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"
45 #include "basic-block.h"
48 #include "target-def.h"
49 #include "langhooks.h"
54 #include "tm-constrs.h"
58 static int x86_builtin_vectorization_cost (bool);
59 static rtx legitimize_dllimport_symbol (rtx, bool);
61 #ifndef CHECK_STACK_LIMIT
62 #define CHECK_STACK_LIMIT (-1)
65 /* Return index of given mode in mult and division cost tables. */
66 #define MODE_INDEX(mode) \
67 ((mode) == QImode ? 0 \
68 : (mode) == HImode ? 1 \
69 : (mode) == SImode ? 2 \
70 : (mode) == DImode ? 3 \
73 /* Processor costs (relative to an add) */
74 /* We assume COSTS_N_INSNS is defined as (N)*4 and an addition is 2 bytes. */
75 #define COSTS_N_BYTES(N) ((N) * 2)
77 #define DUMMY_STRINGOP_ALGS {libcall, {{-1, libcall}}}
80 struct processor_costs ix86_size_cost = {/* costs for tuning for size */
81 COSTS_N_BYTES (2), /* cost of an add instruction */
82 COSTS_N_BYTES (3), /* cost of a lea instruction */
83 COSTS_N_BYTES (2), /* variable shift costs */
84 COSTS_N_BYTES (3), /* constant shift costs */
85 {COSTS_N_BYTES (3), /* cost of starting multiply for QI */
86 COSTS_N_BYTES (3), /* HI */
87 COSTS_N_BYTES (3), /* SI */
88 COSTS_N_BYTES (3), /* DI */
89 COSTS_N_BYTES (5)}, /* other */
90 0, /* cost of multiply per each bit set */
91 {COSTS_N_BYTES (3), /* cost of a divide/mod for QI */
92 COSTS_N_BYTES (3), /* HI */
93 COSTS_N_BYTES (3), /* SI */
94 COSTS_N_BYTES (3), /* DI */
95 COSTS_N_BYTES (5)}, /* other */
96 COSTS_N_BYTES (3), /* cost of movsx */
97 COSTS_N_BYTES (3), /* cost of movzx */
100 2, /* cost for loading QImode using movzbl */
101 {2, 2, 2}, /* cost of loading integer registers
102 in QImode, HImode and SImode.
103 Relative to reg-reg move (2). */
104 {2, 2, 2}, /* cost of storing integer registers */
105 2, /* cost of reg,reg fld/fst */
106 {2, 2, 2}, /* cost of loading fp registers
107 in SFmode, DFmode and XFmode */
108 {2, 2, 2}, /* cost of storing fp registers
109 in SFmode, DFmode and XFmode */
110 3, /* cost of moving MMX register */
111 {3, 3}, /* cost of loading MMX registers
112 in SImode and DImode */
113 {3, 3}, /* cost of storing MMX registers
114 in SImode and DImode */
115 3, /* cost of moving SSE register */
116 {3, 3, 3}, /* cost of loading SSE registers
117 in SImode, DImode and TImode */
118 {3, 3, 3}, /* cost of storing SSE registers
119 in SImode, DImode and TImode */
120 3, /* MMX or SSE register to integer */
121 0, /* size of l1 cache */
122 0, /* size of l2 cache */
123 0, /* size of prefetch block */
124 0, /* number of parallel prefetches */
126 COSTS_N_BYTES (2), /* cost of FADD and FSUB insns. */
127 COSTS_N_BYTES (2), /* cost of FMUL instruction. */
128 COSTS_N_BYTES (2), /* cost of FDIV instruction. */
129 COSTS_N_BYTES (2), /* cost of FABS instruction. */
130 COSTS_N_BYTES (2), /* cost of FCHS instruction. */
131 COSTS_N_BYTES (2), /* cost of FSQRT instruction. */
132 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
133 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
134 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
135 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
136 1, /* scalar_stmt_cost. */
137 1, /* scalar load_cost. */
138 1, /* scalar_store_cost. */
139 1, /* vec_stmt_cost. */
140 1, /* vec_to_scalar_cost. */
141 1, /* scalar_to_vec_cost. */
142 1, /* vec_align_load_cost. */
143 1, /* vec_unalign_load_cost. */
144 1, /* vec_store_cost. */
145 1, /* cond_taken_branch_cost. */
146 1, /* cond_not_taken_branch_cost. */
149 /* Processor costs (relative to an add) */
151 struct processor_costs i386_cost = { /* 386 specific costs */
152 COSTS_N_INSNS (1), /* cost of an add instruction */
153 COSTS_N_INSNS (1), /* cost of a lea instruction */
154 COSTS_N_INSNS (3), /* variable shift costs */
155 COSTS_N_INSNS (2), /* constant shift costs */
156 {COSTS_N_INSNS (6), /* cost of starting multiply for QI */
157 COSTS_N_INSNS (6), /* HI */
158 COSTS_N_INSNS (6), /* SI */
159 COSTS_N_INSNS (6), /* DI */
160 COSTS_N_INSNS (6)}, /* other */
161 COSTS_N_INSNS (1), /* cost of multiply per each bit set */
162 {COSTS_N_INSNS (23), /* cost of a divide/mod for QI */
163 COSTS_N_INSNS (23), /* HI */
164 COSTS_N_INSNS (23), /* SI */
165 COSTS_N_INSNS (23), /* DI */
166 COSTS_N_INSNS (23)}, /* other */
167 COSTS_N_INSNS (3), /* cost of movsx */
168 COSTS_N_INSNS (2), /* cost of movzx */
169 15, /* "large" insn */
171 4, /* cost for loading QImode using movzbl */
172 {2, 4, 2}, /* cost of loading integer registers
173 in QImode, HImode and SImode.
174 Relative to reg-reg move (2). */
175 {2, 4, 2}, /* cost of storing integer registers */
176 2, /* cost of reg,reg fld/fst */
177 {8, 8, 8}, /* cost of loading fp registers
178 in SFmode, DFmode and XFmode */
179 {8, 8, 8}, /* cost of storing fp registers
180 in SFmode, DFmode and XFmode */
181 2, /* cost of moving MMX register */
182 {4, 8}, /* cost of loading MMX registers
183 in SImode and DImode */
184 {4, 8}, /* cost of storing MMX registers
185 in SImode and DImode */
186 2, /* cost of moving SSE register */
187 {4, 8, 16}, /* cost of loading SSE registers
188 in SImode, DImode and TImode */
189 {4, 8, 16}, /* cost of storing SSE registers
190 in SImode, DImode and TImode */
191 3, /* MMX or SSE register to integer */
192 0, /* size of l1 cache */
193 0, /* size of l2 cache */
194 0, /* size of prefetch block */
195 0, /* number of parallel prefetches */
197 COSTS_N_INSNS (23), /* cost of FADD and FSUB insns. */
198 COSTS_N_INSNS (27), /* cost of FMUL instruction. */
199 COSTS_N_INSNS (88), /* cost of FDIV instruction. */
200 COSTS_N_INSNS (22), /* cost of FABS instruction. */
201 COSTS_N_INSNS (24), /* cost of FCHS instruction. */
202 COSTS_N_INSNS (122), /* cost of FSQRT instruction. */
203 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
204 DUMMY_STRINGOP_ALGS},
205 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
206 DUMMY_STRINGOP_ALGS},
207 1, /* scalar_stmt_cost. */
208 1, /* scalar load_cost. */
209 1, /* scalar_store_cost. */
210 1, /* vec_stmt_cost. */
211 1, /* vec_to_scalar_cost. */
212 1, /* scalar_to_vec_cost. */
213 1, /* vec_align_load_cost. */
214 2, /* vec_unalign_load_cost. */
215 1, /* vec_store_cost. */
216 3, /* cond_taken_branch_cost. */
217 1, /* cond_not_taken_branch_cost. */
221 struct processor_costs i486_cost = { /* 486 specific costs */
222 COSTS_N_INSNS (1), /* cost of an add instruction */
223 COSTS_N_INSNS (1), /* cost of a lea instruction */
224 COSTS_N_INSNS (3), /* variable shift costs */
225 COSTS_N_INSNS (2), /* constant shift costs */
226 {COSTS_N_INSNS (12), /* cost of starting multiply for QI */
227 COSTS_N_INSNS (12), /* HI */
228 COSTS_N_INSNS (12), /* SI */
229 COSTS_N_INSNS (12), /* DI */
230 COSTS_N_INSNS (12)}, /* other */
231 1, /* cost of multiply per each bit set */
232 {COSTS_N_INSNS (40), /* cost of a divide/mod for QI */
233 COSTS_N_INSNS (40), /* HI */
234 COSTS_N_INSNS (40), /* SI */
235 COSTS_N_INSNS (40), /* DI */
236 COSTS_N_INSNS (40)}, /* other */
237 COSTS_N_INSNS (3), /* cost of movsx */
238 COSTS_N_INSNS (2), /* cost of movzx */
239 15, /* "large" insn */
241 4, /* cost for loading QImode using movzbl */
242 {2, 4, 2}, /* cost of loading integer registers
243 in QImode, HImode and SImode.
