QI_SI
22192 #define MULTI_ARG_2_DI_CMP V2DI_FTYPE_V2DI_V2DI_CMP
22193 #define MULTI_ARG_2_SI_CMP V4SI_FTYPE_V4SI_V4SI_CMP
22194 #define MULTI_ARG_2_HI_CMP V8HI_FTYPE_V8HI_V8HI_CMP
22195 #define MULTI_ARG_2_QI_CMP V16QI_FTYPE_V16QI_V16QI_CMP
22196 #define MULTI_ARG_2_SF_TF V4SF_FTYPE_V4SF_V4SF_TF
22197 #define MULTI_ARG_2_DF_TF V2DF_FTYPE_V2DF_V2DF_TF
22198 #define MULTI_ARG_2_DI_TF V2DI_FTYPE_V2DI_V2DI_TF
22199 #define MULTI_ARG_2_SI_TF V4SI_FTYPE_V4SI_V4SI_TF
22200 #define MULTI_ARG_2_HI_TF V8HI_FTYPE_V8HI_V8HI_TF
22201 #define MULTI_ARG_2_QI_TF V16QI_FTYPE_V16QI_V16QI_TF
22202 #define MULTI_ARG_1_SF V4SF_FTYPE_V4SF
22203 #define MULTI_ARG_1_DF V2DF_FTYPE_V2DF
22204 #define MULTI_ARG_1_SF2 V8SF_FTYPE_V8SF
22205 #define MULTI_ARG_1_DF2 V4DF_FTYPE_V4DF
22206 #define MULTI_ARG_1_DI V2DI_FTYPE_V2DI
22207 #define MULTI_ARG_1_SI V4SI_FTYPE_V4SI
22208 #define MULTI_ARG_1_HI V8HI_FTYPE_V8HI
22209 #define MULTI_ARG_1_QI V16QI_FTYPE_V16QI
22210 #define MULTI_ARG_1_SI_DI V2DI_FTYPE_V4SI
22211 #define MULTI_ARG_1_HI_DI V2DI_FTYPE_V8HI
22212 #define MULTI_ARG_1_HI_SI V4SI_FTYPE_V8HI
22213 #define MULTI_ARG_1_QI_DI V2DI_FTYPE_V16QI
22214 #define MULTI_ARG_1_QI_SI V4SI_FTYPE_V16QI
22215 #define MULTI_ARG_1_QI_HI V8HI_FTYPE_V16QI
22217 static const struct builtin_description bdesc_multi_arg[] =
22219 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfmaddv4sf4, "__builtin_ia32_vfmaddss", IX86_BUILTIN_VFMADDSS, UNKNOWN, (int)MULTI_ARG_3_SF },
22220 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfmaddv2df4, "__builtin_ia32_vfmaddsd", IX86_BUILTIN_VFMADDSD, UNKNOWN, (int)MULTI_ARG_3_DF },
22221 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddv4sf4, "__builtin_ia32_vfmaddps", IX86_BUILTIN_VFMADDPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22222 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddv2df4, "__builtin_ia32_vfmaddpd", IX86_BUILTIN_VFMADDPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22223 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfmsubv4sf4, "__builtin_ia32_vfmsubss", IX86_BUILTIN_VFMSUBSS, UNKNOWN, (int)MULTI_ARG_3_SF },
22224 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfmsubv2df4, "__builtin_ia32_vfmsubsd", IX86_BUILTIN_VFMSUBSD, UNKNOWN, (int)MULTI_ARG_3_DF },
22225 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubv4sf4, "__builtin_ia32_vfmsubps", IX86_BUILTIN_VFMSUBPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22226 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubv2df4, "__builtin_ia32_vfmsubpd", IX86_BUILTIN_VFMSUBPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22228 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfnmaddv4sf4, "__builtin_ia32_vfnmaddss", IX86_BUILTIN_VFNMADDSS, UNKNOWN, (int)MULTI_ARG_3_SF },
22229 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfnmaddv2df4, "__builtin_ia32_vfnmaddsd", IX86_BUILTIN_VFNMADDSD, UNKNOWN, (int)MULTI_ARG_3_DF },
22230 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmaddv4sf4, "__builtin_ia32_vfnmaddps", IX86_BUILTIN_VFNMADDPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22231 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmaddv2df4, "__builtin_ia32_vfnmaddpd", IX86_BUILTIN_VFNMADDPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22232 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfnmsubv4sf4, "__builtin_ia32_vfnmsubss", IX86_BUILTIN_VFNMSUBSS, UNKNOWN, (int)MULTI_ARG_3_SF },
22233 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_vmfnmsubv2df4, "__builtin_ia32_vfnmsubsd", IX86_BUILTIN_VFNMSUBSD, UNKNOWN, (int)MULTI_ARG_3_DF },
22234 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmsubv4sf4, "__builtin_ia32_vfnmsubps", IX86_BUILTIN_VFNMSUBPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22235 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmsubv2df4, "__builtin_ia32_vfnmsubpd", IX86_BUILTIN_VFNMSUBPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22237 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddsubv4sf4, "__builtin_ia32_vfmaddsubps", IX86_BUILTIN_VFMADDSUBPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22238 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddsubv2df4, "__builtin_ia32_vfmaddsubpd", IX86_BUILTIN_VFMADDSUBPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22239 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubaddv4sf4, "__builtin_ia32_vfmsubaddps", IX86_BUILTIN_VFMSUBADDPS, UNKNOWN, (int)MULTI_ARG_3_SF },
22240 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubaddv2df4, "__builtin_ia32_vfmsubaddpd", IX86_BUILTIN_VFMSUBADDPD, UNKNOWN, (int)MULTI_ARG_3_DF },
22242 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddv8sf4256, "__builtin_ia32_vfmaddps256", IX86_BUILTIN_VFMADDPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22243 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddv4df4256, "__builtin_ia32_vfmaddpd256", IX86_BUILTIN_VFMADDPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22244 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubv8sf4256, "__builtin_ia32_vfmsubps256", IX86_BUILTIN_VFMSUBPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22245 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubv4df4256, "__builtin_ia32_vfmsubpd256", IX86_BUILTIN_VFMSUBPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22247 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmaddv8sf4256, "__builtin_ia32_vfnmaddps256", IX86_BUILTIN_VFNMADDPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22248 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmaddv4df4256, "__builtin_ia32_vfnmaddpd256", IX86_BUILTIN_VFNMADDPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22249 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmsubv8sf4256, "__builtin_ia32_vfnmsubps256", IX86_BUILTIN_VFNMSUBPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22250 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fnmsubv4df4256, "__builtin_ia32_vfnmsubpd256", IX86_BUILTIN_VFNMSUBPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22252 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddsubv8sf4, "__builtin_ia32_vfmaddsubps256", IX86_BUILTIN_VFMADDSUBPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22253 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmaddsubv4df4, "__builtin_ia32_vfmaddsubpd256", IX86_BUILTIN_VFMADDSUBPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22254 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubaddv8sf4, "__builtin_ia32_vfmsubaddps256", IX86_BUILTIN_VFMSUBADDPS256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22255 { OPTION_MASK_ISA_FMA4, CODE_FOR_fma4i_fmsubaddv4df4, "__builtin_ia32_vfmsubaddpd256", IX86_BUILTIN_VFMSUBADDPD256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22257 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v2di, "__builtin_ia32_vpcmov", IX86_BUILTIN_VPCMOV, UNKNOWN, (int)MULTI_ARG_3_DI },
22258 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v2di, "__builtin_ia32_vpcmov_v2di", IX86_BUILTIN_VPCMOV_V2DI, UNKNOWN, (int)MULTI_ARG_3_DI },
22259 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v4si, "__builtin_ia32_vpcmov_v4si", IX86_BUILTIN_VPCMOV_V4SI, UNKNOWN, (int)MULTI_ARG_3_SI },
22260 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v8hi, "__builtin_ia32_vpcmov_v8hi", IX86_BUILTIN_VPCMOV_V8HI, UNKNOWN, (int)MULTI_ARG_3_HI },
22261 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v16qi, "__builtin_ia32_vpcmov_v16qi",IX86_BUILTIN_VPCMOV_V16QI,UNKNOWN, (int)MULTI_ARG_3_QI },
22262 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v2df, "__builtin_ia32_vpcmov_v2df", IX86_BUILTIN_VPCMOV_V2DF, UNKNOWN, (int)MULTI_ARG_3_DF },
22263 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v4sf, "__builtin_ia32_vpcmov_v4sf", IX86_BUILTIN_VPCMOV_V4SF, UNKNOWN, (int)MULTI_ARG_3_SF },
22265 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v4di256, "__builtin_ia32_vpcmov256", IX86_BUILTIN_VPCMOV256, UNKNOWN, (int)MULTI_ARG_3_DI2 },
22266 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v4di256, "__builtin_ia32_vpcmov_v4di256", IX86_BUILTIN_VPCMOV_V4DI256, UNKNOWN, (int)MULTI_ARG_3_DI2 },
22267 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v8si256, "__builtin_ia32_vpcmov_v8si256", IX86_BUILTIN_VPCMOV_V8SI256, UNKNOWN, (int)MULTI_ARG_3_SI2 },
22268 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v16hi256, "__builtin_ia32_vpcmov_v16hi256", IX86_BUILTIN_VPCMOV_V16HI256, UNKNOWN, (int)MULTI_ARG_3_HI2 },
22269 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v32qi256, "__builtin_ia32_vpcmov_v32qi256", IX86_BUILTIN_VPCMOV_V32QI256, UNKNOWN, (int)MULTI_ARG_3_QI2 },
22270 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v4df256, "__builtin_ia32_vpcmov_v4df256", IX86_BUILTIN_VPCMOV_V4DF256, UNKNOWN, (int)MULTI_ARG_3_DF2 },
22271 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcmov_v8sf256, "__builtin_ia32_vpcmov_v8sf256", IX86_BUILTIN_VPCMOV_V8SF256, UNKNOWN, (int)MULTI_ARG_3_SF2 },
22273 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pperm, "__builtin_ia32_vpperm", IX86_BUILTIN_VPPERM, UNKNOWN, (int)MULTI_ARG_3_QI },
22275 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacssww, "__builtin_ia32_vpmacssww", IX86_BUILTIN_VPMACSSWW, UNKNOWN, (int)MULTI_ARG_3_HI },
22276 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacsww, "__builtin_ia32_vpmacsww", IX86_BUILTIN_VPMACSWW, UNKNOWN, (int)MULTI_ARG_3_HI },
22277 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacsswd, "__builtin_ia32_vpmacsswd", IX86_BUILTIN_VPMACSSWD, UNKNOWN, (int)MULTI_ARG_3_HI_SI },
22278 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacswd, "__builtin_ia32_vpmacswd", IX86_BUILTIN_VPMACSWD, UNKNOWN, (int)MULTI_ARG_3_HI_SI },
22279 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacssdd, "__builtin_ia32_vpmacssdd", IX86_BUILTIN_VPMACSSDD, UNKNOWN, (int)MULTI_ARG_3_SI },
22280 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacsdd, "__builtin_ia32_vpmacsdd", IX86_BUILTIN_VPMACSDD, UNKNOWN, (int)MULTI_ARG_3_SI },
22281 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacssdql, "__builtin_ia32_vpmacssdql", IX86_BUILTIN_VPMACSSDQL, UNKNOWN, (int)MULTI_ARG_3_SI_DI },
22282 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacssdqh, "__builtin_ia32_vpmacssdqh", IX86_BUILTIN_VPMACSSDQH, UNKNOWN, (int)MULTI_ARG_3_SI_DI },
22283 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacsdql, "__builtin_ia32_vpmacsdql", IX86_BUILTIN_VPMACSDQL, UNKNOWN, (int)MULTI_ARG_3_SI_DI },
22284 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmacsdqh, "__builtin_ia32_vpmacsdqh", IX86_BUILTIN_VPMACSDQH, UNKNOWN, (int)MULTI_ARG_3_SI_DI },
22285 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmadcsswd, "__builtin_ia32_vpmadcsswd", IX86_BUILTIN_VPMADCSSWD, UNKNOWN, (int)MULTI_ARG_3_HI_SI },
22286 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pmadcswd, "__builtin_ia32_vpmadcswd", IX86_BUILTIN_VPMADCSWD, UNKNOWN, (int)MULTI_ARG_3_HI_SI },
22288 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vrotlv2di3, "__builtin_ia32_vprotq", IX86_BUILTIN_VPROTQ, UNKNOWN, (int)MULTI_ARG_2_DI },
22289 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vrotlv4si3, "__builtin_ia32_vprotd", IX86_BUILTIN_VPROTD, UNKNOWN, (int)MULTI_ARG_2_SI },
22290 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vrotlv8hi3, "__builtin_ia32_vprotw", IX86_BUILTIN_VPROTW, UNKNOWN, (int)MULTI_ARG_2_HI },
22291 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vrotlv16qi3, "__builtin_ia32_vprotb", IX86_BUILTIN_VPROTB, UNKNOWN, (int)MULTI_ARG_2_QI },
22292 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_rotlv2di3, "__builtin_ia32_vprotqi", IX86_BUILTIN_VPROTQ_IMM, UNKNOWN, (int)MULTI_ARG_2_DI_IMM },
22293 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_rotlv4si3, "__builtin_ia32_vprotdi", IX86_BUILTIN_VPROTD_IMM, UNKNOWN, (int)MULTI_ARG_2_SI_IMM },
22294 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_rotlv8hi3, "__builtin_ia32_vprotwi", IX86_BUILTIN_VPROTW_IMM, UNKNOWN, (int)MULTI_ARG_2_HI_IMM },
22295 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_rotlv16qi3, "__builtin_ia32_vprotbi", IX86_BUILTIN_VPROTB_IMM, UNKNOWN, (int)MULTI_ARG_2_QI_IMM },
22296 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_ashlv2di3, "__builtin_ia32_vpshaq", IX86_BUILTIN_VPSHAQ, UNKNOWN, (int)MULTI_ARG_2_DI },
22297 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_ashlv4si3, "__builtin_ia32_vpshad", IX86_BUILTIN_VPSHAD, UNKNOWN, (int)MULTI_ARG_2_SI },
22298 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_ashlv8hi3, "__builtin_ia32_vpshaw", IX86_BUILTIN_VPSHAW, UNKNOWN, (int)MULTI_ARG_2_HI },
22299 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_ashlv16qi3, "__builtin_ia32_vpshab", IX86_BUILTIN_VPSHAB, UNKNOWN, (int)MULTI_ARG_2_QI },
22300 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_lshlv2di3, "__builtin_ia32_vpshlq", IX86_BUILTIN_VPSHLQ, UNKNOWN, (int)MULTI_ARG_2_DI },
22301 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_lshlv4si3, "__builtin_ia32_vpshld", IX86_BUILTIN_VPSHLD, UNKNOWN, (int)MULTI_ARG_2_SI },
22302 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_lshlv8hi3, "__builtin_ia32_vpshlw", IX86_BUILTIN_VPSHLW, UNKNOWN, (int)MULTI_ARG_2_HI },
22303 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_lshlv16qi3, "__builtin_ia32_vpshlb", IX86_BUILTIN_VPSHLB, UNKNOWN, (int)MULTI_ARG_2_QI },
22305 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vmfrczv4sf2, "__builtin_ia32_vfrczss", IX86_BUILTIN_VFRCZSS, UNKNOWN, (int)MULTI_ARG_2_SF },
22306 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_vmfrczv2df2, "__builtin_ia32_vfrczsd", IX86_BUILTIN_VFRCZSD, UNKNOWN, (int)MULTI_ARG_2_DF },
22307 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_frczv4sf2, "__builtin_ia32_vfrczps", IX86_BUILTIN_VFRCZPS, UNKNOWN, (int)MULTI_ARG_1_SF },
22308 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_frczv2df2, "__builtin_ia32_vfrczpd", IX86_BUILTIN_VFRCZPD, UNKNOWN, (int)MULTI_ARG_1_DF },
22309 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_frczv8sf2256, "__builtin_ia32_vfrczps256", IX86_BUILTIN_VFRCZPS256, UNKNOWN, (int)MULTI_ARG_1_SF2 },
22310 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_frczv4df2256, "__builtin_ia32_vfrczpd256", IX86_BUILTIN_VFRCZPD256, UNKNOWN, (int)MULTI_ARG_1_DF2 },
22312 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddbw, "__builtin_ia32_vphaddbw", IX86_BUILTIN_VPHADDBW, UNKNOWN, (int)MULTI_ARG_1_QI_HI },
22313 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddbd, "__builtin_ia32_vphaddbd", IX86_BUILTIN_VPHADDBD, UNKNOWN, (int)MULTI_ARG_1_QI_SI },
22314 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddbq, "__builtin_ia32_vphaddbq", IX86_BUILTIN_VPHADDBQ, UNKNOWN, (int)MULTI_ARG_1_QI_DI },
22315 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddwd, "__builtin_ia32_vphaddwd", IX86_BUILTIN_VPHADDWD, UNKNOWN, (int)MULTI_ARG_1_HI_SI },
22316 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddwq, "__builtin_ia32_vphaddwq", IX86_BUILTIN_VPHADDWQ, UNKNOWN, (int)MULTI_ARG_1_HI_DI },
22317 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phadddq, "__builtin_ia32_vphadddq", IX86_BUILTIN_VPHADDDQ, UNKNOWN, (int)MULTI_ARG_1_SI_DI },
22318 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddubw, "__builtin_ia32_vphaddubw", IX86_BUILTIN_VPHADDUBW, UNKNOWN, (int)MULTI_ARG_1_QI_HI },
22319 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddubd, "__builtin_ia32_vphaddubd", IX86_BUILTIN_VPHADDUBD, UNKNOWN, (int)MULTI_ARG_1_QI_SI },
22320 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddubq, "__builtin_ia32_vphaddubq", IX86_BUILTIN_VPHADDUBQ, UNKNOWN, (int)MULTI_ARG_1_QI_DI },
22321 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phadduwd, "__builtin_ia32_vphadduwd", IX86_BUILTIN_VPHADDUWD, UNKNOWN, (int)MULTI_ARG_1_HI_SI },
22322 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phadduwq, "__builtin_ia32_vphadduwq", IX86_BUILTIN_VPHADDUWQ, UNKNOWN, (int)MULTI_ARG_1_HI_DI },
22323 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phaddudq, "__builtin_ia32_vphaddudq", IX86_BUILTIN_VPHADDUDQ, UNKNOWN, (int)MULTI_ARG_1_SI_DI },
22324 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phsubbw, "__builtin_ia32_vphsubbw", IX86_BUILTIN_VPHSUBBW, UNKNOWN, (int)MULTI_ARG_1_QI_HI },
22325 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phsubwd, "__builtin_ia32_vphsubwd", IX86_BUILTIN_VPHSUBWD, UNKNOWN, (int)MULTI_ARG_1_HI_SI },
22326 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_phsubdq, "__builtin_ia32_vphsubdq", IX86_BUILTIN_VPHSUBDQ, UNKNOWN, (int)MULTI_ARG_1_SI_DI },
22328 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomeqb", IX86_BUILTIN_VPCOMEQB, EQ, (int)MULTI_ARG_2_QI_CMP },
22329 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomneb", IX86_BUILTIN_VPCOMNEB, NE, (int)MULTI_ARG_2_QI_CMP },
22330 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomneqb", IX86_BUILTIN_VPCOMNEB, NE, (int)MULTI_ARG_2_QI_CMP },
22331 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomltb", IX86_BUILTIN_VPCOMLTB, LT, (int)MULTI_ARG_2_QI_CMP },
22332 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomleb", IX86_BUILTIN_VPCOMLEB, LE, (int)MULTI_ARG_2_QI_CMP },
22333 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomgtb", IX86_BUILTIN_VPCOMGTB, GT, (int)MULTI_ARG_2_QI_CMP },
22334 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv16qi3, "__builtin_ia32_vpcomgeb", IX86_BUILTIN_VPCOMGEB, GE, (int)MULTI_ARG_2_QI_CMP },
22336 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomeqw", IX86_BUILTIN_VPCOMEQW, EQ, (int)MULTI_ARG_2_HI_CMP },
22337 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomnew", IX86_BUILTIN_VPCOMNEW, NE, (int)MULTI_ARG_2_HI_CMP },
22338 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomneqw", IX86_BUILTIN_VPCOMNEW, NE, (int)MULTI_ARG_2_HI_CMP },
22339 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomltw", IX86_BUILTIN_VPCOMLTW, LT, (int)MULTI_ARG_2_HI_CMP },
22340 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomlew", IX86_BUILTIN_VPCOMLEW, LE, (int)MULTI_ARG_2_HI_CMP },
22341 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomgtw", IX86_BUILTIN_VPCOMGTW, GT, (int)MULTI_ARG_2_HI_CMP },
22342 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv8hi3, "__builtin_ia32_vpcomgew", IX86_BUILTIN_VPCOMGEW, GE, (int)MULTI_ARG_2_HI_CMP },
22344 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomeqd", IX86_BUILTIN_VPCOMEQD, EQ, (int)MULTI_ARG_2_SI_CMP },
22345 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomned", IX86_BUILTIN_VPCOMNED, NE, (int)MULTI_ARG_2_SI_CMP },
22346 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomneqd", IX86_BUILTIN_VPCOMNED, NE, (int)MULTI_ARG_2_SI_CMP },
22347 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomltd", IX86_BUILTIN_VPCOMLTD, LT, (int)MULTI_ARG_2_SI_CMP },
22348 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomled", IX86_BUILTIN_VPCOMLED, LE, (int)MULTI_ARG_2_SI_CMP },
22349 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomgtd", IX86_BUILTIN_VPCOMGTD, GT, (int)MULTI_ARG_2_SI_CMP },
22350 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv4si3, "__builtin_ia32_vpcomged", IX86_BUILTIN_VPCOMGED, GE, (int)MULTI_ARG_2_SI_CMP },
22352 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomeqq", IX86_BUILTIN_VPCOMEQQ, EQ, (int)MULTI_ARG_2_DI_CMP },
22353 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomneq", IX86_BUILTIN_VPCOMNEQ, NE, (int)MULTI_ARG_2_DI_CMP },
22354 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomneqq", IX86_BUILTIN_VPCOMNEQ, NE, (int)MULTI_ARG_2_DI_CMP },
22355 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomltq", IX86_BUILTIN_VPCOMLTQ, LT, (int)MULTI_ARG_2_DI_CMP },
22356 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomleq", IX86_BUILTIN_VPCOMLEQ, LE, (int)MULTI_ARG_2_DI_CMP },
22357 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomgtq", IX86_BUILTIN_VPCOMGTQ, GT, (int)MULTI_ARG_2_DI_CMP },
22358 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmpv2di3, "__builtin_ia32_vpcomgeq", IX86_BUILTIN_VPCOMGEQ, GE, (int)MULTI_ARG_2_DI_CMP },
22360 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v16qi3,"__builtin_ia32_vpcomequb", IX86_BUILTIN_VPCOMEQUB, EQ, (int)MULTI_ARG_2_QI_CMP },
22361 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v16qi3,"__builtin_ia32_vpcomneub", IX86_BUILTIN_VPCOMNEUB, NE, (int)MULTI_ARG_2_QI_CMP },
22362 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v16qi3,"__builtin_ia32_vpcomnequb", IX86_BUILTIN_VPCOMNEUB, NE, (int)MULTI_ARG_2_QI_CMP },
22363 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv16qi3, "__builtin_ia32_vpcomltub", IX86_BUILTIN_VPCOMLTUB, LTU, (int)MULTI_ARG_2_QI_CMP },
22364 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv16qi3, "__builtin_ia32_vpcomleub", IX86_BUILTIN_VPCOMLEUB, LEU, (int)MULTI_ARG_2_QI_CMP },
22365 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv16qi3, "__builtin_ia32_vpcomgtub", IX86_BUILTIN_VPCOMGTUB, GTU, (int)MULTI_ARG_2_QI_CMP },
22366 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv16qi3, "__builtin_ia32_vpcomgeub", IX86_BUILTIN_VPCOMGEUB, GEU, (int)MULTI_ARG_2_QI_CMP },
22368 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v8hi3, "__builtin_ia32_vpcomequw", IX86_BUILTIN_VPCOMEQUW, EQ, (int)MULTI_ARG_2_HI_CMP },
22369 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v8hi3, "__builtin_ia32_vpcomneuw", IX86_BUILTIN_VPCOMNEUW, NE, (int)MULTI_ARG_2_HI_CMP },
22370 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v8hi3, "__builtin_ia32_vpcomnequw", IX86_BUILTIN_VPCOMNEUW, NE, (int)MULTI_ARG_2_HI_CMP },
22371 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv8hi3, "__builtin_ia32_vpcomltuw", IX86_BUILTIN_VPCOMLTUW, LTU, (int)MULTI_ARG_2_HI_CMP },
22372 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv8hi3, "__builtin_ia32_vpcomleuw", IX86_BUILTIN_VPCOMLEUW, LEU, (int)MULTI_ARG_2_HI_CMP },
22373 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv8hi3, "__builtin_ia32_vpcomgtuw", IX86_BUILTIN_VPCOMGTUW, GTU, (int)MULTI_ARG_2_HI_CMP },
22374 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv8hi3, "__builtin_ia32_vpcomgeuw", IX86_BUILTIN_VPCOMGEUW, GEU, (int)MULTI_ARG_2_HI_CMP },
22376 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v4si3, "__builtin_ia32_vpcomequd", IX86_BUILTIN_VPCOMEQUD, EQ, (int)MULTI_ARG_2_SI_CMP },
22377 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v4si3, "__builtin_ia32_vpcomneud", IX86_BUILTIN_VPCOMNEUD, NE, (int)MULTI_ARG_2_SI_CMP },
22378 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v4si3, "__builtin_ia32_vpcomnequd", IX86_BUILTIN_VPCOMNEUD, NE, (int)MULTI_ARG_2_SI_CMP },
22379 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv4si3, "__builtin_ia32_vpcomltud", IX86_BUILTIN_VPCOMLTUD, LTU, (int)MULTI_ARG_2_SI_CMP },
22380 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv4si3, "__builtin_ia32_vpcomleud", IX86_BUILTIN_VPCOMLEUD, LEU, (int)MULTI_ARG_2_SI_CMP },
22381 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv4si3, "__builtin_ia32_vpcomgtud", IX86_BUILTIN_VPCOMGTUD, GTU, (int)MULTI_ARG_2_SI_CMP },
22382 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv4si3, "__builtin_ia32_vpcomgeud", IX86_BUILTIN_VPCOMGEUD, GEU, (int)MULTI_ARG_2_SI_CMP },
22384 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v2di3, "__builtin_ia32_vpcomequq", IX86_BUILTIN_VPCOMEQUQ, EQ, (int)MULTI_ARG_2_DI_CMP },
22385 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v2di3, "__builtin_ia32_vpcomneuq", IX86_BUILTIN_VPCOMNEUQ, NE, (int)MULTI_ARG_2_DI_CMP },
22386 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_uns2v2di3, "__builtin_ia32_vpcomnequq", IX86_BUILTIN_VPCOMNEUQ, NE, (int)MULTI_ARG_2_DI_CMP },
22387 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv2di3, "__builtin_ia32_vpcomltuq", IX86_BUILTIN_VPCOMLTUQ, LTU, (int)MULTI_ARG_2_DI_CMP },
22388 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv2di3, "__builtin_ia32_vpcomleuq", IX86_BUILTIN_VPCOMLEUQ, LEU, (int)MULTI_ARG_2_DI_CMP },
22389 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv2di3, "__builtin_ia32_vpcomgtuq", IX86_BUILTIN_VPCOMGTUQ, GTU, (int)MULTI_ARG_2_DI_CMP },
22390 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_maskcmp_unsv2di3, "__builtin_ia32_vpcomgeuq", IX86_BUILTIN_VPCOMGEUQ, GEU, (int)MULTI_ARG_2_DI_CMP },
22392 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv16qi3, "__builtin_ia32_vpcomfalseb", IX86_BUILTIN_VPCOMFALSEB, (enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_QI_TF },
22393 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv8hi3, "__builtin_ia32_vpcomfalsew", IX86_BUILTIN_VPCOMFALSEW, (enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_HI_TF },
22394 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv4si3, "__builtin_ia32_vpcomfalsed", IX86_BUILTIN_VPCOMFALSED, (enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_SI_TF },
22395 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv2di3, "__builtin_ia32_vpcomfalseq", IX86_BUILTIN_VPCOMFALSEQ, (enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_DI_TF },
22396 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv16qi3, "__builtin_ia32_vpcomfalseub",IX86_BUILTIN_VPCOMFALSEUB,(enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_QI_TF },
22397 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv8hi3, "__builtin_ia32_vpcomfalseuw",IX86_BUILTIN_VPCOMFALSEUW,(enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_HI_TF },
22398 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv4si3, "__builtin_ia32_vpcomfalseud",IX86_BUILTIN_VPCOMFALSEUD,(enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_SI_TF },
22399 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv2di3, "__builtin_ia32_vpcomfalseuq",IX86_BUILTIN_VPCOMFALSEUQ,(enum rtx_code) PCOM_FALSE, (int)MULTI_ARG_2_DI_TF },
22401 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv16qi3, "__builtin_ia32_vpcomtrueb", IX86_BUILTIN_VPCOMTRUEB, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_QI_TF },
22402 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv8hi3, "__builtin_ia32_vpcomtruew", IX86_BUILTIN_VPCOMTRUEW, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_HI_TF },
22403 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv4si3, "__builtin_ia32_vpcomtrued", IX86_BUILTIN_VPCOMTRUED, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_SI_TF },
22404 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv2di3, "__builtin_ia32_vpcomtrueq", IX86_BUILTIN_VPCOMTRUEQ, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_DI_TF },
22405 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv16qi3, "__builtin_ia32_vpcomtrueub", IX86_BUILTIN_VPCOMTRUEUB, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_QI_TF },
22406 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv8hi3, "__builtin_ia32_vpcomtrueuw", IX86_BUILTIN_VPCOMTRUEUW, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_HI_TF },
22407 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv4si3, "__builtin_ia32_vpcomtrueud", IX86_BUILTIN_VPCOMTRUEUD, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_SI_TF },
22408 { OPTION_MASK_ISA_XOP, CODE_FOR_xop_pcom_tfv2di3, "__builtin_ia32_vpcomtrueuq", IX86_BUILTIN_VPCOMTRUEUQ, (enum rtx_code) PCOM_TRUE, (int)MULTI_ARG_2_DI_TF },
22412 /* Set up all the MMX/SSE builtins, even builtins for instructions that are not
22413 in the current target ISA to allow the user to compile particular modules
22414 with different target specific options that differ from the command line
22417 ix86_init_mmx_sse_builtins (void)
22419 const struct builtin_description * d;
22420 enum ix86_builtin_func_type ftype;
22423 /* Add all special builtins with variable number of operands. */
22424 for (i = 0, d = bdesc_special_args;
22425 i < ARRAY_SIZE (bdesc_special_args);
22431 ftype = (enum ix86_builtin_func_type) d->flag;
22432 def_builtin (d->mask, d->name, ftype, d->code);
22435 /* Add all builtins with variable number of operands. */
22436 for (i = 0, d = bdesc_args;
22437 i < ARRAY_SIZE (bdesc_args);
22443 ftype = (enum ix86_builtin_func_type) d->flag;