244 Relative to reg-reg move (2). */
245 {2, 4, 2}, /* cost of storing integer registers */
246 2, /* cost of reg,reg fld/fst */
247 {8, 8, 8}, /* cost of loading fp registers
248 in SFmode, DFmode and XFmode */
249 {8, 8, 8}, /* cost of storing fp registers
250 in SFmode, DFmode and XFmode */
251 2, /* cost of moving MMX register */
252 {4, 8}, /* cost of loading MMX registers
253 in SImode and DImode */
254 {4, 8}, /* cost of storing MMX registers
255 in SImode and DImode */
256 2, /* cost of moving SSE register */
257 {4, 8, 16}, /* cost of loading SSE registers
258 in SImode, DImode and TImode */
259 {4, 8, 16}, /* cost of storing SSE registers
260 in SImode, DImode and TImode */
261 3, /* MMX or SSE register to integer */
262 4, /* size of l1 cache. 486 has 8kB cache
263 shared for code and data, so 4kB is
264 not really precise. */
265 4, /* size of l2 cache */
266 0, /* size of prefetch block */
267 0, /* number of parallel prefetches */
269 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
270 COSTS_N_INSNS (16), /* cost of FMUL instruction. */
271 COSTS_N_INSNS (73), /* cost of FDIV instruction. */
272 COSTS_N_INSNS (3), /* cost of FABS instruction. */
273 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
274 COSTS_N_INSNS (83), /* cost of FSQRT instruction. */
275 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
276 DUMMY_STRINGOP_ALGS},
277 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
278 DUMMY_STRINGOP_ALGS},
279 1, /* scalar_stmt_cost. */
280 1, /* scalar load_cost. */
281 1, /* scalar_store_cost. */
282 1, /* vec_stmt_cost. */
283 1, /* vec_to_scalar_cost. */
284 1, /* scalar_to_vec_cost. */
285 1, /* vec_align_load_cost. */
286 2, /* vec_unalign_load_cost. */
287 1, /* vec_store_cost. */
288 3, /* cond_taken_branch_cost. */
289 1, /* cond_not_taken_branch_cost. */
293 struct processor_costs pentium_cost = {
294 COSTS_N_INSNS (1), /* cost of an add instruction */
295 COSTS_N_INSNS (1), /* cost of a lea instruction */
296 COSTS_N_INSNS (4), /* variable shift costs */
297 COSTS_N_INSNS (1), /* constant shift costs */
298 {COSTS_N_INSNS (11), /* cost of starting multiply for QI */
299 COSTS_N_INSNS (11), /* HI */
300 COSTS_N_INSNS (11), /* SI */
301 COSTS_N_INSNS (11), /* DI */
302 COSTS_N_INSNS (11)}, /* other */
303 0, /* cost of multiply per each bit set */
304 {COSTS_N_INSNS (25), /* cost of a divide/mod for QI */
305 COSTS_N_INSNS (25), /* HI */
306 COSTS_N_INSNS (25), /* SI */
307 COSTS_N_INSNS (25), /* DI */
308 COSTS_N_INSNS (25)}, /* other */
309 COSTS_N_INSNS (3), /* cost of movsx */
310 COSTS_N_INSNS (2), /* cost of movzx */
311 8, /* "large" insn */
313 6, /* cost for loading QImode using movzbl */
314 {2, 4, 2}, /* cost of loading integer registers
315 in QImode, HImode and SImode.
316 Relative to reg-reg move (2). */
317 {2, 4, 2}, /* cost of storing integer registers */
318 2, /* cost of reg,reg fld/fst */
319 {2, 2, 6}, /* cost of loading fp registers
320 in SFmode, DFmode and XFmode */
321 {4, 4, 6}, /* cost of storing fp registers
322 in SFmode, DFmode and XFmode */
323 8, /* cost of moving MMX register */
324 {8, 8}, /* cost of loading MMX registers
325 in SImode and DImode */
326 {8, 8}, /* cost of storing MMX registers
327 in SImode and DImode */
328 2, /* cost of moving SSE register */
329 {4, 8, 16}, /* cost of loading SSE registers
330 in SImode, DImode and TImode */
331 {4, 8, 16}, /* cost of storing SSE registers
332 in SImode, DImode and TImode */
333 3, /* MMX or SSE register to integer */
334 8, /* size of l1 cache. */
335 8, /* size of l2 cache */
336 0, /* size of prefetch block */
337 0, /* number of parallel prefetches */
339 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
340 COSTS_N_INSNS (3), /* cost of FMUL instruction. */
341 COSTS_N_INSNS (39), /* cost of FDIV instruction. */
342 COSTS_N_INSNS (1), /* cost of FABS instruction. */
343 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
344 COSTS_N_INSNS (70), /* cost of FSQRT instruction. */
345 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
346 DUMMY_STRINGOP_ALGS},
347 {{libcall, {{-1, rep_prefix_4_byte}}},
348 DUMMY_STRINGOP_ALGS},
349 1, /* scalar_stmt_cost. */
350 1, /* scalar load_cost. */
351 1, /* scalar_store_cost. */
352 1, /* vec_stmt_cost. */
353 1, /* vec_to_scalar_cost. */
354 1, /* scalar_to_vec_cost. */
355 1, /* vec_align_load_cost. */
356 2, /* vec_unalign_load_cost. */
357 1, /* vec_store_cost. */
358 3, /* cond_taken_branch_cost. */
359 1, /* cond_not_taken_branch_cost. */
363 struct processor_costs pentiumpro_cost = {
364 COSTS_N_INSNS (1), /* cost of an add instruction */
365 COSTS_N_INSNS (1), /* cost of a lea instruction */
366 COSTS_N_INSNS (1), /* variable shift costs */
367 COSTS_N_INSNS (1), /* constant shift costs */
368 {COSTS_N_INSNS (4), /* cost of starting multiply for QI */
369 COSTS_N_INSNS (4), /* HI */
370 COSTS_N_INSNS (4), /* SI */
371 COSTS_N_INSNS (4), /* DI */
372 COSTS_N_INSNS (4)}, /* other */
373 0, /* cost of multiply per each bit set */
374 {COSTS_N_INSNS (17), /* cost of a divide/mod for QI */
375 COSTS_N_INSNS (17), /* HI */
376 COSTS_N_INSNS (17), /* SI */
377 COSTS_N_INSNS (17), /* DI */
378 COSTS_N_INSNS (17)}, /* other */
379 COSTS_N_INSNS (1), /* cost of movsx */
380 COSTS_N_INSNS (1), /* cost of movzx */
381 8, /* "large" insn */
383 2, /* cost for loading QImode using movzbl */
384 {4, 4, 4}, /* cost of loading integer registers
385 in QImode, HImode and SImode.
386 Relative to reg-reg move (2). */
387 {2, 2, 2}, /* cost of storing integer registers */
388 2, /* cost of reg,reg fld/fst */
389 {2, 2, 6}, /* cost of loading fp registers
390 in SFmode, DFmode and XFmode */
391 {4, 4, 6}, /* cost of storing fp registers
392 in SFmode, DFmode and XFmode */
393 2, /* cost of moving MMX register */
394 {2, 2}, /* cost of loading MMX registers
395 in SImode and DImode */
396 {2, 2}, /* cost of storing MMX registers
397 in SImode and DImode */
398 2, /* cost of moving SSE register */
399 {2, 2, 8}, /* cost of loading SSE registers
400 in SImode, DImode and TImode */
401 {2, 2, 8}, /* cost of storing SSE registers
402 in SImode, DImode and TImode */
403 3, /* MMX or SSE register to integer */
404 8, /* size of l1 cache. */
405 256, /* size of l2 cache */
406 32, /* size of prefetch block */
407 6, /* number of parallel prefetches */
409 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
410 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
411 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
412 COSTS_N_INSNS (2), /* cost of FABS instruction. */
413 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
414 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
415 /* PentiumPro has optimized rep instructions for blocks aligned by 8 bytes (we ensure
416 the alignment). For small blocks inline loop is still a noticeable win, for bigger
417 blocks either rep movsl or rep movsb is way to go. Rep movsb has apparently
418 more expensive startup time in CPU, but after 4K the difference is down in the noise.
420 {{rep_prefix_4_byte, {{128, loop}, {1024, unrolled_loop},
421 {8192, rep_prefix_4_byte}, {-1, rep_prefix_1_byte}}},
422 DUMMY_STRINGOP_ALGS},
423 {{rep_prefix_4_byte, {{1024, unrolled_loop},
424 {8192, rep_prefix_4_byte}, {-1, libcall}}},
425 DUMMY_STRINGOP_ALGS},
426 1, /* scalar_stmt_cost. */
427 1, /* scalar load_cost. */
428 1, /* scalar_store_cost. */
429 1, /* vec_stmt_cost. */
430 1, /* vec_to_scalar_cost. */
431 1, /* scalar_to_vec_cost. */
432 1, /* vec_align_load_cost. */
433 2, /* vec_unalign_load_cost. */
434 1, /* vec_store_cost. */
435 3, /* cond_taken_branch_cost. */
436 1, /* cond_not_taken_branch_cost. */
440 struct processor_costs geode_cost = {
441 COSTS_N_INSNS (1), /* cost of an add instruction */
442 COSTS_N_INSNS (1), /* cost of a lea instruction */
443 COSTS_N_INSNS (2), /* variable shift costs */
444 COSTS_N_INSNS (1), /* constant shift costs */
445 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
446 COSTS_N_INSNS (4), /* HI */
447 COSTS_N_INSNS (7), /* SI */
448 COSTS_N_INSNS (7), /* DI */
449 COSTS_N_INSNS (7)}, /* other */
450 0, /* cost of multiply per each bit set */
451 {COSTS_N_INSNS (15), /* cost of a divide/mod for QI */
452 COSTS_N_INSNS (23), /* HI */
453 COSTS_N_INSNS (39), /* SI */
454 COSTS_N_INSNS (39), /* DI */
455 COSTS_N_INSNS (39)}, /* other */
456 COSTS_N_INSNS (1), /* cost of movsx */
457 COSTS_N_INSNS (1), /* cost of movzx */
458 8, /* "large" insn */
460 1, /* cost for loading QImode using movzbl */
461 {1, 1, 1}, /* cost of loading integer registers
462 in QImode, HImode and SImode.
463 Relative to reg-reg move (2). */
464 {1, 1, 1}, /* cost of storing integer registers */
465 1, /* cost of reg,reg fld/fst */
466 {1, 1, 1}, /* cost of loading fp registers
467 in SFmode, DFmode and XFmode */
468 {4, 6, 6}, /* cost of storing fp registers
469 in SFmode, DFmode and XFmode */
471 1, /* cost of moving MMX register */
472 {1, 1}, /* cost of loading MMX registers
473 in SImode and DImode */
474 {1, 1}, /* cost of storing MMX registers
475 in SImode and DImode */
476 1, /* cost of moving SSE register */
477 {1, 1, 1}, /* cost of loading SSE registers
478 in SImode, DImode and TImode */
479 {1, 1, 1}, /* cost of storing SSE registers
480 in SImode, DImode and TImode */
481 1, /* MMX or SSE register to integer */
482 64, /* size of l1 cache. */
483 128, /* size of l2 cache. */
484 32, /* size of prefetch block */
485 1, /* number of parallel prefetches */
487 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
488 COSTS_N_INSNS (11), /* cost of FMUL instruction. */
489 COSTS_N_INSNS (47), /* cost of FDIV instruction. */
490 COSTS_N_INSNS (1), /* cost of FABS instruction. */
491 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
492 COSTS_N_INSNS (54), /* cost of FSQRT instruction. */
493 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
494 DUMMY_STRINGOP_ALGS},
495 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
496 DUMMY_STRINGOP_ALGS},
497 1, /* scalar_stmt_cost. */
498 1, /* scalar load_cost. */
499 1, /* scalar_store_cost. */
500 1, /* vec_stmt_cost. */
501 1, /* vec_to_scalar_cost. */
502 1, /* scalar_to_vec_cost. */
503 1, /* vec_align_load_cost. */
504 2, /* vec_unalign_load_cost. */
505 1, /* vec_store_cost. */
506 3, /* cond_taken_branch_cost. */
507 1, /* cond_not_taken_branch_cost. */
511 struct processor_costs k6_cost = {
512 COSTS_N_INSNS (1), /* cost of an add instruction */
513 COSTS_N_INSNS (2), /* cost of a lea instruction */
514 COSTS_N_INSNS (1), /* variable shift costs */
515 COSTS_N_INSNS (1), /* constant shift costs */
516 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
517 COSTS_N_INSNS (3), /* HI */
518 COSTS_N_INSNS (3), /* SI */
519 COSTS_N_INSNS (3), /* DI */
520 COSTS_N_INSNS (3)}, /* other */
521 0, /* cost of multiply per each bit set */
522 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
523 COSTS_N_INSNS (18), /* HI */
524 COSTS_N_INSNS (18), /* SI */
525 COSTS_N_INSNS (18), /* DI */
526 COSTS_N_INSNS (18)}, /* other */
527 COSTS_N_INSNS (2), /* cost of movsx */
528 COSTS_N_INSNS (2), /* cost of movzx */
529 8, /* "large" insn */
531 3, /* cost for loading QImode using movzbl */
532 {4, 5, 4}, /* cost of loading integer registers
533 in QImode, HImode and SImode.
534 Relative to reg-reg move (2). */
535 {2, 3, 2}, /* cost of storing integer registers */
536 4, /* cost of reg,reg fld/fst */
537 {6, 6, 6}, /* cost of loading fp registers
538 in SFmode, DFmode and XFmode */
539 {4, 4, 4}, /* cost of storing fp registers
540 in SFmode, DFmode and XFmode */
541 2, /* cost of moving MMX register */
542 {2, 2}, /* cost of loading MMX registers
543 in SImode and DImode */
544 {2, 2}, /* cost of storing MMX registers
545 in SImode and DImode */
546 2, /* cost of moving SSE register */
547 {2, 2, 8}, /* cost of loading SSE registers
548 in SImode, DImode and TImode */
549 {2, 2, 8}, /* cost of storing SSE registers
550 in SImode, DImode and TImode */
551 6, /* MMX or SSE register to integer */
552 32, /* size of l1 cache. */
553 32, /* size of l2 cache. Some models
554 have integrated l2 cache, but
555 optimizing for k6 is not important
556 enough to worry about that. */
557 32, /* size of prefetch block */
558 1, /* number of parallel prefetches */
560 COSTS_N_INSNS (2), /* cost of FADD and FSUB insns. */
561 COSTS_N_INSNS (2), /* cost of FMUL instruction. */
562 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
563 COSTS_N_INSNS (2), /* cost of FABS instruction. */
564 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
565 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
566 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
567 DUMMY_STRINGOP_ALGS},
568 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
569 DUMMY_STRINGOP_ALGS},
570 1, /* scalar_stmt_cost. */
571 1, /* scalar load_cost. */
572 1, /* scalar_store_cost. */
573 1, /* vec_stmt_cost. */
574 1, /* vec_to_scalar_cost. */
575 1, /* scalar_to_vec_cost. */
576 1, /* vec_align_load_cost. */
577 2, /* vec_unalign_load_cost. */
578 1, /* vec_store_cost. */
579 3, /* cond_taken_branch_cost. */
580 1, /* cond_not_taken_branch_cost. */
584 struct processor_costs athlon_cost = {
585 COSTS_N_INSNS (1), /* cost of an add instruction */
586 COSTS_N_INSNS (2), /* cost of a lea instruction */
587 COSTS_N_INSNS (1), /* variable shift costs */
588 COSTS_N_INSNS (1), /* constant shift costs */
589 {COSTS_N_INSNS (5), /* cost of starting multiply for QI */
590 COSTS_N_INSNS (5), /* HI */
591 COSTS_N_INSNS (5), /* SI */
592 COSTS_N_INSNS (5), /* DI */
593 COSTS_N_INSNS (5)}, /* other */
594 0, /* cost of multiply per each bit set */
595 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
596 COSTS_N_INSNS (26), /* HI */
597 COSTS_N_INSNS (42), /* SI */
598 COSTS_N_INSNS (74), /* DI */
599 COSTS_N_INSNS (74)}, /* other */
600 COSTS_N_INSNS (1), /* cost of movsx */
601 COSTS_N_INSNS (1), /* cost of movzx */
602 8, /* "large" insn */
604 4, /* cost for loading QImode using movzbl */
605 {3, 4, 3}, /* cost of loading integer registers
606 in QImode, HImode and SImode.