22444 def_builtin_const (d->mask, d->name, ftype, d->code);
22447 /* pcmpestr[im] insns. */
22448 for (i = 0, d = bdesc_pcmpestr;
22449 i < ARRAY_SIZE (bdesc_pcmpestr);
22452 if (d->code == IX86_BUILTIN_PCMPESTRM128)
22453 ftype = V16QI_FTYPE_V16QI_INT_V16QI_INT_INT;
22455 ftype = INT_FTYPE_V16QI_INT_V16QI_INT_INT;
22456 def_builtin_const (d->mask, d->name, ftype, d->code);
22459 /* pcmpistr[im] insns. */
22460 for (i = 0, d = bdesc_pcmpistr;
22461 i < ARRAY_SIZE (bdesc_pcmpistr);
22464 if (d->code == IX86_BUILTIN_PCMPISTRM128)
22465 ftype = V16QI_FTYPE_V16QI_V16QI_INT;
22467 ftype = INT_FTYPE_V16QI_V16QI_INT;
22468 def_builtin_const (d->mask, d->name, ftype, d->code);
22471 /* comi/ucomi insns. */
22472 for (i = 0, d = bdesc_comi; i < ARRAY_SIZE (bdesc_comi); i++, d++)
22474 if (d->mask == OPTION_MASK_ISA_SSE2)
22475 ftype = INT_FTYPE_V2DF_V2DF;
22477 ftype = INT_FTYPE_V4SF_V4SF;
22478 def_builtin_const (d->mask, d->name, ftype, d->code);
22482 def_builtin (OPTION_MASK_ISA_SSE, "__builtin_ia32_ldmxcsr",
22483 VOID_FTYPE_UNSIGNED, IX86_BUILTIN_LDMXCSR);
22484 def_builtin (OPTION_MASK_ISA_SSE, "__builtin_ia32_stmxcsr",
22485 UNSIGNED_FTYPE_VOID, IX86_BUILTIN_STMXCSR);
22487 /* SSE or 3DNow!A */
22488 def_builtin (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A,
22489 "__builtin_ia32_maskmovq", VOID_FTYPE_V8QI_V8QI_PCHAR,
22490 IX86_BUILTIN_MASKMOVQ);
22493 def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_maskmovdqu",
22494 VOID_FTYPE_V16QI_V16QI_PCHAR, IX86_BUILTIN_MASKMOVDQU);
22496 def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_clflush",
22497 VOID_FTYPE_PCVOID, IX86_BUILTIN_CLFLUSH);
22498 x86_mfence = def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_mfence",
22499 VOID_FTYPE_VOID, IX86_BUILTIN_MFENCE);
22502 def_builtin (OPTION_MASK_ISA_SSE3, "__builtin_ia32_monitor",
22503 VOID_FTYPE_PCVOID_UNSIGNED_UNSIGNED, IX86_BUILTIN_MONITOR);
22504 def_builtin (OPTION_MASK_ISA_SSE3, "__builtin_ia32_mwait",
22505 VOID_FTYPE_UNSIGNED_UNSIGNED, IX86_BUILTIN_MWAIT);
22508 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesenc128",
22509 V2DI_FTYPE_V2DI_V2DI, IX86_BUILTIN_AESENC128);
22510 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesenclast128",
22511 V2DI_FTYPE_V2DI_V2DI, IX86_BUILTIN_AESENCLAST128);
22512 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesdec128",
22513 V2DI_FTYPE_V2DI_V2DI, IX86_BUILTIN_AESDEC128);
22514 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesdeclast128",
22515 V2DI_FTYPE_V2DI_V2DI, IX86_BUILTIN_AESDECLAST128);
22516 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesimc128",
22517 V2DI_FTYPE_V2DI, IX86_BUILTIN_AESIMC128);
22518 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aeskeygenassist128",
22519 V2DI_FTYPE_V2DI_INT, IX86_BUILTIN_AESKEYGENASSIST128);
22522 def_builtin_const (OPTION_MASK_ISA_PCLMUL, "__builtin_ia32_pclmulqdq128",
22523 V2DI_FTYPE_V2DI_V2DI_INT, IX86_BUILTIN_PCLMULQDQ128);
22525 /* MMX access to the vec_init patterns. */
22526 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v2si",
22527 V2SI_FTYPE_INT_INT, IX86_BUILTIN_VEC_INIT_V2SI);
22529 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v4hi",
22530 V4HI_FTYPE_HI_HI_HI_HI,
22531 IX86_BUILTIN_VEC_INIT_V4HI);
22533 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v8qi",
22534 V8QI_FTYPE_QI_QI_QI_QI_QI_QI_QI_QI,
22535 IX86_BUILTIN_VEC_INIT_V8QI);
22537 /* Access to the vec_extract patterns. */
22538 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v2df",
22539 DOUBLE_FTYPE_V2DF_INT, IX86_BUILTIN_VEC_EXT_V2DF);
22540 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v2di",
22541 DI_FTYPE_V2DI_INT, IX86_BUILTIN_VEC_EXT_V2DI);
22542 def_builtin_const (OPTION_MASK_ISA_SSE, "__builtin_ia32_vec_ext_v4sf",
22543 FLOAT_FTYPE_V4SF_INT, IX86_BUILTIN_VEC_EXT_V4SF);
22544 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v4si",
22545 SI_FTYPE_V4SI_INT, IX86_BUILTIN_VEC_EXT_V4SI);
22546 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v8hi",
22547 HI_FTYPE_V8HI_INT, IX86_BUILTIN_VEC_EXT_V8HI);
22549 def_builtin_const (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A,
22550 "__builtin_ia32_vec_ext_v4hi",
22551 HI_FTYPE_V4HI_INT, IX86_BUILTIN_VEC_EXT_V4HI);
22553 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_ext_v2si",
22554 SI_FTYPE_V2SI_INT, IX86_BUILTIN_VEC_EXT_V2SI);
22556 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v16qi",
22557 QI_FTYPE_V16QI_INT, IX86_BUILTIN_VEC_EXT_V16QI);
22559 /* Access to the vec_set patterns. */
22560 def_builtin_const (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_64BIT,
22561 "__builtin_ia32_vec_set_v2di",
22562 V2DI_FTYPE_V2DI_DI_INT, IX86_BUILTIN_VEC_SET_V2DI);
22564 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v4sf",
22565 V4SF_FTYPE_V4SF_FLOAT_INT, IX86_BUILTIN_VEC_SET_V4SF);
22567 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v4si",
22568 V4SI_FTYPE_V4SI_SI_INT, IX86_BUILTIN_VEC_SET_V4SI);
22570 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_set_v8hi",
22571 V8HI_FTYPE_V8HI_HI_INT, IX86_BUILTIN_VEC_SET_V8HI);
22573 def_builtin_const (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A,
22574 "__builtin_ia32_vec_set_v4hi",
22575 V4HI_FTYPE_V4HI_HI_INT, IX86_BUILTIN_VEC_SET_V4HI);
22577 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v16qi",
22578 V16QI_FTYPE_V16QI_QI_INT, IX86_BUILTIN_VEC_SET_V16QI);
22580 /* Add FMA4 multi-arg argument instructions */
22581 for (i = 0, d = bdesc_multi_arg; i < ARRAY_SIZE (bdesc_multi_arg); i++, d++)
22586 ftype = (enum ix86_builtin_func_type) d->flag;
22587 def_builtin_const (d->mask, d->name, ftype, d->code);
22591 /* Internal method for ix86_init_builtins. */
22594 ix86_init_builtins_va_builtins_abi (void)
22596 tree ms_va_ref, sysv_va_ref;
22597 tree fnvoid_va_end_ms, fnvoid_va_end_sysv;
22598 tree fnvoid_va_start_ms, fnvoid_va_start_sysv;
22599 tree fnvoid_va_copy_ms, fnvoid_va_copy_sysv;
22600 tree fnattr_ms = NULL_TREE, fnattr_sysv = NULL_TREE;
22604 fnattr_ms = build_tree_list (get_identifier ("ms_abi"), NULL_TREE);
22605 fnattr_sysv = build_tree_list (get_identifier ("sysv_abi"), NULL_TREE);
22606 ms_va_ref = build_reference_type (ms_va_list_type_node);
22608 build_pointer_type (TREE_TYPE (sysv_va_list_type_node));
22611 build_function_type_list (void_type_node, ms_va_ref, NULL_TREE);
22612 fnvoid_va_start_ms =
22613 build_varargs_function_type_list (void_type_node, ms_va_ref, NULL_TREE);
22614 fnvoid_va_end_sysv =
22615 build_function_type_list (void_type_node, sysv_va_ref, NULL_TREE);
22616 fnvoid_va_start_sysv =
22617 build_varargs_function_type_list (void_type_node, sysv_va_ref,
22619 fnvoid_va_copy_ms =
22620 build_function_type_list (void_type_node, ms_va_ref, ms_va_list_type_node,
22622 fnvoid_va_copy_sysv =
22623 build_function_type_list (void_type_node, sysv_va_ref,
22624 sysv_va_ref, NULL_TREE);
22626 add_builtin_function ("__builtin_ms_va_start", fnvoid_va_start_ms,
22627 BUILT_IN_VA_START, BUILT_IN_NORMAL, NULL, fnattr_ms);
22628 add_builtin_function ("__builtin_ms_va_end", fnvoid_va_end_ms,
22629 BUILT_IN_VA_END, BUILT_IN_NORMAL, NULL, fnattr_ms);
22630 add_builtin_function ("__builtin_ms_va_copy", fnvoid_va_copy_ms,
22631 BUILT_IN_VA_COPY, BUILT_IN_NORMAL, NULL, fnattr_ms);
22632 add_builtin_function ("__builtin_sysv_va_start", fnvoid_va_start_sysv,
22633 BUILT_IN_VA_START, BUILT_IN_NORMAL, NULL, fnattr_sysv);
22634 add_builtin_function ("__builtin_sysv_va_end", fnvoid_va_end_sysv,
22635 BUILT_IN_VA_END, BUILT_IN_NORMAL, NULL, fnattr_sysv);
22636 add_builtin_function ("__builtin_sysv_va_copy", fnvoid_va_copy_sysv,
22637 BUILT_IN_VA_COPY, BUILT_IN_NORMAL, NULL, fnattr_sysv);
22641 ix86_init_builtin_types (void)
22643 tree float128_type_node, float80_type_node;
22645 /* The __float80 type. */
22646 float80_type_node = long_double_type_node;
22647 if (TYPE_MODE (float80_type_node) != XFmode)
22649 /* The __float80 type. */
22650 float80_type_node = make_node (REAL_TYPE);
22652 TYPE_PRECISION (float80_type_node) = 80;
22653 layout_type (float80_type_node);
22655 (*lang_hooks.types.register_builtin_type) (float80_type_node, "__float80");
22657 /* The __float128 type. */
22658 float128_type_node = make_node (REAL_TYPE);
22659 TYPE_PRECISION (float128_type_node) = 128;
22660 layout_type (float128_type_node);
22661 (*lang_hooks.types.register_builtin_type) (float128_type_node, "__float128");
22663 /* This macro is built by i386-builtin-types.awk. */
22664 DEFINE_BUILTIN_PRIMITIVE_TYPES;
22668 ix86_init_builtins (void)
22672 ix86_init_builtin_types ();
22674 /* TFmode support builtins. */
22675 def_builtin_const (0, "__builtin_infq",
22676 FLOAT128_FTYPE_VOID, IX86_BUILTIN_INFQ);
22677 def_builtin_const (0, "__builtin_huge_valq",
22678 FLOAT128_FTYPE_VOID, IX86_BUILTIN_HUGE_VALQ);
22680 /* We will expand them to normal call if SSE2 isn't available since
22681 they are used by libgcc. */
22682 t = ix86_get_builtin_func_type (FLOAT128_FTYPE_FLOAT128);
22683 t = add_builtin_function ("__builtin_fabsq", t, IX86_BUILTIN_FABSQ,
22684 BUILT_IN_MD, "__fabstf2", NULL_TREE);
22685 TREE_READONLY (t) = 1;
22686 ix86_builtins[(int) IX86_BUILTIN_FABSQ] = t;
22688 t = ix86_get_builtin_func_type (FLOAT128_FTYPE_FLOAT128_FLOAT128);
22689 t = add_builtin_function ("__builtin_copysignq", t, IX86_BUILTIN_COPYSIGNQ,
22690 BUILT_IN_MD, "__copysigntf3", NULL_TREE);
22691 TREE_READONLY (t) = 1;
22692 ix86_builtins[(int) IX86_BUILTIN_COPYSIGNQ] = t;
22694 ix86_init_mmx_sse_builtins ();
22697 ix86_init_builtins_va_builtins_abi ();
22700 /* Return the ix86 builtin for CODE. */
22703 ix86_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED)
22705 if (code >= IX86_BUILTIN_MAX)
22706 return error_mark_node;
22708 return ix86_builtins[code];
22711 /* Errors in the source file can cause expand_expr to return const0_rtx
22712 where we expect a vector. To avoid crashing, use one of the vector
22713 clear instructions. */
22715 safe_vector_operand (rtx x, enum machine_mode mode)
22717 if (x == const0_rtx)
22718 x = CONST0_RTX (mode);
22722 /* Subroutine of ix86_expand_builtin to take care of binop insns. */
22725 ix86_expand_binop_builtin (enum insn_code icode, tree exp, rtx target)
22728 tree arg0 = CALL_EXPR_ARG (exp, 0);
22729 tree arg1 = CALL_EXPR_ARG (exp, 1);
22730 rtx op0 = expand_normal (arg0);
22731 rtx op1 = expand_normal (arg1);
22732 enum machine_mode tmode = insn_data[icode].operand[0].mode;
22733 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
22734 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
22736 if (VECTOR_MODE_P (mode0))
22737 op0 = safe_vector_operand (op0, mode0);
22738 if (VECTOR_MODE_P (mode1))
22739 op1 = safe_vector_operand (op1, mode1);
22741 if (optimize || !target
22742 || GET_MODE (target) != tmode
22743 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
22744 target = gen_reg_rtx (tmode);
22746 if (GET_MODE (op1) == SImode && mode1 == TImode)
22748 rtx x = gen_reg_rtx (V4SImode);
22749 emit_insn (gen_sse2_loadd (x, op1));
22750 op1 = gen_lowpart (TImode, x);
22753 if (!(*insn_data[icode].operand[1].predicate) (op0, mode0))
22754 op0 = copy_to_mode_reg (mode0, op0);
22755 if (!(*insn_data[icode].operand[2].predicate) (op1, mode1))
22756 op1 = copy_to_mode_reg (mode1, op1);
22758 pat = GEN_FCN (icode) (target, op0, op1);
22767 /* Subroutine of ix86_expand_builtin to take care of 2-4 argument insns. */
22770 ix86_expand_multi_arg_builtin (enum insn_code icode, tree exp, rtx target,
22771 enum ix86_builtin_func_type m_type,
22772 enum rtx_code sub_code)
22777 bool comparison_p = false;
22779 bool last_arg_constant = false;
22780 int num_memory = 0;
22783 enum machine_mode mode;
22786 enum machine_mode tmode = insn_data[icode].operand[0].mode;
22790 case MULTI_ARG_3_SF:
22791 case MULTI_ARG_3_DF:
22792 case MULTI_ARG_3_SF2:
22793 case MULTI_ARG_3_DF2:
22794 case MULTI_ARG_3_DI:
22795 case MULTI_ARG_3_SI:
22796 case MULTI_ARG_3_SI_DI:
22797 case MULTI_ARG_3_HI:
22798 case MULTI_ARG_3_HI_SI:
22799 case MULTI_ARG_3_QI:
22800 case MULTI_ARG_3_DI2:
22801 case MULTI_ARG_3_SI2:
22802 case MULTI_ARG_3_HI2:
22803 case MULTI_ARG_3_QI2:
22807 case MULTI_ARG_2_SF:
22808 case MULTI_ARG_2_DF:
22809 case MULTI_ARG_2_DI:
22810 case MULTI_ARG_2_SI:
22811 case MULTI_ARG_2_HI:
22812 case MULTI_ARG_2_QI:
22816 case MULTI_ARG_2_DI_IMM:
22817 case MULTI_ARG_2_SI_IMM:
22818 case MULTI_ARG_2_HI_IMM:
22819 case MULTI_ARG_2_QI_IMM:
22821 last_arg_constant = true;
22824 case MULTI_ARG_1_SF:
22825 case MULTI_ARG_1_DF:
22826 case MULTI_ARG_1_SF2:
22827 case MULTI_ARG_1_DF2:
22828 case MULTI_ARG_1_DI:
22829 case MULTI_ARG_1_SI:
22830 case MULTI_ARG_1_HI:
22831 case MULTI_ARG_1_QI:
22832 case MULTI_ARG_1_SI_DI:
22833 case MULTI_ARG_1_HI_DI:
22834 case MULTI_ARG_1_HI_SI:
22835 case MULTI_ARG_1_QI_DI:
22836 case MULTI_ARG_1_QI_SI:
22837 case MULTI_ARG_1_QI_HI:
22841 case MULTI_ARG_2_DI_CMP:
22842 case MULTI_ARG_2_SI_CMP:
22843 case MULTI_ARG_2_HI_CMP:
22844 case MULTI_ARG_2_QI_CMP:
22846 comparison_p = true;
22849 case MULTI_ARG_2_SF_TF:
22850 case MULTI_ARG_2_DF_TF:
22851 case MULTI_ARG_2_DI_TF:
22852 case MULTI_ARG_2_SI_TF:
22853 case MULTI_ARG_2_HI_TF:
22854 case MULTI_ARG_2_QI_TF:
22860 gcc_unreachable ();
22863 if (optimize || !target
22864 || GET_MODE (target) != tmode
22865 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
22866 target = gen_reg_rtx (tmode);
22868 gcc_assert (nargs <= 4);
22870 for (i = 0; i < nargs; i++)
22872 tree arg = CALL_EXPR_ARG (exp, i);
22873 rtx op = expand_normal (arg);
22874 int adjust = (comparison_p) ? 1 : 0;
22875 enum machine_mode mode = insn_data[icode].operand[i+adjust+1].mode;
22877 if (last_arg_constant && i == nargs-1)
22879 if (!CONST_INT_P (op))
22881 error ("last argument must be an immediate");
22882 return gen_reg_rtx (tmode);
22887 if (VECTOR_MODE_P (mode))
22888 op = safe_vector_operand (op, mode);
22890 /* If we aren't optimizing, only allow one memory operand to be
22892 if (memory_operand (op, mode))
22895 gcc_assert (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode);
22898 || ! (*insn_data[icode].operand[i+adjust+1].predicate) (op, mode)
22900 op = force_reg (mode, op);
22904 args[i].mode = mode;
22910 pat = GEN_FCN (icode) (target, args[0].op);
22915 pat = GEN_FCN (icode) (target, args[0].op, args[1].op,
22916 GEN_INT ((int)sub_code));
22917 else if (! comparison_p)
22918 pat = GEN_FCN (icode) (target, args[0].op, args[1].op);
22921 rtx cmp_op = gen_rtx_fmt_ee (sub_code, GET_MODE (target),
22925 pat = GEN_FCN (icode) (target, cmp_op, args[0].op, args[1].op);
22930 pat = GEN_FCN (icode) (target, args[0].op, args[1].op, args[2].op);
22934 gcc_unreachable ();
22944 /* Subroutine of ix86_expand_args_builtin to take care of scalar unop
22945 insns with vec_merge. */
22948 ix86_expand_unop_vec_merge_builtin (enum insn_code icode, tree exp,
22952 tree arg0 = CALL_EXPR_ARG (exp, 0);
22953 rtx op1, op0 = expand_normal (arg0);
22954 enum machine_mode tmode = insn_data[icode].operand[0].mode;
22955 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
22957 if (optimize || !target
22958 || GET_MODE (target) != tmode
22959 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
22960 target = gen_reg_rtx (tmode);
22962 if (VECTOR_MODE_P (mode0))
22963 op0 = safe_vector_operand (op0, mode0);
22965 if ((optimize && !register_operand (op0, mode0))
22966 || ! (*insn_data[icode].operand[1].predicate) (op0, mode0))
22967 op0 = copy_to_mode_reg (mode0, op0);
22970 if (! (*insn_data[icode].operand[2].predicate) (op1, mode0))
22971 op1 = copy_to_mode_reg (mode0, op1);
22973 pat = GEN_FCN (icode) (target, op0, op1);
22980 /* Subroutine of ix86_expand_builtin to take care of comparison insns. */
22983 ix86_expand_sse_compare (const struct builtin_description *d,
22984 tree exp, rtx target, bool swap)
22987 tree arg0 = CALL_EXPR_ARG (exp, 0);
22988 tree arg1 = CALL_EXPR_ARG (exp, 1);
22989 rtx op0 = expand_normal (arg0);
22990 rtx op1 = expand_normal (arg1);
22992 enum machine_mode tmode = insn_data[d->icode].operand[0].mode;
22993 enum machine_mode mode0 = insn_data[d->icode].operand[1].mode;
22994 enum machine_mode mode1 = insn_data[d->icode].operand[2].mode;
22995 enum rtx_code comparison = d->comparison;
22997 if (VECTOR_MODE_P (mode0))
22998 op0 = safe_vector_operand (op0, mode0);
22999 if (VECTOR_MODE_P (mode1))
23000 op1 = safe_vector_operand (op1, mode1);
23002 /* Swap operands if we have a comparison that isn't available in
23006 rtx tmp = gen_reg_rtx (mode1);
23007 emit_move_insn (tmp, op1);
23012 if (optimize || !target
23013 || GET_MODE (target) != tmode
23014 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode))
23015 target = gen_reg_rtx (tmode);
23017 if ((optimize && !register_operand (op0, mode0))
23018 || ! (*insn_data[d->icode].operand[1].predicate) (op0, mode0))
23019 op0 = copy_to_mode_reg (mode0, op0);
23020 if ((optimize && !register_operand (op1, mode1))
23021 || ! (*insn_data[d->icode].operand[2].predicate) (op1, mode1))
23022 op1 = copy_to_mode_reg (mode1, op1);
23024 op2 = gen_rtx_fmt_ee (comparison, mode0, op0, op1);
23025 pat = GEN_FCN (d->icode) (target, op0, op1, op2);
23032 /* Subroutine of ix86_expand_builtin to take care of comi insns. */
23035 ix86_expand_sse_comi (const struct builtin_description *d, tree exp,
23039 tree arg0 = CALL_EXPR_ARG (exp, 0);
23040 tree arg1 = CALL_EXPR_ARG (exp, 1);
23041 rtx op0 = expand_normal (arg0);
23042 rtx op1 = expand_normal (arg1);
23043 enum machine_mode mode0 = insn_data[d->icode].operand[0].mode;
23044 enum machine_mode mode1 = insn_data[d->icode].operand[1].mode;
23045 enum rtx_code comparison = d->comparison;
23047 if (VECTOR_MODE_P (mode0))
23048 op0 = safe_vector_operand (op0, mode0);
23049 if (VECTOR_MODE_P (mode1))
23050 op1 = safe_vector_operand (op1, mode1);
23052 /* Swap operands if we have a comparison that isn't available in
23054 if (d->flag & BUILTIN_DESC_SWAP_OPERANDS)
23061 target = gen_reg_rtx (SImode);
23062 emit_move_insn (target, const0_rtx);
23063 target = gen_rtx_SUBREG (QImode, target, 0);
23065 if ((optimize && !register_operand (op0, mode0))
23066 || !(*insn_data[d->icode].operand[0].predicate) (op0, mode0))
23067 op0 = copy_to_mode_reg (mode0, op0);
23068 if ((optimize && !register_operand (op1, mode1))
23069 || !(*insn_data[d->icode].operand[1].predicate) (op1, mode1))
23070 op1 = copy_to_mode_reg (mode1, op1);
23072 pat = GEN_FCN (d->icode) (op0, op1);
23076 emit_insn (gen_rtx_SET (VOIDmode,
23077 gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23078 gen_rtx_fmt_ee (comparison, QImode,
23082 return SUBREG_REG (target);
23085 /* Subroutine of ix86_expand_builtin to take care of ptest insns. */
23088 ix86_expand_sse_ptest (const struct builtin_description *d, tree exp,
23092 tree arg0 = CALL_EXPR_ARG (exp, 0);
23093 tree arg1 = CALL_EXPR_ARG (exp, 1);
23094 rtx op0 = expand_normal (arg0);
23095 rtx op1 = expand_normal (arg1);
23096 enum machine_mode mode0 = insn_data[d->icode].operand[0].mode;
23097 enum machine_mode mode1 = insn_data[d->icode].operand[1].mode;
23098 enum rtx_code comparison = d->comparison;
23100 if (VECTOR_MODE_P (mode0))
23101 op0 = safe_vector_operand (op0, mode0);
23102 if (VECTOR_MODE_P (mode1))
23103 op1 = safe_vector_operand (op1, mode1);
23105 target = gen_reg_rtx (SImode);
23106 emit_move_insn (target, const0_rtx);
23107 target = gen_rtx_SUBREG (QImode, target, 0);
23109 if ((optimize && !register_operand (op0, mode0))
23110 || !(*insn_data[d->icode].operand[0].predicate) (op0, mode0))
23111 op0 = copy_to_mode_reg (mode0, op0);
23112 if ((optimize && !register_operand (op1, mode1))
23113 || !(*insn_data[d->icode].operand[1].predicate) (op1, mode1))
23114 op1 = copy_to_mode_reg (mode1, op1);
23116 pat = GEN_FCN (d->icode) (op0, op1);
23120 emit_insn (gen_rtx_SET (VOIDmode,
23121 gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23122 gen_rtx_fmt_ee (comparison, QImode,
23126 return SUBREG_REG (target);
23129 /* Subroutine of ix86_expand_builtin to take care of pcmpestr[im] insns. */
23132 ix86_expand_sse_pcmpestr (const struct builtin_description *d,
23133 tree exp, rtx target)
23136 tree arg0 = CALL_EXPR_ARG (exp, 0);
23137 tree arg1 = CALL_EXPR_ARG (exp, 1);
23138 tree arg2 = CALL_EXPR_ARG (exp, 2);
23139 tree arg3 = CALL_EXPR_ARG (exp, 3);
23140 tree arg4 = CALL_EXPR_ARG (exp, 4);
23141 rtx scratch0, scratch1;
23142 rtx op0 = expand_normal (arg0);
23143 rtx op1 = expand_normal (arg1);
23144 rtx op2 = expand_normal (arg2);
23145 rtx op3 = expand_normal (arg3);
23146 rtx op4 = expand_normal (arg4);
23147 enum machine_mode tmode0, tmode1, modev2, modei3, modev4, modei5, modeimm;
23149 tmode0 = insn_data[d->icode].operand[0].mode;
23150 tmode1 = insn_data[d->icode].operand[1].mode;
23151 modev2 = insn_data[d->icode].operand[2].mode;
23152 modei3 = insn_data[d->icode].operand[3].mode;
23153 modev4 = insn_data[d->icode].operand[4].mode;
23154 modei5 = insn_data[d->icode].operand[5].mode;
23155 modeimm = insn_data[d->icode].operand[6].mode;
23157 if (VECTOR_MODE_P (modev2))
23158 op0 = safe_vector_operand (op0, modev2);
23159 if (VECTOR_MODE_P (modev4))
23160 op2 = safe_vector_operand (op2, modev4);
23162 if (! (*insn_data[d->icode].operand[2].predicate) (op0, modev2))
23163 op0 = copy_to_mode_reg (modev2, op0);
23164 if (! (*insn_data[d->icode].operand[3].predicate) (op1, modei3))
23165 op1 = copy_to_mode_reg (modei3, op1);
23166 if ((optimize && !register_operand (op2, modev4))
23167 || !(*insn_data[d->icode].operand[4].predicate) (op2, modev4))
23168 op2 = copy_to_mode_reg (modev4, op2);
23169 if (! (*insn_data[d->icode].operand[5].predicate) (op3, modei5))
23170 op3 = copy_to_mode_reg (modei5, op3);
23172 if (! (*insn_data[d->icode].operand[6].predicate) (op4, modeimm))
23174 error ("the fifth argument must be a 8-bit immediate");
23178 if (d->code == IX86_BUILTIN_PCMPESTRI128)
23180 if (optimize || !target
23181 || GET_MODE (target) != tmode0
23182 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode0))
23183 target = gen_reg_rtx (tmode0);
23185 scratch1 = gen_reg_rtx (tmode1);
23187 pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2, op3, op4);
23189 else if (d->code == IX86_BUILTIN_PCMPESTRM128)
23191 if (optimize || !target
23192 || GET_MODE (target) != tmode1
23193 || ! (*insn_data[d->icode].operand[1].predicate) (target, tmode1))
23194 target = gen_reg_rtx (tmode1);
23196 scratch0 = gen_reg_rtx (tmode0);
23198 pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2, op3, op4);
23202 gcc_assert (d->flag);
23204 scratch0 = gen_reg_rtx (tmode0);
23205 scratch1 = gen_reg_rtx (tmode1);
23207 pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2, op3, op4);
23217 target = gen_reg_rtx (SImode);
23218 emit_move_insn (target, const0_rtx);
23219 target = gen_rtx_SUBREG (QImode, target, 0);
23222 (gen_rtx_SET (VOIDmode, gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23223 gen_rtx_fmt_ee (EQ, QImode,
23224 gen_rtx_REG ((enum machine_mode) d->flag,
23227 return SUBREG_REG (target);
23234 /* Subroutine of ix86_expand_builtin to take care of pcmpistr[im] insns. */
23237 ix86_expand_sse_pcmpistr (const struct builtin_description *d,
23238 tree exp, rtx target)
23241 tree arg0 = CALL_EXPR_ARG (exp, 0);
23242 tree arg1 = CALL_EXPR_ARG (exp, 1);
23243 tree arg2 = CALL_EXPR_ARG (exp, 2);
23244 rtx scratch0, scratch1;
23245 rtx op0 = expand_normal (arg0);
23246 rtx op1 = expand_normal (arg1);
23247 rtx op2 = expand_normal (arg2);
23248 enum machine_mode tmode0, tmode1, modev2, modev3, modeimm;
23250 tmode0 = insn_data[d->icode].operand[0].mode;
23251 tmode1 = insn_data[d->icode].operand[1].mode;
23252 modev2 = insn_data[d->icode].operand[2].mode;
23253 modev3 = insn_data[d->icode].operand[3].mode;
23254 modeimm = insn_data[d->icode].operand[4].mode;
23256 if (VECTOR_MODE_P (modev2))
23257 op0 = safe_vector_operand (op0, modev2);
23258 if (VECTOR_MODE_P (modev3))
23259 op1 = safe_vector_operand (op1, modev3);
23261 if (! (*insn_data[d->icode].operand[2].predicate) (op0, modev2))
23262 op0 = copy_to_mode_reg (modev2, op0);
23263 if ((optimize && !register_operand (op1, modev3))
23264 || !(*insn_data[d->icode].operand[3].predicate) (op1, modev3))
23265 op1 = copy_to_mode_reg (modev3, op1);
23267 if (! (*insn_data[d->icode].operand[4].predicate) (op2, modeimm))
23269 error ("the third argument must be a 8-bit immediate");
23273 if (d->code == IX86_BUILTIN_PCMPISTRI128)
23275 if (optimize || !target
23276 || GET_MODE (target) != tmode0
23277 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode0))
23278 target = gen_reg_rtx (tmode0);
23280 scratch1 = gen_reg_rtx (tmode1);
23282 pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2);
23284 else if (d->code == IX86_BUILTIN_PCMPISTRM128)
23286 if (optimize || !target
23287 || GET_MODE (target) != tmode1
23288 || ! (*insn_data[d->icode].operand[1].predicate) (target, tmode1))
23289 target = gen_reg_rtx (tmode1);
23291 scratch0 = gen_reg_rtx (tmode0);
23293 pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2);
23297 gcc_assert (d->flag);
23299 scratch0 = gen_reg_rtx (tmode0);
23300 scratch1 = gen_reg_rtx (tmode1);
23302 pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2);
23312 target = gen_reg_rtx (SImode);
23313 emit_move_insn (target, const0_rtx);
23314 target = gen_rtx_SUBREG (QImode, target, 0);
23317 (gen_rtx_SET (VOIDmode, gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23318 gen_rtx_fmt_ee (EQ, QImode,
23319 gen_rtx_REG ((enum machine_mode) d->flag,
23322 return SUBREG_REG (target);
23328 /* Subroutine of ix86_expand_builtin to take care of insns with
23329 variable number of operands. */
23332 ix86_expand_args_builtin (const struct builtin_description *d,
23333 tree exp, rtx target)
23335 rtx pat, real_target;
23336 unsigned int i, nargs;
23337 unsigned int nargs_constant = 0;
23338 int num_memory = 0;
23342 enum machine_mode mode;
23344 bool last_arg_count = false;
23345 enum insn_code icode = d->icode;
23346 const struct insn_data *insn_p = &insn_data[icode];
23347 enum machine_mode tmode = insn_p->operand[0].mode;
23348 enum machine_mode rmode = VOIDmode;
23350 enum rtx_code comparison = d->comparison;
23352 switch ((enum ix86_builtin_func_type) d->flag)
23354 case INT_FTYPE_V8SF_V8SF_PTEST:
23355 case INT_FTYPE_V4DI_V4DI_PTEST:
23356 case INT_FTYPE_V4DF_V4DF_PTEST:
23357 case INT_FTYPE_V4SF_V4SF_PTEST:
23358 case INT_FTYPE_V2DI_V2DI_PTEST:
23359 case INT_FTYPE_V2DF_V2DF_PTEST:
23360 return ix86_expand_sse_ptest (d, exp, target);
23361 case FLOAT128_FTYPE_FLOAT128:
23362 case FLOAT_FTYPE_FLOAT:
23363 case INT_FTYPE_INT:
23364 case UINT64_FTYPE_INT:
23365 case UINT16_FTYPE_UINT16:
23366 case INT64_FTYPE_INT64:
23367 case INT64_FTYPE_V4SF:
23368 case INT64_FTYPE_V2DF:
23369 case INT_FTYPE_V16QI:
23370 case INT_FTYPE_V8QI:
23371 case INT_FTYPE_V8SF:
23372 case INT_FTYPE_V4DF:
23373 case INT_FTYPE_V4SF:
23374 case INT_FTYPE_V2DF:
23375 case V16QI_FTYPE_V16QI:
23376 case V8SI_FTYPE_V8SF:
23377 case V8SI_FTYPE_V4SI:
23378 case V8HI_FTYPE_V8HI:
23379 case V8HI_FTYPE_V16QI:
23380 case V8QI_FTYPE_V8QI:
23381 case V8SF_FTYPE_V8SF:
23382 case V8SF_FTYPE_V8SI:
23383 case V8SF_FTYPE_V4SF:
23384 case V4SI_FTYPE_V4SI:
23385 case V4SI_FTYPE_V16QI:
23386 case V4SI_FTYPE_V4SF:
23387 case V4SI_FTYPE_V8SI:
23388 case V4SI_FTYPE_V8HI:
23389 case V4SI_FTYPE_V4DF:
23390 case V4SI_FTYPE_V2DF:
23391 case V4HI_FTYPE_V4HI:
23392 case V4DF_FTYPE_V4DF:
23393 case V4DF_FTYPE_V4SI:
23394 case V4DF_FTYPE_V4SF:
23395 case V4DF_FTYPE_V2DF:
23396 case V4SF_FTYPE_V4SF:
23397 case V4SF_FTYPE_V4SI:
23398 case V4SF_FTYPE_V8SF:
23399 case V4SF_FTYPE_V4DF:
23400 case V4SF_FTYPE_V2DF:
23401 case V2DI_FTYPE_V2DI:
23402 case V2DI_FTYPE_V16QI:
23403 case V2DI_FTYPE_V8HI:
23404 case V2DI_FTYPE_V4SI:
23405 case V2DF_FTYPE_V2DF:
23406 case V2DF_FTYPE_V4SI:
23407 case V2DF_FTYPE_V4DF:
23408 case V2DF_FTYPE_V4SF:
23409 case V2DF_FTYPE_V2SI:
23410 case V2SI_FTYPE_V2SI:
23411 case V2SI_FTYPE_V4SF:
23412 case V2SI_FTYPE_V2SF:
23413 case V2SI_FTYPE_V2DF:
23414 case V2SF_FTYPE_V2SF:
23415 case V2SF_FTYPE_V2SI:
23418 case V4SF_FTYPE_V4SF_VEC_MERGE:
23419 case V2DF_FTYPE_V2DF_VEC_MERGE:
23420 return ix86_expand_unop_vec_merge_builtin (icode, exp, target);
23421 case FLOAT128_FTYPE_FLOAT128_FLOAT128:
23422 case V16QI_FTYPE_V16QI_V16QI:
23423 case V16QI_FTYPE_V8HI_V8HI:
23424 case V8QI_FTYPE_V8QI_V8QI:
23425 case V8QI_FTYPE_V4HI_V4HI:
23426 case V8HI_FTYPE_V8HI_V8HI:
23427 case V8HI_FTYPE_V16QI_V16QI:
23428 case V8HI_FTYPE_V4SI_V4SI:
23429 case V8SF_FTYPE_V8SF_V8SF:
23430 case V8SF_FTYPE_V8SF_V8SI:
23431 case V4SI_FTYPE_V4SI_V4SI:
23432 case V4SI_FTYPE_V8HI_V8HI:
23433 case V4SI_FTYPE_V4SF_V4SF:
23434 case V4SI_FTYPE_V2DF_V2DF:
23435 case V4HI_FTYPE_V4HI_V4HI:
23436 case V4HI_FTYPE_V8QI_V8QI:
23437 case V4HI_FTYPE_V2SI_V2SI:
23438 case V4DF_FTYPE_V4DF_V4DF:
23439 case V4DF_FTYPE_V4DF_V4DI:
23440 case V4SF_FTYPE_V4SF_V4SF:
23441 case V4SF_FTYPE_V4SF_V4SI:
23442 case V4SF_FTYPE_V4SF_V2SI:
23443 case V4SF_FTYPE_V4SF_V2DF:
23444 case V4SF_FTYPE_V4SF_DI:
23445 case V4SF_FTYPE_V4SF_SI:
23446 case V2DI_FTYPE_V2DI_V2DI:
23447 case V2DI_FTYPE_V16QI_V16QI:
23448 case V2DI_FTYPE_V4SI_V4SI:
23449 case V2DI_FTYPE_V2DI_V16QI:
23450 case V2DI_FTYPE_V2DF_V2DF:
23451 case V2SI_FTYPE_V2SI_V2SI:
23452 case V2SI_FTYPE_V4HI_V4HI:
23453 case V2SI_FTYPE_V2SF_V2SF:
23454 case V2DF_FTYPE_V2DF_V2DF:
23455 case V2DF_FTYPE_V2DF_V4SF:
23456 case V2DF_FTYPE_V2DF_V2DI:
23457 case V2DF_FTYPE_V2DF_DI:
23458 case V2DF_FTYPE_V2DF_SI:
23459 case V2SF_FTYPE_V2SF_V2SF:
23460 case V1DI_FTYPE_V1DI_V1DI:
23461 case V1DI_FTYPE_V8QI_V8QI:
23462 case V1DI_FTYPE_V2SI_V2SI:
23463 if (comparison == UNKNOWN)
23464 return ix86_expand_binop_builtin (icode, exp, target);
23467 case V4SF_FTYPE_V4SF_V4SF_SWAP:
23468 case V2DF_FTYPE_V2DF_V2DF_SWAP:
23469 gcc_assert (comparison != UNKNOWN);
23473 case V8HI_FTYPE_V8HI_V8HI_COUNT:
23474 case V8HI_FTYPE_V8HI_SI_COUNT:
23475 case V4SI_FTYPE_V4SI_V4SI_COUNT:
23476 case V4SI_FTYPE_V4SI_SI_COUNT:
23477 case V4HI_FTYPE_V4HI_V4HI_COUNT:
23478 case V4HI_FTYPE_V4HI_SI_COUNT:
23479 case V2DI_FTYPE_V2DI_V2DI_COUNT:
23480 case V2DI_FTYPE_V2DI_SI_COUNT:
23481 case V2SI_FTYPE_V2SI_V2SI_COUNT:
23482 case V2SI_FTYPE_V2SI_SI_COUNT:
23483 case V1DI_FTYPE_V1DI_V1DI_COUNT:
23484 case V1DI_FTYPE_V1DI_SI_COUNT:
23486 last_arg_count = true;
23488 case UINT64_FTYPE_UINT64_UINT64:
23489 case UINT_FTYPE_UINT_UINT:
23490 case UINT_FTYPE_UINT_USHORT:
23491 case UINT_FTYPE_UINT_UCHAR:
23492 case UINT16_FTYPE_UINT16_INT:
23493 case UINT8_FTYPE_UINT8_INT:
23496 case V2DI_FTYPE_V2DI_INT_CONVERT:
23499 nargs_constant = 1;
23501 case V8HI_FTYPE_V8HI_INT:
23502 case V8SF_FTYPE_V8SF_INT:
23503 case V4SI_FTYPE_V4SI_INT:
23504 case V4SI_FTYPE_V8SI_INT:
23505 case V4HI_FTYPE_V4HI_INT:
23506 case V4DF_FTYPE_V4DF_INT:
23507 case V4SF_FTYPE_V4SF_INT:
23508 case V4SF_FTYPE_V8SF_INT:
23509 case V2DI_FTYPE_V2DI_INT:
23510 case V2DF_FTYPE_V2DF_INT:
23511 case V2DF_FTYPE_V4DF_INT:
23513 nargs_constant = 1;
23515 case V16QI_FTYPE_V16QI_V16QI_V16QI:
23516 case V8SF_FTYPE_V8SF_V8SF_V8SF:
23517 case V4DF_FTYPE_V4DF_V4DF_V4DF:
23518 case V4SF_FTYPE_V4SF_V4SF_V4SF:
23519 case V2DF_FTYPE_V2DF_V2DF_V2DF:
23522 case V16QI_FTYPE_V16QI_V16QI_INT:
23523 case V8HI_FTYPE_V8HI_V8HI_INT:
23524 case V8SI_FTYPE_V8SI_V8SI_INT:
23525 case V8SI_FTYPE_V8SI_V4SI_INT:
23526 case V8SF_FTYPE_V8SF_V8SF_INT:
23527 case V8SF_FTYPE_V8SF_V4SF_INT:
23528 case V4SI_FTYPE_V4SI_V4SI_INT:
23529 case V4DF_FTYPE_V4DF_V4DF_INT:
23530 case V4DF_FTYPE_V4DF_V2DF_INT:
23531 case V4SF_FTYPE_V4SF_V4SF_INT:
23532 case V2DI_FTYPE_V2DI_V2DI_INT:
23533 case V2DF_FTYPE_V2DF_V2DF_INT:
23535 nargs_constant = 1;
23537 case V2DI_FTYPE_V2DI_V2DI_INT_CONVERT:
23540 nargs_constant = 1;
23542 case V1DI_FTYPE_V1DI_V1DI_INT_CONVERT:
23545 nargs_constant = 1;
23547 case V2DI_FTYPE_V2DI_UINT_UINT:
23549 nargs_constant = 2;
23551 case V2DI_FTYPE_V2DI_V2DI_UINT_UINT:
23553 nargs_constant = 2;
23556 gcc_unreachable ();
23559 gcc_assert (nargs <= ARRAY_SIZE (args));
23561 if (comparison != UNKNOWN)
23563 gcc_assert (nargs == 2);
23564 return ix86_expand_sse_compare (d, exp, target, swap);
23567 if (rmode == VOIDmode || rmode == tmode)
23571 || GET_MODE (target) != tmode
23572 || ! (*insn_p->operand[0].predicate) (target, tmode))
23573 target = gen_reg_rtx (tmode);
23574 real_target = target;
23578 target = gen_reg_rtx (rmode);
23579 real_target = simplify_gen_subreg (tmode, target, rmode, 0);
23582 for (i = 0; i < nargs; i++)
23584 tree arg = CALL_EXPR_ARG (exp, i);
23585 rtx op = expand_normal (arg);
23586 enum machine_mode mode = insn_p->operand[i + 1].mode;
23587 bool match = (*insn_p->operand[i + 1].predicate) (op, mode);
23589 if (last_arg_count && (i + 1) == nargs)
23591 /* SIMD shift insns take either an 8-bit immediate or
23592 register as count. But builtin functions take int as
23593 count. If count doesn't match, we put it in register. */
23596 op = simplify_gen_subreg (SImode, op, GET_MODE (op), 0);
23597 if (!(*insn_p->operand[i + 1].predicate) (op, mode))
23598 op = copy_to_reg (op);
23601 else if ((nargs - i) <= nargs_constant)
23606 case CODE_FOR_sse4_1_roundpd:
23607 case CODE_FOR_sse4_1_roundps:
23608 case CODE_FOR_sse4_1_roundsd:
23609 case CODE_FOR_sse4_1_roundss:
23610 case CODE_FOR_sse4_1_blendps:
23611 case CODE_FOR_avx_blendpd256:
23612 case CODE_FOR_avx_vpermilv4df:
23613 case CODE_FOR_avx_roundpd256:
23614 case CODE_FOR_avx_roundps256:
23615 error ("the last argument must be a 4-bit immediate");
23618 case CODE_FOR_sse4_1_blendpd:
23619 case CODE_FOR_avx_vpermilv2df:
23620 error ("the last argument must be a 2-bit immediate");
23623 case CODE_FOR_avx_vextractf128v4df:
23624 case CODE_FOR_avx_vextractf128v8sf:
23625 case CODE_FOR_avx_vextractf128v8si:
23626 case CODE_FOR_avx_vinsertf128v4df:
23627 case CODE_FOR_avx_vinsertf128v8sf:
23628 case CODE_FOR_avx_vinsertf128v8si:
23629 error ("the last argument must be a 1-bit immediate");
23632 case CODE_FOR_avx_cmpsdv2df3:
23633 case CODE_FOR_avx_cmpssv4sf3:
23634 case CODE_FOR_avx_cmppdv2df3:
23635 case CODE_FOR_avx_cmppsv4sf3:
23636 case CODE_FOR_avx_cmppdv4df3:
23637 case CODE_FOR_avx_cmppsv8sf3:
23638 error ("the last argument must be a 5-bit immediate");
23642 switch (nargs_constant)
23645 if ((nargs - i) == nargs_constant)
23647 error ("the next to last argument must be an 8-bit immediate");
23651 error ("the last argument must be an 8-bit immediate");
23654 gcc_unreachable ();
23661 if (VECTOR_MODE_P (mode))
23662 op = safe_vector_operand (op, mode);
23664 /* If we aren't optimizing, only allow one memory operand to
23666 if (memory_operand (op, mode))
23669 if (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
23671 if (optimize || !match || num_memory > 1)
23672 op = copy_to_mode_reg (mode, op);
23676 op = copy_to_reg (op);
23677 op = simplify_gen_subreg (mode, op, GET_MODE (op), 0);
23682 args[i].mode = mode;
23688 pat = GEN_FCN (icode) (real_target, args[0].op);
23691 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op);
23694 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op,
23698 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op,
23699 args[2].op, args[3].op);
23702 gcc_unreachable ();
23712 /* Subroutine of ix86_expand_builtin to take care of special insns
23713 with variable number of operands. */
23716 ix86_expand_special_args_builtin (const struct builtin_description *d,
23717 tree exp, rtx target)
23721 unsigned int i, nargs, arg_adjust, memory;
23725 enum machine_mode mode;
23727 enum insn_code icode = d->icode;
23728 bool last_arg_constant = false;
23729 const struct insn_data *insn_p = &insn_data[icode];
23730 enum machine_mode tmode = insn_p->operand[0].mode;
23731 enum { load, store } klass;
23733 switch ((enum ix86_builtin_func_type) d->flag)
23735 case VOID_FTYPE_VOID:
23736 emit_insn (GEN_FCN (icode) (target));
23738 case UINT64_FTYPE_VOID:
23743 case UINT64_FTYPE_PUNSIGNED:
23744 case V2DI_FTYPE_PV2DI:
23745 case V32QI_FTYPE_PCCHAR:
23746 case V16QI_FTYPE_PCCHAR:
23747 case V8SF_FTYPE_PCV4SF:
23748 case V8SF_FTYPE_PCFLOAT:
23749 case V4SF_FTYPE_PCFLOAT:
23750 case V4DF_FTYPE_PCV2DF:
23751 case V4DF_FTYPE_PCDOUBLE:
23752 case V2DF_FTYPE_PCDOUBLE:
23753 case VOID_FTYPE_PVOID:
23758 case VOID_FTYPE_PV2SF_V4SF:
23759 case VOID_FTYPE_PV4DI_V4DI:
23760 case VOID_FTYPE_PV2DI_V2DI:
23761 case VOID_FTYPE_PCHAR_V32QI:
23762 case VOID_FTYPE_PCHAR_V16QI:
23763 case VOID_FTYPE_PFLOAT_V8SF:
23764 case VOID_FTYPE_PFLOAT_V4SF:
23765 case VOID_FTYPE_PDOUBLE_V4DF:
23766 case VOID_FTYPE_PDOUBLE_V2DF:
23767 case VOID_FTYPE_PULONGLONG_ULONGLONG:
23768 case VOID_FTYPE_PINT_INT:
23771 /* Reserve memory operand for target. */
23772 memory = ARRAY_SIZE (args);
23774 case V4SF_FTYPE_V4SF_PCV2SF:
23775 case V2DF_FTYPE_V2DF_PCDOUBLE:
23780 case V8SF_FTYPE_PCV8SF_V8SF:
23781 case V4DF_FTYPE_PCV4DF_V4DF:
23782 case V4SF_FTYPE_PCV4SF_V4SF:
23783 case V2DF_FTYPE_PCV2DF_V2DF:
23788 case VOID_FTYPE_PV8SF_V8SF_V8SF:
23789 case VOID_FTYPE_PV4DF_V4DF_V4DF:
23790 case VOID_FTYPE_PV4SF_V4SF_V4SF:
23791 case VOID_FTYPE_PV2DF_V2DF_V2DF:
23794 /* Reserve memory operand for target. */
23795 memory = ARRAY_SIZE (args);
23797 case VOID_FTYPE_UINT_UINT_UINT:
23798 case VOID_FTYPE_UINT64_UINT_UINT:
23799 case UCHAR_FTYPE_UINT_UINT_UINT:
23800 case UCHAR_FTYPE_UINT64_UINT_UINT:
23803 memory = ARRAY_SIZE (args);
23804 last_arg_constant = true;
23807 gcc_unreachable ();
23810 gcc_assert (nargs <= ARRAY_SIZE (args));
23812 if (klass == store)
23814 arg = CALL_EXPR_ARG (exp, 0);
23815 op = expand_normal (arg);
23816 gcc_assert (target == 0);
23817 target = gen_rtx_MEM (tmode, copy_to_mode_reg (Pmode, op));
23825 || GET_MODE (target) != tmode
23826 || ! (*insn_p->operand[0].predicate) (target, tmode))
23827 target = gen_reg_rtx (tmode);
23830 for (i = 0; i < nargs; i++)
23832 enum machine_mode mode = insn_p->operand[i + 1].mode;
23835 arg = CALL_EXPR_ARG (exp, i + arg_adjust);
23836 op = expand_normal (arg);
23837 match = (*insn_p->operand[i + 1].predicate) (op, mode);
23839 if (last_arg_constant && (i + 1) == nargs)
23843 if (icode == CODE_FOR_lwp_lwpvalsi3
23844 || icode == CODE_FOR_lwp_lwpinssi3
23845 || icode == CODE_FOR_lwp_lwpvaldi3
23846 || icode == CODE_FOR_lwp_lwpinsdi3)
23847 error ("the last argument must be a 32-bit immediate");
23849 error ("the last argument must be an 8-bit immediate");
23857 /* This must be the memory operand. */
23858 op = gen_rtx_MEM (mode, copy_to_mode_reg (Pmode, op));
23859 gcc_assert (GET_MODE (op) == mode
23860 || GET_MODE (op) == VOIDmode);
23864 /* This must be register. */
23865 if (VECTOR_MODE_P (mode))
23866 op = safe_vector_operand (op, mode);
23868 gcc_assert (GET_MODE (op) == mode
23869 || GET_MODE (op) == VOIDmode);
23870 op = copy_to_mode_reg (mode, op);
23875 args[i].mode = mode;
23881 pat = GEN_FCN (icode) (target);
23884 pat = GEN_FCN (icode) (target, args[0].op);
23887 pat = GEN_FCN (icode) (target, args[0].op, args[1].op);
23890 pat = GEN_FCN (icode) (target, args[0].op, args[1].op, args[2].op);
23893 gcc_unreachable ();
23899 return klass == store ? 0 : target;
23902 /* Return the integer constant in ARG. Constrain it to be in the range
23903 of the subparts of VEC_TYPE; issue an error if not. */
23906 get_element_number (tree vec_type, tree arg)
23908 unsigned HOST_WIDE_INT elt, max = TYPE_VECTOR_SUBPARTS (vec_type) - 1;
23910 if (!host_integerp (arg, 1)
23911 || (elt = tree_low_cst (arg, 1), elt > max))
23913 error ("selector must be an integer constant in the range 0..%wi", max);
23920 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
23921 ix86_expand_vector_init. We DO have language-level syntax for this, in
23922 the form of (type){ init-list }. Except that since we can't place emms
23923 instructions from inside the compiler, we can't allow the use of MMX
23924 registers unless the user explicitly asks for it. So we do *not* define
23925 vec_set/vec_extract/vec_init patterns for MMX modes in mmx.md. Instead
23926 we have builtins invoked by mmintrin.h that gives us license to emit
23927 these sorts of instructions. */
23930 ix86_expand_vec_init_builtin (tree type, tree exp, rtx target)
23932 enum machine_mode tmode = TYPE_MODE (type);
23933 enum machine_mode inner_mode = GET_MODE_INNER (tmode);
23934 int i, n_elt = GET_MODE_NUNITS (tmode);
23935 rtvec v = rtvec_alloc (n_elt);
23937 gcc_assert (VECTOR_MODE_P (tmode));
23938 gcc_assert (call_expr_nargs (exp) == n_elt);
23940 for (i = 0; i < n_elt; ++i)
23942 rtx x = expand_normal (CALL_EXPR_ARG (exp, i));
23943 RTVEC_ELT (v, i) = gen_lowpart (inner_mode, x);
23946 if (!target || !register_operand (target, tmode))
23947 target = gen_reg_rtx (tmode);
23949 ix86_expand_vector_init (true, target, gen_rtx_PARALLEL (tmode, v));
23953 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
23954 ix86_expand_vector_extract. They would be redundant (for non-MMX) if we
23955 had a language-level syntax for referencing vector elements. */
23958 ix86_expand_vec_ext_builtin (tree exp, rtx target)
23960 enum machine_mode tmode, mode0;
23965 arg0 = CALL_EXPR_ARG (exp, 0);
23966 arg1 = CALL_EXPR_ARG (exp, 1);
23968 op0 = expand_normal (arg0);
23969 elt = get_element_number (TREE_TYPE (arg0), arg1);
23971 tmode = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
23972 mode0 = TYPE_MODE (TREE_TYPE (arg0));
23973 gcc_assert (VECTOR_MODE_P (mode0));
23975 op0 = force_reg (mode0, op0);
23977 if (optimize || !target || !register_operand (target, tmode))
23978 target = gen_reg_rtx (tmode);
23980 ix86_expand_vector_extract (true, target, op0, elt);
23985 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
23986 ix86_expand_vector_set. They would be redundant (for non-MMX) if we had
23987 a language-level syntax for referencing vector elements. */
23990 ix86_expand_vec_set_builtin (tree exp)
23992 enum machine_mode tmode, mode1;
23993 tree arg0, arg1, arg2;
23995 rtx op0, op1, target;
23997 arg0 = CALL_EXPR_ARG (exp, 0);
23998 arg1 = CALL_EXPR_ARG (exp, 1);
23999 arg2 = CALL_EXPR_ARG (exp, 2);
24001 tmode = TYPE_MODE (TREE_TYPE (arg0));
24002 mode1 = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
24003 gcc_assert (VECTOR_MODE_P (tmode));
24005 op0 = expand_expr (arg0, NULL_RTX, tmode, EXPAND_NORMAL);
24006 op1 = expand_expr (arg1, NULL_RTX, mode1, EXPAND_NORMAL);
24007 elt = get_element_number (TREE_TYPE (arg0), arg2);
24009 if (GET_MODE (op1) != mode1 && GET_MODE (op1) != VOIDmode)
24010 op1 = convert_modes (mode1, GET_MODE (op1), op1, true);
24012 op0 = force_reg (tmode, op0);
24013 op1 = force_reg (mode1, op1);
24015 /* OP0 is the source of these builtin functions and shouldn't be
24016 modified. Create a copy, use it and return it as target. */
24017 target = gen_reg_rtx (tmode);
24018 emit_move_insn (target, op0);
24019 ix86_expand_vector_set (true, target, op1, elt);
24024 /* Expand an expression EXP that calls a built-in function,
24025 with result going to TARGET if that's convenient
24026 (and in mode MODE if that's convenient).
24027 SUBTARGET may be used as the target for computing one of EXP's operands.
24028 IGNORE is nonzero if the value is to be ignored. */
24031 ix86_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
24032 enum machine_mode mode ATTRIBUTE_UNUSED,
24033 int ignore ATTRIBUTE_UNUSED)
24035 const struct builtin_description *d;
24037 enum insn_code icode;
24038 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
24039 tree arg0, arg1, arg2;
24040 rtx op0, op1, op2, pat;
24041 enum machine_mode mode0, mode1, mode2;
24042 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
24044 /* Determine whether the builtin function is available under the current ISA.
24045 Originally the builtin was not created if it wasn't applicable to the
24046 current ISA based on the command line switches. With function specific
24047 options, we need to check in the context of the function making the call
24048 whether it is supported. */
24049 if (ix86_builtins_isa[fcode].isa
24050 && !(ix86_builtins_isa[fcode].isa & ix86_isa_flags))
24052 char *opts = ix86_target_string (ix86_builtins_isa[fcode].isa, 0, NULL,
24053 NULL, NULL, false);
24056 error ("%qE needs unknown isa option", fndecl);
24059 gcc_assert (opts != NULL);
24060 error ("%qE needs isa option %s", fndecl, opts);
24068 case IX86_BUILTIN_MASKMOVQ:
24069 case IX86_BUILTIN_MASKMOVDQU:
24070 icode = (fcode == IX86_BUILTIN_MASKMOVQ
24071 ? CODE_FOR_mmx_maskmovq
24072 : CODE_FOR_sse2_maskmovdqu);
24073 /* Note the arg order is different from the operand order. */
24074 arg1 = CALL_EXPR_ARG (exp, 0);
24075 arg2 = CALL_EXPR_ARG (exp, 1);
24076 arg0 = CALL_EXPR_ARG (exp, 2);
24077 op0 = expand_normal (arg0);
24078 op1 = expand_normal (arg1);
24079 op2 = expand_normal (arg2);
24080 mode0 = insn_data[icode].operand[0].mode;
24081 mode1 = insn_data[icode].operand[1].mode;
24082 mode2 = insn_data[icode].operand[2].mode;
24084 op0 = force_reg (Pmode, op0);
24085 op0 = gen_rtx_MEM (mode1, op0);
24087 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
24088 op0 = copy_to_mode_reg (mode0, op0);
24089 if (! (*insn_data[icode].operand[1].predicate) (op1, mode1))
24090 op1 = copy_to_mode_reg (mode1, op1);
24091 if (! (*insn_data[icode].operand[2].predicate) (op2, mode2))
24092 op2 = copy_to_mode_reg (mode2, op2);
24093 pat = GEN_FCN (icode) (op0, op1, op2);
24099 case IX86_BUILTIN_LDMXCSR:
24100 op0 = expand_normal (CALL_EXPR_ARG (exp, 0));
24101 target = assign_386_stack_local (SImode, SLOT_VIRTUAL);
24102 emit_move_insn (target, op0);
24103 emit_insn (gen_sse_ldmxcsr (target));
24106 case IX86_BUILTIN_STMXCSR:
24107 target = assign_386_stack_local (SImode, SLOT_VIRTUAL);
24108 emit_insn (gen_sse_stmxcsr (target));
24109 return copy_to_mode_reg (SImode, target);
24111 case IX86_BUILTIN_CLFLUSH:
24112 arg0 = CALL_EXPR_ARG (exp, 0);
24113 op0 = expand_normal (arg0);
24114 icode = CODE_FOR_sse2_clflush;
24115 if (! (*insn_data[icode].operand[0].predicate) (op0, Pmode))
24116 op0 = copy_to_mode_reg (Pmode, op0);
24118 emit_insn (gen_sse2_clflush (op0));
24121 case IX86_BUILTIN_MONITOR:
24122 arg0 = CALL_EXPR_ARG (exp, 0);
24123 arg1 = CALL_EXPR_ARG (exp, 1);
24124 arg2 = CALL_EXPR_ARG (exp, 2);
24125 op0 = expand_normal (arg0);
24126 op1 = expand_normal (arg1);
24127 op2 = expand_normal (arg2);
24129 op0 = copy_to_mode_reg (Pmode, op0);
24131 op1 = copy_to_mode_reg (SImode, op1);
24133 op2 = copy_to_mode_reg (SImode, op2);
24134 emit_insn ((*ix86_gen_monitor) (op0, op1, op2));
24137 case IX86_BUILTIN_MWAIT:
24138 arg0 = CALL_EXPR_ARG (exp, 0);
24139 arg1 = CALL_EXPR_ARG (exp, 1);
24140 op0 = expand_normal (arg0);
24141 op1 = expand_normal (arg1);
24143 op0 = copy_to_mode_reg (SImode, op0);
24145 op1 = copy_to_mode_reg (SImode, op1);
24146 emit_insn (gen_sse3_mwait (op0, op1));
24149 case IX86_BUILTIN_VEC_INIT_V2SI:
24150 case IX86_BUILTIN_VEC_INIT_V4HI:
24151 case IX86_BUILTIN_VEC_INIT_V8QI:
24152 return ix86_expand_vec_init_builtin (TREE_TYPE (exp), exp, target);
24154 case IX86_BUILTIN_VEC_EXT_V2DF:
24155 case IX86_BUILTIN_VEC_EXT_V2DI:
24156 case IX86_BUILTIN_VEC_EXT_V4SF:
24157 case IX86_BUILTIN_VEC_EXT_V4SI:
24158 case IX86_BUILTIN_VEC_EXT_V8HI:
24159 case IX86_BUILTIN_VEC_EXT_V2SI:
24160 case IX86_BUILTIN_VEC_EXT_V4HI:
24161 case IX86_BUILTIN_VEC_EXT_V16QI:
24162 return ix86_expand_vec_ext_builtin (exp, target);
24164 case IX86_BUILTIN_VEC_SET_V2DI:
24165 case IX86_BUILTIN_VEC_SET_V4SF:
24166 case IX86_BUILTIN_VEC_SET_V4SI:
24167 case IX86_BUILTIN_VEC_SET_V8HI:
24168 case IX86_BUILTIN_VEC_SET_V4HI:
24169 case IX86_BUILTIN_VEC_SET_V16QI:
24170 return ix86_expand_vec_set_builtin (exp);
24172 case IX86_BUILTIN_VEC_PERM_V2DF:
24173 case IX86_BUILTIN_VEC_PERM_V4SF:
24174 case IX86_BUILTIN_VEC_PERM_V2DI:
24175 case IX86_BUILTIN_VEC_PERM_V4SI:
24176 case IX86_BUILTIN_VEC_PERM_V8HI:
24177 case IX86_BUILTIN_VEC_PERM_V16QI:
24178 case IX86_BUILTIN_VEC_PERM_V2DI_U:
24179 case IX86_BUILTIN_VEC_PERM_V4SI_U:
24180 case IX86_BUILTIN_VEC_PERM_V8HI_U:
24181 case IX86_BUILTIN_VEC_PERM_V16QI_U:
24182 case IX86_BUILTIN_VEC_PERM_V4DF:
24183 case IX86_BUILTIN_VEC_PERM_V8SF:
24184 return ix86_expand_vec_perm_builtin (exp);
24186 case IX86_BUILTIN_INFQ:
24187 case IX86_BUILTIN_HUGE_VALQ:
24189 REAL_VALUE_TYPE inf;
24193 tmp = CONST_DOUBLE_FROM_REAL_VALUE (inf, mode);
24195 tmp = validize_mem (force_const_mem (mode, tmp));
24198 target = gen_reg_rtx (mode);
24200 emit_move_insn (target, tmp);
24204 case IX86_BUILTIN_LLWPCB:
24205 arg0 = CALL_EXPR_ARG (exp, 0);
24206 op0 = expand_normal (arg0);
24207 icode = CODE_FOR_lwp_llwpcb;
24208 if (! (*insn_data[icode].operand[0].predicate) (op0, Pmode))
24209 op0 = copy_to_mode_reg (Pmode, op0);
24210 emit_insn (gen_lwp_llwpcb (op0));
24213 case IX86_BUILTIN_SLWPCB:
24214 icode = CODE_FOR_lwp_slwpcb;
24216 || ! (*insn_data[icode].operand[0].predicate) (target, Pmode))
24217 target = gen_reg_rtx (Pmode);
24218 emit_insn (gen_lwp_slwpcb (target));
24225 for (i = 0, d = bdesc_special_args;
24226 i < ARRAY_SIZE (bdesc_special_args);
24228 if (d->code == fcode)
24229 return ix86_expand_special_args_builtin (d, exp, target);
24231 for (i = 0, d = bdesc_args;
24232 i < ARRAY_SIZE (bdesc_args);
24234 if (d->code == fcode)
24237 case IX86_BUILTIN_FABSQ:
24238 case IX86_BUILTIN_COPYSIGNQ:
24240 /* Emit a normal call if SSE2 isn't available. */
24241 return expand_call (exp, target, ignore);
24243 return ix86_expand_args_builtin (d, exp, target);
24246 for (i = 0, d = bdesc_comi; i < ARRAY_SIZE (bdesc_comi); i++, d++)
24247 if (d->code == fcode)
24248 return ix86_expand_sse_comi (d, exp, target);
24250 for (i = 0, d = bdesc_pcmpestr;
24251 i < ARRAY_SIZE (bdesc_pcmpestr);
24253 if (d->code == fcode)
24254 return ix86_expand_sse_pcmpestr (d, exp, target);
24256 for (i = 0, d = bdesc_pcmpistr;
24257 i < ARRAY_SIZE (bdesc_pcmpistr);
24259 if (d->code == fcode)
24260 return ix86_expand_sse_pcmpistr (d, exp, target);
24262 for (i = 0, d = bdesc_multi_arg; i < ARRAY_SIZE (bdesc_multi_arg); i++, d++)
24263 if (d->code == fcode)
24264 return ix86_expand_multi_arg_builtin (d->icode, exp, target,
24265 (enum ix86_builtin_func_type)
24266 d->flag, d->comparison);
24268 gcc_unreachable ();
24271 /* Returns a function decl for a vectorized version of the builtin function
24272 with builtin function code FN and the result vector type TYPE, or NULL_TREE
24273 if it is not available. */
24276 ix86_builtin_vectorized_function (unsigned int fn, tree type_out,
24279 enum machine_mode in_mode, out_mode;
24282 if (TREE_CODE (type_out) != VECTOR_TYPE
24283 || TREE_CODE (type_in) != VECTOR_TYPE)
24286 out_mode = TYPE_MODE (TREE_TYPE (type_out));
24287 out_n = TYPE_VECTOR_SUBPARTS (type_out);
24288 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24289 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24293 case BUILT_IN_SQRT:
24294 if (out_mode == DFmode && out_n == 2
24295 && in_mode == DFmode && in_n == 2)
24296 return ix86_builtins[IX86_BUILTIN_SQRTPD];
24299 case BUILT_IN_SQRTF:
24300 if (out_mode == SFmode && out_n == 4
24301 && in_mode == SFmode && in_n == 4)
24302 return ix86_builtins[IX86_BUILTIN_SQRTPS_NR];
24305 case BUILT_IN_LRINT:
24306 if (out_mode == SImode && out_n == 4
24307 && in_mode == DFmode && in_n == 2)
24308 return ix86_builtins[IX86_BUILTIN_VEC_PACK_SFIX];
24311 case BUILT_IN_LRINTF:
24312 if (out_mode == SImode && out_n == 4
24313 && in_mode == SFmode && in_n == 4)
24314 return ix86_builtins[IX86_BUILTIN_CVTPS2DQ];
24317 case BUILT_IN_COPYSIGN:
24318 if (out_mode == DFmode && out_n == 2
24319 && in_mode == DFmode && in_n == 2)
24320 return ix86_builtins[IX86_BUILTIN_CPYSGNPD];
24323 case BUILT_IN_COPYSIGNF:
24324 if (out_mode == SFmode && out_n == 4
24325 && in_mode == SFmode && in_n == 4)
24326 return ix86_builtins[IX86_BUILTIN_CPYSGNPS];
24333 /* Dispatch to a handler for a vectorization library. */
24334 if (ix86_veclib_handler)
24335 return (*ix86_veclib_handler) ((enum built_in_function) fn, type_out,
24341 /* Handler for an SVML-style interface to
24342 a library with vectorized intrinsics. */
24345 ix86_veclibabi_svml (enum built_in_function fn, tree type_out, tree type_in)
24348 tree fntype, new_fndecl, args;
24351 enum machine_mode el_mode, in_mode;
24354 /* The SVML is suitable for unsafe math only. */
24355 if (!flag_unsafe_math_optimizations)
24358 el_mode = TYPE_MODE (TREE_TYPE (type_out));
24359 n = TYPE_VECTOR_SUBPARTS (type_out);
24360 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24361 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24362 if (el_mode != in_mode
24370 case BUILT_IN_LOG10:
24372 case BUILT_IN_TANH:
24374 case BUILT_IN_ATAN:
24375 case BUILT_IN_ATAN2:
24376 case BUILT_IN_ATANH:
24377 case BUILT_IN_CBRT:
24378 case BUILT_IN_SINH:
24380 case BUILT_IN_ASINH:
24381 case BUILT_IN_ASIN:
24382 case BUILT_IN_COSH:
24384 case BUILT_IN_ACOSH:
24385 case BUILT_IN_ACOS:
24386 if (el_mode != DFmode || n != 2)
24390 case BUILT_IN_EXPF:
24391 case BUILT_IN_LOGF:
24392 case BUILT_IN_LOG10F:
24393 case BUILT_IN_POWF:
24394 case BUILT_IN_TANHF:
24395 case BUILT_IN_TANF:
24396 case BUILT_IN_ATANF:
24397 case BUILT_IN_ATAN2F:
24398 case BUILT_IN_ATANHF:
24399 case BUILT_IN_CBRTF:
24400 case BUILT_IN_SINHF:
24401 case BUILT_IN_SINF:
24402 case BUILT_IN_ASINHF:
24403 case BUILT_IN_ASINF:
24404 case BUILT_IN_COSHF:
24405 case BUILT_IN_COSF:
24406 case BUILT_IN_ACOSHF:
24407 case BUILT_IN_ACOSF:
24408 if (el_mode != SFmode || n != 4)
24416 bname = IDENTIFIER_POINTER (DECL_NAME (implicit_built_in_decls[fn]));
24418 if (fn == BUILT_IN_LOGF)
24419 strcpy (name, "vmlsLn4");
24420 else if (fn == BUILT_IN_LOG)
24421 strcpy (name, "vmldLn2");
24424 sprintf (name, "vmls%s", bname+10);
24425 name[strlen (name)-1] = '4';
24428 sprintf (name, "vmld%s2", bname+10);
24430 /* Convert to uppercase. */
24434 for (args = DECL_ARGUMENTS (implicit_built_in_decls[fn]); args;
24435 args = TREE_CHAIN (args))
24439 fntype = build_function_type_list (type_out, type_in, NULL);
24441 fntype = build_function_type_list (type_out, type_in, type_in, NULL);
24443 /* Build a function declaration for the vectorized function. */
24444 new_fndecl = build_decl (BUILTINS_LOCATION,
24445 FUNCTION_DECL, get_identifier (name), fntype);
24446 TREE_PUBLIC (new_fndecl) = 1;
24447 DECL_EXTERNAL (new_fndecl) = 1;
24448 DECL_IS_NOVOPS (new_fndecl) = 1;
24449 TREE_READONLY (new_fndecl) = 1;
24454 /* Handler for an ACML-style interface to
24455 a library with vectorized intrinsics. */
24458 ix86_veclibabi_acml (enum built_in_function fn, tree type_out, tree type_in)
24460 char name[20] = "__vr.._";
24461 tree fntype, new_fndecl, args;
24464 enum machine_mode el_mode, in_mode;
24467 /* The ACML is 64bits only and suitable for unsafe math only as
24468 it does not correctly support parts of IEEE with the required
24469 precision such as denormals. */
24471 || !flag_unsafe_math_optimizations)
24474 el_mode = TYPE_MODE (TREE_TYPE (type_out));
24475 n = TYPE_VECTOR_SUBPARTS (type_out);
24476 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24477 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24478 if (el_mode != in_mode
24488 case BUILT_IN_LOG2:
24489 case BUILT_IN_LOG10:
24492 if (el_mode != DFmode
24497 case BUILT_IN_SINF:
24498 case BUILT_IN_COSF:
24499 case BUILT_IN_EXPF:
24500 case BUILT_IN_POWF:
24501 case BUILT_IN_LOGF:
24502 case BUILT_IN_LOG2F:
24503 case BUILT_IN_LOG10F:
24506 if (el_mode != SFmode
24515 bname = IDENTIFIER_POINTER (DECL_NAME (implicit_built_in_decls[fn]));
24516 sprintf (name + 7, "%s", bname+10);
24519 for (args = DECL_ARGUMENTS (implicit_built_in_decls[fn]); args;
24520 args = TREE_CHAIN (args))
24524 fntype = build_function_type_list (type_out, type_in, NULL);
24526 fntype = build_function_type_list (type_out, type_in, type_in, NULL);
24528 /* Build a function declaration for the vectorized function. */
24529 new_fndecl = build_decl (BUILTINS_LOCATION,
24530 FUNCTION_DECL, get_identifier (name), fntype);
24531 TREE_PUBLIC (new_fndecl) = 1;
24532 DECL_EXTERNAL (new_fndecl) = 1;
24533 DECL_IS_NOVOPS (new_fndecl) = 1;
24534 TREE_READONLY (new_fndecl) = 1;
24540 /* Returns a decl of a function that implements conversion of an integer vector
24541 into a floating-point vector, or vice-versa. TYPE is the type of the integer
24542 side of the conversion.