607 Relative to reg-reg move (2). */
608 {3, 4, 3}, /* cost of storing integer registers */
609 4, /* cost of reg,reg fld/fst */
610 {4, 4, 12}, /* cost of loading fp registers
611 in SFmode, DFmode and XFmode */
612 {6, 6, 8}, /* cost of storing fp registers
613 in SFmode, DFmode and XFmode */
614 2, /* cost of moving MMX register */
615 {4, 4}, /* cost of loading MMX registers
616 in SImode and DImode */
617 {4, 4}, /* cost of storing MMX registers
618 in SImode and DImode */
619 2, /* cost of moving SSE register */
620 {4, 4, 6}, /* cost of loading SSE registers
621 in SImode, DImode and TImode */
622 {4, 4, 5}, /* cost of storing SSE registers
623 in SImode, DImode and TImode */
624 5, /* MMX or SSE register to integer */
625 64, /* size of l1 cache. */
626 256, /* size of l2 cache. */
627 64, /* size of prefetch block */
628 6, /* number of parallel prefetches */
630 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
631 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
632 COSTS_N_INSNS (24), /* cost of FDIV instruction. */
633 COSTS_N_INSNS (2), /* cost of FABS instruction. */
634 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
635 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
636 /* For some reason, Athlon deals better with REP prefix (relative to loops)
637 compared to K8. Alignment becomes important after 8 bytes for memcpy and
638 128 bytes for memset. */
639 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
640 DUMMY_STRINGOP_ALGS},
641 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
642 DUMMY_STRINGOP_ALGS},
643 1, /* scalar_stmt_cost. */
644 1, /* scalar load_cost. */
645 1, /* scalar_store_cost. */
646 1, /* vec_stmt_cost. */
647 1, /* vec_to_scalar_cost. */
648 1, /* scalar_to_vec_cost. */
649 1, /* vec_align_load_cost. */
650 2, /* vec_unalign_load_cost. */
651 1, /* vec_store_cost. */
652 3, /* cond_taken_branch_cost. */
653 1, /* cond_not_taken_branch_cost. */
657 struct processor_costs k8_cost = {
658 COSTS_N_INSNS (1), /* cost of an add instruction */
659 COSTS_N_INSNS (2), /* cost of a lea instruction */
660 COSTS_N_INSNS (1), /* variable shift costs */
661 COSTS_N_INSNS (1), /* constant shift costs */
662 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
663 COSTS_N_INSNS (4), /* HI */
664 COSTS_N_INSNS (3), /* SI */
665 COSTS_N_INSNS (4), /* DI */
666 COSTS_N_INSNS (5)}, /* other */
667 0, /* cost of multiply per each bit set */
668 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
669 COSTS_N_INSNS (26), /* HI */
670 COSTS_N_INSNS (42), /* SI */
671 COSTS_N_INSNS (74), /* DI */
672 COSTS_N_INSNS (74)}, /* other */
673 COSTS_N_INSNS (1), /* cost of movsx */
674 COSTS_N_INSNS (1), /* cost of movzx */
675 8, /* "large" insn */
677 4, /* cost for loading QImode using movzbl */
678 {3, 4, 3}, /* cost of loading integer registers
679 in QImode, HImode and SImode.
680 Relative to reg-reg move (2). */
681 {3, 4, 3}, /* cost of storing integer registers */
682 4, /* cost of reg,reg fld/fst */
683 {4, 4, 12}, /* cost of loading fp registers
684 in SFmode, DFmode and XFmode */
685 {6, 6, 8}, /* cost of storing fp registers
686 in SFmode, DFmode and XFmode */
687 2, /* cost of moving MMX register */
688 {3, 3}, /* cost of loading MMX registers
689 in SImode and DImode */
690 {4, 4}, /* cost of storing MMX registers
691 in SImode and DImode */
692 2, /* cost of moving SSE register */
693 {4, 3, 6}, /* cost of loading SSE registers
694 in SImode, DImode and TImode */
695 {4, 4, 5}, /* cost of storing SSE registers
696 in SImode, DImode and TImode */
697 5, /* MMX or SSE register to integer */
698 64, /* size of l1 cache. */
699 512, /* size of l2 cache. */
700 64, /* size of prefetch block */
701 /* New AMD processors never drop prefetches; if they cannot be performed
702 immediately, they are queued. We set number of simultaneous prefetches
703 to a large constant to reflect this (it probably is not a good idea not
704 to limit number of prefetches at all, as their execution also takes some
706 100, /* number of parallel prefetches */
708 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
709 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
710 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
711 COSTS_N_INSNS (2), /* cost of FABS instruction. */
712 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
713 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
714 /* K8 has optimized REP instruction for medium sized blocks, but for very small
715 blocks it is better to use loop. For large blocks, libcall can do
716 nontemporary accesses and beat inline considerably. */
717 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
718 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
719 {{libcall, {{8, loop}, {24, unrolled_loop},
720 {2048, rep_prefix_4_byte}, {-1, libcall}}},
721 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
722 4, /* scalar_stmt_cost. */
723 2, /* scalar load_cost. */
724 2, /* scalar_store_cost. */
725 5, /* vec_stmt_cost. */
726 0, /* vec_to_scalar_cost. */
727 2, /* scalar_to_vec_cost. */
728 2, /* vec_align_load_cost. */
729 3, /* vec_unalign_load_cost. */
730 3, /* vec_store_cost. */
731 3, /* cond_taken_branch_cost. */
732 2, /* cond_not_taken_branch_cost. */
735 struct processor_costs amdfam10_cost = {
736 COSTS_N_INSNS (1), /* cost of an add instruction */
737 COSTS_N_INSNS (2), /* cost of a lea instruction */
738 COSTS_N_INSNS (1), /* variable shift costs */
739 COSTS_N_INSNS (1), /* constant shift costs */
740 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
741 COSTS_N_INSNS (4), /* HI */
742 COSTS_N_INSNS (3), /* SI */
743 COSTS_N_INSNS (4), /* DI */
744 COSTS_N_INSNS (5)}, /* other */
745 0, /* cost of multiply per each bit set */
746 {COSTS_N_INSNS (19), /* cost of a divide/mod for QI */
747 COSTS_N_INSNS (35), /* HI */
748 COSTS_N_INSNS (51), /* SI */
749 COSTS_N_INSNS (83), /* DI */
750 COSTS_N_INSNS (83)}, /* other */
751 COSTS_N_INSNS (1), /* cost of movsx */
752 COSTS_N_INSNS (1), /* cost of movzx */
753 8, /* "large" insn */
755 4, /* cost for loading QImode using movzbl */
756 {3, 4, 3}, /* cost of loading integer registers
757 in QImode, HImode and SImode.
758 Relative to reg-reg move (2). */
759 {3, 4, 3}, /* cost of storing integer registers */
760 4, /* cost of reg,reg fld/fst */
761 {4, 4, 12}, /* cost of loading fp registers
762 in SFmode, DFmode and XFmode */
763 {6, 6, 8}, /* cost of storing fp registers
764 in SFmode, DFmode and XFmode */
765 2, /* cost of moving MMX register */
766 {3, 3}, /* cost of loading MMX registers
767 in SImode and DImode */
768 {4, 4}, /* cost of storing MMX registers
769 in SImode and DImode */
770 2, /* cost of moving SSE register */
771 {4, 4, 3}, /* cost of loading SSE registers
772 in SImode, DImode and TImode */
773 {4, 4, 5}, /* cost of storing SSE registers
774 in SImode, DImode and TImode */
775 3, /* MMX or SSE register to integer */
777 MOVD reg64, xmmreg Double FSTORE 4
778 MOVD reg32, xmmreg Double FSTORE 4
780 MOVD reg64, xmmreg Double FADD 3
782 MOVD reg32, xmmreg Double FADD 3
784 64, /* size of l1 cache. */
785 512, /* size of l2 cache. */
786 64, /* size of prefetch block */
787 /* New AMD processors never drop prefetches; if they cannot be performed
788 immediately, they are queued. We set number of simultaneous prefetches
789 to a large constant to reflect this (it probably is not a good idea not
790 to limit number of prefetches at all, as their execution also takes some
792 100, /* number of parallel prefetches */
794 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
795 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
796 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
797 COSTS_N_INSNS (2), /* cost of FABS instruction. */
798 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
799 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
801 /* AMDFAM10 has optimized REP instruction for medium sized blocks, but for
802 very small blocks it is better to use loop. For large blocks, libcall can
803 do nontemporary accesses and beat inline considerably. */
804 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
805 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
806 {{libcall, {{8, loop}, {24, unrolled_loop},
807 {2048, rep_prefix_4_byte}, {-1, libcall}}},
808 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
809 4, /* scalar_stmt_cost. */
810 2, /* scalar load_cost. */
811 2, /* scalar_store_cost. */
812 6, /* vec_stmt_cost. */
813 0, /* vec_to_scalar_cost. */
814 2, /* scalar_to_vec_cost. */
815 2, /* vec_align_load_cost. */
816 2, /* vec_unalign_load_cost. */
817 2, /* vec_store_cost. */
818 2, /* cond_taken_branch_cost. */
819 1, /* cond_not_taken_branch_cost. */
823 struct processor_costs pentium4_cost = {
824 COSTS_N_INSNS (1), /* cost of an add instruction */
825 COSTS_N_INSNS (3), /* cost of a lea instruction */
826 COSTS_N_INSNS (4), /* variable shift costs */
827 COSTS_N_INSNS (4), /* constant shift costs */
828 {COSTS_N_INSNS (15), /* cost of starting multiply for QI */
829 COSTS_N_INSNS (15), /* HI */
830 COSTS_N_INSNS (15), /* SI */
831 COSTS_N_INSNS (15), /* DI */
832 COSTS_N_INSNS (15)}, /* other */
833 0, /* cost of multiply per each bit set */
834 {COSTS_N_INSNS (56), /* cost of a divide/mod for QI */
835 COSTS_N_INSNS (56), /* HI */
836 COSTS_N_INSNS (56), /* SI */
837 COSTS_N_INSNS (56), /* DI */
838 COSTS_N_INSNS (56)}, /* other */
839 COSTS_N_INSNS (1), /* cost of movsx */
840 COSTS_N_INSNS (1), /* cost of movzx */
841 16, /* "large" insn */
843 2, /* cost for loading QImode using movzbl */
844 {4, 5, 4}, /* cost of loading integer registers
845 in QImode, HImode and SImode.
846 Relative to reg-reg move (2). */
847 {2, 3, 2}, /* cost of storing integer registers */
848 2, /* cost of reg,reg fld/fst */
849 {2, 2, 6}, /* cost of loading fp registers
850 in SFmode, DFmode and XFmode */
851 {4, 4, 6}, /* cost of storing fp registers
852 in SFmode, DFmode and XFmode */
853 2, /* cost of moving MMX register */
854 {2, 2}, /* cost of loading MMX registers
855 in SImode and DImode */
856 {2, 2}, /* cost of storing MMX registers
857 in SImode and DImode */
858 12, /* cost of moving SSE register */
859 {12, 12, 12}, /* cost of loading SSE registers
860 in SImode, DImode and TImode */
861 {2, 2, 8}, /* cost of storing SSE registers
862 in SImode, DImode and TImode */
863 10, /* MMX or SSE register to integer */
864 8, /* size of l1 cache. */
865 256, /* size of l2 cache. */
866 64, /* size of prefetch block */
867 6, /* number of parallel prefetches */
869 COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */
870 COSTS_N_INSNS (7), /* cost of FMUL instruction. */
871 COSTS_N_INSNS (43), /* cost of FDIV instruction. */
872 COSTS_N_INSNS (2), /* cost of FABS instruction. */
873 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
874 COSTS_N_INSNS (43), /* cost of FSQRT instruction. */
875 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
876 DUMMY_STRINGOP_ALGS},
877 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
879 DUMMY_STRINGOP_ALGS},
880 1, /* scalar_stmt_cost. */
881 1, /* scalar load_cost. */
882 1, /* scalar_store_cost. */
883 1, /* vec_stmt_cost. */
884 1, /* vec_to_scalar_cost. */
885 1, /* scalar_to_vec_cost. */
886 1, /* vec_align_load_cost. */
887 2, /* vec_unalign_load_cost. */
888 1, /* vec_store_cost. */
889 3, /* cond_taken_branch_cost. */
890 1, /* cond_not_taken_branch_cost. */
894 struct processor_costs nocona_cost = {
895 COSTS_N_INSNS (1), /* cost of an add instruction */
896 COSTS_N_INSNS (1), /* cost of a lea instruction */
897 COSTS_N_INSNS (1), /* variable shift costs */
898 COSTS_N_INSNS (1), /* constant shift costs */
899 {COSTS_N_INSNS (10), /* cost of starting multiply for QI */
900 COSTS_N_INSNS (10), /* HI */
901 COSTS_N_INSNS (10), /* SI */
902 COSTS_N_INSNS (10), /* DI */
903 COSTS_N_INSNS (10)}, /* other */
904 0, /* cost of multiply per each bit set */
905 {COSTS_N_INSNS (66), /* cost of a divide/mod for QI */
906 COSTS_N_INSNS (66), /* HI */
907 COSTS_N_INSNS (66), /* SI */
908 COSTS_N_INSNS (66), /* DI */
909 COSTS_N_INSNS (66)}, /* other */
910 COSTS_N_INSNS (1), /* cost of movsx */
911 COSTS_N_INSNS (1), /* cost of movzx */
912 16, /* "large" insn */
914 4, /* cost for loading QImode using movzbl */
915 {4, 4, 4}, /* cost of loading integer registers
916 in QImode, HImode and SImode.
917 Relative to reg-reg move (2). */
918 {4, 4, 4}, /* cost of storing integer registers */
919 3, /* cost of reg,reg fld/fst */
920 {12, 12, 12}, /* cost of loading fp registers
921 in SFmode, DFmode and XFmode */
922 {4, 4, 4}, /* cost of storing fp registers
923 in SFmode, DFmode and XFmode */
924 6, /* cost of moving MMX register */
925 {12, 12}, /* cost of loading MMX registers
926 in SImode and DImode */
927 {12, 12}, /* cost of storing MMX registers
928 in SImode and DImode */
929 6, /* cost of moving SSE register */
930 {12, 12, 12}, /* cost of loading SSE registers
931 in SImode, DImode and TImode */
932 {12, 12, 12}, /* cost of storing SSE registers
933 in SImode, DImode and TImode */
934 8, /* MMX or SSE register to integer */
935 8, /* size of l1 cache. */
936 1024, /* size of l2 cache. */
937 128, /* size of prefetch block */
938 8, /* number of parallel prefetches */
940 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
941 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
942 COSTS_N_INSNS (40), /* cost of FDIV instruction. */
943 COSTS_N_INSNS (3), /* cost of FABS instruction. */
944 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
945 COSTS_N_INSNS (44), /* cost of FSQRT instruction. */
946 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
947 {libcall, {{32, loop}, {20000, rep_prefix_8_byte},
948 {100000, unrolled_loop}, {-1, libcall}}}},
949 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
951 {libcall, {{24, loop}, {64, unrolled_loop},
952 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
953 1, /* scalar_stmt_cost. */
954 1, /* scalar load_cost. */
955 1, /* scalar_store_cost. */
956 1, /* vec_stmt_cost. */
957 1, /* vec_to_scalar_cost. */
958 1, /* scalar_to_vec_cost. */
959 1, /* vec_align_load_cost. */
960 2, /* vec_unalign_load_cost. */
961 1, /* vec_store_cost. */
962 3, /* cond_taken_branch_cost. */
963 1, /* cond_not_taken_branch_cost. */
967 struct processor_costs core2_cost = {
968 COSTS_N_INSNS (1), /* cost of an add instruction */
969 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
970 COSTS_N_INSNS (1), /* variable shift costs */
971 COSTS_N_INSNS (1), /* constant shift costs */
972 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
973 COSTS_N_INSNS (3), /* HI */
974 COSTS_N_INSNS (3), /* SI */
975 COSTS_N_INSNS (3), /* DI */
976 COSTS_N_INSNS (3)}, /* other */
977 0, /* cost of multiply per each bit set */
978 {COSTS_N_INSNS (22), /* cost of a divide/mod for QI */
979 COSTS_N_INSNS (22), /* HI */
980 COSTS_N_INSNS (22), /* SI */
981 COSTS_N_INSNS (22), /* DI */
982 COSTS_N_INSNS (22)}, /* other */
983 COSTS_N_INSNS (1), /* cost of movsx */
984 COSTS_N_INSNS (1), /* cost of movzx */
985 8, /* "large" insn */
987 2, /* cost for loading QImode using movzbl */
988 {6, 6, 6}, /* cost of loading integer registers
989 in QImode, HImode and SImode.