24543 Return NULL_TREE if it is not available. */
24546 ix86_vectorize_builtin_conversion (unsigned int code, tree type)
24548 if (! (TARGET_SSE2 && TREE_CODE (type) == VECTOR_TYPE))
24554 switch (TYPE_MODE (type))
24557 return TYPE_UNSIGNED (type)
24558 ? ix86_builtins[IX86_BUILTIN_CVTUDQ2PS]
24559 : ix86_builtins[IX86_BUILTIN_CVTDQ2PS];
24564 case FIX_TRUNC_EXPR:
24565 switch (TYPE_MODE (type))
24568 return TYPE_UNSIGNED (type)
24570 : ix86_builtins[IX86_BUILTIN_CVTTPS2DQ];
24580 /* Returns a code for a target-specific builtin that implements
24581 reciprocal of the function, or NULL_TREE if not available. */
24584 ix86_builtin_reciprocal (unsigned int fn, bool md_fn,
24585 bool sqrt ATTRIBUTE_UNUSED)
24587 if (! (TARGET_SSE_MATH && !optimize_insn_for_size_p ()
24588 && flag_finite_math_only && !flag_trapping_math
24589 && flag_unsafe_math_optimizations))
24593 /* Machine dependent builtins. */
24596 /* Vectorized version of sqrt to rsqrt conversion. */
24597 case IX86_BUILTIN_SQRTPS_NR:
24598 return ix86_builtins[IX86_BUILTIN_RSQRTPS_NR];
24604 /* Normal builtins. */
24607 /* Sqrt to rsqrt conversion. */
24608 case BUILT_IN_SQRTF:
24609 return ix86_builtins[IX86_BUILTIN_RSQRTF];
24616 /* Helper for avx_vpermilps256_operand et al. This is also used by
24617 the expansion functions to turn the parallel back into a mask.
24618 The return value is 0 for no match and the imm8+1 for a match. */
24621 avx_vpermilp_parallel (rtx par, enum machine_mode mode)
24623 unsigned i, nelt = GET_MODE_NUNITS (mode);
24625 unsigned char ipar[8];
24627 if (XVECLEN (par, 0) != (int) nelt)
24630 /* Validate that all of the elements are constants, and not totally
24631 out of range. Copy the data into an integral array to make the
24632 subsequent checks easier. */
24633 for (i = 0; i < nelt; ++i)
24635 rtx er = XVECEXP (par, 0, i);
24636 unsigned HOST_WIDE_INT ei;
24638 if (!CONST_INT_P (er))
24641 if (ei >= 2 * nelt)
24649 /* In the 256-bit DFmode case, we can only move elements within
24651 for (i = 0; i < 2; ++i)
24655 mask |= ipar[i] << i;
24657 for (i = 2; i < 4; ++i)
24661 mask |= (ipar[i] - 2) << i;
24666 /* In the 256-bit SFmode case, we have full freedom of movement
24667 within the low 128-bit lane, but the high 128-bit lane must
24668 mirror the exact same pattern. */
24669 for (i = 0; i < 4; ++i)
24670 if (ipar[i] + 4 != ipar[i + 4])
24677 /* In the 128-bit case, we've full freedom in the placement of
24678 the elements from the source operand. */
24679 for (i = 0; i < nelt; ++i)
24680 mask |= ipar[i] << (i * (nelt / 2));
24684 gcc_unreachable ();
24687 /* Make sure success has a non-zero value by adding one. */
24691 /* Helper for avx_vperm2f128_v4df_operand et al. This is also used by
24692 the expansion functions to turn the parallel back into a mask.
24693 The return value is 0 for no match and the imm8+1 for a match. */
24696 avx_vperm2f128_parallel (rtx par, enum machine_mode mode)
24698 unsigned i, nelt = GET_MODE_NUNITS (mode), nelt2 = nelt / 2;
24700 unsigned char ipar[8];
24702 if (XVECLEN (par, 0) != (int) nelt)
24705 /* Validate that all of the elements are constants, and not totally
24706 out of range. Copy the data into an integral array to make the
24707 subsequent checks easier. */
24708 for (i = 0; i < nelt; ++i)
24710 rtx er = XVECEXP (par, 0, i);
24711 unsigned HOST_WIDE_INT ei;
24713 if (!CONST_INT_P (er))
24716 if (ei >= 2 * nelt)
24721 /* Validate that the halves of the permute are halves. */
24722 for (i = 0; i < nelt2 - 1; ++i)
24723 if (ipar[i] + 1 != ipar[i + 1])
24725 for (i = nelt2; i < nelt - 1; ++i)
24726 if (ipar[i] + 1 != ipar[i + 1])
24729 /* Reconstruct the mask. */
24730 for (i = 0; i < 2; ++i)
24732 unsigned e = ipar[i * nelt2];
24736 mask |= e << (i * 4);
24739 /* Make sure success has a non-zero value by adding one. */
24744 /* Store OPERAND to the memory after reload is completed. This means
24745 that we can't easily use assign_stack_local. */
24747 ix86_force_to_memory (enum machine_mode mode, rtx operand)
24751 gcc_assert (reload_completed);
24752 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE)
24754 result = gen_rtx_MEM (mode,
24755 gen_rtx_PLUS (Pmode,
24757 GEN_INT (-RED_ZONE_SIZE)));
24758 emit_move_insn (result, operand);
24760 else if ((TARGET_64BIT_MS_ABI || !TARGET_RED_ZONE) && TARGET_64BIT)
24766 operand = gen_lowpart (DImode, operand);
24770 gen_rtx_SET (VOIDmode,
24771 gen_rtx_MEM (DImode,
24772 gen_rtx_PRE_DEC (DImode,
24773 stack_pointer_rtx)),
24777 gcc_unreachable ();
24779 result = gen_rtx_MEM (mode, stack_pointer_rtx);
24788 split_di (&operand, 1, operands, operands + 1);
24790 gen_rtx_SET (VOIDmode,
24791 gen_rtx_MEM (SImode,
24792 gen_rtx_PRE_DEC (Pmode,
24793 stack_pointer_rtx)),
24796 gen_rtx_SET (VOIDmode,
24797 gen_rtx_MEM (SImode,
24798 gen_rtx_PRE_DEC (Pmode,
24799 stack_pointer_rtx)),
24804 /* Store HImodes as SImodes. */
24805 operand = gen_lowpart (SImode, operand);
24809 gen_rtx_SET (VOIDmode,
24810 gen_rtx_MEM (GET_MODE (operand),
24811 gen_rtx_PRE_DEC (SImode,
24812 stack_pointer_rtx)),
24816 gcc_unreachable ();
24818 result = gen_rtx_MEM (mode, stack_pointer_rtx);
24823 /* Free operand from the memory. */
24825 ix86_free_from_memory (enum machine_mode mode)
24827 if (!TARGET_RED_ZONE || TARGET_64BIT_MS_ABI)
24831 if (mode == DImode || TARGET_64BIT)
24835 /* Use LEA to deallocate stack space. In peephole2 it will be converted
24836 to pop or add instruction if registers are available. */
24837 emit_insn (gen_rtx_SET (VOIDmode, stack_pointer_rtx,
24838 gen_rtx_PLUS (Pmode, stack_pointer_rtx,
24843 /* Implement TARGET_IRA_COVER_CLASSES. If -mfpmath=sse, we prefer
24844 SSE_REGS to FLOAT_REGS if their costs for a pseudo are the
24846 static const enum reg_class *
24847 i386_ira_cover_classes (void)
24849 static const enum reg_class sse_fpmath_classes[] = {
24850 GENERAL_REGS, SSE_REGS, MMX_REGS, FLOAT_REGS, LIM_REG_CLASSES
24852 static const enum reg_class no_sse_fpmath_classes[] = {
24853 GENERAL_REGS, FLOAT_REGS, MMX_REGS, SSE_REGS, LIM_REG_CLASSES
24856 return TARGET_SSE_MATH ? sse_fpmath_classes : no_sse_fpmath_classes;
24859 /* Put float CONST_DOUBLE in the constant pool instead of fp regs.
24860 QImode must go into class Q_REGS.
24861 Narrow ALL_REGS to GENERAL_REGS. This supports allowing movsf and
24862 movdf to do mem-to-mem moves through integer regs. */
24864 ix86_preferred_reload_class (rtx x, enum reg_class regclass)
24866 enum machine_mode mode = GET_MODE (x);
24868 /* We're only allowed to return a subclass of CLASS. Many of the
24869 following checks fail for NO_REGS, so eliminate that early. */
24870 if (regclass == NO_REGS)
24873 /* All classes can load zeros. */
24874 if (x == CONST0_RTX (mode))
24877 /* Force constants into memory if we are loading a (nonzero) constant into
24878 an MMX or SSE register. This is because there are no MMX/SSE instructions
24879 to load from a constant. */
24881 && (MAYBE_MMX_CLASS_P (regclass) || MAYBE_SSE_CLASS_P (regclass)))
24884 /* Prefer SSE regs only, if we can use them for math. */
24885 if (TARGET_SSE_MATH && !TARGET_MIX_SSE_I387 && SSE_FLOAT_MODE_P (mode))
24886 return SSE_CLASS_P (regclass) ? regclass : NO_REGS;
24888 /* Floating-point constants need more complex checks. */
24889 if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) != VOIDmode)
24891 /* General regs can load everything. */
24892 if (reg_class_subset_p (regclass, GENERAL_REGS))
24895 /* Floats can load 0 and 1 plus some others. Note that we eliminated
24896 zero above. We only want to wind up preferring 80387 registers if
24897 we plan on doing computation with them. */
24899 && standard_80387_constant_p (x))
24901 /* Limit class to non-sse. */
24902 if (regclass == FLOAT_SSE_REGS)
24904 if (regclass == FP_TOP_SSE_REGS)
24906 if (regclass == FP_SECOND_SSE_REGS)
24907 return FP_SECOND_REG;
24908 if (regclass == FLOAT_INT_REGS || regclass == FLOAT_REGS)
24915 /* Generally when we see PLUS here, it's the function invariant
24916 (plus soft-fp const_int). Which can only be computed into general
24918 if (GET_CODE (x) == PLUS)
24919 return reg_class_subset_p (regclass, GENERAL_REGS) ? regclass : NO_REGS;
24921 /* QImode constants are easy to load, but non-constant QImode data
24922 must go into Q_REGS. */
24923 if (GET_MODE (x) == QImode && !CONSTANT_P (x))
24925 if (reg_class_subset_p (regclass, Q_REGS))
24927 if (reg_class_subset_p (Q_REGS, regclass))
24935 /* Discourage putting floating-point values in SSE registers unless
24936 SSE math is being used, and likewise for the 387 registers. */
24938 ix86_preferred_output_reload_class (rtx x, enum reg_class regclass)
24940 enum machine_mode mode = GET_MODE (x);
24942 /* Restrict the output reload class to the register bank that we are doing
24943 math on. If we would like not to return a subset of CLASS, reject this
24944 alternative: if reload cannot do this, it will still use its choice. */
24945 mode = GET_MODE (x);
24946 if (TARGET_SSE_MATH && SSE_FLOAT_MODE_P (mode))
24947 return MAYBE_SSE_CLASS_P (regclass) ? SSE_REGS : NO_REGS;
24949 if (X87_FLOAT_MODE_P (mode))
24951 if (regclass == FP_TOP_SSE_REGS)
24953 else if (regclass == FP_SECOND_SSE_REGS)
24954 return FP_SECOND_REG;
24956 return FLOAT_CLASS_P (regclass) ? regclass : NO_REGS;
24962 static enum reg_class
24963 ix86_secondary_reload (bool in_p, rtx x, enum reg_class rclass,
24964 enum machine_mode mode,
24965 secondary_reload_info *sri ATTRIBUTE_UNUSED)
24967 /* QImode spills from non-QI registers require
24968 intermediate register on 32bit targets. */
24969 if (!in_p && mode == QImode && !TARGET_64BIT
24970 && (rclass == GENERAL_REGS
24971 || rclass == LEGACY_REGS
24972 || rclass == INDEX_REGS))
24981 if (regno >= FIRST_PSEUDO_REGISTER || GET_CODE (x) == SUBREG)
24982 regno = true_regnum (x);
24984 /* Return Q_REGS if the operand is in memory. */
24992 /* If we are copying between general and FP registers, we need a memory
24993 location. The same is true for SSE and MMX registers.
24995 To optimize register_move_cost performance, allow inline variant.
24997 The macro can't work reliably when one of the CLASSES is class containing
24998 registers from multiple units (SSE, MMX, integer). We avoid this by never
24999 combining those units in single alternative in the machine description.
25000 Ensure that this constraint holds to avoid unexpected surprises.
25002 When STRICT is false, we are being called from REGISTER_MOVE_COST, so do not
25003 enforce these sanity checks. */
25006 inline_secondary_memory_needed (enum reg_class class1, enum reg_class class2,
25007 enum machine_mode mode, int strict)
25009 if (MAYBE_FLOAT_CLASS_P (class1) != FLOAT_CLASS_P (class1)
25010 || MAYBE_FLOAT_CLASS_P (class2) != FLOAT_CLASS_P (class2)
25011 || MAYBE_SSE_CLASS_P (class1) != SSE_CLASS_P (class1)
25012 || MAYBE_SSE_CLASS_P (class2) != SSE_CLASS_P (class2)
25013 || MAYBE_MMX_CLASS_P (class1) != MMX_CLASS_P (class1)
25014 || MAYBE_MMX_CLASS_P (class2) != MMX_CLASS_P (class2))
25016 gcc_assert (!strict);
25020 if (FLOAT_CLASS_P (class1) != FLOAT_CLASS_P (class2))
25023 /* ??? This is a lie. We do have moves between mmx/general, and for
25024 mmx/sse2. But by saying we need secondary memory we discourage the
25025 register allocator from using the mmx registers unless needed. */
25026 if (MMX_CLASS_P (class1) != MMX_CLASS_P (class2))
25029 if (SSE_CLASS_P (class1) != SSE_CLASS_P (class2))
25031 /* SSE1 doesn't have any direct moves from other classes. */
25035 /* If the target says that inter-unit moves are more expensive
25036 than moving through memory, then don't generate them. */
25037 if (!TARGET_INTER_UNIT_MOVES)
25040 /* Between SSE and general, we have moves no larger than word size. */
25041 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD)
25049 ix86_secondary_memory_needed (enum reg_class class1, enum reg_class class2,
25050 enum machine_mode mode, int strict)
25052 return inline_secondary_memory_needed (class1, class2, mode, strict);
25055 /* Return true if the registers in CLASS cannot represent the change from
25056 modes FROM to TO. */
25059 ix86_cannot_change_mode_class (enum machine_mode from, enum machine_mode to,
25060 enum reg_class regclass)
25065 /* x87 registers can't do subreg at all, as all values are reformatted
25066 to extended precision. */
25067 if (MAYBE_FLOAT_CLASS_P (regclass))
25070 if (MAYBE_SSE_CLASS_P (regclass) || MAYBE_MMX_CLASS_P (regclass))
25072 /* Vector registers do not support QI or HImode loads. If we don't
25073 disallow a change to these modes, reload will assume it's ok to
25074 drop the subreg from (subreg:SI (reg:HI 100) 0). This affects
25075 the vec_dupv4hi pattern. */
25076 if (GET_MODE_SIZE (from) < 4)
25079 /* Vector registers do not support subreg with nonzero offsets, which
25080 are otherwise valid for integer registers. Since we can't see
25081 whether we have a nonzero offset from here, prohibit all
25082 nonparadoxical subregs changing size. */
25083 if (GET_MODE_SIZE (to) < GET_MODE_SIZE (from))
25090 /* Return the cost of moving data of mode M between a
25091 register and memory. A value of 2 is the default; this cost is
25092 relative to those in `REGISTER_MOVE_COST'.
25094 This function is used extensively by register_move_cost that is used to
25095 build tables at startup. Make it inline in this case.
25096 When IN is 2, return maximum of in and out move cost.
25098 If moving between registers and memory is more expensive than
25099 between two registers, you should define this macro to express the
25102 Model also increased moving costs of QImode registers in non
25106 inline_memory_move_cost (enum machine_mode mode, enum reg_class regclass,
25110 if (FLOAT_CLASS_P (regclass))
25128 return MAX (ix86_cost->fp_load [index], ix86_cost->fp_store [index]);
25129 return in ? ix86_cost->fp_load [index] : ix86_cost->fp_store [index];
25131 if (SSE_CLASS_P (regclass))
25134 switch (GET_MODE_SIZE (mode))
25149 return MAX (ix86_cost->sse_load [index], ix86_cost->sse_store [index]);
25150 return in ? ix86_cost->sse_load [index] : ix86_cost->sse_store [index];
25152 if (MMX_CLASS_P (regclass))
25155 switch (GET_MODE_SIZE (mode))
25167 return MAX (ix86_cost->mmx_load [index], ix86_cost->mmx_store [index]);
25168 return in ? ix86_cost->mmx_load [index] : ix86_cost->mmx_store [index];
25170 switch (GET_MODE_SIZE (mode))
25173 if (Q_CLASS_P (regclass) || TARGET_64BIT)
25176 return ix86_cost->int_store[0];
25177 if (TARGET_PARTIAL_REG_DEPENDENCY
25178 && optimize_function_for_speed_p (cfun))
25179 cost = ix86_cost->movzbl_load;
25181 cost = ix86_cost->int_load[0];
25183 return MAX (cost, ix86_cost->int_store[0]);
25189 return MAX (ix86_cost->movzbl_load, ix86_cost->int_store[0] + 4);
25191 return ix86_cost->movzbl_load;
25193 return ix86_cost->int_store[0] + 4;
25198 return MAX (ix86_cost->int_load[1], ix86_cost->int_store[1]);
25199 return in ? ix86_cost->int_load[1] : ix86_cost->int_store[1];
25201 /* Compute number of 32bit moves needed. TFmode is moved as XFmode. */
25202 if (mode == TFmode)
25205 cost = MAX (ix86_cost->int_load[2] , ix86_cost->int_store[2]);
25207 cost = ix86_cost->int_load[2];
25209 cost = ix86_cost->int_store[2];
25210 return (cost * (((int) GET_MODE_SIZE (mode)
25211 + UNITS_PER_WORD - 1) / UNITS_PER_WORD));
25216 ix86_memory_move_cost (enum machine_mode mode, enum reg_class regclass, int in)
25218 return inline_memory_move_cost (mode, regclass, in);
25222 /* Return the cost of moving data from a register in class CLASS1 to
25223 one in class CLASS2.
25225 It is not required that the cost always equal 2 when FROM is the same as TO;
25226 on some machines it is expensive to move between registers if they are not
25227 general registers. */
25230 ix86_register_move_cost (enum machine_mode mode, enum reg_class class1,
25231 enum reg_class class2)
25233 /* In case we require secondary memory, compute cost of the store followed
25234 by load. In order to avoid bad register allocation choices, we need
25235 for this to be *at least* as high as the symmetric MEMORY_MOVE_COST. */
25237 if (inline_secondary_memory_needed (class1, class2, mode, 0))
25241 cost += inline_memory_move_cost (mode, class1, 2);
25242 cost += inline_memory_move_cost (mode, class2, 2);
25244 /* In case of copying from general_purpose_register we may emit multiple
25245 stores followed by single load causing memory size mismatch stall.
25246 Count this as arbitrarily high cost of 20. */
25247 if (CLASS_MAX_NREGS (class1, mode) > CLASS_MAX_NREGS (class2, mode))
25250 /* In the case of FP/MMX moves, the registers actually overlap, and we
25251 have to switch modes in order to treat them differently. */
25252 if ((MMX_CLASS_P (class1) && MAYBE_FLOAT_CLASS_P (class2))
25253 || (MMX_CLASS_P (class2) && MAYBE_FLOAT_CLASS_P (class1)))
25259 /* Moves between SSE/MMX and integer unit are expensive. */
25260 if (MMX_CLASS_P (class1) != MMX_CLASS_P (class2)
25261 || SSE_CLASS_P (class1) != SSE_CLASS_P (class2))
25263 /* ??? By keeping returned value relatively high, we limit the number
25264 of moves between integer and MMX/SSE registers for all targets.
25265 Additionally, high value prevents problem with x86_modes_tieable_p(),
25266 where integer modes in MMX/SSE registers are not tieable
25267 because of missing QImode and HImode moves to, from or between
25268 MMX/SSE registers. */
25269 return MAX (8, ix86_cost->mmxsse_to_integer);
25271 if (MAYBE_FLOAT_CLASS_P (class1))
25272 return ix86_cost->fp_move;
25273 if (MAYBE_SSE_CLASS_P (class1))
25274 return ix86_cost->sse_move;
25275 if (MAYBE_MMX_CLASS_P (class1))
25276 return ix86_cost->mmx_move;
25280 /* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
25283 ix86_hard_regno_mode_ok (int regno, enum machine_mode mode)
25285 /* Flags and only flags can only hold CCmode values. */
25286 if (CC_REGNO_P (regno))
25287 return GET_MODE_CLASS (mode) == MODE_CC;
25288 if (GET_MODE_CLASS (mode) == MODE_CC
25289 || GET_MODE_CLASS (mode) == MODE_RANDOM
25290 || GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
25292 if (FP_REGNO_P (regno))
25293 return VALID_FP_MODE_P (mode);
25294 if (SSE_REGNO_P (regno))
25296 /* We implement the move patterns for all vector modes into and
25297 out of SSE registers, even when no operation instructions
25298 are available. OImode move is available only when AVX is
25300 return ((TARGET_AVX && mode == OImode)
25301 || VALID_AVX256_REG_MODE (mode)
25302 || VALID_SSE_REG_MODE (mode)
25303 || VALID_SSE2_REG_MODE (mode)
25304 || VALID_MMX_REG_MODE (mode)
25305 || VALID_MMX_REG_MODE_3DNOW (mode));
25307 if (MMX_REGNO_P (regno))
25309 /* We implement the move patterns for 3DNOW modes even in MMX mode,
25310 so if the register is available at all, then we can move data of
25311 the given mode into or out of it. */
25312 return (VALID_MMX_REG_MODE (mode)
25313 || VALID_MMX_REG_MODE_3DNOW (mode));
25316 if (mode == QImode)
25318 /* Take care for QImode values - they can be in non-QI regs,
25319 but then they do cause partial register stalls. */
25320 if (regno <= BX_REG || TARGET_64BIT)
25322 if (!TARGET_PARTIAL_REG_STALL)
25324 return reload_in_progress || reload_completed;
25326 /* We handle both integer and floats in the general purpose registers. */
25327 else if (VALID_INT_MODE_P (mode))
25329 else if (VALID_FP_MODE_P (mode))
25331 else if (VALID_DFP_MODE_P (mode))
25333 /* Lots of MMX code casts 8 byte vector modes to DImode. If we then go
25334 on to use that value in smaller contexts, this can easily force a
25335 pseudo to be allocated to GENERAL_REGS. Since this is no worse than
25336 supporting DImode, allow it. */
25337 else if (VALID_MMX_REG_MODE_3DNOW (mode) || VALID_MMX_REG_MODE (mode))
25343 /* A subroutine of ix86_modes_tieable_p. Return true if MODE is a
25344 tieable integer mode. */
25347 ix86_tieable_integer_mode_p (enum machine_mode mode)
25356 return TARGET_64BIT || !TARGET_PARTIAL_REG_STALL;
25359 return TARGET_64BIT;
25366 /* Return true if MODE1 is accessible in a register that can hold MODE2
25367 without copying. That is, all register classes that can hold MODE2
25368 can also hold MODE1. */
25371 ix86_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
25373 if (mode1 == mode2)
25376 if (ix86_tieable_integer_mode_p (mode1)
25377 && ix86_tieable_integer_mode_p (mode2))
25380 /* MODE2 being XFmode implies fp stack or general regs, which means we
25381 can tie any smaller floating point modes to it. Note that we do not
25382 tie this with TFmode. */
25383 if (mode2 == XFmode)
25384 return mode1 == SFmode || mode1 == DFmode;
25386 /* MODE2 being DFmode implies fp stack, general or sse regs, which means
25387 that we can tie it with SFmode. */
25388 if (mode2 == DFmode)
25389 return mode1 == SFmode;
25391 /* If MODE2 is only appropriate for an SSE register, then tie with
25392 any other mode acceptable to SSE registers. */
25393 if (GET_MODE_SIZE (mode2) == 16
25394 && ix86_hard_regno_mode_ok (FIRST_SSE_REG, mode2))
25395 return (GET_MODE_SIZE (mode1) == 16
25396 && ix86_hard_regno_mode_ok (FIRST_SSE_REG, mode1));
25398 /* If MODE2 is appropriate for an MMX register, then tie
25399 with any other mode acceptable to MMX registers. */
25400 if (GET_MODE_SIZE (mode2) == 8
25401 && ix86_hard_regno_mode_ok (FIRST_MMX_REG, mode2))
25402 return (GET_MODE_SIZE (mode1) == 8
25403 && ix86_hard_regno_mode_ok (FIRST_MMX_REG, mode1));
25408 /* Compute a (partial) cost for rtx X. Return true if the complete
25409 cost has been computed, and false if subexpressions should be
25410 scanned. In either case, *TOTAL contains the cost result. */
25413 ix86_rtx_costs (rtx x, int code, int outer_code_i, int *total, bool speed)
25415 enum rtx_code outer_code = (enum rtx_code) outer_code_i;
25416 enum machine_mode mode = GET_MODE (x);
25417 const struct processor_costs *cost = speed ? ix86_cost : &ix86_size_cost;
25425 if (TARGET_64BIT && !x86_64_immediate_operand (x, VOIDmode))
25427 else if (TARGET_64BIT && !x86_64_zext_immediate_operand (x, VOIDmode))
25429 else if (flag_pic && SYMBOLIC_CONST (x)
25431 || (!GET_CODE (x) != LABEL_REF
25432 && (GET_CODE (x) != SYMBOL_REF
25433 || !SYMBOL_REF_LOCAL_P (x)))))
25440 if (mode == VOIDmode)
25443 switch (standard_80387_constant_p (x))
25448 default: /* Other constants */
25453 /* Start with (MEM (SYMBOL_REF)), since that's where
25454 it'll probably end up. Add a penalty for size. */
25455 *total = (COSTS_N_INSNS (1)
25456 + (flag_pic != 0 && !TARGET_64BIT)
25457 + (mode == SFmode ? 0 : mode == DFmode ? 1 : 2));
25463 /* The zero extensions is often completely free on x86_64, so make
25464 it as cheap as possible. */
25465 if (TARGET_64BIT && mode == DImode
25466 && GET_MODE (XEXP (x, 0)) == SImode)
25468 else if (TARGET_ZERO_EXTEND_WITH_AND)
25469 *total = cost->add;
25471 *total = cost->movzx;
25475 *total = cost->movsx;
25479 if (CONST_INT_P (XEXP (x, 1))
25480 && (GET_MODE (XEXP (x, 0)) != DImode || TARGET_64BIT))
25482 HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
25485 *total = cost->add;
25488 if ((value == 2 || value == 3)
25489 && cost->lea <= cost->shift_const)
25491 *total = cost->lea;
25501 if (!TARGET_64BIT && GET_MODE (XEXP (x, 0)) == DImode)
25503 if (CONST_INT_P (XEXP (x, 1)))
25505 if (INTVAL (XEXP (x, 1)) > 32)
25506 *total = cost->shift_const + COSTS_N_INSNS (2);
25508 *total = cost->shift_const * 2;
25512 if (GET_CODE (XEXP (x, 1)) == AND)
25513 *total = cost->shift_var * 2;
25515 *total = cost->shift_var * 6 + COSTS_N_INSNS (2);
25520 if (CONST_INT_P (XEXP (x, 1)))
25521 *total = cost->shift_const;
25523 *total = cost->shift_var;
25528 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25530 /* ??? SSE scalar cost should be used here. */
25531 *total = cost->fmul;
25534 else if (X87_FLOAT_MODE_P (mode))
25536 *total = cost->fmul;
25539 else if (FLOAT_MODE_P (mode))
25541 /* ??? SSE vector cost should be used here. */
25542 *total = cost->fmul;
25547 rtx op0 = XEXP (x, 0);
25548 rtx op1 = XEXP (x, 1);
25550 if (CONST_INT_P (XEXP (x, 1)))
25552 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
25553 for (nbits = 0; value != 0; value &= value - 1)
25557 /* This is arbitrary. */
25560 /* Compute costs correctly for widening multiplication. */
25561 if ((GET_CODE (op0) == SIGN_EXTEND || GET_CODE (op0) == ZERO_EXTEND)
25562 && GET_MODE_SIZE (GET_MODE (XEXP (op0, 0))) * 2
25563 == GET_MODE_SIZE (mode))
25565 int is_mulwiden = 0;
25566 enum machine_mode inner_mode = GET_MODE (op0);
25568 if (GET_CODE (op0) == GET_CODE (op1))
25569 is_mulwiden = 1, op1 = XEXP (op1, 0);
25570 else if (CONST_INT_P (op1))
25572 if (GET_CODE (op0) == SIGN_EXTEND)
25573 is_mulwiden = trunc_int_for_mode (INTVAL (op1), inner_mode)
25576 is_mulwiden = !(INTVAL (op1) & ~GET_MODE_MASK (inner_mode));
25580 op0 = XEXP (op0, 0), mode = GET_MODE (op0);
25583 *total = (cost->mult_init[MODE_INDEX (mode)]
25584 + nbits * cost->mult_bit
25585 + rtx_cost (op0, outer_code, speed) + rtx_cost (op1, outer_code, speed));
25594 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25595 /* ??? SSE cost should be used here. */
25596 *total = cost->fdiv;
25597 else if (X87_FLOAT_MODE_P (mode))
25598 *total = cost->fdiv;
25599 else if (FLOAT_MODE_P (mode))
25600 /* ??? SSE vector cost should be used here. */
25601 *total = cost->fdiv;
25603 *total = cost->divide[MODE_INDEX (mode)];
25607 if (GET_MODE_CLASS (mode) == MODE_INT
25608 && GET_MODE_BITSIZE (mode) <= GET_MODE_BITSIZE (Pmode))
25610 if (GET_CODE (XEXP (x, 0)) == PLUS
25611 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
25612 && CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 0), 1))
25613 && CONSTANT_P (XEXP (x, 1)))
25615 HOST_WIDE_INT val = INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1));
25616 if (val == 2 || val == 4 || val == 8)
25618 *total = cost->lea;
25619 *total += rtx_cost (XEXP (XEXP (x, 0), 1), outer_code, speed);
25620 *total += rtx_cost (XEXP (XEXP (XEXP (x, 0), 0), 0),
25621 outer_code, speed);
25622 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25626 else if (GET_CODE (XEXP (x, 0)) == MULT
25627 && CONST_INT_P (XEXP (XEXP (x, 0), 1)))
25629 HOST_WIDE_INT val = INTVAL (XEXP (XEXP (x, 0), 1));
25630 if (val == 2 || val == 4 || val == 8)
25632 *total = cost->lea;
25633 *total += rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed);
25634 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25638 else if (GET_CODE (XEXP (x, 0)) == PLUS)
25640 *total = cost->lea;
25641 *total += rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed);
25642 *total += rtx_cost (XEXP (XEXP (x, 0), 1), outer_code, speed);
25643 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25650 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25652 /* ??? SSE cost should be used here. */
25653 *total = cost->fadd;
25656 else if (X87_FLOAT_MODE_P (mode))
25658 *total = cost->fadd;
25661 else if (FLOAT_MODE_P (mode))
25663 /* ??? SSE vector cost should be used here. */
25664 *total = cost->fadd;
25672 if (!TARGET_64BIT && mode == DImode)
25674 *total = (cost->add * 2
25675 + (rtx_cost (XEXP (x, 0), outer_code, speed)
25676 << (GET_MODE (XEXP (x, 0)) != DImode))
25677 + (rtx_cost (XEXP (x, 1), outer_code, speed)
25678 << (GET_MODE (XEXP (x, 1)) != DImode)));
25684 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25686 /* ??? SSE cost should be used here. */
25687 *total = cost->fchs;
25690 else if (X87_FLOAT_MODE_P (mode))
25692 *total = cost->fchs;
25695 else if (FLOAT_MODE_P (mode))
25697 /* ??? SSE vector cost should be used here. */
25698 *total = cost->fchs;
25704 if (!TARGET_64BIT && mode == DImode)
25705 *total = cost->add * 2;
25707 *total = cost->add;
25711 if (GET_CODE (XEXP (x, 0)) == ZERO_EXTRACT
25712 && XEXP (XEXP (x, 0), 1) == const1_rtx
25713 && CONST_INT_P (XEXP (XEXP (x, 0), 2))
25714 && XEXP (x, 1) == const0_rtx)
25716 /* This kind of construct is implemented using test[bwl].