990 Relative to reg-reg move (2). */
991 {4, 4, 4}, /* cost of storing integer registers */
992 2, /* cost of reg,reg fld/fst */
993 {6, 6, 6}, /* cost of loading fp registers
994 in SFmode, DFmode and XFmode */
995 {4, 4, 4}, /* cost of storing fp registers
996 in SFmode, DFmode and XFmode */
997 2, /* cost of moving MMX register */
998 {6, 6}, /* cost of loading MMX registers
999 in SImode and DImode */
1000 {4, 4}, /* cost of storing MMX registers
1001 in SImode and DImode */
1002 2, /* cost of moving SSE register */
1003 {6, 6, 6}, /* cost of loading SSE registers
1004 in SImode, DImode and TImode */
1005 {4, 4, 4}, /* cost of storing SSE registers
1006 in SImode, DImode and TImode */
1007 2, /* MMX or SSE register to integer */
1008 32, /* size of l1 cache. */
1009 2048, /* size of l2 cache. */
1010 128, /* size of prefetch block */
1011 8, /* number of parallel prefetches */
1012 3, /* Branch cost */
1013 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
1014 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
1015 COSTS_N_INSNS (32), /* cost of FDIV instruction. */
1016 COSTS_N_INSNS (1), /* cost of FABS instruction. */
1017 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
1018 COSTS_N_INSNS (58), /* cost of FSQRT instruction. */
1019 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1020 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1021 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1022 {{libcall, {{8, loop}, {15, unrolled_loop},
1023 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1024 {libcall, {{24, loop}, {32, unrolled_loop},
1025 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1026 1, /* scalar_stmt_cost. */
1027 1, /* scalar load_cost. */
1028 1, /* scalar_store_cost. */
1029 1, /* vec_stmt_cost. */
1030 1, /* vec_to_scalar_cost. */
1031 1, /* scalar_to_vec_cost. */
1032 1, /* vec_align_load_cost. */
1033 2, /* vec_unalign_load_cost. */
1034 1, /* vec_store_cost. */
1035 3, /* cond_taken_branch_cost. */
1036 1, /* cond_not_taken_branch_cost. */
1039 /* Generic64 should produce code tuned for Nocona and K8. */
1041 struct processor_costs generic64_cost = {
1042 COSTS_N_INSNS (1), /* cost of an add instruction */
1043 /* On all chips taken into consideration lea is 2 cycles and more. With
1044 this cost however our current implementation of synth_mult results in
1045 use of unnecessary temporary registers causing regression on several
1046 SPECfp benchmarks. */
1047 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1048 COSTS_N_INSNS (1), /* variable shift costs */
1049 COSTS_N_INSNS (1), /* constant shift costs */
1050 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1051 COSTS_N_INSNS (4), /* HI */
1052 COSTS_N_INSNS (3), /* SI */
1053 COSTS_N_INSNS (4), /* DI */
1054 COSTS_N_INSNS (2)}, /* other */
1055 0, /* cost of multiply per each bit set */
1056 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1057 COSTS_N_INSNS (26), /* HI */
1058 COSTS_N_INSNS (42), /* SI */
1059 COSTS_N_INSNS (74), /* DI */
1060 COSTS_N_INSNS (74)}, /* other */
1061 COSTS_N_INSNS (1), /* cost of movsx */
1062 COSTS_N_INSNS (1), /* cost of movzx */
1063 8, /* "large" insn */
1064 17, /* MOVE_RATIO */
1065 4, /* cost for loading QImode using movzbl */
1066 {4, 4, 4}, /* cost of loading integer registers
1067 in QImode, HImode and SImode.
1068 Relative to reg-reg move (2). */
1069 {4, 4, 4}, /* cost of storing integer registers */
1070 4, /* cost of reg,reg fld/fst */
1071 {12, 12, 12}, /* cost of loading fp registers
1072 in SFmode, DFmode and XFmode */
1073 {6, 6, 8}, /* cost of storing fp registers
1074 in SFmode, DFmode and XFmode */
1075 2, /* cost of moving MMX register */
1076 {8, 8}, /* cost of loading MMX registers
1077 in SImode and DImode */
1078 {8, 8}, /* cost of storing MMX registers
1079 in SImode and DImode */
1080 2, /* cost of moving SSE register */
1081 {8, 8, 8}, /* cost of loading SSE registers
1082 in SImode, DImode and TImode */
1083 {8, 8, 8}, /* cost of storing SSE registers
1084 in SImode, DImode and TImode */
1085 5, /* MMX or SSE register to integer */
1086 32, /* size of l1 cache. */
1087 512, /* size of l2 cache. */
1088 64, /* size of prefetch block */
1089 6, /* number of parallel prefetches */
1090 /* Benchmarks shows large regressions on K8 sixtrack benchmark when this value
1091 is increased to perhaps more appropriate value of 5. */
1092 3, /* Branch cost */
1093 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1094 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1095 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1096 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1097 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1098 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1099 {DUMMY_STRINGOP_ALGS,
1100 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1101 {DUMMY_STRINGOP_ALGS,
1102 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1103 1, /* scalar_stmt_cost. */
1104 1, /* scalar load_cost. */
1105 1, /* scalar_store_cost. */
1106 1, /* vec_stmt_cost. */
1107 1, /* vec_to_scalar_cost. */
1108 1, /* scalar_to_vec_cost. */
1109 1, /* vec_align_load_cost. */
1110 2, /* vec_unalign_load_cost. */
1111 1, /* vec_store_cost. */
1112 3, /* cond_taken_branch_cost. */
1113 1, /* cond_not_taken_branch_cost. */
1116 /* Generic32 should produce code tuned for Athlon, PPro, Pentium4, Nocona and K8. */
1118 struct processor_costs generic32_cost = {
1119 COSTS_N_INSNS (1), /* cost of an add instruction */
1120 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1121 COSTS_N_INSNS (1), /* variable shift costs */
1122 COSTS_N_INSNS (1), /* constant shift costs */
1123 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1124 COSTS_N_INSNS (4), /* HI */
1125 COSTS_N_INSNS (3), /* SI */
1126 COSTS_N_INSNS (4), /* DI */
1127 COSTS_N_INSNS (2)}, /* other */
1128 0, /* cost of multiply per each bit set */
1129 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1130 COSTS_N_INSNS (26), /* HI */
1131 COSTS_N_INSNS (42), /* SI */
1132 COSTS_N_INSNS (74), /* DI */
1133 COSTS_N_INSNS (74)}, /* other */
1134 COSTS_N_INSNS (1), /* cost of movsx */
1135 COSTS_N_INSNS (1), /* cost of movzx */
1136 8, /* "large" insn */
1137 17, /* MOVE_RATIO */
1138 4, /* cost for loading QImode using movzbl */
1139 {4, 4, 4}, /* cost of loading integer registers
1140 in QImode, HImode and SImode.