25717 Treat it as if we had an AND. */
25718 *total = (cost->add
25719 + rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed)
25720 + rtx_cost (const1_rtx, outer_code, speed));
25726 if (!(SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH))
25731 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25732 /* ??? SSE cost should be used here. */
25733 *total = cost->fabs;
25734 else if (X87_FLOAT_MODE_P (mode))
25735 *total = cost->fabs;
25736 else if (FLOAT_MODE_P (mode))
25737 /* ??? SSE vector cost should be used here. */
25738 *total = cost->fabs;
25742 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25743 /* ??? SSE cost should be used here. */
25744 *total = cost->fsqrt;
25745 else if (X87_FLOAT_MODE_P (mode))
25746 *total = cost->fsqrt;
25747 else if (FLOAT_MODE_P (mode))
25748 /* ??? SSE vector cost should be used here. */
25749 *total = cost->fsqrt;
25753 if (XINT (x, 1) == UNSPEC_TP)
25760 case VEC_DUPLICATE:
25761 /* ??? Assume all of these vector manipulation patterns are
25762 recognizable. In which case they all pretty much have the
25764 *total = COSTS_N_INSNS (1);
25774 static int current_machopic_label_num;
25776 /* Given a symbol name and its associated stub, write out the
25777 definition of the stub. */
25780 machopic_output_stub (FILE *file, const char *symb, const char *stub)
25782 unsigned int length;
25783 char *binder_name, *symbol_name, lazy_ptr_name[32];
25784 int label = ++current_machopic_label_num;
25786 /* For 64-bit we shouldn't get here. */
25787 gcc_assert (!TARGET_64BIT);
25789 /* Lose our funky encoding stuff so it doesn't contaminate the stub. */
25790 symb = (*targetm.strip_name_encoding) (symb);
25792 length = strlen (stub);
25793 binder_name = XALLOCAVEC (char, length + 32);
25794 GEN_BINDER_NAME_FOR_STUB (binder_name, stub, length);
25796 length = strlen (symb);
25797 symbol_name = XALLOCAVEC (char, length + 32);
25798 GEN_SYMBOL_NAME_FOR_SYMBOL (symbol_name, symb, length);
25800 sprintf (lazy_ptr_name, "L%d$lz", label);
25803 switch_to_section (darwin_sections[machopic_picsymbol_stub_section]);
25805 switch_to_section (darwin_sections[machopic_symbol_stub_section]);
25807 fprintf (file, "%s:\n", stub);
25808 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
25812 fprintf (file, "\tcall\tLPC$%d\nLPC$%d:\tpopl\t%%eax\n", label, label);
25813 fprintf (file, "\tmovl\t%s-LPC$%d(%%eax),%%edx\n", lazy_ptr_name, label);
25814 fprintf (file, "\tjmp\t*%%edx\n");
25817 fprintf (file, "\tjmp\t*%s\n", lazy_ptr_name);
25819 fprintf (file, "%s:\n", binder_name);
25823 fprintf (file, "\tlea\t%s-LPC$%d(%%eax),%%eax\n", lazy_ptr_name, label);
25824 fputs ("\tpushl\t%eax\n", file);
25827 fprintf (file, "\tpushl\t$%s\n", lazy_ptr_name);
25829 fputs ("\tjmp\tdyld_stub_binding_helper\n", file);
25831 switch_to_section (darwin_sections[machopic_lazy_symbol_ptr_section]);
25832 fprintf (file, "%s:\n", lazy_ptr_name);
25833 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
25834 fprintf (file, ASM_LONG "%s\n", binder_name);
25838 darwin_x86_file_end (void)
25840 darwin_file_end ();
25843 #endif /* TARGET_MACHO */
25845 /* Order the registers for register allocator. */
25848 x86_order_regs_for_local_alloc (void)
25853 /* First allocate the local general purpose registers. */
25854 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
25855 if (GENERAL_REGNO_P (i) && call_used_regs[i])
25856 reg_alloc_order [pos++] = i;
25858 /* Global general purpose registers. */
25859 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
25860 if (GENERAL_REGNO_P (i) && !call_used_regs[i])
25861 reg_alloc_order [pos++] = i;
25863 /* x87 registers come first in case we are doing FP math
25865 if (!TARGET_SSE_MATH)
25866 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++)
25867 reg_alloc_order [pos++] = i;
25869 /* SSE registers. */
25870 for (i = FIRST_SSE_REG; i <= LAST_SSE_REG; i++)
25871 reg_alloc_order [pos++] = i;
25872 for (i = FIRST_REX_SSE_REG; i <= LAST_REX_SSE_REG; i++)
25873 reg_alloc_order [pos++] = i;
25875 /* x87 registers. */
25876 if (TARGET_SSE_MATH)
25877 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++)
25878 reg_alloc_order [pos++] = i;
25880 for (i = FIRST_MMX_REG; i <= LAST_MMX_REG; i++)
25881 reg_alloc_order [pos++] = i;
25883 /* Initialize the rest of array as we do not allocate some registers
25885 while (pos < FIRST_PSEUDO_REGISTER)
25886 reg_alloc_order [pos++] = 0;
25889 /* Handle a "ms_abi" or "sysv" attribute; arguments as in
25890 struct attribute_spec.handler. */
25892 ix86_handle_abi_attribute (tree *node, tree name,
25893 tree args ATTRIBUTE_UNUSED,
25894 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
25896 if (TREE_CODE (*node) != FUNCTION_TYPE
25897 && TREE_CODE (*node) != METHOD_TYPE
25898 && TREE_CODE (*node) != FIELD_DECL
25899 && TREE_CODE (*node) != TYPE_DECL)
25901 warning (OPT_Wattributes, "%qE attribute only applies to functions",
25903 *no_add_attrs = true;
25908 warning (OPT_Wattributes, "%qE attribute only available for 64-bit",
25910 *no_add_attrs = true;
25914 /* Can combine regparm with all attributes but fastcall. */
25915 if (is_attribute_p ("ms_abi", name))
25917 if (lookup_attribute ("sysv_abi", TYPE_ATTRIBUTES (*node)))
25919 error ("ms_abi and sysv_abi attributes are not compatible");
25924 else if (is_attribute_p ("sysv_abi", name))
25926 if (lookup_attribute ("ms_abi", TYPE_ATTRIBUTES (*node)))
25928 error ("ms_abi and sysv_abi attributes are not compatible");
25937 /* Handle a "ms_struct" or "gcc_struct" attribute; arguments as in
25938 struct attribute_spec.handler. */
25940 ix86_handle_struct_attribute (tree *node, tree name,
25941 tree args ATTRIBUTE_UNUSED,
25942 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
25945 if (DECL_P (*node))
25947 if (TREE_CODE (*node) == TYPE_DECL)
25948 type = &TREE_TYPE (*node);
25953 if (!(type && (TREE_CODE (*type) == RECORD_TYPE
25954 || TREE_CODE (*type) == UNION_TYPE)))
25956 warning (OPT_Wattributes, "%qE attribute ignored",
25958 *no_add_attrs = true;
25961 else if ((is_attribute_p ("ms_struct", name)
25962 && lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (*type)))
25963 || ((is_attribute_p ("gcc_struct", name)
25964 && lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (*type)))))
25966 warning (OPT_Wattributes, "%qE incompatible attribute ignored",
25968 *no_add_attrs = true;
25975 ix86_handle_fndecl_attribute (tree *node, tree name,
25976 tree args ATTRIBUTE_UNUSED,
25977 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
25979 if (TREE_CODE (*node) != FUNCTION_DECL)
25981 warning (OPT_Wattributes, "%qE attribute only applies to functions",
25983 *no_add_attrs = true;
25989 warning (OPT_Wattributes, "%qE attribute only available for 32-bit",
25994 #ifndef HAVE_AS_IX86_SWAP
25995 sorry ("ms_hook_prologue attribute needs assembler swap suffix support");
26002 ix86_ms_bitfield_layout_p (const_tree record_type)
26004 return (TARGET_MS_BITFIELD_LAYOUT &&
26005 !lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (record_type)))
26006 || lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (record_type));
26009 /* Returns an expression indicating where the this parameter is
26010 located on entry to the FUNCTION. */
26013 x86_this_parameter (tree function)
26015 tree type = TREE_TYPE (function);
26016 bool aggr = aggregate_value_p (TREE_TYPE (type), type) != 0;
26021 const int *parm_regs;
26023 if (ix86_function_type_abi (type) == MS_ABI)
26024 parm_regs = x86_64_ms_abi_int_parameter_registers;
26026 parm_regs = x86_64_int_parameter_registers;
26027 return gen_rtx_REG (DImode, parm_regs[aggr]);
26030 nregs = ix86_function_regparm (type, function);
26032 if (nregs > 0 && !stdarg_p (type))
26036 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type)))
26037 regno = aggr ? DX_REG : CX_REG;
26045 return gen_rtx_MEM (SImode,
26046 plus_constant (stack_pointer_rtx, 4));
26049 return gen_rtx_REG (SImode, regno);
26052 return gen_rtx_MEM (SImode, plus_constant (stack_pointer_rtx, aggr ? 8 : 4));
26055 /* Determine whether x86_output_mi_thunk can succeed. */
26058 x86_can_output_mi_thunk (const_tree thunk ATTRIBUTE_UNUSED,
26059 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
26060 HOST_WIDE_INT vcall_offset, const_tree function)
26062 /* 64-bit can handle anything. */
26066 /* For 32-bit, everything's fine if we have one free register. */
26067 if (ix86_function_regparm (TREE_TYPE (function), function) < 3)
26070 /* Need a free register for vcall_offset. */
26074 /* Need a free register for GOT references. */
26075 if (flag_pic && !(*targetm.binds_local_p) (function))
26078 /* Otherwise ok. */
26082 /* Output the assembler code for a thunk function. THUNK_DECL is the
26083 declaration for the thunk function itself, FUNCTION is the decl for
26084 the target function. DELTA is an immediate constant offset to be
26085 added to THIS. If VCALL_OFFSET is nonzero, the word at
26086 *(*this + vcall_offset) should be added to THIS. */
26089 x86_output_mi_thunk (FILE *file ATTRIBUTE_UNUSED,
26090 tree thunk ATTRIBUTE_UNUSED, HOST_WIDE_INT delta,
26091 HOST_WIDE_INT vcall_offset, tree function)
26094 rtx this_param = x86_this_parameter (function);
26097 /* If VCALL_OFFSET, we'll need THIS in a register. Might as well
26098 pull it in now and let DELTA benefit. */
26099 if (REG_P (this_param))
26100 this_reg = this_param;
26101 else if (vcall_offset)
26103 /* Put the this parameter into %eax. */
26104 xops[0] = this_param;
26105 xops[1] = this_reg = gen_rtx_REG (Pmode, AX_REG);
26106 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26109 this_reg = NULL_RTX;
26111 /* Adjust the this parameter by a fixed constant. */
26114 /* Make things pretty and `subl $4,%eax' rather than `addl $-4,%eax'.
26115 Exceptions: -128 encodes smaller than 128, so swap sign and op. */
26116 bool sub = delta < 0 || delta == 128;
26117 xops[0] = GEN_INT (sub ? -delta : delta);
26118 xops[1] = this_reg ? this_reg : this_param;
26121 if (!x86_64_general_operand (xops[0], DImode))
26123 tmp = gen_rtx_REG (DImode, R10_REG);
26125 output_asm_insn ("mov{q}\t{%1, %0|%0, %1}", xops);
26127 xops[1] = this_param;
26130 output_asm_insn ("sub{q}\t{%0, %1|%1, %0}", xops);
26132 output_asm_insn ("add{q}\t{%0, %1|%1, %0}", xops);
26135 output_asm_insn ("sub{l}\t{%0, %1|%1, %0}", xops);
26137 output_asm_insn ("add{l}\t{%0, %1|%1, %0}", xops);
26140 /* Adjust the this parameter by a value stored in the vtable. */
26144 tmp = gen_rtx_REG (DImode, R10_REG);
26147 int tmp_regno = CX_REG;
26148 if (lookup_attribute ("fastcall",
26149 TYPE_ATTRIBUTES (TREE_TYPE (function))))
26150 tmp_regno = AX_REG;
26151 tmp = gen_rtx_REG (SImode, tmp_regno);
26154 xops[0] = gen_rtx_MEM (Pmode, this_reg);
26156 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26158 /* Adjust the this parameter. */
26159 xops[0] = gen_rtx_MEM (Pmode, plus_constant (tmp, vcall_offset));
26160 if (TARGET_64BIT && !memory_operand (xops[0], Pmode))
26162 rtx tmp2 = gen_rtx_REG (DImode, R11_REG);
26163 xops[0] = GEN_INT (vcall_offset);
26165 output_asm_insn ("mov{q}\t{%0, %1|%1, %0}", xops);
26166 xops[0] = gen_rtx_MEM (Pmode, gen_rtx_PLUS (Pmode, tmp, tmp2));
26168 xops[1] = this_reg;
26169 output_asm_insn ("add%z1\t{%0, %1|%1, %0}", xops);
26172 /* If necessary, drop THIS back to its stack slot. */
26173 if (this_reg && this_reg != this_param)
26175 xops[0] = this_reg;
26176 xops[1] = this_param;
26177 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26180 xops[0] = XEXP (DECL_RTL (function), 0);
26183 if (!flag_pic || (*targetm.binds_local_p) (function))
26184 output_asm_insn ("jmp\t%P0", xops);
26185 /* All thunks should be in the same object as their target,
26186 and thus binds_local_p should be true. */
26187 else if (TARGET_64BIT && cfun->machine->call_abi == MS_ABI)
26188 gcc_unreachable ();
26191 tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, xops[0]), UNSPEC_GOTPCREL);
26192 tmp = gen_rtx_CONST (Pmode, tmp);
26193 tmp = gen_rtx_MEM (QImode, tmp);
26195 output_asm_insn ("jmp\t%A0", xops);
26200 if (!flag_pic || (*targetm.binds_local_p) (function))
26201 output_asm_insn ("jmp\t%P0", xops);
26206 rtx sym_ref = XEXP (DECL_RTL (function), 0);
26207 tmp = (gen_rtx_SYMBOL_REF
26209 machopic_indirection_name (sym_ref, /*stub_p=*/true)));
26210 tmp = gen_rtx_MEM (QImode, tmp);
26212 output_asm_insn ("jmp\t%0", xops);
26215 #endif /* TARGET_MACHO */
26217 tmp = gen_rtx_REG (SImode, CX_REG);
26218 output_set_got (tmp, NULL_RTX);
26221 output_asm_insn ("mov{l}\t{%0@GOT(%1), %1|%1, %0@GOT[%1]}", xops);
26222 output_asm_insn ("jmp\t{*}%1", xops);
26228 x86_file_start (void)
26230 default_file_start ();
26232 darwin_file_start ();
26234 if (X86_FILE_START_VERSION_DIRECTIVE)
26235 fputs ("\t.version\t\"01.01\"\n", asm_out_file);
26236 if (X86_FILE_START_FLTUSED)
26237 fputs ("\t.global\t__fltused\n", asm_out_file);
26238 if (ix86_asm_dialect == ASM_INTEL)
26239 fputs ("\t.intel_syntax noprefix\n", asm_out_file);
26243 x86_field_alignment (tree field, int computed)
26245 enum machine_mode mode;
26246 tree type = TREE_TYPE (field);
26248 if (TARGET_64BIT || TARGET_ALIGN_DOUBLE)
26250 mode = TYPE_MODE (strip_array_types (type));
26251 if (mode == DFmode || mode == DCmode
26252 || GET_MODE_CLASS (mode) == MODE_INT
26253 || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT)
26254 return MIN (32, computed);
26258 /* Output assembler code to FILE to increment profiler label # LABELNO
26259 for profiling a function entry. */
26261 x86_function_profiler (FILE *file, int labelno ATTRIBUTE_UNUSED)
26265 #ifndef NO_PROFILE_COUNTERS
26266 fprintf (file, "\tleaq\t" LPREFIX "P%d@(%%rip),%%r11\n", labelno);
26269 if (DEFAULT_ABI == SYSV_ABI && flag_pic)
26270 fputs ("\tcall\t*" MCOUNT_NAME "@GOTPCREL(%rip)\n", file);
26272 fputs ("\tcall\t" MCOUNT_NAME "\n", file);
26276 #ifndef NO_PROFILE_COUNTERS
26277 fprintf (file, "\tleal\t" LPREFIX "P%d@GOTOFF(%%ebx),%%" PROFILE_COUNT_REGISTER "\n",
26280 fputs ("\tcall\t*" MCOUNT_NAME "@GOT(%ebx)\n", file);
26284 #ifndef NO_PROFILE_COUNTERS
26285 fprintf (file, "\tmovl\t$" LPREFIX "P%d,%%" PROFILE_COUNT_REGISTER "\n",
26288 fputs ("\tcall\t" MCOUNT_NAME "\n", file);
26292 #ifdef ASM_OUTPUT_MAX_SKIP_PAD
26293 /* We don't have exact information about the insn sizes, but we may assume
26294 quite safely that we are informed about all 1 byte insns and memory
26295 address sizes. This is enough to eliminate unnecessary padding in
26299 min_insn_size (rtx insn)
26303 if (!INSN_P (insn) || !active_insn_p (insn))
26306 /* Discard alignments we've emit and jump instructions. */
26307 if (GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
26308 && XINT (PATTERN (insn), 1) == UNSPECV_ALIGN)
26310 if (JUMP_TABLE_DATA_P (insn))
26313 /* Important case - calls are always 5 bytes.
26314 It is common to have many calls in the row. */
26316 && symbolic_reference_mentioned_p (PATTERN (insn))
26317 && !SIBLING_CALL_P (insn))
26319 len = get_attr_length (insn);
26323 /* For normal instructions we rely on get_attr_length being exact,
26324 with a few exceptions. */
26325 if (!JUMP_P (insn))
26327 enum attr_type type = get_attr_type (insn);
26332 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
26333 || asm_noperands (PATTERN (insn)) >= 0)
26340 /* Otherwise trust get_attr_length. */
26344 l = get_attr_length_address (insn);
26345 if (l < 4 && symbolic_reference_mentioned_p (PATTERN (insn)))
26354 /* AMD K8 core mispredicts jumps when there are more than 3 jumps in 16 byte
26358 ix86_avoid_jump_mispredicts (void)
26360 rtx insn, start = get_insns ();
26361 int nbytes = 0, njumps = 0;
26364 /* Look for all minimal intervals of instructions containing 4 jumps.
26365 The intervals are bounded by START and INSN. NBYTES is the total
26366 size of instructions in the interval including INSN and not including
26367 START. When the NBYTES is smaller than 16 bytes, it is possible
26368 that the end of START and INSN ends up in the same 16byte page.
26370 The smallest offset in the page INSN can start is the case where START
26371 ends on the offset 0. Offset of INSN is then NBYTES - sizeof (INSN).
26372 We add p2align to 16byte window with maxskip 15 - NBYTES + sizeof (INSN).
26374 for (insn = start; insn; insn = NEXT_INSN (insn))
26378 if (LABEL_P (insn))
26380 int align = label_to_alignment (insn);
26381 int max_skip = label_to_max_skip (insn);
26385 /* If align > 3, only up to 16 - max_skip - 1 bytes can be
26386 already in the current 16 byte page, because otherwise
26387 ASM_OUTPUT_MAX_SKIP_ALIGN could skip max_skip or fewer
26388 bytes to reach 16 byte boundary. */
26390 || (align <= 3 && max_skip != (1 << align) - 1))
26393 fprintf (dump_file, "Label %i with max_skip %i\n",
26394 INSN_UID (insn), max_skip);
26397 while (nbytes + max_skip >= 16)
26399 start = NEXT_INSN (start);
26400 if ((JUMP_P (start)
26401 && GET_CODE (PATTERN (start)) != ADDR_VEC
26402 && GET_CODE (PATTERN (start)) != ADDR_DIFF_VEC)
26404 njumps--, isjump = 1;
26407 nbytes -= min_insn_size (start);
26413 min_size = min_insn_size (insn);
26414 nbytes += min_size;
26416 fprintf (dump_file, "Insn %i estimated to %i bytes\n",
26417 INSN_UID (insn), min_size);
26419 && GET_CODE (PATTERN (insn)) != ADDR_VEC
26420 && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)
26428 start = NEXT_INSN (start);
26429 if ((JUMP_P (start)
26430 && GET_CODE (PATTERN (start)) != ADDR_VEC
26431 && GET_CODE (PATTERN (start)) != ADDR_DIFF_VEC)
26433 njumps--, isjump = 1;
26436 nbytes -= min_insn_size (start);
26438 gcc_assert (njumps >= 0);
26440 fprintf (dump_file, "Interval %i to %i has %i bytes\n",
26441 INSN_UID (start), INSN_UID (insn), nbytes);
26443 if (njumps == 3 && isjump && nbytes < 16)
26445 int padsize = 15 - nbytes + min_insn_size (insn);
26448 fprintf (dump_file, "Padding insn %i by %i bytes!\n",
26449 INSN_UID (insn), padsize);
26450 emit_insn_before (gen_pad (GEN_INT (padsize)), insn);
26456 /* AMD Athlon works faster
26457 when RET is not destination of conditional jump or directly preceded
26458 by other jump instruction. We avoid the penalty by inserting NOP just
26459 before the RET instructions in such cases. */
26461 ix86_pad_returns (void)
26466 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
26468 basic_block bb = e->src;
26469 rtx ret = BB_END (bb);
26471 bool replace = false;
26473 if (!JUMP_P (ret) || GET_CODE (PATTERN (ret)) != RETURN
26474 || optimize_bb_for_size_p (bb))
26476 for (prev = PREV_INSN (ret); prev; prev = PREV_INSN (prev))
26477 if (active_insn_p (prev) || LABEL_P (prev))
26479 if (prev && LABEL_P (prev))
26484 FOR_EACH_EDGE (e, ei, bb->preds)
26485 if (EDGE_FREQUENCY (e) && e->src->index >= 0
26486 && !(e->flags & EDGE_FALLTHRU))
26491 prev = prev_active_insn (ret);
26493 && ((JUMP_P (prev) && any_condjump_p (prev))
26496 /* Empty functions get branch mispredict even when the jump destination
26497 is not visible to us. */
26498 if (!prev && cfun->function_frequency > FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
26503 emit_jump_insn_before (gen_return_internal_long (), ret);
26509 /* Implement machine specific optimizations. We implement padding of returns
26510 for K8 CPUs and pass to avoid 4 jumps in the single 16 byte window. */
26514 if (optimize && optimize_function_for_speed_p (cfun))
26516 if (TARGET_PAD_RETURNS)
26517 ix86_pad_returns ();
26518 #ifdef ASM_OUTPUT_MAX_SKIP_PAD
26519 if (TARGET_FOUR_JUMP_LIMIT)
26520 ix86_avoid_jump_mispredicts ();
26525 /* Return nonzero when QImode register that must be represented via REX prefix
26528 x86_extended_QIreg_mentioned_p (rtx insn)
26531 extract_insn_cached (insn);
26532 for (i = 0; i < recog_data.n_operands; i++)
26533 if (REG_P (recog_data.operand[i])
26534 && REGNO (recog_data.operand[i]) > BX_REG)
26539 /* Return nonzero when P points to register encoded via REX prefix.
26540 Called via for_each_rtx. */
26542 extended_reg_mentioned_1 (rtx *p, void *data ATTRIBUTE_UNUSED)
26544 unsigned int regno;
26547 regno = REGNO (*p);
26548 return REX_INT_REGNO_P (regno) || REX_SSE_REGNO_P (regno);
26551 /* Return true when INSN mentions register that must be encoded using REX
26554 x86_extended_reg_mentioned_p (rtx insn)
26556 return for_each_rtx (INSN_P (insn) ? &PATTERN (insn) : &insn,
26557 extended_reg_mentioned_1, NULL);
26560 /* Generate an unsigned DImode/SImode to FP conversion. This is the same code
26561 optabs would emit if we didn't have TFmode patterns. */
26564 x86_emit_floatuns (rtx operands[2])
26566 rtx neglab, donelab, i0, i1, f0, in, out;
26567 enum machine_mode mode, inmode;
26569 inmode = GET_MODE (operands[1]);
26570 gcc_assert (inmode == SImode || inmode == DImode);
26573 in = force_reg (inmode, operands[1]);
26574 mode = GET_MODE (out);
26575 neglab = gen_label_rtx ();
26576 donelab = gen_label_rtx ();
26577 f0 = gen_reg_rtx (mode);
26579 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, inmode, 0, neglab);
26581 expand_float (out, in, 0);
26583 emit_jump_insn (gen_jump (donelab));
26586 emit_label (neglab);
26588 i0 = expand_simple_binop (inmode, LSHIFTRT, in, const1_rtx, NULL,
26590 i1 = expand_simple_binop (inmode, AND, in, const1_rtx, NULL,
26592 i0 = expand_simple_binop (inmode, IOR, i0, i1, i0, 1, OPTAB_DIRECT);
26594 expand_float (f0, i0, 0);
26596 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
26598 emit_label (donelab);
26601 /* AVX does not support 32-byte integer vector operations,
26602 thus the longest vector we are faced with is V16QImode. */
26603 #define MAX_VECT_LEN 16
26605 struct expand_vec_perm_d
26607 rtx target, op0, op1;
26608 unsigned char perm[MAX_VECT_LEN];
26609 enum machine_mode vmode;
26610 unsigned char nelt;
26614 static bool expand_vec_perm_1 (struct expand_vec_perm_d *d);
26615 static bool expand_vec_perm_broadcast_1 (struct expand_vec_perm_d *d);
26617 /* Get a vector mode of the same size as the original but with elements
26618 twice as wide. This is only guaranteed to apply to integral vectors. */
26620 static inline enum machine_mode
26621 get_mode_wider_vector (enum machine_mode o)
26623 /* ??? Rely on the ordering that genmodes.c gives to vectors. */
26624 enum machine_mode n = GET_MODE_WIDER_MODE (o);
26625 gcc_assert (GET_MODE_NUNITS (o) == GET_MODE_NUNITS (n) * 2);
26626 gcc_assert (GET_MODE_SIZE (o) == GET_MODE_SIZE (n));
26630 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
26631 with all elements equal to VAR. Return true if successful. */
26634 ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
26635 rtx target, rtx val)
26658 /* First attempt to recognize VAL as-is. */
26659 dup = gen_rtx_VEC_DUPLICATE (mode, val);
26660 insn = emit_insn (gen_rtx_SET (VOIDmode, target, dup));
26661 if (recog_memoized (insn) < 0)
26663 /* If that fails, force VAL into a register. */
26664 XEXP (dup, 0) = force_reg (GET_MODE_INNER (mode), val);
26665 ok = recog_memoized (insn) >= 0;
26674 if (TARGET_SSE || TARGET_3DNOW_A)
26678 val = gen_lowpart (SImode, val);
26679 x = gen_rtx_TRUNCATE (HImode, val);
26680 x = gen_rtx_VEC_DUPLICATE (mode, x);
26681 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26694 struct expand_vec_perm_d dperm;
26698 memset (&dperm, 0, sizeof (dperm));
26699 dperm.target = target;
26700 dperm.vmode = mode;
26701 dperm.nelt = GET_MODE_NUNITS (mode);
26702 dperm.op0 = dperm.op1 = gen_reg_rtx (mode);
26704 /* Extend to SImode using a paradoxical SUBREG. */
26705 tmp1 = gen_reg_rtx (SImode);
26706 emit_move_insn (tmp1, gen_lowpart (SImode, val));
26708 /* Insert the SImode value as low element of a V4SImode vector. */
26709 tmp2 = gen_lowpart (V4SImode, dperm.op0);
26710 emit_insn (gen_vec_setv4si_0 (tmp2, CONST0_RTX (V4SImode), tmp1));
26712 ok = (expand_vec_perm_1 (&dperm)
26713 || expand_vec_perm_broadcast_1 (&dperm));
26725 /* Replicate the value once into the next wider mode and recurse. */
26727 enum machine_mode smode, wsmode, wvmode;
26730 smode = GET_MODE_INNER (mode);
26731 wvmode = get_mode_wider_vector (mode);
26732 wsmode = GET_MODE_INNER (wvmode);
26734 val = convert_modes (wsmode, smode, val, true);
26735 x = expand_simple_binop (wsmode, ASHIFT, val,
26736 GEN_INT (GET_MODE_BITSIZE (smode)),
26737 NULL_RTX, 1, OPTAB_LIB_WIDEN);
26738 val = expand_simple_binop (wsmode, IOR, val, x, x, 1, OPTAB_LIB_WIDEN);
26740 x = gen_lowpart (wvmode, target);
26741 ok = ix86_expand_vector_init_duplicate (mmx_ok, wvmode, x, val);
26749 enum machine_mode hvmode = (mode == V16HImode ? V8HImode : V16QImode);
26750 rtx x = gen_reg_rtx (hvmode);
26752 ok = ix86_expand_vector_init_duplicate (false, hvmode, x, val);
26755 x = gen_rtx_VEC_CONCAT (mode, x, x);
26756 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26765 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
26766 whose ONE_VAR element is VAR, and other elements are zero. Return true
26770 ix86_expand_vector_init_one_nonzero (bool mmx_ok, enum machine_mode mode,
26771 rtx target, rtx var, int one_var)
26773 enum machine_mode vsimode;
26776 bool use_vector_set = false;
26781 /* For SSE4.1, we normally use vector set. But if the second
26782 element is zero and inter-unit moves are OK, we use movq
26784 use_vector_set = (TARGET_64BIT
26786 && !(TARGET_INTER_UNIT_MOVES
26792 use_vector_set = TARGET_SSE4_1;
26795 use_vector_set = TARGET_SSE2;
26798 use_vector_set = TARGET_SSE || TARGET_3DNOW_A;
26805 use_vector_set = TARGET_AVX;
26808 /* Use ix86_expand_vector_set in 64bit mode only. */
26809 use_vector_set = TARGET_AVX && TARGET_64BIT;
26815 if (use_vector_set)
26817 emit_insn (gen_rtx_SET (VOIDmode, target, CONST0_RTX (mode)));
26818 var = force_reg (GET_MODE_INNER (mode), var);
26819 ix86_expand_vector_set (mmx_ok, target, var, one_var);
26835 var = force_reg (GET_MODE_INNER (mode), var);
26836 x = gen_rtx_VEC_CONCAT (mode, var, CONST0_RTX (GET_MODE_INNER (mode)));
26837 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26842 if (!REG_P (target) || REGNO (target) < FIRST_PSEUDO_REGISTER)
26843 new_target = gen_reg_rtx (mode);
26845 new_target = target;
26846 var = force_reg (GET_MODE_INNER (mode), var);
26847 x = gen_rtx_VEC_DUPLICATE (mode, var);
26848 x = gen_rtx_VEC_MERGE (mode, x, CONST0_RTX (mode), const1_rtx);
26849 emit_insn (gen_rtx_SET (VOIDmode, new_target, x));
26852 /* We need to shuffle the value to the correct position, so
26853 create a new pseudo to store the intermediate result. */
26855 /* With SSE2, we can use the integer shuffle insns. */
26856 if (mode != V4SFmode && TARGET_SSE2)
26858 emit_insn (gen_sse2_pshufd_1 (new_target, new_target,
26860 GEN_INT (one_var == 1 ? 0 : 1),
26861 GEN_INT (one_var == 2 ? 0 : 1),
26862 GEN_INT (one_var == 3 ? 0 : 1)));
26863 if (target != new_target)
26864 emit_move_insn (target, new_target);
26868 /* Otherwise convert the intermediate result to V4SFmode and
26869 use the SSE1 shuffle instructions. */
26870 if (mode != V4SFmode)
26872 tmp = gen_reg_rtx (V4SFmode);
26873 emit_move_insn (tmp, gen_lowpart (V4SFmode, new_target));
26878 emit_insn (gen_sse_shufps_v4sf (tmp, tmp, tmp,
26880 GEN_INT (one_var == 1 ? 0 : 1),
26881 GEN_INT (one_var == 2 ? 0+4 : 1+4),
26882 GEN_INT (one_var == 3 ? 0+4 : 1+4)));
26884 if (mode != V4SFmode)
26885 emit_move_insn (target, gen_lowpart (V4SImode, tmp));
26886 else if (tmp != target)
26887 emit_move_insn (target, tmp);
26889 else if (target != new_target)
26890 emit_move_insn (target, new_target);
26895 vsimode = V4SImode;
26901 vsimode = V2SImode;
26907 /* Zero extend the variable element to SImode and recurse. */
26908 var = convert_modes (SImode, GET_MODE_INNER (mode), var, true);
26910 x = gen_reg_rtx (vsimode);
26911 if (!ix86_expand_vector_init_one_nonzero (mmx_ok, vsimode, x,
26913 gcc_unreachable ();
26915 emit_move_insn (target, gen_lowpart (mode, x));
26923 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
26924 consisting of the values in VALS. It is known that all elements
26925 except ONE_VAR are constants. Return true if successful. */
26928 ix86_expand_vector_init_one_var (bool mmx_ok, enum machine_mode mode,
26929 rtx target, rtx vals, int one_var)
26931 rtx var = XVECEXP (vals, 0, one_var);
26932 enum machine_mode wmode;
26935 const_vec = copy_rtx (vals);
26936 XVECEXP (const_vec, 0, one_var) = CONST0_RTX (GET_MODE_INNER (mode));
26937 const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (const_vec, 0));
26945 /* For the two element vectors, it's just as easy to use
26946 the general case. */
26950 /* Use ix86_expand_vector_set in 64bit mode only. */
26973 /* There's no way to set one QImode entry easily. Combine
26974 the variable value with its adjacent constant value, and
26975 promote to an HImode set. */
26976 x = XVECEXP (vals, 0, one_var ^ 1);
26979 var = convert_modes (HImode, QImode, var, true);
26980 var = expand_simple_binop (HImode, ASHIFT, var, GEN_INT (8),
26981 NULL_RTX, 1, OPTAB_LIB_WIDEN);
26982 x = GEN_INT (INTVAL (x) & 0xff);
26986 var = convert_modes (HImode, QImode, var, true);
26987 x = gen_int_mode (INTVAL (x) << 8, HImode);
26989 if (x != const0_rtx)
26990 var = expand_simple_binop (HImode, IOR, var, x, var,
26991 1, OPTAB_LIB_WIDEN);
26993 x = gen_reg_rtx (wmode);
26994 emit_move_insn (x, gen_lowpart (wmode, const_vec));
26995 ix86_expand_vector_set (mmx_ok, x, var, one_var >> 1);
26997 emit_move_insn (target, gen_lowpart (mode, x));
27004 emit_move_insn (target, const_vec);
27005 ix86_expand_vector_set (mmx_ok, target, var, one_var);
27009 /* A subroutine of ix86_expand_vector_init_general. Use vector
27010 concatenate to handle the most general case: all values variable,
27011 and none identical. */
27014 ix86_expand_vector_init_concat (enum machine_mode mode,
27015 rtx target, rtx *ops, int n)
27017 enum machine_mode cmode, hmode = VOIDmode;
27018 rtx first[8], second[4];
27058 gcc_unreachable ();
27061 if (!register_operand (ops[1], cmode))
27062 ops[1] = force_reg (cmode, ops[1]);
27063 if (!register_operand (ops[0], cmode))
27064 ops[0] = force_reg (cmode, ops[0]);
27065 emit_insn (gen_rtx_SET (VOIDmode, target,
27066 gen_rtx_VEC_CONCAT (mode, ops[0],
27086 gcc_unreachable ();
27102 gcc_unreachable ();
27107 /* FIXME: We process inputs backward to help RA. PR 36222. */
27110 for (; i > 0; i -= 2, j--)
27112 first[j] = gen_reg_rtx (cmode);
27113 v = gen_rtvec (2, ops[i - 1], ops[i]);
27114 ix86_expand_vector_init (false, first[j],
27115 gen_rtx_PARALLEL (cmode, v));
27121 gcc_assert (hmode != VOIDmode);
27122 for (i = j = 0; i < n; i += 2, j++)
27124 second[j] = gen_reg_rtx (hmode);
27125 ix86_expand_vector_init_concat (hmode, second [j],
27129 ix86_expand_vector_init_concat (mode, target, second, n);
27132 ix86_expand_vector_init_concat (mode, target, first, n);
27136 gcc_unreachable ();
27140 /* A subroutine of ix86_expand_vector_init_general. Use vector
27141 interleave to handle the most general case: all values variable,
27142 and none identical. */
27145 ix86_expand_vector_init_interleave (enum machine_mode mode,
27146 rtx target, rtx *ops, int n)
27148 enum machine_mode first_imode, second_imode, third_imode, inner_mode;
27151 rtx (*gen_load_even) (rtx, rtx, rtx);
27152 rtx (*gen_interleave_first_low) (rtx, rtx, rtx);
27153 rtx (*gen_interleave_second_low) (rtx, rtx, rtx);
27158 gen_load_even = gen_vec_setv8hi;
27159 gen_interleave_first_low = gen_vec_interleave_lowv4si;
27160 gen_interleave_second_low = gen_vec_interleave_lowv2di;
27161 inner_mode = HImode;
27162 first_imode = V4SImode;
27163 second_imode = V2DImode;
27164 third_imode = VOIDmode;
27167 gen_load_even = gen_vec_setv16qi;
27168 gen_interleave_first_low = gen_vec_interleave_lowv8hi;
27169 gen_interleave_second_low = gen_vec_interleave_lowv4si;
27170 inner_mode = QImode;
27171 first_imode = V8HImode;
27172 second_imode = V4SImode;
27173 third_imode = V2DImode;
27176 gcc_unreachable ();
27179 for (i = 0; i < n; i++)
27181 /* Extend the odd elment to SImode using a paradoxical SUBREG. */
27182 op0 = gen_reg_rtx (SImode);
27183 emit_move_insn (op0, gen_lowpart (SImode, ops [i + i]));
27185 /* Insert the SImode value as low element of V4SImode vector. */
27186 op1 = gen_reg_rtx (V4SImode);
27187 op0 = gen_rtx_VEC_MERGE (V4SImode,
27188 gen_rtx_VEC_DUPLICATE (V4SImode,
27190 CONST0_RTX (V4SImode),
27192 emit_insn (gen_rtx_SET (VOIDmode, op1, op0));
27194 /* Cast the V4SImode vector back to a vector in orignal mode. */
27195 op0 = gen_reg_rtx (mode);
27196 emit_move_insn (op0, gen_lowpart (mode, op1));
27198 /* Load even elements into the second positon. */
27199 emit_insn ((*gen_load_even) (op0,
27200 force_reg (inner_mode,
27204 /* Cast vector to FIRST_IMODE vector. */
27205 ops[i] = gen_reg_rtx (first_imode);
27206 emit_move_insn (ops[i], gen_lowpart (first_imode, op0));
27209 /* Interleave low FIRST_IMODE vectors. */
27210 for (i = j = 0; i < n; i += 2, j++)
27212 op0 = gen_reg_rtx (first_imode);
27213 emit_insn ((*gen_interleave_first_low) (op0, ops[i], ops[i + 1]));
27215 /* Cast FIRST_IMODE vector to SECOND_IMODE vector. */
27216 ops[j] = gen_reg_rtx (second_imode);
27217 emit_move_insn (ops[j], gen_lowpart (second_imode, op0));
27220 /* Interleave low SECOND_IMODE vectors. */
27221 switch (second_imode)
27224 for (i = j = 0; i < n / 2; i += 2, j++)
27226 op0 = gen_reg_rtx (second_imode);
27227 emit_insn ((*gen_interleave_second_low) (op0, ops[i],
27230 /* Cast the SECOND_IMODE vector to the THIRD_IMODE
27232 ops[j] = gen_reg_rtx (third_imode);
27233 emit_move_insn (ops[j], gen_lowpart (third_imode, op0));
27235 second_imode = V2DImode;
27236 gen_interleave_second_low = gen_vec_interleave_lowv2di;
27240 op0 = gen_reg_rtx (second_imode);
27241 emit_insn ((*gen_interleave_second_low) (op0, ops[0],
27244 /* Cast the SECOND_IMODE vector back to a vector on original
27246 emit_insn (gen_rtx_SET (VOIDmode, target,
27247 gen_lowpart (mode, op0)));
27251 gcc_unreachable ();
27255 /* A subroutine of ix86_expand_vector_init. Handle the most general case:
27256 all values variable, and none identical. */
27259 ix86_expand_vector_init_general (bool mmx_ok, enum machine_mode mode,
27260 rtx target, rtx vals)
27262 rtx ops[32], op0, op1;
27263 enum machine_mode half_mode = VOIDmode;
27270 if (!mmx_ok && !TARGET_SSE)
27282 n = GET_MODE_NUNITS (mode);
27283 for (i = 0; i < n; i++)
27284 ops[i] = XVECEXP (vals, 0, i);
27285 ix86_expand_vector_init_concat (mode, target, ops, n);
27289 half_mode = V16QImode;
27293 half_mode = V8HImode;
27297 n = GET_MODE_NUNITS (mode);
27298 for (i = 0; i < n; i++)
27299 ops[i] = XVECEXP (vals, 0, i);
27300 op0 = gen_reg_rtx (half_mode);
27301 op1 = gen_reg_rtx (half_mode);
27302 ix86_expand_vector_init_interleave (half_mode, op0, ops,
27304 ix86_expand_vector_init_interleave (half_mode, op1,
27305 &ops [n >> 1], n >> 2);
27306 emit_insn (gen_rtx_SET (VOIDmode, target,
27307 gen_rtx_VEC_CONCAT (mode, op0, op1)));
27311 if (!TARGET_SSE4_1)
27319 /* Don't use ix86_expand_vector_init_interleave if we can't
27320 move from GPR to SSE register directly. */
27321 if (!TARGET_INTER_UNIT_MOVES)
27324 n = GET_MODE_NUNITS (mode);
27325 for (i = 0; i < n; i++)
27326 ops[i] = XVECEXP (vals, 0, i);
27327 ix86_expand_vector_init_interleave (mode, target, ops, n >> 1);
27335 gcc_unreachable ();
27339 int i, j, n_elts, n_words, n_elt_per_word;
27340 enum machine_mode inner_mode;
27341 rtx words[4], shift;
27343 inner_mode = GET_MODE_INNER (mode);
27344 n_elts = GET_MODE_NUNITS (mode);
27345 n_words = GET_MODE_SIZE (mode) / UNITS_PER_WORD;
27346 n_elt_per_word = n_elts / n_words;
27347 shift = GEN_INT (GET_MODE_BITSIZE (inner_mode));
27349 for (i = 0; i < n_words; ++i)
27351 rtx word = NULL_RTX;
27353 for (j = 0; j < n_elt_per_word; ++j)
27355 rtx elt = XVECEXP (vals, 0, (i+1)*n_elt_per_word - j - 1);
27356 elt = convert_modes (word_mode, inner_mode, elt, true);
27362 word = expand_simple_binop (word_mode, ASHIFT, word, shift,
27363 word, 1, OPTAB_LIB_WIDEN);
27364 word = expand_simple_binop (word_mode, IOR, word, elt,
27365 word, 1, OPTAB_LIB_WIDEN);
27373 emit_move_insn (target, gen_lowpart (mode, words[0]));
27374 else if (n_words == 2)
27376 rtx tmp = gen_reg_rtx (mode);
27377 emit_clobber (tmp);
27378 emit_move_insn (gen_lowpart (word_mode, tmp), words[0]);
27379 emit_move_insn (gen_highpart (word_mode, tmp), words[1]);
27380 emit_move_insn (target, tmp);
27382 else if (n_words == 4)
27384 rtx tmp = gen_reg_rtx (V4SImode);
27385 gcc_assert (word_mode == SImode);
27386 vals = gen_rtx_PARALLEL (V4SImode, gen_rtvec_v (4, words));
27387 ix86_expand_vector_init_general (false, V4SImode, tmp, vals);
27388 emit_move_insn (target, gen_lowpart (mode, tmp));
27391 gcc_unreachable ();
27395 /* Initialize vector TARGET via VALS. Suppress the use of MMX
27396 instructions unless MMX_OK is true. */
27399 ix86_expand_vector_init (bool mmx_ok, rtx target, rtx vals)
27401 enum machine_mode mode = GET_MODE (target);
27402 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27403 int n_elts = GET_MODE_NUNITS (mode);
27404 int n_var = 0, one_var = -1;
27405 bool all_same = true, all_const_zero = true;
27409 for (i = 0; i < n_elts; ++i)
27411 x = XVECEXP (vals, 0, i);
27412 if (!(CONST_INT_P (x)
27413 || GET_CODE (x) == CONST_DOUBLE
27414 || GET_CODE (x) == CONST_FIXED))
27415 n_var++, one_var = i;
27416 else if (x != CONST0_RTX (inner_mode))
27417 all_const_zero = false;
27418 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
27422 /* Constants are best loaded from the constant pool. */
27425 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
27429 /* If all values are identical, broadcast the value. */
27431 && ix86_expand_vector_init_duplicate (mmx_ok, mode, target,
27432 XVECEXP (vals, 0, 0)))
27435 /* Values where only one field is non-constant are best loaded from
27436 the pool and overwritten via move later. */
27440 && ix86_expand_vector_init_one_nonzero (mmx_ok, mode, target,
27441 XVECEXP (vals, 0, one_var),
27445 if (ix86_expand_vector_init_one_var (mmx_ok, mode, target, vals, one_var))
27449 ix86_expand_vector_init_general (mmx_ok, mode, target, vals);
27453 ix86_expand_vector_set (bool mmx_ok, rtx target, rtx val, int elt)
27455 enum machine_mode mode = GET_MODE (target);
27456 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27457 enum machine_mode half_mode;
27458 bool use_vec_merge = false;
27460 static rtx (*gen_extract[6][2]) (rtx, rtx)
27462 { gen_vec_extract_lo_v32qi, gen_vec_extract_hi_v32qi },
27463 { gen_vec_extract_lo_v16hi, gen_vec_extract_hi_v16hi },
27464 { gen_vec_extract_lo_v8si, gen_vec_extract_hi_v8si },
27465 { gen_vec_extract_lo_v4di, gen_vec_extract_hi_v4di },
27466 { gen_vec_extract_lo_v8sf, gen_vec_extract_hi_v8sf },
27467 { gen_vec_extract_lo_v4df, gen_vec_extract_hi_v4df }
27469 static rtx (*gen_insert[6][2]) (rtx, rtx, rtx)
27471 { gen_vec_set_lo_v32qi, gen_vec_set_hi_v32qi },
27472 { gen_vec_set_lo_v16hi, gen_vec_set_hi_v16hi },
27473 { gen_vec_set_lo_v8si, gen_vec_set_hi_v8si },
27474 { gen_vec_set_lo_v4di, gen_vec_set_hi_v4di },
27475 { gen_vec_set_lo_v8sf, gen_vec_set_hi_v8sf },
27476 { gen_vec_set_lo_v4df, gen_vec_set_hi_v4df }
27486 tmp = gen_reg_rtx (GET_MODE_INNER (mode));
27487 ix86_expand_vector_extract (true, tmp, target, 1 - elt);
27489 tmp = gen_rtx_VEC_CONCAT (mode, tmp, val);
27491 tmp = gen_rtx_VEC_CONCAT (mode, val, tmp);
27492 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27498 use_vec_merge = TARGET_SSE4_1;
27506 /* For the two element vectors, we implement a VEC_CONCAT with
27507 the extraction of the other element. */
27509 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (1 - elt)));
27510 tmp = gen_rtx_VEC_SELECT (inner_mode, target, tmp);
27513 op0 = val, op1 = tmp;
27515 op0 = tmp, op1 = val;
27517 tmp = gen_rtx_VEC_CONCAT (mode, op0, op1);
27518 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27523 use_vec_merge = TARGET_SSE4_1;
27530 use_vec_merge = true;
27534 /* tmp = target = A B C D */
27535 tmp = copy_to_reg (target);
27536 /* target = A A B B */
27537 emit_insn (gen_vec_interleave_lowv4sf (target, target, target));
27538 /* target = X A B B */
27539 ix86_expand_vector_set (false, target, val, 0);
27540 /* target = A X C D */
27541 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27542 const1_rtx, const0_rtx,
27543 GEN_INT (2+4), GEN_INT (3+4)));
27547 /* tmp = target = A B C D */
27548 tmp = copy_to_reg (target);
27549 /* tmp = X B C D */
27550 ix86_expand_vector_set (false, tmp, val, 0);
27551 /* target = A B X D */
27552 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27553 const0_rtx, const1_rtx,
27554 GEN_INT (0+4), GEN_INT (3+4)));
27558 /* tmp = target = A B C D */
27559 tmp = copy_to_reg (target);
27560 /* tmp = X B C D */
27561 ix86_expand_vector_set (false, tmp, val, 0);
27562 /* target = A B X D */
27563 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27564 const0_rtx, const1_rtx,
27565 GEN_INT (2+4), GEN_INT (0+4)));
27569 gcc_unreachable ();
27574 use_vec_merge = TARGET_SSE4_1;
27578 /* Element 0 handled by vec_merge below. */
27581 use_vec_merge = true;
27587 /* With SSE2, use integer shuffles to swap element 0 and ELT,
27588 store into element 0, then shuffle them back. */
27592 order[0] = GEN_INT (elt);
27593 order[1] = const1_rtx;
27594 order[2] = const2_rtx;
27595 order[3] = GEN_INT (3);
27596 order[elt] = const0_rtx;
27598 emit_insn (gen_sse2_pshufd_1 (target, target, order[0],
27599 order[1], order[2], order[3]));
27601 ix86_expand_vector_set (false, target, val, 0);
27603 emit_insn (gen_sse2_pshufd_1 (target, target, order[0],
27604 order[1], order[2], order[3]));
27608 /* For SSE1, we have to reuse the V4SF code. */
27609 ix86_expand_vector_set (false, gen_lowpart (V4SFmode, target),
27610 gen_lowpart (SFmode, val), elt);
27615 use_vec_merge = TARGET_SSE2;
27618 use_vec_merge = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A);
27622 use_vec_merge = TARGET_SSE4_1;
27629 half_mode = V16QImode;
27635 half_mode = V8HImode;
27641 half_mode = V4SImode;
27647 half_mode = V2DImode;
27653 half_mode = V4SFmode;
27659 half_mode = V2DFmode;
27665 /* Compute offset. */
27669 gcc_assert (i <= 1);
27671 /* Extract the half. */
27672 tmp = gen_reg_rtx (half_mode);
27673 emit_insn ((*gen_extract[j][i]) (tmp, target));
27675 /* Put val in tmp at elt. */
27676 ix86_expand_vector_set (false, tmp, val, elt);
27679 emit_insn ((*gen_insert[j][i]) (target, target, tmp));
27688 tmp = gen_rtx_VEC_DUPLICATE (mode, val);
27689 tmp = gen_rtx_VEC_MERGE (mode, tmp, target, GEN_INT (1 << elt));
27690 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27694 rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), false);
27696 emit_move_insn (mem, target);
27698 tmp = adjust_address (mem, inner_mode, elt*GET_MODE_SIZE (inner_mode));
27699 emit_move_insn (tmp, val);
27701 emit_move_insn (target, mem);
27706 ix86_expand_vector_extract (bool mmx_ok, rtx target, rtx vec, int elt)
27708 enum machine_mode mode = GET_MODE (vec);
27709 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27710 bool use_vec_extr = false;
27723 use_vec_extr = true;
27727 use_vec_extr = TARGET_SSE4_1;
27739 tmp = gen_reg_rtx (mode);
27740 emit_insn (gen_sse_shufps_v4sf (tmp, vec, vec,
27741 GEN_INT (elt), GEN_INT (elt),
27742 GEN_INT (elt+4), GEN_INT (elt+4)));
27746 tmp = gen_reg_rtx (mode);
27747 emit_insn (gen_vec_interleave_highv4sf (tmp, vec, vec));
27751 gcc_unreachable ();
27754 use_vec_extr = true;
27759 use_vec_extr = TARGET_SSE4_1;
27773 tmp = gen_reg_rtx (mode);
27774 emit_insn (gen_sse2_pshufd_1 (tmp, vec,
27775 GEN_INT (elt), GEN_INT (elt),
27776 GEN_INT (elt), GEN_INT (elt)));
27780 tmp = gen_reg_rtx (mode);
27781 emit_insn (gen_vec_interleave_highv4si (tmp, vec, vec));
27785 gcc_unreachable ();
27788 use_vec_extr = true;
27793 /* For SSE1, we have to reuse the V4SF code. */
27794 ix86_expand_vector_extract (false, gen_lowpart (SFmode, target),
27795 gen_lowpart (V4SFmode, vec), elt);
27801 use_vec_extr = TARGET_SSE2;
27804 use_vec_extr = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A);
27808 use_vec_extr = TARGET_SSE4_1;
27812 /* ??? Could extract the appropriate HImode element and shift. */
27819 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (elt)));
27820 tmp = gen_rtx_VEC_SELECT (inner_mode, vec, tmp);
27822 /* Let the rtl optimizers know about the zero extension performed. */
27823 if (inner_mode == QImode || inner_mode == HImode)
27825 tmp = gen_rtx_ZERO_EXTEND (SImode, tmp);
27826 target = gen_lowpart (SImode, target);
27829 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27833 rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), false);
27835 emit_move_insn (mem, vec);
27837 tmp = adjust_address (mem, inner_mode, elt*GET_MODE_SIZE (inner_mode));
27838 emit_move_insn (target, tmp);
27842 /* Expand a vector reduction on V4SFmode for SSE1. FN is the binary
27843 pattern to reduce; DEST is the destination; IN is the input vector. */
27846 ix86_expand_reduc_v4sf (rtx (*fn) (rtx, rtx, rtx), rtx dest, rtx in)
27848 rtx tmp1, tmp2, tmp3;
27850 tmp1 = gen_reg_rtx (V4SFmode);
27851 tmp2 = gen_reg_rtx (V4SFmode);
27852 tmp3 = gen_reg_rtx (V4SFmode);
27854 emit_insn (gen_sse_movhlps (tmp1, in, in));
27855 emit_insn (fn (tmp2, tmp1, in));
27857 emit_insn (gen_sse_shufps_v4sf (tmp3, tmp2, tmp2,
27858 const1_rtx, const1_rtx,
27859 GEN_INT (1+4), GEN_INT (1+4)));
27860 emit_insn (fn (dest, tmp2, tmp3));
27863 /* Target hook for scalar_mode_supported_p. */
27865 ix86_scalar_mode_supported_p (enum machine_mode mode)
27867 if (DECIMAL_FLOAT_MODE_P (mode))
27868 return default_decimal_float_supported_p ();
27869 else if (mode == TFmode)
27872 return default_scalar_mode_supported_p (mode);
27875 /* Implements target hook vector_mode_supported_p. */
27877 ix86_vector_mode_supported_p (enum machine_mode mode)
27879 if (TARGET_SSE && VALID_SSE_REG_MODE (mode))
27881 if (TARGET_SSE2 && VALID_SSE2_REG_MODE (mode))
27883 if (TARGET_AVX && VALID_AVX256_REG_MODE (mode))
27885 if (TARGET_MMX && VALID_MMX_REG_MODE (mode))
27887 if (TARGET_3DNOW && VALID_MMX_REG_MODE_3DNOW (mode))
27892 /* Target hook for c_mode_for_suffix. */
27893 static enum machine_mode
27894 ix86_c_mode_for_suffix (char suffix)
27904 /* Worker function for TARGET_MD_ASM_CLOBBERS.
27906 We do this in the new i386 backend to maintain source compatibility
27907 with the old cc0-based compiler. */
27910 ix86_md_asm_clobbers (tree outputs ATTRIBUTE_UNUSED,
27911 tree inputs ATTRIBUTE_UNUSED,
27914 clobbers = tree_cons (NULL_TREE, build_string (5, "flags"),
27916 clobbers = tree_cons (NULL_TREE, build_string (4, "fpsr"),
27921 /* Implements target vector targetm.asm.encode_section_info. This
27922 is not used by netware. */
27924 static void ATTRIBUTE_UNUSED
27925 ix86_encode_section_info (tree decl, rtx rtl, int first)
27927 default_encode_section_info (decl, rtl, first);
27929 if (TREE_CODE (decl) == VAR_DECL
27930 && (TREE_STATIC (decl) || DECL_EXTERNAL (decl))
27931 && ix86_in_large_data_p (decl))
27932 SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FAR_ADDR;
27935 /* Worker function for REVERSE_CONDITION. */
27938 ix86_reverse_condition (enum rtx_code code, enum machine_mode mode)
27940 return (mode != CCFPmode && mode != CCFPUmode
27941 ? reverse_condition (code)
27942 : reverse_condition_maybe_unordered (code));
27945 /* Output code to perform an x87 FP register move, from OPERANDS[1]
27949 output_387_reg_move (rtx insn, rtx *operands)
27951 if (REG_P (operands[0]))
27953 if (REG_P (operands[1])
27954 && find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
27956 if (REGNO (operands[0]) == FIRST_STACK_REG)
27957 return output_387_ffreep (operands, 0);
27958 return "fstp\t%y0";
27960 if (STACK_TOP_P (operands[0]))
27961 return "fld%Z1\t%y1";
27964 else if (MEM_P (operands[0]))
27966 gcc_assert (REG_P (operands[1]));
27967 if (find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
27968 return "fstp%Z0\t%y0";
27971 /* There is no non-popping store to memory for XFmode.
27972 So if we need one, follow the store with a load. */
27973 if (GET_MODE (operands[0]) == XFmode)
27974 return "fstp%Z0\t%y0\n\tfld%Z0\t%y0";
27976 return "fst%Z0\t%y0";
27983 /* Output code to perform a conditional jump to LABEL, if C2 flag in
27984 FP status register is set. */
27987 ix86_emit_fp_unordered_jump (rtx label)
27989 rtx reg = gen_reg_rtx (HImode);
27992 emit_insn (gen_x86_fnstsw_1 (reg));
27994 if (TARGET_SAHF && (TARGET_USE_SAHF || optimize_insn_for_size_p ()))
27996 emit_insn (gen_x86_sahf_1 (reg));
27998 temp = gen_rtx_REG (CCmode, FLAGS_REG);
27999 temp = gen_rtx_UNORDERED (VOIDmode, temp, const0_rtx);
28003 emit_insn (gen_testqi_ext_ccno_0 (reg, GEN_INT (0x04)));
28005 temp = gen_rtx_REG (CCNOmode, FLAGS_REG);
28006 temp = gen_rtx_NE (VOIDmode, temp, const0_rtx);
28009 temp = gen_rtx_IF_THEN_ELSE (VOIDmode, temp,
28010 gen_rtx_LABEL_REF (VOIDmode, label),
28012 temp = gen_rtx_SET (VOIDmode, pc_rtx, temp);
28014 emit_jump_insn (temp);
28015 predict_jump (REG_BR_PROB_BASE * 10 / 100);
28018 /* Output code to perform a log1p XFmode calculation. */
28020 void ix86_emit_i387_log1p (rtx op0, rtx op1)
28022 rtx label1 = gen_label_rtx ();
28023 rtx label2 = gen_label_rtx ();
28025 rtx tmp = gen_reg_rtx (XFmode);
28026 rtx tmp2 = gen_reg_rtx (XFmode);
28029 emit_insn (gen_absxf2 (tmp, op1));
28030 test = gen_rtx_GE (VOIDmode, tmp,
28031 CONST_DOUBLE_FROM_REAL_VALUE (
28032 REAL_VALUE_ATOF ("0.29289321881345247561810596348408353", XFmode),
28034 emit_jump_insn (gen_cbranchxf4 (test, XEXP (test, 0), XEXP (test, 1), label1));
28036 emit_move_insn (tmp2, standard_80387_constant_rtx (4)); /* fldln2 */
28037 emit_insn (gen_fyl2xp1xf3_i387 (op0, op1, tmp2));
28038 emit_jump (label2);
28040 emit_label (label1);
28041 emit_move_insn (tmp, CONST1_RTX (XFmode));
28042 emit_insn (gen_addxf3 (tmp, op1, tmp));
28043 emit_move_insn (tmp2, standard_80387_constant_rtx (4)); /* fldln2 */
28044 emit_insn (gen_fyl2xxf3_i387 (op0, tmp, tmp2));
28046 emit_label (label2);
28049 /* Output code to perform a Newton-Rhapson approximation of a single precision
28050 floating point divide [http://en.wikipedia.org/wiki/N-th_root_algorithm]. */
28052 void ix86_emit_swdivsf (rtx res, rtx a, rtx b, enum machine_mode mode)
28054 rtx x0, x1, e0, e1, two;
28056 x0 = gen_reg_rtx (mode);
28057 e0 = gen_reg_rtx (mode);
28058 e1 = gen_reg_rtx (mode);
28059 x1 = gen_reg_rtx (mode);
28061 two = CONST_DOUBLE_FROM_REAL_VALUE (dconst2, SFmode);
28063 if (VECTOR_MODE_P (mode))
28064 two = ix86_build_const_vector (SFmode, true, two);
28066 two = force_reg (mode, two);
28068 /* a / b = a * rcp(b) * (2.0 - b * rcp(b)) */
28070 /* x0 = rcp(b) estimate */
28071 emit_insn (gen_rtx_SET (VOIDmode, x0,
28072 gen_rtx_UNSPEC (mode, gen_rtvec (1, b),
28075 emit_insn (gen_rtx_SET (VOIDmode, e0,
28076 gen_rtx_MULT (mode, x0, a)));
28078 emit_insn (gen_rtx_SET (VOIDmode, e1,
28079 gen_rtx_MULT (mode, x0, b)));
28081 emit_insn (gen_rtx_SET (VOIDmode, x1,
28082 gen_rtx_MINUS (mode, two, e1)));
28083 /* res = e0 * x1 */
28084 emit_insn (gen_rtx_SET (VOIDmode, res,
28085 gen_rtx_MULT (mode, e0, x1)));
28088 /* Output code to perform a Newton-Rhapson approximation of a
28089 single precision floating point [reciprocal] square root. */
28091 void ix86_emit_swsqrtsf (rtx res, rtx a, enum machine_mode mode,
28094 rtx x0, e0, e1, e2, e3, mthree, mhalf;
28097 x0 = gen_reg_rtx (mode);
28098 e0 = gen_reg_rtx (mode);
28099 e1 = gen_reg_rtx (mode);
28100 e2 = gen_reg_rtx (mode);
28101 e3 = gen_reg_rtx (mode);
28103 real_from_integer (&r, VOIDmode, -3, -1, 0);
28104 mthree = CONST_DOUBLE_FROM_REAL_VALUE (r, SFmode);
28106 real_arithmetic (&r, NEGATE_EXPR, &dconsthalf, NULL);
28107 mhalf = CONST_DOUBLE_FROM_REAL_VALUE (r, SFmode);
28109 if (VECTOR_MODE_P (mode))
28111 mthree = ix86_build_const_vector (SFmode, true, mthree);
28112 mhalf = ix86_build_const_vector (SFmode, true, mhalf);
28115 /* sqrt(a) = -0.5 * a * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0)
28116 rsqrt(a) = -0.5 * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0) */
28118 /* x0 = rsqrt(a) estimate */
28119 emit_insn (gen_rtx_SET (VOIDmode, x0,
28120 gen_rtx_UNSPEC (mode, gen_rtvec (1, a),
28123 /* If (a == 0.0) Filter out infinity to prevent NaN for sqrt(0.0). */
28128 zero = gen_reg_rtx (mode);
28129 mask = gen_reg_rtx (mode);
28131 zero = force_reg (mode, CONST0_RTX(mode));
28132 emit_insn (gen_rtx_SET (VOIDmode, mask,
28133 gen_rtx_NE (mode, zero, a)));
28135 emit_insn (gen_rtx_SET (VOIDmode, x0,
28136 gen_rtx_AND (mode, x0, mask)));
28140 emit_insn (gen_rtx_SET (VOIDmode, e0,
28141 gen_rtx_MULT (mode, x0, a)));
28143 emit_insn (gen_rtx_SET (VOIDmode, e1,
28144 gen_rtx_MULT (mode, e0, x0)));
28147 mthree = force_reg (mode, mthree);
28148 emit_insn (gen_rtx_SET (VOIDmode, e2,
28149 gen_rtx_PLUS (mode, e1, mthree)));
28151 mhalf = force_reg (mode, mhalf);
28153 /* e3 = -.5 * x0 */
28154 emit_insn (gen_rtx_SET (VOIDmode, e3,
28155 gen_rtx_MULT (mode, x0, mhalf)));
28157 /* e3 = -.5 * e0 */
28158 emit_insn (gen_rtx_SET (VOIDmode, e3,
28159 gen_rtx_MULT (mode, e0, mhalf)));
28160 /* ret = e2 * e3 */
28161 emit_insn (gen_rtx_SET (VOIDmode, res,
28162 gen_rtx_MULT (mode, e2, e3)));
28165 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
28167 static void ATTRIBUTE_UNUSED
28168 i386_solaris_elf_named_section (const char *name, unsigned int flags,
28171 /* With Binutils 2.15, the "@unwind" marker must be specified on
28172 every occurrence of the ".eh_frame" section, not just the first
28175 && strcmp (name, ".eh_frame") == 0)
28177 fprintf (asm_out_file, "\t.section\t%s,\"%s\",@unwind\n", name,
28178 flags & SECTION_WRITE ? "aw" : "a");
28181 default_elf_asm_named_section (name, flags, decl);
28184 /* Return the mangling of TYPE if it is an extended fundamental type. */
28186 static const char *
28187 ix86_mangle_type (const_tree type)
28189 type = TYPE_MAIN_VARIANT (type);
28191 if (TREE_CODE (type) != VOID_TYPE && TREE_CODE (type) != BOOLEAN_TYPE
28192 && TREE_CODE (type) != INTEGER_TYPE && TREE_CODE (type) != REAL_TYPE)
28195 switch (TYPE_MODE (type))
28198 /* __float128 is "g". */
28201 /* "long double" or __float80 is "e". */
28208 /* For 32-bit code we can save PIC register setup by using
28209 __stack_chk_fail_local hidden function instead of calling
28210 __stack_chk_fail directly. 64-bit code doesn't need to setup any PIC
28211 register, so it is better to call __stack_chk_fail directly. */
28214 ix86_stack_protect_fail (void)
28216 return TARGET_64BIT
28217 ? default_external_stack_protect_fail ()
28218 : default_hidden_stack_protect_fail ();
28221 /* Select a format to encode pointers in exception handling data. CODE
28222 is 0 for data, 1 for code labels, 2 for function pointers. GLOBAL is
28223 true if the symbol may be affected by dynamic relocations.
28225 ??? All x86 object file formats are capable of representing this.