1141 Relative to reg-reg move (2). */
1142 {4, 4, 4}, /* cost of storing integer registers */
1143 4, /* cost of reg,reg fld/fst */
1144 {12, 12, 12}, /* cost of loading fp registers
1145 in SFmode, DFmode and XFmode */
1146 {6, 6, 8}, /* cost of storing fp registers
1147 in SFmode, DFmode and XFmode */
1148 2, /* cost of moving MMX register */
1149 {8, 8}, /* cost of loading MMX registers
1150 in SImode and DImode */
1151 {8, 8}, /* cost of storing MMX registers
1152 in SImode and DImode */
1153 2, /* cost of moving SSE register */
1154 {8, 8, 8}, /* cost of loading SSE registers
1155 in SImode, DImode and TImode */
1156 {8, 8, 8}, /* cost of storing SSE registers
1157 in SImode, DImode and TImode */
1158 5, /* MMX or SSE register to integer */
1159 32, /* size of l1 cache. */
1160 256, /* size of l2 cache. */
1161 64, /* size of prefetch block */
1162 6, /* number of parallel prefetches */
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 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1171 DUMMY_STRINGOP_ALGS},
1172 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1173 DUMMY_STRINGOP_ALGS},
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 const struct processor_costs *ix86_cost = &pentium_cost;
1189 /* Processor feature/optimization bitmasks. */
1190 #define m_386 (1<<PROCESSOR_I386)
1191 #define m_486 (1<<PROCESSOR_I486)
1192 #define m_PENT (1<<PROCESSOR_PENTIUM)
1193 #define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
1194 #define m_PENT4 (1<<PROCESSOR_PENTIUM4)
1195 #define m_NOCONA (1<<PROCESSOR_NOCONA)
1196 #define m_CORE2 (1<<PROCESSOR_CORE2)
1198 #define m_GEODE (1<<PROCESSOR_GEODE)
1199 #define m_K6 (1<<PROCESSOR_K6)
1200 #define m_K6_GEODE (m_K6 | m_GEODE)
1201 #define m_K8 (1<<PROCESSOR_K8)
1202 #define m_ATHLON (1<<PROCESSOR_ATHLON)
1203 #define m_ATHLON_K8 (m_K8 | m_ATHLON)
1204 #define m_AMDFAM10 (1<<PROCESSOR_AMDFAM10)
1205 #define m_AMD_MULTIPLE (m_K8 | m_ATHLON | m_AMDFAM10)
1207 #define m_GENERIC32 (1<<PROCESSOR_GENERIC32)
1208 #define m_GENERIC64 (1<<PROCESSOR_GENERIC64)
1210 /* Generic instruction choice should be common subset of supported CPUs
1211 (PPro/PENT4/NOCONA/CORE2/Athlon/K8). */
1212 #define m_GENERIC (m_GENERIC32 | m_GENERIC64)
1214 /* Feature tests against the various tunings. */
1215 unsigned char ix86_tune_features[X86_TUNE_LAST];
1217 /* Feature tests against the various tunings used to create ix86_tune_features
1218 based on the processor mask. */
1219 static unsigned int initial_ix86_tune_features[X86_TUNE_LAST] = {
1220 /* X86_TUNE_USE_LEAVE: Leave does not affect Nocona SPEC2000 results
1221 negatively, so enabling for Generic64 seems like good code size
1222 tradeoff. We can't enable it for 32bit generic because it does not
1223 work well with PPro base chips. */
1224 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_CORE2 | m_GENERIC64,
1226 /* X86_TUNE_PUSH_MEMORY */
1227 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4
1228 | m_NOCONA | m_CORE2 | m_GENERIC,
1230 /* X86_TUNE_ZERO_EXTEND_WITH_AND */
1233 /* X86_TUNE_UNROLL_STRLEN */
1234 m_486 | m_PENT | m_PPRO | m_AMD_MULTIPLE | m_K6 | m_CORE2 | m_GENERIC,
1236 /* X86_TUNE_DEEP_BRANCH_PREDICTION */
1237 m_PPRO | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4 | m_GENERIC,
1239 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
1240 on simulation result. But after P4 was made, no performance benefit
1241 was observed with branch hints. It also increases the code size.
1242 As a result, icc never generates branch hints. */
1245 /* X86_TUNE_DOUBLE_WITH_ADD */
1248 /* X86_TUNE_USE_SAHF */
1249 m_PPRO | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_PENT4
1250 | m_NOCONA | m_CORE2 | m_GENERIC,
1252 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
1253 partial dependencies. */
1254 m_AMD_MULTIPLE | m_PPRO | m_PENT4 | m_NOCONA
1255 | m_CORE2 | m_GENERIC | m_GEODE /* m_386 | m_K6 */,
1257 /* X86_TUNE_PARTIAL_REG_STALL: We probably ought to watch for partial
1258 register stalls on Generic32 compilation setting as well. However
1259 in current implementation the partial register stalls are not eliminated
1260 very well - they can be introduced via subregs synthesized by combine
1261 and can happen in caller/callee saving sequences. Because this option
1262 pays back little on PPro based chips and is in conflict with partial reg
1263 dependencies used by Athlon/P4 based chips, it is better to leave it off
1264 for generic32 for now. */
1267 /* X86_TUNE_PARTIAL_FLAG_REG_STALL */
1268 m_CORE2 | m_GENERIC,
1270 /* X86_TUNE_USE_HIMODE_FIOP */
1271 m_386 | m_486 | m_K6_GEODE,
1273 /* X86_TUNE_USE_SIMODE_FIOP */
1274 ~(m_PPRO | m_AMD_MULTIPLE | m_PENT | m_CORE2 | m_GENERIC),
1276 /* X86_TUNE_USE_MOV0 */
1279 /* X86_TUNE_USE_CLTD */
1280 ~(m_PENT | m_K6 | m_CORE2 | m_GENERIC),
1282 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
1285 /* X86_TUNE_SPLIT_LONG_MOVES */
1288 /* X86_TUNE_READ_MODIFY_WRITE */
1291 /* X86_TUNE_READ_MODIFY */
1294 /* X86_TUNE_PROMOTE_QIMODE */
1295 m_K6_GEODE | m_PENT | m_386 | m_486 | m_AMD_MULTIPLE | m_CORE2
1296 | m_GENERIC /* | m_PENT4 ? */,
1298 /* X86_TUNE_FAST_PREFIX */
1299 ~(m_PENT | m_486 | m_386),
1301 /* X86_TUNE_SINGLE_STRINGOP */
1302 m_386 | m_PENT4 | m_NOCONA,
1304 /* X86_TUNE_QIMODE_MATH */
1307 /* X86_TUNE_HIMODE_MATH: On PPro this flag is meant to avoid partial
1308 register stalls. Just like X86_TUNE_PARTIAL_REG_STALL this option
1309 might be considered for Generic32 if our scheme for avoiding partial
1310 stalls was more effective. */
1313 /* X86_TUNE_PROMOTE_QI_REGS */
1316 /* X86_TUNE_PROMOTE_HI_REGS */
1319 /* X86_TUNE_ADD_ESP_4: Enable if add/sub is preferred over 1/2 push/pop. */
1320 m_AMD_MULTIPLE | m_K6_GEODE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1322 /* X86_TUNE_ADD_ESP_8 */
1323 m_AMD_MULTIPLE | m_PPRO | m_K6_GEODE | m_386
1324 | m_486 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1326 /* X86_TUNE_SUB_ESP_4 */
1327 m_AMD_MULTIPLE | m_PPRO | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1329 /* X86_TUNE_SUB_ESP_8 */
1330 m_AMD_MULTIPLE | m_PPRO | m_386 | m_486
1331 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1333 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
1334 for DFmode copies */
1335 ~(m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1336 | m_GENERIC | m_GEODE),
1338 /* X86_TUNE_PARTIAL_REG_DEPENDENCY */
1339 m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1341 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: In the Generic model we have a
1342 conflict here in between PPro/Pentium4 based chips that thread 128bit
1343 SSE registers as single units versus K8 based chips that divide SSE
1344 registers to two 64bit halves. This knob promotes all store destinations
1345 to be 128bit to allow register renaming on 128bit SSE units, but usually
1346 results in one extra microop on 64bit SSE units. Experimental results
1347 shows that disabling this option on P4 brings over 20% SPECfp regression,
1348 while enabling it on K8 brings roughly 2.4% regression that can be partly
1349 masked by careful scheduling of moves. */
1350 m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC | m_AMDFAM10,
1352 /* X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL */
1355 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
1356 are resolved on SSE register parts instead of whole registers, so we may
1357 maintain just lower part of scalar values in proper format leaving the
1358 upper part undefined. */
1361 /* X86_TUNE_SSE_TYPELESS_STORES */
1364 /* X86_TUNE_SSE_LOAD0_BY_PXOR */
1365 m_PPRO | m_PENT4 | m_NOCONA,
1367 /* X86_TUNE_MEMORY_MISMATCH_STALL */
1368 m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1370 /* X86_TUNE_PROLOGUE_USING_MOVE */