28226 After all, the relocation needed is the same as for the call insn.
28227 Whether or not a particular assembler allows us to enter such, I
28228 guess we'll have to see. */
28230 asm_preferred_eh_data_format (int code, int global)
28234 int type = DW_EH_PE_sdata8;
28236 || ix86_cmodel == CM_SMALL_PIC
28237 || (ix86_cmodel == CM_MEDIUM_PIC && (global || code)))
28238 type = DW_EH_PE_sdata4;
28239 return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type;
28241 if (ix86_cmodel == CM_SMALL
28242 || (ix86_cmodel == CM_MEDIUM && code))
28243 return DW_EH_PE_udata4;
28244 return DW_EH_PE_absptr;
28247 /* Expand copysign from SIGN to the positive value ABS_VALUE
28248 storing in RESULT. If MASK is non-null, it shall be a mask to mask out
28251 ix86_sse_copysign_to_positive (rtx result, rtx abs_value, rtx sign, rtx mask)
28253 enum machine_mode mode = GET_MODE (sign);
28254 rtx sgn = gen_reg_rtx (mode);
28255 if (mask == NULL_RTX)
28257 mask = ix86_build_signbit_mask (mode, VECTOR_MODE_P (mode), false);
28258 if (!VECTOR_MODE_P (mode))
28260 /* We need to generate a scalar mode mask in this case. */
28261 rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx));
28262 tmp = gen_rtx_VEC_SELECT (mode, mask, tmp);
28263 mask = gen_reg_rtx (mode);
28264 emit_insn (gen_rtx_SET (VOIDmode, mask, tmp));
28268 mask = gen_rtx_NOT (mode, mask);
28269 emit_insn (gen_rtx_SET (VOIDmode, sgn,
28270 gen_rtx_AND (mode, mask, sign)));
28271 emit_insn (gen_rtx_SET (VOIDmode, result,
28272 gen_rtx_IOR (mode, abs_value, sgn)));
28275 /* Expand fabs (OP0) and return a new rtx that holds the result. The
28276 mask for masking out the sign-bit is stored in *SMASK, if that is
28279 ix86_expand_sse_fabs (rtx op0, rtx *smask)
28281 enum machine_mode mode = GET_MODE (op0);
28284 xa = gen_reg_rtx (mode);
28285 mask = ix86_build_signbit_mask (mode, VECTOR_MODE_P (mode), true);
28286 if (!VECTOR_MODE_P (mode))
28288 /* We need to generate a scalar mode mask in this case. */
28289 rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx));
28290 tmp = gen_rtx_VEC_SELECT (mode, mask, tmp);
28291 mask = gen_reg_rtx (mode);
28292 emit_insn (gen_rtx_SET (VOIDmode, mask, tmp));
28294 emit_insn (gen_rtx_SET (VOIDmode, xa,
28295 gen_rtx_AND (mode, op0, mask)));
28303 /* Expands a comparison of OP0 with OP1 using comparison code CODE,
28304 swapping the operands if SWAP_OPERANDS is true. The expanded
28305 code is a forward jump to a newly created label in case the
28306 comparison is true. The generated label rtx is returned. */
28308 ix86_expand_sse_compare_and_jump (enum rtx_code code, rtx op0, rtx op1,
28309 bool swap_operands)
28320 label = gen_label_rtx ();
28321 tmp = gen_rtx_REG (CCFPUmode, FLAGS_REG);
28322 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28323 gen_rtx_COMPARE (CCFPUmode, op0, op1)));
28324 tmp = gen_rtx_fmt_ee (code, VOIDmode, tmp, const0_rtx);
28325 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
28326 gen_rtx_LABEL_REF (VOIDmode, label), pc_rtx);
28327 tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
28328 JUMP_LABEL (tmp) = label;
28333 /* Expand a mask generating SSE comparison instruction comparing OP0 with OP1
28334 using comparison code CODE. Operands are swapped for the comparison if
28335 SWAP_OPERANDS is true. Returns a rtx for the generated mask. */
28337 ix86_expand_sse_compare_mask (enum rtx_code code, rtx op0, rtx op1,
28338 bool swap_operands)
28340 enum machine_mode mode = GET_MODE (op0);
28341 rtx mask = gen_reg_rtx (mode);
28350 if (mode == DFmode)
28351 emit_insn (gen_sse2_maskcmpdf3 (mask, op0, op1,
28352 gen_rtx_fmt_ee (code, mode, op0, op1)));
28354 emit_insn (gen_sse_maskcmpsf3 (mask, op0, op1,
28355 gen_rtx_fmt_ee (code, mode, op0, op1)));
28360 /* Generate and return a rtx of mode MODE for 2**n where n is the number
28361 of bits of the mantissa of MODE, which must be one of DFmode or SFmode. */
28363 ix86_gen_TWO52 (enum machine_mode mode)
28365 REAL_VALUE_TYPE TWO52r;
28368 real_ldexp (&TWO52r, &dconst1, mode == DFmode ? 52 : 23);
28369 TWO52 = const_double_from_real_value (TWO52r, mode);
28370 TWO52 = force_reg (mode, TWO52);
28375 /* Expand SSE sequence for computing lround from OP1 storing
28378 ix86_expand_lround (rtx op0, rtx op1)
28380 /* C code for the stuff we're doing below:
28381 tmp = op1 + copysign (nextafter (0.5, 0.0), op1)
28384 enum machine_mode mode = GET_MODE (op1);
28385 const struct real_format *fmt;
28386 REAL_VALUE_TYPE pred_half, half_minus_pred_half;
28389 /* load nextafter (0.5, 0.0) */
28390 fmt = REAL_MODE_FORMAT (mode);
28391 real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode);
28392 REAL_ARITHMETIC (pred_half, MINUS_EXPR, dconsthalf, half_minus_pred_half);
28394 /* adj = copysign (0.5, op1) */
28395 adj = force_reg (mode, const_double_from_real_value (pred_half, mode));
28396 ix86_sse_copysign_to_positive (adj, adj, force_reg (mode, op1), NULL_RTX);
28398 /* adj = op1 + adj */
28399 adj = expand_simple_binop (mode, PLUS, adj, op1, NULL_RTX, 0, OPTAB_DIRECT);
28401 /* op0 = (imode)adj */
28402 expand_fix (op0, adj, 0);
28405 /* Expand SSE2 sequence for computing lround from OPERAND1 storing
28408 ix86_expand_lfloorceil (rtx op0, rtx op1, bool do_floor)
28410 /* C code for the stuff we're doing below (for do_floor):
28412 xi -= (double)xi > op1 ? 1 : 0;
28415 enum machine_mode fmode = GET_MODE (op1);
28416 enum machine_mode imode = GET_MODE (op0);
28417 rtx ireg, freg, label, tmp;
28419 /* reg = (long)op1 */
28420 ireg = gen_reg_rtx (imode);
28421 expand_fix (ireg, op1, 0);
28423 /* freg = (double)reg */
28424 freg = gen_reg_rtx (fmode);
28425 expand_float (freg, ireg, 0);
28427 /* ireg = (freg > op1) ? ireg - 1 : ireg */
28428 label = ix86_expand_sse_compare_and_jump (UNLE,
28429 freg, op1, !do_floor);
28430 tmp = expand_simple_binop (imode, do_floor ? MINUS : PLUS,
28431 ireg, const1_rtx, NULL_RTX, 0, OPTAB_DIRECT);
28432 emit_move_insn (ireg, tmp);
28434 emit_label (label);
28435 LABEL_NUSES (label) = 1;
28437 emit_move_insn (op0, ireg);
28440 /* Expand rint (IEEE round to nearest) rounding OPERAND1 and storing the
28441 result in OPERAND0. */
28443 ix86_expand_rint (rtx operand0, rtx operand1)
28445 /* C code for the stuff we're doing below:
28446 xa = fabs (operand1);
28447 if (!isless (xa, 2**52))
28449 xa = xa + 2**52 - 2**52;
28450 return copysign (xa, operand1);
28452 enum machine_mode mode = GET_MODE (operand0);
28453 rtx res, xa, label, TWO52, mask;
28455 res = gen_reg_rtx (mode);
28456 emit_move_insn (res, operand1);
28458 /* xa = abs (operand1) */
28459 xa = ix86_expand_sse_fabs (res, &mask);
28461 /* if (!isless (xa, TWO52)) goto label; */
28462 TWO52 = ix86_gen_TWO52 (mode);
28463 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28465 xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28466 xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT);
28468 ix86_sse_copysign_to_positive (res, xa, res, mask);
28470 emit_label (label);
28471 LABEL_NUSES (label) = 1;
28473 emit_move_insn (operand0, res);
28476 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
28479 ix86_expand_floorceildf_32 (rtx operand0, rtx operand1, bool do_floor)
28481 /* C code for the stuff we expand below.
28482 double xa = fabs (x), x2;
28483 if (!isless (xa, TWO52))
28485 xa = xa + TWO52 - TWO52;
28486 x2 = copysign (xa, x);
28495 enum machine_mode mode = GET_MODE (operand0);
28496 rtx xa, TWO52, tmp, label, one, res, mask;
28498 TWO52 = ix86_gen_TWO52 (mode);
28500 /* Temporary for holding the result, initialized to the input
28501 operand to ease control flow. */
28502 res = gen_reg_rtx (mode);
28503 emit_move_insn (res, operand1);
28505 /* xa = abs (operand1) */
28506 xa = ix86_expand_sse_fabs (res, &mask);
28508 /* if (!isless (xa, TWO52)) goto label; */
28509 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28511 /* xa = xa + TWO52 - TWO52; */
28512 xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28513 xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT);
28515 /* xa = copysign (xa, operand1) */
28516 ix86_sse_copysign_to_positive (xa, xa, res, mask);
28518 /* generate 1.0 or -1.0 */
28519 one = force_reg (mode,
28520 const_double_from_real_value (do_floor
28521 ? dconst1 : dconstm1, mode));
28523 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
28524 tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor);
28525 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28526 gen_rtx_AND (mode, one, tmp)));
28527 /* We always need to subtract here to preserve signed zero. */
28528 tmp = expand_simple_binop (mode, MINUS,
28529 xa, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28530 emit_move_insn (res, tmp);
28532 emit_label (label);
28533 LABEL_NUSES (label) = 1;
28535 emit_move_insn (operand0, res);
28538 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
28541 ix86_expand_floorceil (rtx operand0, rtx operand1, bool do_floor)
28543 /* C code for the stuff we expand below.
28544 double xa = fabs (x), x2;
28545 if (!isless (xa, TWO52))
28547 x2 = (double)(long)x;
28554 if (HONOR_SIGNED_ZEROS (mode))
28555 return copysign (x2, x);
28558 enum machine_mode mode = GET_MODE (operand0);
28559 rtx xa, xi, TWO52, tmp, label, one, res, mask;
28561 TWO52 = ix86_gen_TWO52 (mode);
28563 /* Temporary for holding the result, initialized to the input
28564 operand to ease control flow. */
28565 res = gen_reg_rtx (mode);
28566 emit_move_insn (res, operand1);
28568 /* xa = abs (operand1) */
28569 xa = ix86_expand_sse_fabs (res, &mask);
28571 /* if (!isless (xa, TWO52)) goto label; */
28572 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28574 /* xa = (double)(long)x */
28575 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28576 expand_fix (xi, res, 0);
28577 expand_float (xa, xi, 0);
28580 one = force_reg (mode, const_double_from_real_value (dconst1, mode));
28582 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
28583 tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor);
28584 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28585 gen_rtx_AND (mode, one, tmp)));
28586 tmp = expand_simple_binop (mode, do_floor ? MINUS : PLUS,
28587 xa, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28588 emit_move_insn (res, tmp);
28590 if (HONOR_SIGNED_ZEROS (mode))
28591 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), mask);
28593 emit_label (label);
28594 LABEL_NUSES (label) = 1;
28596 emit_move_insn (operand0, res);
28599 /* Expand SSE sequence for computing round from OPERAND1 storing
28600 into OPERAND0. Sequence that works without relying on DImode truncation
28601 via cvttsd2siq that is only available on 64bit targets. */
28603 ix86_expand_rounddf_32 (rtx operand0, rtx operand1)
28605 /* C code for the stuff we expand below.
28606 double xa = fabs (x), xa2, x2;
28607 if (!isless (xa, TWO52))
28609 Using the absolute value and copying back sign makes
28610 -0.0 -> -0.0 correct.
28611 xa2 = xa + TWO52 - TWO52;
28616 else if (dxa > 0.5)
28618 x2 = copysign (xa2, x);
28621 enum machine_mode mode = GET_MODE (operand0);
28622 rtx xa, xa2, dxa, TWO52, tmp, label, half, mhalf, one, res, mask;
28624 TWO52 = ix86_gen_TWO52 (mode);
28626 /* Temporary for holding the result, initialized to the input
28627 operand to ease control flow. */
28628 res = gen_reg_rtx (mode);
28629 emit_move_insn (res, operand1);
28631 /* xa = abs (operand1) */
28632 xa = ix86_expand_sse_fabs (res, &mask);
28634 /* if (!isless (xa, TWO52)) goto label; */
28635 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28637 /* xa2 = xa + TWO52 - TWO52; */
28638 xa2 = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28639 xa2 = expand_simple_binop (mode, MINUS, xa2, TWO52, xa2, 0, OPTAB_DIRECT);
28641 /* dxa = xa2 - xa; */
28642 dxa = expand_simple_binop (mode, MINUS, xa2, xa, NULL_RTX, 0, OPTAB_DIRECT);
28644 /* generate 0.5, 1.0 and -0.5 */
28645 half = force_reg (mode, const_double_from_real_value (dconsthalf, mode));
28646 one = expand_simple_binop (mode, PLUS, half, half, NULL_RTX, 0, OPTAB_DIRECT);
28647 mhalf = expand_simple_binop (mode, MINUS, half, one, NULL_RTX,
28651 tmp = gen_reg_rtx (mode);
28652 /* xa2 = xa2 - (dxa > 0.5 ? 1 : 0) */
28653 tmp = ix86_expand_sse_compare_mask (UNGT, dxa, half, false);
28654 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28655 gen_rtx_AND (mode, one, tmp)));
28656 xa2 = expand_simple_binop (mode, MINUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28657 /* xa2 = xa2 + (dxa <= -0.5 ? 1 : 0) */
28658 tmp = ix86_expand_sse_compare_mask (UNGE, mhalf, dxa, false);
28659 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28660 gen_rtx_AND (mode, one, tmp)));
28661 xa2 = expand_simple_binop (mode, PLUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28663 /* res = copysign (xa2, operand1) */
28664 ix86_sse_copysign_to_positive (res, xa2, force_reg (mode, operand1), mask);
28666 emit_label (label);
28667 LABEL_NUSES (label) = 1;
28669 emit_move_insn (operand0, res);
28672 /* Expand SSE sequence for computing trunc from OPERAND1 storing
28675 ix86_expand_trunc (rtx operand0, rtx operand1)
28677 /* C code for SSE variant we expand below.
28678 double xa = fabs (x), x2;
28679 if (!isless (xa, TWO52))
28681 x2 = (double)(long)x;
28682 if (HONOR_SIGNED_ZEROS (mode))
28683 return copysign (x2, x);
28686 enum machine_mode mode = GET_MODE (operand0);
28687 rtx xa, xi, TWO52, label, res, mask;
28689 TWO52 = ix86_gen_TWO52 (mode);
28691 /* Temporary for holding the result, initialized to the input
28692 operand to ease control flow. */
28693 res = gen_reg_rtx (mode);
28694 emit_move_insn (res, operand1);
28696 /* xa = abs (operand1) */
28697 xa = ix86_expand_sse_fabs (res, &mask);
28699 /* if (!isless (xa, TWO52)) goto label; */
28700 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28702 /* x = (double)(long)x */
28703 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28704 expand_fix (xi, res, 0);
28705 expand_float (res, xi, 0);
28707 if (HONOR_SIGNED_ZEROS (mode))
28708 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), mask);
28710 emit_label (label);
28711 LABEL_NUSES (label) = 1;
28713 emit_move_insn (operand0, res);
28716 /* Expand SSE sequence for computing trunc from OPERAND1 storing
28719 ix86_expand_truncdf_32 (rtx operand0, rtx operand1)
28721 enum machine_mode mode = GET_MODE (operand0);
28722 rtx xa, mask, TWO52, label, one, res, smask, tmp;
28724 /* C code for SSE variant we expand below.
28725 double xa = fabs (x), x2;
28726 if (!isless (xa, TWO52))
28728 xa2 = xa + TWO52 - TWO52;
28732 x2 = copysign (xa2, x);
28736 TWO52 = ix86_gen_TWO52 (mode);
28738 /* Temporary for holding the result, initialized to the input
28739 operand to ease control flow. */
28740 res = gen_reg_rtx (mode);
28741 emit_move_insn (res, operand1);
28743 /* xa = abs (operand1) */
28744 xa = ix86_expand_sse_fabs (res, &smask);
28746 /* if (!isless (xa, TWO52)) goto label; */
28747 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28749 /* res = xa + TWO52 - TWO52; */
28750 tmp = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28751 tmp = expand_simple_binop (mode, MINUS, tmp, TWO52, tmp, 0, OPTAB_DIRECT);
28752 emit_move_insn (res, tmp);
28755 one = force_reg (mode, const_double_from_real_value (dconst1, mode));
28757 /* Compensate: res = xa2 - (res > xa ? 1 : 0) */
28758 mask = ix86_expand_sse_compare_mask (UNGT, res, xa, false);
28759 emit_insn (gen_rtx_SET (VOIDmode, mask,
28760 gen_rtx_AND (mode, mask, one)));
28761 tmp = expand_simple_binop (mode, MINUS,
28762 res, mask, NULL_RTX, 0, OPTAB_DIRECT);
28763 emit_move_insn (res, tmp);
28765 /* res = copysign (res, operand1) */
28766 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), smask);
28768 emit_label (label);
28769 LABEL_NUSES (label) = 1;
28771 emit_move_insn (operand0, res);
28774 /* Expand SSE sequence for computing round from OPERAND1 storing
28777 ix86_expand_round (rtx operand0, rtx operand1)
28779 /* C code for the stuff we're doing below:
28780 double xa = fabs (x);
28781 if (!isless (xa, TWO52))
28783 xa = (double)(long)(xa + nextafter (0.5, 0.0));
28784 return copysign (xa, x);
28786 enum machine_mode mode = GET_MODE (operand0);
28787 rtx res, TWO52, xa, label, xi, half, mask;
28788 const struct real_format *fmt;
28789 REAL_VALUE_TYPE pred_half, half_minus_pred_half;
28791 /* Temporary for holding the result, initialized to the input
28792 operand to ease control flow. */
28793 res = gen_reg_rtx (mode);
28794 emit_move_insn (res, operand1);
28796 TWO52 = ix86_gen_TWO52 (mode);
28797 xa = ix86_expand_sse_fabs (res, &mask);
28798 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28800 /* load nextafter (0.5, 0.0) */
28801 fmt = REAL_MODE_FORMAT (mode);
28802 real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode);
28803 REAL_ARITHMETIC (pred_half, MINUS_EXPR, dconsthalf, half_minus_pred_half);
28805 /* xa = xa + 0.5 */
28806 half = force_reg (mode, const_double_from_real_value (pred_half, mode));
28807 xa = expand_simple_binop (mode, PLUS, xa, half, NULL_RTX, 0, OPTAB_DIRECT);
28809 /* xa = (double)(int64_t)xa */
28810 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28811 expand_fix (xi, xa, 0);
28812 expand_float (xa, xi, 0);
28814 /* res = copysign (xa, operand1) */
28815 ix86_sse_copysign_to_positive (res, xa, force_reg (mode, operand1), mask);
28817 emit_label (label);
28818 LABEL_NUSES (label) = 1;
28820 emit_move_insn (operand0, res);
28824 /* Table of valid machine attributes. */
28825 static const struct attribute_spec ix86_attribute_table[] =
28827 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
28828 /* Stdcall attribute says callee is responsible for popping arguments
28829 if they are not variable. */
28830 { "stdcall", 0, 0, false, true, true, ix86_handle_cconv_attribute },
28831 /* Fastcall attribute says callee is responsible for popping arguments
28832 if they are not variable. */
28833 { "fastcall", 0, 0, false, true, true, ix86_handle_cconv_attribute },
28834 /* Cdecl attribute says the callee is a normal C declaration */
28835 { "cdecl", 0, 0, false, true, true, ix86_handle_cconv_attribute },
28836 /* Regparm attribute specifies how many integer arguments are to be
28837 passed in registers. */
28838 { "regparm", 1, 1, false, true, true, ix86_handle_cconv_attribute },
28839 /* Sseregparm attribute says we are using x86_64 calling conventions
28840 for FP arguments. */
28841 { "sseregparm", 0, 0, false, true, true, ix86_handle_cconv_attribute },
28842 /* force_align_arg_pointer says this function realigns the stack at entry. */
28843 { (const char *)&ix86_force_align_arg_pointer_string, 0, 0,
28844 false, true, true, ix86_handle_cconv_attribute },
28845 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
28846 { "dllimport", 0, 0, false, false, false, handle_dll_attribute },
28847 { "dllexport", 0, 0, false, false, false, handle_dll_attribute },
28848 { "shared", 0, 0, true, false, false, ix86_handle_shared_attribute },
28850 { "ms_struct", 0, 0, false, false, false, ix86_handle_struct_attribute },
28851 { "gcc_struct", 0, 0, false, false, false, ix86_handle_struct_attribute },
28852 #ifdef SUBTARGET_ATTRIBUTE_TABLE
28853 SUBTARGET_ATTRIBUTE_TABLE,
28855 /* ms_abi and sysv_abi calling convention function attributes. */
28856 { "ms_abi", 0, 0, false, true, true, ix86_handle_abi_attribute },
28857 { "sysv_abi", 0, 0, false, true, true, ix86_handle_abi_attribute },
28858 { "ms_hook_prologue", 0, 0, true, false, false, ix86_handle_fndecl_attribute },
28860 { NULL, 0, 0, false, false, false, NULL }
28863 /* Implement targetm.vectorize.builtin_vectorization_cost. */
28865 ix86_builtin_vectorization_cost (bool runtime_test)
28867 /* If the branch of the runtime test is taken - i.e. - the vectorized
28868 version is skipped - this incurs a misprediction cost (because the
28869 vectorized version is expected to be the fall-through). So we subtract
28870 the latency of a mispredicted branch from the costs that are incured
28871 when the vectorized version is executed.
28873 TODO: The values in individual target tables have to be tuned or new
28874 fields may be needed. For eg. on K8, the default branch path is the
28875 not-taken path. If the taken path is predicted correctly, the minimum
28876 penalty of going down the taken-path is 1 cycle. If the taken-path is
28877 not predicted correctly, then the minimum penalty is 10 cycles. */
28881 return (-(ix86_cost->cond_taken_branch_cost));
28887 /* Implement targetm.vectorize.builtin_vec_perm. */
28890 ix86_vectorize_builtin_vec_perm (tree vec_type, tree *mask_type)
28892 tree itype = TREE_TYPE (vec_type);
28893 bool u = TYPE_UNSIGNED (itype);
28894 enum machine_mode vmode = TYPE_MODE (vec_type);
28895 enum ix86_builtins fcode;
28896 bool ok = TARGET_SSE2;
28902 fcode = IX86_BUILTIN_VEC_PERM_V4DF;
28905 fcode = IX86_BUILTIN_VEC_PERM_V2DF;
28907 itype = ix86_get_builtin_type (IX86_BT_DI);
28912 fcode = IX86_BUILTIN_VEC_PERM_V8SF;
28916 fcode = IX86_BUILTIN_VEC_PERM_V4SF;
28918 itype = ix86_get_builtin_type (IX86_BT_SI);
28922 fcode = u ? IX86_BUILTIN_VEC_PERM_V2DI_U : IX86_BUILTIN_VEC_PERM_V2DI;
28925 fcode = u ? IX86_BUILTIN_VEC_PERM_V4SI_U : IX86_BUILTIN_VEC_PERM_V4SI;
28928 fcode = u ? IX86_BUILTIN_VEC_PERM_V8HI_U : IX86_BUILTIN_VEC_PERM_V8HI;
28931 fcode = u ? IX86_BUILTIN_VEC_PERM_V16QI_U : IX86_BUILTIN_VEC_PERM_V16QI;
28941 *mask_type = itype;
28942 return ix86_builtins[(int) fcode];
28945 /* Return a vector mode with twice as many elements as VMODE. */
28946 /* ??? Consider moving this to a table generated by genmodes.c. */
28948 static enum machine_mode
28949 doublesize_vector_mode (enum machine_mode vmode)
28953 case V2SFmode: return V4SFmode;
28954 case V1DImode: return V2DImode;
28955 case V2SImode: return V4SImode;
28956 case V4HImode: return V8HImode;
28957 case V8QImode: return V16QImode;
28959 case V2DFmode: return V4DFmode;
28960 case V4SFmode: return V8SFmode;
28961 case V2DImode: return V4DImode;
28962 case V4SImode: return V8SImode;
28963 case V8HImode: return V16HImode;
28964 case V16QImode: return V32QImode;
28966 case V4DFmode: return V8DFmode;
28967 case V8SFmode: return V16SFmode;
28968 case V4DImode: return V8DImode;
28969 case V8SImode: return V16SImode;
28970 case V16HImode: return V32HImode;
28971 case V32QImode: return V64QImode;
28974 gcc_unreachable ();
28978 /* Construct (set target (vec_select op0 (parallel perm))) and
28979 return true if that's a valid instruction in the active ISA. */
28982 expand_vselect (rtx target, rtx op0, const unsigned char *perm, unsigned nelt)
28984 rtx rperm[MAX_VECT_LEN], x;
28987 for (i = 0; i < nelt; ++i)
28988 rperm[i] = GEN_INT (perm[i]);
28990 x = gen_rtx_PARALLEL (VOIDmode, gen_rtvec_v (nelt, rperm));
28991 x = gen_rtx_VEC_SELECT (GET_MODE (target), op0, x);
28992 x = gen_rtx_SET (VOIDmode, target, x);
28995 if (recog_memoized (x) < 0)
29003 /* Similar, but generate a vec_concat from op0 and op1 as well. */
29006 expand_vselect_vconcat (rtx target, rtx op0, rtx op1,
29007 const unsigned char *perm, unsigned nelt)
29009 enum machine_mode v2mode;
29012 v2mode = doublesize_vector_mode (GET_MODE (op0));
29013 x = gen_rtx_VEC_CONCAT (v2mode, op0, op1);
29014 return expand_vselect (target, x, perm, nelt);
29017 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to implement D
29018 in terms of blendp[sd] / pblendw / pblendvb. */
29021 expand_vec_perm_blend (struct expand_vec_perm_d *d)
29023 enum machine_mode vmode = d->vmode;
29024 unsigned i, mask, nelt = d->nelt;
29025 rtx target, op0, op1, x;
29027 if (!TARGET_SSE4_1 || d->op0 == d->op1)
29029 if (!(GET_MODE_SIZE (vmode) == 16 || vmode == V4DFmode || vmode == V8SFmode))
29032 /* This is a blend, not a permute. Elements must stay in their
29033 respective lanes. */
29034 for (i = 0; i < nelt; ++i)
29036 unsigned e = d->perm[i];
29037 if (!(e == i || e == i + nelt))
29044 /* ??? Without SSE4.1, we could implement this with and/andn/or. This
29045 decision should be extracted elsewhere, so that we only try that
29046 sequence once all budget==3 options have been tried. */
29048 /* For bytes, see if bytes move in pairs so we can use pblendw with
29049 an immediate argument, rather than pblendvb with a vector argument. */
29050 if (vmode == V16QImode)
29052 bool pblendw_ok = true;
29053 for (i = 0; i < 16 && pblendw_ok; i += 2)
29054 pblendw_ok = (d->perm[i] + 1 == d->perm[i + 1]);
29058 rtx rperm[16], vperm;
29060 for (i = 0; i < nelt; ++i)
29061 rperm[i] = (d->perm[i] < nelt ? const0_rtx : constm1_rtx);
29063 vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm));
29064 vperm = force_reg (V16QImode, vperm);
29066 emit_insn (gen_sse4_1_pblendvb (d->target, d->op0, d->op1, vperm));
29071 target = d->target;
29083 for (i = 0; i < nelt; ++i)
29084 mask |= (d->perm[i] >= nelt) << i;
29088 for (i = 0; i < 2; ++i)
29089 mask |= (d->perm[i] >= 2 ? 15 : 0) << (i * 4);
29093 for (i = 0; i < 4; ++i)
29094 mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2);
29098 for (i = 0; i < 8; ++i)
29099 mask |= (d->perm[i * 2] >= 16) << i;
29103 target = gen_lowpart (vmode, target);
29104 op0 = gen_lowpart (vmode, target);
29105 op1 = gen_lowpart (vmode, target);
29109 gcc_unreachable ();
29112 /* This matches five different patterns with the different modes. */
29113 x = gen_rtx_VEC_MERGE (vmode, op0, op1, GEN_INT (mask));
29114 x = gen_rtx_SET (VOIDmode, target, x);
29120 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to implement D
29121 in terms of the variable form of vpermilps.
29123 Note that we will have already failed the immediate input vpermilps,
29124 which requires that the high and low part shuffle be identical; the
29125 variable form doesn't require that. */
29128 expand_vec_perm_vpermil (struct expand_vec_perm_d *d)
29130 rtx rperm[8], vperm;
29133 if (!TARGET_AVX || d->vmode != V8SFmode || d->op0 != d->op1)
29136 /* We can only permute within the 128-bit lane. */
29137 for (i = 0; i < 8; ++i)
29139 unsigned e = d->perm[i];
29140 if (i < 4 ? e >= 4 : e < 4)
29147 for (i = 0; i < 8; ++i)
29149 unsigned e = d->perm[i];
29151 /* Within each 128-bit lane, the elements of op0 are numbered
29152 from 0 and the elements of op1 are numbered from 4. */
29158 rperm[i] = GEN_INT (e);
29161 vperm = gen_rtx_CONST_VECTOR (V8SImode, gen_rtvec_v (8, rperm));
29162 vperm = force_reg (V8SImode, vperm);
29163 emit_insn (gen_avx_vpermilvarv8sf3 (d->target, d->op0, vperm));
29168 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to implement D
29169 in terms of pshufb or vpperm. */
29172 expand_vec_perm_pshufb (struct expand_vec_perm_d *d)
29174 unsigned i, nelt, eltsz;
29175 rtx rperm[16], vperm, target, op0, op1;
29177 if (!(d->op0 == d->op1 ? TARGET_SSSE3 : TARGET_XOP))
29179 if (GET_MODE_SIZE (d->vmode) != 16)
29186 eltsz = GET_MODE_SIZE (GET_MODE_INNER (d->vmode));
29188 for (i = 0; i < nelt; ++i)
29190 unsigned j, e = d->perm[i];
29191 for (j = 0; j < eltsz; ++j)
29192 rperm[i * eltsz + j] = GEN_INT (e * eltsz + j);
29195 vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm));
29196 vperm = force_reg (V16QImode, vperm);
29198 target = gen_lowpart (V16QImode, d->target);
29199 op0 = gen_lowpart (V16QImode, d->op0);
29200 if (d->op0 == d->op1)
29201 emit_insn (gen_ssse3_pshufbv16qi3 (target, op0, vperm));
29204 op1 = gen_lowpart (V16QImode, d->op1);
29205 emit_insn (gen_xop_pperm (target, op0, op1, vperm));
29211 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to instantiate D
29212 in a single instruction. */
29215 expand_vec_perm_1 (struct expand_vec_perm_d *d)
29217 unsigned i, nelt = d->nelt;
29218 unsigned char perm2[MAX_VECT_LEN];
29220 /* Check plain VEC_SELECT first, because AVX has instructions that could
29221 match both SEL and SEL+CONCAT, but the plain SEL will allow a memory
29222 input where SEL+CONCAT may not. */
29223 if (d->op0 == d->op1)
29225 if (expand_vselect (d->target, d->op0, d->perm, nelt))
29228 /* There are plenty of patterns in sse.md that are written for
29229 SEL+CONCAT and are not replicated for a single op. Perhaps
29230 that should be changed, to avoid the nastiness here. */
29232 /* Recognize interleave style patterns, which means incrementing
29233 every other permutation operand. */
29234 for (i = 0; i < nelt; i += 2)
29236 perm2[i] = d->perm[i];
29237 perm2[i+1] = d->perm[i+1] + nelt;
29239 if (expand_vselect_vconcat (d->target, d->op0, d->op0, perm2, nelt))
29242 /* Recognize shufps, which means adding {0, 0, nelt, nelt}. */
29245 memcpy (perm2, d->perm, nelt);
29246 for (i = 2; i < nelt; i += 4)
29248 perm2[i+0] += nelt;
29249 perm2[i+1] += nelt;
29252 if (expand_vselect_vconcat (d->target, d->op0, d->op0, perm2, nelt))
29257 /* Finally, try the fully general two operand permute. */
29258 if (expand_vselect_vconcat (d->target, d->op0, d->op1, d->perm, nelt))
29261 /* Recognize interleave style patterns with reversed operands. */
29262 if (d->op0 != d->op1)
29264 for (i = 0; i < nelt; ++i)
29266 unsigned e = d->perm[i];
29274 if (expand_vselect_vconcat (d->target, d->op1, d->op0, perm2, nelt))
29278 /* Try the SSE4.1 blend variable merge instructions. */
29279 if (expand_vec_perm_blend (d))
29282 /* Try one of the AVX vpermil variable permutations. */
29283 if (expand_vec_perm_vpermil (d))
29286 /* Try the SSSE3 pshufb or XOP vpperm variable permutation. */
29287 if (expand_vec_perm_pshufb (d))
29293 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to implement D
29294 in terms of a pair of pshuflw + pshufhw instructions. */
29297 expand_vec_perm_pshuflw_pshufhw (struct expand_vec_perm_d *d)
29299 unsigned char perm2[MAX_VECT_LEN];
29303 if (d->vmode != V8HImode || d->op0 != d->op1)
29306 /* The two permutations only operate in 64-bit lanes. */
29307 for (i = 0; i < 4; ++i)
29308 if (d->perm[i] >= 4)
29310 for (i = 4; i < 8; ++i)
29311 if (d->perm[i] < 4)
29317 /* Emit the pshuflw. */
29318 memcpy (perm2, d->perm, 4);
29319 for (i = 4; i < 8; ++i)
29321 ok = expand_vselect (d->target, d->op0, perm2, 8);
29324 /* Emit the pshufhw. */
29325 memcpy (perm2 + 4, d->perm + 4, 4);
29326 for (i = 0; i < 4; ++i)
29328 ok = expand_vselect (d->target, d->target, perm2, 8);
29334 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to simplify
29335 the permutation using the SSSE3 palignr instruction. This succeeds
29336 when all of the elements in PERM fit within one vector and we merely
29337 need to shift them down so that a single vector permutation has a
29338 chance to succeed. */
29341 expand_vec_perm_palignr (struct expand_vec_perm_d *d)
29343 unsigned i, nelt = d->nelt;
29348 /* Even with AVX, palignr only operates on 128-bit vectors. */
29349 if (!TARGET_SSSE3 || GET_MODE_SIZE (d->vmode) != 16)
29352 min = nelt, max = 0;
29353 for (i = 0; i < nelt; ++i)
29355 unsigned e = d->perm[i];
29361 if (min == 0 || max - min >= nelt)
29364 /* Given that we have SSSE3, we know we'll be able to implement the
29365 single operand permutation after the palignr with pshufb. */
29369 shift = GEN_INT (min * GET_MODE_BITSIZE (GET_MODE_INNER (d->vmode)));
29370 emit_insn (gen_ssse3_palignrti (gen_lowpart (TImode, d->target),
29371 gen_lowpart (TImode, d->op1),
29372 gen_lowpart (TImode, d->op0), shift));
29374 d->op0 = d->op1 = d->target;
29377 for (i = 0; i < nelt; ++i)
29379 unsigned e = d->perm[i] - min;
29385 /* Test for the degenerate case where the alignment by itself
29386 produces the desired permutation. */
29390 ok = expand_vec_perm_1 (d);
29396 /* A subroutine of ix86_expand_vec_perm_builtin_1. Try to simplify
29397 a two vector permutation into a single vector permutation by using
29398 an interleave operation to merge the vectors. */
29401 expand_vec_perm_interleave2 (struct expand_vec_perm_d *d)
29403 struct expand_vec_perm_d dremap, dfinal;
29404 unsigned i, nelt = d->nelt, nelt2 = nelt / 2;
29405 unsigned contents, h1, h2, h3, h4;
29406 unsigned char remap[2 * MAX_VECT_LEN];
29410 if (d->op0 == d->op1)
29413 /* The 256-bit unpck[lh]p[sd] instructions only operate within the 128-bit
29414 lanes. We can use similar techniques with the vperm2f128 instruction,
29415 but it requires slightly different logic. */
29416 if (GET_MODE_SIZE (d->vmode) != 16)
29419 /* Examine from whence the elements come. */
29421 for (i = 0; i < nelt; ++i)
29422 contents |= 1u << d->perm[i];
29424 /* Split the two input vectors into 4 halves. */
29425 h1 = (1u << nelt2) - 1;
29430 memset (remap, 0xff, sizeof (remap));
29433 /* If the elements from the low halves use interleave low, and similarly
29434 for interleave high. If the elements are from mis-matched halves, we
29435 can use shufps for V4SF/V4SI or do a DImode shuffle. */
29436 if ((contents & (h1 | h3)) == contents)
29438 for (i = 0; i < nelt2; ++i)
29441 remap[i + nelt] = i * 2 + 1;
29442 dremap.perm[i * 2] = i;
29443 dremap.perm[i * 2 + 1] = i + nelt;
29446 else if ((contents & (h2 | h4)) == contents)
29448 for (i = 0; i < nelt2; ++i)
29450 remap[i + nelt2] = i * 2;
29451 remap[i + nelt + nelt2] = i * 2 + 1;
29452 dremap.perm[i * 2] = i + nelt2;
29453 dremap.perm[i * 2 + 1] = i + nelt + nelt2;
29456 else if ((contents & (h1 | h4)) == contents)
29458 for (i = 0; i < nelt2; ++i)
29461 remap[i + nelt + nelt2] = i + nelt2;
29462 dremap.perm[i] = i;
29463 dremap.perm[i + nelt2] = i + nelt + nelt2;
29467 dremap.vmode = V2DImode;
29469 dremap.perm[0] = 0;
29470 dremap.perm[1] = 3;
29473 else if ((contents & (h2 | h3)) == contents)
29475 for (i = 0; i < nelt2; ++i)
29477 remap[i + nelt2] = i;
29478 remap[i + nelt] = i + nelt2;
29479 dremap.perm[i] = i + nelt2;
29480 dremap.perm[i + nelt2] = i + nelt;
29484 dremap.vmode = V2DImode;
29486 dremap.perm[0] = 1;
29487 dremap.perm[1] = 2;
29493 /* Use the remapping array set up above to move the elements from their
29494 swizzled locations into their final destinations. */
29496 for (i = 0; i < nelt; ++i)
29498 unsigned e = remap[d->perm[i]];
29499 gcc_assert (e < nelt);
29500 dfinal.perm[i] = e;
29502 dfinal.op0 = gen_reg_rtx (dfinal.vmode);
29503 dfinal.op1 = dfinal.op0;
29504 dremap.target = dfinal.op0;
29506 /* Test if the final remap can be done with a single insn. For V4SFmode or
29507 V4SImode this *will* succeed. For V8HImode or V16QImode it may not. */
29509 ok = expand_vec_perm_1 (&dfinal);
29510 seq = get_insns ();
29516 if (dremap.vmode != dfinal.vmode)
29518 dremap.target = gen_lowpart (dremap.vmode, dremap.target);
29519 dremap.op0 = gen_lowpart (dremap.vmode, dremap.op0);
29520 dremap.op1 = gen_lowpart (dremap.vmode, dremap.op1);
29523 ok = expand_vec_perm_1 (&dremap);
29530 /* A subroutine of expand_vec_perm_even_odd_1. Implement the double-word
29531 permutation with two pshufb insns and an ior. We should have already
29532 failed all two instruction sequences. */
29535 expand_vec_perm_pshufb2 (struct expand_vec_perm_d *d)
29537 rtx rperm[2][16], vperm, l, h, op, m128;
29538 unsigned int i, nelt, eltsz;
29540 if (!TARGET_SSSE3 || GET_MODE_SIZE (d->vmode) != 16)
29542 gcc_assert (d->op0 != d->op1);
29545 eltsz = GET_MODE_SIZE (GET_MODE_INNER (d->vmode));
29547 /* Generate two permutation masks. If the required element is within
29548 the given vector it is shuffled into the proper lane. If the required
29549 element is in the other vector, force a zero into the lane by setting
29550 bit 7 in the permutation mask. */
29551 m128 = GEN_INT (-128);
29552 for (i = 0; i < nelt; ++i)
29554 unsigned j, e = d->perm[i];
29555 unsigned which = (e >= nelt);
29559 for (j = 0; j < eltsz; ++j)
29561 rperm[which][i*eltsz + j] = GEN_INT (e*eltsz + j);
29562 rperm[1-which][i*eltsz + j] = m128;
29566 vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm[0]));
29567 vperm = force_reg (V16QImode, vperm);
29569 l = gen_reg_rtx (V16QImode);
29570 op = gen_lowpart (V16QImode, d->op0);
29571 emit_insn (gen_ssse3_pshufbv16qi3 (l, op, vperm));
29573 vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm[1]));
29574 vperm = force_reg (V16QImode, vperm);
29576 h = gen_reg_rtx (V16QImode);
29577 op = gen_lowpart (V16QImode, d->op1);
29578 emit_insn (gen_ssse3_pshufbv16qi3 (h, op, vperm));
29580 op = gen_lowpart (V16QImode, d->target);
29581 emit_insn (gen_iorv16qi3 (op, l, h));
29586 /* A subroutine of ix86_expand_vec_perm_builtin_1. Implement extract-even
29587 and extract-odd permutations. */
29590 expand_vec_perm_even_odd_1 (struct expand_vec_perm_d *d, unsigned odd)
29592 rtx t1, t2, t3, t4;
29597 t1 = gen_reg_rtx (V4DFmode);
29598 t2 = gen_reg_rtx (V4DFmode);
29600 /* Shuffle the lanes around into { 0 1 4 5 } and { 2 3 6 7 }. */
29601 emit_insn (gen_avx_vperm2f128v4df3 (t1, d->op0, d->op1, GEN_INT (0x20)));
29602 emit_insn (gen_avx_vperm2f128v4df3 (t2, d->op0, d->op1, GEN_INT (0x31)));
29604 /* Now an unpck[lh]pd will produce the result required. */
29606 t3 = gen_avx_unpckhpd256 (d->target, t1, t2);
29608 t3 = gen_avx_unpcklpd256 (d->target, t1, t2);
29614 static const unsigned char perm1[8] = { 0, 2, 1, 3, 5, 6, 5, 7 };
29615 static const unsigned char perme[8] = { 0, 1, 8, 9, 4, 5, 12, 13 };
29616 static const unsigned char permo[8] = { 2, 3, 10, 11, 6, 7, 14, 15 };
29618 t1 = gen_reg_rtx (V8SFmode);
29619 t2 = gen_reg_rtx (V8SFmode);
29620 t3 = gen_reg_rtx (V8SFmode);
29621 t4 = gen_reg_rtx (V8SFmode);
29623 /* Shuffle within the 128-bit lanes to produce:
29624 { 0 2 1 3 4 6 5 7 } and { 8 a 9 b c e d f }. */
29625 expand_vselect (t1, d->op0, perm1, 8);
29626 expand_vselect (t2, d->op1, perm1, 8);
29628 /* Shuffle the lanes around to produce:
29629 { 0 2 1 3 8 a 9 b } and { 4 6 5 7 c e d f }. */
29630 emit_insn (gen_avx_vperm2f128v8sf3 (t3, t1, t2, GEN_INT (0x20)));
29631 emit_insn (gen_avx_vperm2f128v8sf3 (t4, t1, t2, GEN_INT (0x31)));
29633 /* Now a vpermil2p will produce the result required. */
29634 /* ??? The vpermil2p requires a vector constant. Another option
29635 is a unpck[lh]ps to merge the two vectors to produce
29636 { 0 4 2 6 8 c a e } or { 1 5 3 7 9 d b f }. Then use another
29637 vpermilps to get the elements into the final order. */
29640 memcpy (d->perm, odd ? permo: perme, 8);
29641 expand_vec_perm_vpermil (d);
29649 /* These are always directly implementable by expand_vec_perm_1. */
29650 gcc_unreachable ();
29654 return expand_vec_perm_pshufb2 (d);
29657 /* We need 2*log2(N)-1 operations to achieve odd/even
29658 with interleave. */
29659 t1 = gen_reg_rtx (V8HImode);
29660 t2 = gen_reg_rtx (V8HImode);
29661 emit_insn (gen_vec_interleave_highv8hi (t1, d->op0, d->op1));
29662 emit_insn (gen_vec_interleave_lowv8hi (d->target, d->op0, d->op1));
29663 emit_insn (gen_vec_interleave_highv8hi (t2, d->target, t1));
29664 emit_insn (gen_vec_interleave_lowv8hi (d->target, d->target, t1));
29666 t3 = gen_vec_interleave_highv8hi (d->target, d->target, t2);
29668 t3 = gen_vec_interleave_lowv8hi (d->target, d->target, t2);
29675 return expand_vec_perm_pshufb2 (d);
29678 t1 = gen_reg_rtx (V16QImode);
29679 t2 = gen_reg_rtx (V16QImode);
29680 t3 = gen_reg_rtx (V16QImode);
29681 emit_insn (gen_vec_interleave_highv16qi (t1, d->op0, d->op1));
29682 emit_insn (gen_vec_interleave_lowv16qi (d->target, d->op0, d->op1));
29683 emit_insn (gen_vec_interleave_highv16qi (t2, d->target, t1));
29684 emit_insn (gen_vec_interleave_lowv16qi (d->target, d->target, t1));
29685 emit_insn (gen_vec_interleave_highv16qi (t3, d->target, t2));
29686 emit_insn (gen_vec_interleave_lowv16qi (d->target, d->target, t2));
29688 t3 = gen_vec_interleave_highv16qi (d->target, d->target, t3);
29690 t3 = gen_vec_interleave_lowv16qi (d->target, d->target, t3);
29696 gcc_unreachable ();
29702 /* A subroutine of ix86_expand_vec_perm_builtin_1. Pattern match
29703 extract-even and extract-odd permutations. */
29706 expand_vec_perm_even_odd (struct expand_vec_perm_d *d)
29708 unsigned i, odd, nelt = d->nelt;
29711 if (odd != 0 && odd != 1)
29714 for (i = 1; i < nelt; ++i)
29715 if (d->perm[i] != 2 * i + odd)
29718 return expand_vec_perm_even_odd_1 (d, odd);
29721 /* A subroutine of ix86_expand_vec_perm_builtin_1. Implement broadcast
29722 permutations. We assume that expand_vec_perm_1 has already failed. */
29725 expand_vec_perm_broadcast_1 (struct expand_vec_perm_d *d)
29727 unsigned elt = d->perm[0], nelt2 = d->nelt / 2;
29728 enum machine_mode vmode = d->vmode;
29729 unsigned char perm2[4];
29737 /* These are special-cased in sse.md so that we can optionally
29738 use the vbroadcast instruction. They expand to two insns
29739 if the input happens to be in a register. */
29740 gcc_unreachable ();
29746 /* These are always implementable using standard shuffle patterns. */
29747 gcc_unreachable ();
29751 /* These can be implemented via interleave. We save one insn by
29752 stopping once we have promoted to V4SImode and then use pshufd. */
29755 optab otab = vec_interleave_low_optab;
29759 otab = vec_interleave_high_optab;
29764 op0 = expand_binop (vmode, otab, op0, op0, NULL, 0, OPTAB_DIRECT);
29765 vmode = get_mode_wider_vector (vmode);
29766 op0 = gen_lowpart (vmode, op0);
29768 while (vmode != V4SImode);
29770 memset (perm2, elt, 4);
29771 ok = expand_vselect (gen_lowpart (V4SImode, d->target), op0, perm2, 4);
29776 gcc_unreachable ();
29780 /* A subroutine of ix86_expand_vec_perm_builtin_1. Pattern match
29781 broadcast permutations. */
29784 expand_vec_perm_broadcast (struct expand_vec_perm_d *d)
29786 unsigned i, elt, nelt = d->nelt;
29788 if (d->op0 != d->op1)
29792 for (i = 1; i < nelt; ++i)
29793 if (d->perm[i] != elt)
29796 return expand_vec_perm_broadcast_1 (d);
29799 /* The guts of ix86_expand_vec_perm_builtin, also used by the ok hook.
29800 With all of the interface bits taken care of, perform the expansion
29801 in D and return true on success. */
29804 ix86_expand_vec_perm_builtin_1 (struct expand_vec_perm_d *d)
29806 /* Try a single instruction expansion. */
29807 if (expand_vec_perm_1 (d))
29810 /* Try sequences of two instructions. */
29812 if (expand_vec_perm_pshuflw_pshufhw (d))
29815 if (expand_vec_perm_palignr (d))
29818 if (expand_vec_perm_interleave2 (d))
29821 if (expand_vec_perm_broadcast (d))
29824 /* Try sequences of three instructions. */
29826 if (expand_vec_perm_pshufb2 (d))
29829 /* ??? Look for narrow permutations whose element orderings would
29830 allow the promotion to a wider mode. */
29832 /* ??? Look for sequences of interleave or a wider permute that place
29833 the data into the correct lanes for a half-vector shuffle like
29834 pshuf[lh]w or vpermilps. */
29836 /* ??? Look for sequences of interleave that produce the desired results.
29837 The combinatorics of punpck[lh] get pretty ugly... */
29839 if (expand_vec_perm_even_odd (d))
29845 /* Extract the values from the vector CST into the permutation array in D.
29846 Return 0 on error, 1 if all values from the permutation come from the
29847 first vector, 2 if all values from the second vector, and 3 otherwise. */
29850 extract_vec_perm_cst (struct expand_vec_perm_d *d, tree cst)
29852 tree list = TREE_VECTOR_CST_ELTS (cst);
29853 unsigned i, nelt = d->nelt;
29856 for (i = 0; i < nelt; ++i, list = TREE_CHAIN (list))
29858 unsigned HOST_WIDE_INT e;
29860 if (!host_integerp (TREE_VALUE (list), 1))
29862 e = tree_low_cst (TREE_VALUE (list), 1);
29866 ret |= (e < nelt ? 1 : 2);
29869 gcc_assert (list == NULL);
29871 /* For all elements from second vector, fold the elements to first. */
29873 for (i = 0; i < nelt; ++i)
29874 d->perm[i] -= nelt;
29880 ix86_expand_vec_perm_builtin (tree exp)
29882 struct expand_vec_perm_d d;
29883 tree arg0, arg1, arg2;
29885 arg0 = CALL_EXPR_ARG (exp, 0);
29886 arg1 = CALL_EXPR_ARG (exp, 1);
29887 arg2 = CALL_EXPR_ARG (exp, 2);
29889 d.vmode = TYPE_MODE (TREE_TYPE (arg0));
29890 d.nelt = GET_MODE_NUNITS (d.vmode);
29891 d.testing_p = false;
29892 gcc_assert (VECTOR_MODE_P (d.vmode));
29894 if (TREE_CODE (arg2) != VECTOR_CST)
29896 error_at (EXPR_LOCATION (exp),
29897 "vector permutation requires vector constant");
29901 switch (extract_vec_perm_cst (&d, arg2))
29907 error_at (EXPR_LOCATION (exp), "invalid vector permutation constant");
29911 if (!operand_equal_p (arg0, arg1, 0))
29913 d.op0 = expand_expr (arg0, NULL_RTX, d.vmode, EXPAND_NORMAL);
29914 d.op0 = force_reg (d.vmode, d.op0);
29915 d.op1 = expand_expr (arg1, NULL_RTX, d.vmode, EXPAND_NORMAL);
29916 d.op1 = force_reg (d.vmode, d.op1);
29920 /* The elements of PERM do not suggest that only the first operand
29921 is used, but both operands are identical. Allow easier matching
29922 of the permutation by folding the permutation into the single
29925 unsigned i, nelt = d.nelt;
29926 for (i = 0; i < nelt; ++i)
29927 if (d.perm[i] >= nelt)
29933 d.op0 = expand_expr (arg0, NULL_RTX, d.vmode, EXPAND_NORMAL);
29934 d.op0 = force_reg (d.vmode, d.op0);
29939 d.op0 = expand_expr (arg1, NULL_RTX, d.vmode, EXPAND_NORMAL);
29940 d.op0 = force_reg (d.vmode, d.op0);
29945 d.target = gen_reg_rtx (d.vmode);
29946 if (ix86_expand_vec_perm_builtin_1 (&d))
29949 /* For compiler generated permutations, we should never got here, because
29950 the compiler should also be checking the ok hook. But since this is a
29951 builtin the user has access too, so don't abort. */
29955 sorry ("vector permutation (%d %d)", d.perm[0], d.perm[1]);
29958 sorry ("vector permutation (%d %d %d %d)",
29959 d.perm[0], d.perm[1], d.perm[2], d.perm[3]);
29962 sorry ("vector permutation (%d %d %d %d %d %d %d %d)",
29963 d.perm[0], d.perm[1], d.perm[2], d.perm[3],
29964 d.perm[4], d.perm[5], d.perm[6], d.perm[7]);
29967 sorry ("vector permutation "
29968 "(%d %d %d %d %d %d %d %d %d %d %d %d %d %d %d %d)",
29969 d.perm[0], d.perm[1], d.perm[2], d.perm[3],
29970 d.perm[4], d.perm[5], d.perm[6], d.perm[7],
29971 d.perm[8], d.perm[9], d.perm[10], d.perm[11],
29972 d.perm[12], d.perm[13], d.perm[14], d.perm[15]);
29975 gcc_unreachable ();
29978 return CONST0_RTX (d.vmode);
29981 /* Implement targetm.vectorize.builtin_vec_perm_ok. */
29984 ix86_vectorize_builtin_vec_perm_ok (tree vec_type, tree mask)
29986 struct expand_vec_perm_d d;
29990 d.vmode = TYPE_MODE (vec_type);
29991 d.nelt = GET_MODE_NUNITS (d.vmode);
29992 d.testing_p = true;
29994 /* Given sufficient ISA support we can just return true here
29995 for selected vector modes. */
29996 if (GET_MODE_SIZE (d.vmode) == 16)
29998 /* All implementable with a single vpperm insn. */
30001 /* All implementable with 2 pshufb + 1 ior. */
30004 /* All implementable with shufpd or unpck[lh]pd. */
30009 vec_mask = extract_vec_perm_cst (&d, mask);
30011 /* This hook is cannot be called in response to something that the
30012 user does (unlike the builtin expander) so we shouldn't ever see
30013 an error generated from the extract. */
30014 gcc_assert (vec_mask > 0 && vec_mask <= 3);
30015 one_vec = (vec_mask != 3);
30017 /* Implementable with shufps or pshufd. */
30018 if (one_vec && (d.vmode == V4SFmode || d.vmode == V4SImode))
30021 /* Otherwise we have to go through the motions and see if we can
30022 figure out how to generate the requested permutation. */
30023 d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1);
30024 d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2);
30026 d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3);
30029 ret = ix86_expand_vec_perm_builtin_1 (&d);
30036 ix86_expand_vec_extract_even_odd (rtx targ, rtx op0, rtx op1, unsigned odd)
30038 struct expand_vec_perm_d d;
30044 d.vmode = GET_MODE (targ);
30045 d.nelt = nelt = GET_MODE_NUNITS (d.vmode);
30046 d.testing_p = false;
30048 for (i = 0; i < nelt; ++i)
30049 d.perm[i] = i * 2 + odd;
30051 /* We'll either be able to implement the permutation directly... */
30052 if (expand_vec_perm_1 (&d))
30055 /* ... or we use the special-case patterns. */
30056 expand_vec_perm_even_odd_1 (&d, odd);
30059 /* This function returns the calling abi specific va_list type node.
30060 It returns the FNDECL specific va_list type. */
30063 ix86_fn_abi_va_list (tree fndecl)
30066 return va_list_type_node;
30067 gcc_assert (fndecl != NULL_TREE);
30069 if (ix86_function_abi ((const_tree) fndecl) == MS_ABI)
30070 return ms_va_list_type_node;
30072 return sysv_va_list_type_node;
30075 /* Returns the canonical va_list type specified by TYPE. If there
30076 is no valid TYPE provided, it return NULL_TREE. */
30079 ix86_canonical_va_list_type (tree type)
30083 /* Resolve references and pointers to va_list type. */
30084 if (INDIRECT_REF_P (type))
30085 type = TREE_TYPE (type);
30086 else if (POINTER_TYPE_P (type) && POINTER_TYPE_P (TREE_TYPE(type)))
30087 type = TREE_TYPE (type);
30091 wtype = va_list_type_node;
30092 gcc_assert (wtype != NULL_TREE);
30094 if (TREE_CODE (wtype) == ARRAY_TYPE)
30096 /* If va_list is an array type, the argument may have decayed
30097 to a pointer type, e.g. by being passed to another function.
30098 In that case, unwrap both types so that we can compare the
30099 underlying records. */
30100 if (TREE_CODE (htype) == ARRAY_TYPE
30101 || POINTER_TYPE_P (htype))
30103 wtype = TREE_TYPE (wtype);
30104 htype = TREE_TYPE (htype);
30107 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
30108 return va_list_type_node;
30109 wtype = sysv_va_list_type_node;
30110 gcc_assert (wtype != NULL_TREE);
30112 if (TREE_CODE (wtype) == ARRAY_TYPE)
30114 /* If va_list is an array type, the argument may have decayed
30115 to a pointer type, e.g. by being passed to another function.
30116 In that case, unwrap both types so that we can compare the
30117 underlying records. */
30118 if (TREE_CODE (htype) == ARRAY_TYPE
30119 || POINTER_TYPE_P (htype))
30121 wtype = TREE_TYPE (wtype);
30122 htype = TREE_TYPE (htype);
30125 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
30126 return sysv_va_list_type_node;
30127 wtype = ms_va_list_type_node;
30128 gcc_assert (wtype != NULL_TREE);
30130 if (TREE_CODE (wtype) == ARRAY_TYPE)
30132 /* If va_list is an array type, the argument may have decayed
30133 to a pointer type, e.g. by being passed to another function.
30134 In that case, unwrap both types so that we can compare the
30135 underlying records. */
30136 if (TREE_CODE (htype) == ARRAY_TYPE
30137 || POINTER_TYPE_P (htype))
30139 wtype = TREE_TYPE (wtype);
30140 htype = TREE_TYPE (htype);
30143 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
30144 return ms_va_list_type_node;
30147 return std_canonical_va_list_type (type);
30150 /* Iterate through the target-specific builtin types for va_list.
30151 IDX denotes the iterator, *PTREE is set to the result type of
30152 the va_list builtin, and *PNAME to its internal type.
30153 Returns zero if there is no element for this index, otherwise
30154 IDX should be increased upon the next call.
30155 Note, do not iterate a base builtin's name like __builtin_va_list.
30156 Used from c_common_nodes_and_builtins. */
30159 ix86_enum_va_list (int idx, const char **pname, tree *ptree)
30165 *ptree = ms_va_list_type_node;
30166 *pname = "__builtin_ms_va_list";
30169 *ptree = sysv_va_list_type_node;
30170 *pname = "__builtin_sysv_va_list";
30178 /* Initialize the GCC target structure. */
30179 #undef TARGET_RETURN_IN_MEMORY
30180 #define TARGET_RETURN_IN_MEMORY ix86_return_in_memory
30182 #undef TARGET_LEGITIMIZE_ADDRESS
30183 #define TARGET_LEGITIMIZE_ADDRESS ix86_legitimize_address
30185 #undef TARGET_ATTRIBUTE_TABLE
30186 #define TARGET_ATTRIBUTE_TABLE ix86_attribute_table
30187 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
30188 # undef TARGET_MERGE_DECL_ATTRIBUTES
30189 # define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes
30192 #undef TARGET_COMP_TYPE_ATTRIBUTES
30193 #define TARGET_COMP_TYPE_ATTRIBUTES ix86_comp_type_attributes
30195 #undef TARGET_INIT_BUILTINS
30196 #define TARGET_INIT_BUILTINS ix86_init_builtins
30197 #undef TARGET_BUILTIN_DECL
30198 #define TARGET_BUILTIN_DECL ix86_builtin_decl
30199 #undef TARGET_EXPAND_BUILTIN
30200 #define TARGET_EXPAND_BUILTIN ix86_expand_builtin
30202 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
30203 #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
30204 ix86_builtin_vectorized_function
30206 #undef TARGET_VECTORIZE_BUILTIN_CONVERSION
30207 #define TARGET_VECTORIZE_BUILTIN_CONVERSION ix86_vectorize_builtin_conversion
30209 #undef TARGET_BUILTIN_RECIPROCAL
30210 #define TARGET_BUILTIN_RECIPROCAL ix86_builtin_reciprocal
30212 #undef TARGET_ASM_FUNCTION_EPILOGUE
30213 #define TARGET_ASM_FUNCTION_EPILOGUE ix86_output_function_epilogue
30215 #undef TARGET_ENCODE_SECTION_INFO
30216 #ifndef SUBTARGET_ENCODE_SECTION_INFO
30217 #define TARGET_ENCODE_SECTION_INFO ix86_encode_section_info
30219 #define TARGET_ENCODE_SECTION_INFO SUBTARGET_ENCODE_SECTION_INFO
30222 #undef TARGET_ASM_OPEN_PAREN
30223 #define TARGET_ASM_OPEN_PAREN ""
30224 #undef TARGET_ASM_CLOSE_PAREN
30225 #define TARGET_ASM_CLOSE_PAREN ""
30227 #undef TARGET_ASM_BYTE_OP
30228 #define TARGET_ASM_BYTE_OP ASM_BYTE
30230 #undef TARGET_ASM_ALIGNED_HI_OP
30231 #define TARGET_ASM_ALIGNED_HI_OP ASM_SHORT
30232 #undef TARGET_ASM_ALIGNED_SI_OP
30233 #define TARGET_ASM_ALIGNED_SI_OP ASM_LONG
30235 #undef TARGET_ASM_ALIGNED_DI_OP
30236 #define TARGET_ASM_ALIGNED_DI_OP ASM_QUAD
30239 #undef TARGET_ASM_UNALIGNED_HI_OP
30240 #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP
30241 #undef TARGET_ASM_UNALIGNED_SI_OP
30242 #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP
30243 #undef TARGET_ASM_UNALIGNED_DI_OP
30244 #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP
30246 #undef TARGET_SCHED_ADJUST_COST
30247 #define TARGET_SCHED_ADJUST_COST ix86_adjust_cost
30248 #undef TARGET_SCHED_ISSUE_RATE
30249 #define TARGET_SCHED_ISSUE_RATE ix86_issue_rate
30250 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
30251 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
30252 ia32_multipass_dfa_lookahead
30254 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
30255 #define TARGET_FUNCTION_OK_FOR_SIBCALL ix86_function_ok_for_sibcall
30258 #undef TARGET_HAVE_TLS
30259 #define TARGET_HAVE_TLS true
30261 #undef TARGET_CANNOT_FORCE_CONST_MEM
30262 #define TARGET_CANNOT_FORCE_CONST_MEM ix86_cannot_force_const_mem
30263 #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
30264 #define TARGET_USE_BLOCKS_FOR_CONSTANT_P hook_bool_mode_const_rtx_true
30266 #undef TARGET_DELEGITIMIZE_ADDRESS
30267 #define TARGET_DELEGITIMIZE_ADDRESS ix86_delegitimize_address
30269 #undef TARGET_MS_BITFIELD_LAYOUT_P
30270 #define TARGET_MS_BITFIELD_LAYOUT_P ix86_ms_bitfield_layout_p
30273 #undef TARGET_BINDS_LOCAL_P
30274 #define TARGET_BINDS_LOCAL_P darwin_binds_local_p
30276 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
30277 #undef TARGET_BINDS_LOCAL_P
30278 #define TARGET_BINDS_LOCAL_P i386_pe_binds_local_p
30281 #undef TARGET_ASM_OUTPUT_MI_THUNK
30282 #define TARGET_ASM_OUTPUT_MI_THUNK x86_output_mi_thunk
30283 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
30284 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK x86_can_output_mi_thunk
30286 #undef TARGET_ASM_FILE_START
30287 #define TARGET_ASM_FILE_START x86_file_start
30289 #undef TARGET_DEFAULT_TARGET_FLAGS
30290 #define TARGET_DEFAULT_TARGET_FLAGS \
30292 | TARGET_SUBTARGET_DEFAULT \
30293 | TARGET_TLS_DIRECT_SEG_REFS_DEFAULT \
30296 #undef TARGET_HANDLE_OPTION
30297 #define TARGET_HANDLE_OPTION ix86_handle_option
30299 #undef TARGET_RTX_COSTS
30300 #define TARGET_RTX_COSTS ix86_rtx_costs
30301 #undef TARGET_ADDRESS_COST
30302 #define TARGET_ADDRESS_COST ix86_address_cost
30304 #undef TARGET_FIXED_CONDITION_CODE_REGS
30305 #define TARGET_FIXED_CONDITION_CODE_REGS ix86_fixed_condition_code_regs
30306 #undef TARGET_CC_MODES_COMPATIBLE
30307 #define TARGET_CC_MODES_COMPATIBLE ix86_cc_modes_compatible
30309 #undef TARGET_MACHINE_DEPENDENT_REORG
30310 #define TARGET_MACHINE_DEPENDENT_REORG ix86_reorg
30312 #undef TARGET_BUILTIN_SETJMP_FRAME_VALUE
30313 #define TARGET_BUILTIN_SETJMP_FRAME_VALUE ix86_builtin_setjmp_frame_value
30315 #undef TARGET_BUILD_BUILTIN_VA_LIST
30316 #define TARGET_BUILD_BUILTIN_VA_LIST ix86_build_builtin_va_list
30318 #undef TARGET_FN_ABI_VA_LIST
30319 #define TARGET_FN_ABI_VA_LIST ix86_fn_abi_va_list
30321 #undef TARGET_CANONICAL_VA_LIST_TYPE
30322 #define TARGET_CANONICAL_VA_LIST_TYPE ix86_canonical_va_list_type
30324 #undef TARGET_EXPAND_BUILTIN_VA_START
30325 #define TARGET_EXPAND_BUILTIN_VA_START ix86_va_start
30327 #undef TARGET_MD_ASM_CLOBBERS
30328 #define TARGET_MD_ASM_CLOBBERS ix86_md_asm_clobbers
30330 #undef TARGET_PROMOTE_PROTOTYPES
30331 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
30332 #undef TARGET_STRUCT_VALUE_RTX
30333 #define TARGET_STRUCT_VALUE_RTX ix86_struct_value_rtx
30334 #undef TARGET_SETUP_INCOMING_VARARGS
30335 #define TARGET_SETUP_INCOMING_VARARGS ix86_setup_incoming_varargs
30336 #undef TARGET_MUST_PASS_IN_STACK
30337 #define TARGET_MUST_PASS_IN_STACK ix86_must_pass_in_stack
30338 #undef TARGET_PASS_BY_REFERENCE
30339 #define TARGET_PASS_BY_REFERENCE ix86_pass_by_reference
30340 #undef TARGET_INTERNAL_ARG_POINTER
30341 #define TARGET_INTERNAL_ARG_POINTER ix86_internal_arg_pointer
30342 #undef TARGET_UPDATE_STACK_BOUNDARY
30343 #define TARGET_UPDATE_STACK_BOUNDARY ix86_update_stack_boundary
30344 #undef TARGET_GET_DRAP_RTX
30345 #define TARGET_GET_DRAP_RTX ix86_get_drap_rtx
30346 #undef TARGET_STRICT_ARGUMENT_NAMING
30347 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
30348 #undef TARGET_STATIC_CHAIN
30349 #define TARGET_STATIC_CHAIN ix86_static_chain
30350 #undef TARGET_TRAMPOLINE_INIT
30351 #define TARGET_TRAMPOLINE_INIT ix86_trampoline_init
30353 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
30354 #define TARGET_GIMPLIFY_VA_ARG_EXPR ix86_gimplify_va_arg
30356 #undef TARGET_SCALAR_MODE_SUPPORTED_P
30357 #define TARGET_SCALAR_MODE_SUPPORTED_P ix86_scalar_mode_supported_p
30359 #undef TARGET_VECTOR_MODE_SUPPORTED_P
30360 #define TARGET_VECTOR_MODE_SUPPORTED_P ix86_vector_mode_supported_p
30362 #undef TARGET_C_MODE_FOR_SUFFIX
30363 #define TARGET_C_MODE_FOR_SUFFIX ix86_c_mode_for_suffix
30366 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
30367 #define TARGET_ASM_OUTPUT_DWARF_DTPREL i386_output_dwarf_dtprel
30370 #ifdef SUBTARGET_INSERT_ATTRIBUTES
30371 #undef TARGET_INSERT_ATTRIBUTES
30372 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
30375 #undef TARGET_MANGLE_TYPE
30376 #define TARGET_MANGLE_TYPE ix86_mangle_type
30378 #undef TARGET_STACK_PROTECT_FAIL
30379 #define TARGET_STACK_PROTECT_FAIL ix86_stack_protect_fail
30381 #undef TARGET_FUNCTION_VALUE
30382 #define TARGET_FUNCTION_VALUE ix86_function_value
30384 #undef TARGET_SECONDARY_RELOAD
30385 #define TARGET_SECONDARY_RELOAD ix86_secondary_reload
30387 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST
30388 #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \
30389 ix86_builtin_vectorization_cost
30390 #undef TARGET_VECTORIZE_BUILTIN_VEC_PERM
30391 #define TARGET_VECTORIZE_BUILTIN_VEC_PERM \
30392 ix86_vectorize_builtin_vec_perm
30393 #undef TARGET_VECTORIZE_BUILTIN_VEC_PERM_OK
30394 #define TARGET_VECTORIZE_BUILTIN_VEC_PERM_OK \
30395 ix86_vectorize_builtin_vec_perm_ok
30397 #undef TARGET_SET_CURRENT_FUNCTION
30398 #define TARGET_SET_CURRENT_FUNCTION ix86_set_current_function
30400 #undef TARGET_OPTION_VALID_ATTRIBUTE_P
30401 #define TARGET_OPTION_VALID_ATTRIBUTE_P ix86_valid_target_attribute_p
30403 #undef TARGET_OPTION_SAVE
30404 #define TARGET_OPTION_SAVE ix86_function_specific_save
30406 #undef TARGET_OPTION_RESTORE
30407 #define TARGET_OPTION_RESTORE ix86_function_specific_restore
30409 #undef TARGET_OPTION_PRINT
30410 #define TARGET_OPTION_PRINT ix86_function_specific_print
30412 #undef TARGET_CAN_INLINE_P
30413 #define TARGET_CAN_INLINE_P ix86_can_inline_p
30415 #undef TARGET_EXPAND_TO_RTL_HOOK
30416 #define TARGET_EXPAND_TO_RTL_HOOK ix86_maybe_switch_abi
30418 #undef TARGET_LEGITIMATE_ADDRESS_P
30419 #define TARGET_LEGITIMATE_ADDRESS_P ix86_legitimate_address_p
30421 #undef TARGET_IRA_COVER_CLASSES
30422 #define TARGET_IRA_COVER_CLASSES i386_ira_cover_classes
30424 #undef TARGET_FRAME_POINTER_REQUIRED
30425 #define TARGET_FRAME_POINTER_REQUIRED ix86_frame_pointer_required
30427 #undef TARGET_CAN_ELIMINATE
30428 #define TARGET_CAN_ELIMINATE ix86_can_eliminate
30430 struct gcc_target targetm = TARGET_INITIALIZER;
30432 #include "gt-i386.h"
OSDN Git Service