1371 m_ATHLON_K8 | m_PPRO | m_CORE2 | m_GENERIC,
1373 /* X86_TUNE_EPILOGUE_USING_MOVE */
1374 m_ATHLON_K8 | m_PPRO | m_CORE2 | m_GENERIC,
1376 /* X86_TUNE_SHIFT1 */
1379 /* X86_TUNE_USE_FFREEP */
1382 /* X86_TUNE_INTER_UNIT_MOVES */
1383 ~(m_AMD_MULTIPLE | m_GENERIC),
1385 /* X86_TUNE_INTER_UNIT_CONVERSIONS */
1388 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
1389 than 4 branch instructions in the 16 byte window. */
1390 m_PPRO | m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1392 /* X86_TUNE_SCHEDULE */
1393 m_PPRO | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT | m_CORE2 | m_GENERIC,
1395 /* X86_TUNE_USE_BT */
1396 m_AMD_MULTIPLE | m_CORE2 | m_GENERIC,
1398 /* X86_TUNE_USE_INCDEC */
1399 ~(m_PENT4 | m_NOCONA | m_GENERIC),
1401 /* X86_TUNE_PAD_RETURNS */
1402 m_AMD_MULTIPLE | m_CORE2 | m_GENERIC,
1404 /* X86_TUNE_EXT_80387_CONSTANTS */
1405 m_K6_GEODE | m_ATHLON_K8 | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC,
1407 /* X86_TUNE_SHORTEN_X87_SSE */
1410 /* X86_TUNE_AVOID_VECTOR_DECODE */
1413 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
1414 and SImode multiply, but 386 and 486 do HImode multiply faster. */
1417 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
1418 vector path on AMD machines. */
1419 m_K8 | m_GENERIC64 | m_AMDFAM10,
1421 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
1423 m_K8 | m_GENERIC64 | m_AMDFAM10,
1425 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
1429 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
1430 but one byte longer. */
1433 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
1434 operand that cannot be represented using a modRM byte. The XOR
1435 replacement is long decoded, so this split helps here as well. */
1438 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
1440 m_AMDFAM10 | m_GENERIC,
1442 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
1443 from integer to FP. */
1446 /* X86_TUNE_FUSE_CMP_AND_BRANCH: Fuse a compare or test instruction
1447 with a subsequent conditional jump instruction into a single
1448 compare-and-branch uop. */
1452 /* Feature tests against the various architecture variations. */
1453 unsigned char ix86_arch_features[X86_ARCH_LAST];
1455 /* Feature tests against the various architecture variations, used to create
1456 ix86_arch_features based on the processor mask. */
1457 static unsigned int initial_ix86_arch_features[X86_ARCH_LAST] = {
1458 /* X86_ARCH_CMOVE: Conditional move was added for pentiumpro. */
1459 ~(m_386 | m_486 | m_PENT | m_K6),
1461 /* X86_ARCH_CMPXCHG: Compare and exchange was added for 80486. */
1464 /* X86_ARCH_CMPXCHG8B: Compare and exchange 8 bytes was added for pentium. */
1467 /* X86_ARCH_XADD: Exchange and add was added for 80486. */
1470 /* X86_ARCH_BSWAP: Byteswap was added for 80486. */
1474 static const unsigned int x86_accumulate_outgoing_args
1475 = m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC;
1477 static const unsigned int x86_arch_always_fancy_math_387
1478 = m_PENT | m_PPRO | m_AMD_MULTIPLE | m_PENT4
1479 | m_NOCONA | m_CORE2 | m_GENERIC;
1481 static enum stringop_alg stringop_alg = no_stringop;
1483 /* In case the average insn count for single function invocation is
1484 lower than this constant, emit fast (but longer) prologue and
1486 #define FAST_PROLOGUE_INSN_COUNT 20
1488 /* Names for 8 (low), 8 (high), and 16-bit registers, respectively. */
1489 static const char *const qi_reg_name[] = QI_REGISTER_NAMES;
1490 static const char *const qi_high_reg_name[] = QI_HIGH_REGISTER_NAMES;
1491 static const char *const hi_reg_name[] = HI_REGISTER_NAMES;
1493 /* Array of the smallest class containing reg number REGNO, indexed by
1494 REGNO. Used by REGNO_REG_CLASS in i386.h. */
1496 enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER] =
1498 /* ax, dx, cx, bx */
1499 AREG, DREG, CREG, BREG,
1500 /* si, di, bp, sp */
1501 SIREG, DIREG, NON_Q_REGS, NON_Q_REGS,
1503 FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS,
1504 FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
1507 /* flags, fpsr, fpcr, frame */
1508 NO_REGS, NO_REGS, NO_REGS, NON_Q_REGS,
1510 SSE_FIRST_REG, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1513 MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS,
1516 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1517 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1518 /* SSE REX registers */
1519 SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1523 /* The "default" register map used in 32bit mode. */
1525 int const dbx_register_map[FIRST_PSEUDO_REGISTER] =
1527 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */
1528 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */
1529 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1530 21, 22, 23, 24, 25, 26, 27, 28, /* SSE */
1531 29, 30, 31, 32, 33, 34, 35, 36, /* MMX */
1532 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1533 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1536 static int const x86_64_int_parameter_registers[6] =
1538 5 /*RDI*/, 4 /*RSI*/, 1 /*RDX*/, 2 /*RCX*/,
1539 FIRST_REX_INT_REG /*R8 */, FIRST_REX_INT_REG + 1 /*R9 */
1542 static int const x86_64_ms_abi_int_parameter_registers[4] =
1544 2 /*RCX*/, 1 /*RDX*/,
1545 FIRST_REX_INT_REG /*R8 */, FIRST_REX_INT_REG + 1 /*R9 */
1548 static int const x86_64_int_return_registers[4] =
1550 0 /*RAX*/, 1 /*RDX*/, 5 /*RDI*/, 4 /*RSI*/
1553 /* The "default" register map used in 64bit mode. */
1554 int const dbx64_register_map[FIRST_PSEUDO_REGISTER] =
1556 0, 1, 2, 3, 4, 5, 6, 7, /* general regs */
1557 33, 34, 35, 36, 37, 38, 39, 40, /* fp regs */
1558 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1559 17, 18, 19, 20, 21, 22, 23, 24, /* SSE */
1560 41, 42, 43, 44, 45, 46, 47, 48, /* MMX */
1561 8,9,10,11,12,13,14,15, /* extended integer registers */
1562 25, 26, 27, 28, 29, 30, 31, 32, /* extended SSE registers */
1565 /* Define the register numbers to be used in Dwarf debugging information.
1566 The SVR4 reference port C compiler uses the following register numbers
1567 in its Dwarf output code:
1568 0 for %eax (gcc regno = 0)
1569 1 for %ecx (gcc regno = 2)
1570 2 for %edx (gcc regno = 1)
1571 3 for %ebx (gcc regno = 3)
1572 4 for %esp (gcc regno = 7)
1573 5 for %ebp (gcc regno = 6)
1574 6 for %esi (gcc regno = 4)
1575 7 for %edi (gcc regno = 5)
1576 The following three DWARF register numbers are never generated by
1577 the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4
1578 believes these numbers have these meanings.
1579 8 for %eip (no gcc equivalent)
1580 9 for %eflags (gcc regno = 17)
1581 10 for %trapno (no gcc equivalent)
1582 It is not at all clear how we should number the FP stack registers
1583 for the x86 architecture. If the version of SDB on x86/svr4 were
1584 a bit less brain dead with respect to floating-point then we would
1585 have a precedent to follow with respect to DWARF register numbers
1586 for x86 FP registers, but the SDB on x86/svr4 is so completely
1587 broken with respect to FP registers that it is hardly worth thinking
1588 of it as something to strive for compatibility with.
1589 The version of x86/svr4 SDB I have at the moment does (partially)
1590 seem to believe that DWARF register number 11 is associated with
1591 the x86 register %st(0), but that's about all. Higher DWARF
1592 register numbers don't seem to be associated with anything in
1593 particular, and even for DWARF regno 11, SDB only seems to under-
1594 stand that it should say that a variable lives in %st(0) (when
1595 asked via an `=' command) if we said it was in DWARF regno 11,
1596 but SDB still prints garbage when asked for the value of the
1597 variable in question (via a `/' command).
1598 (Also note that the labels SDB prints for various FP stack regs
1599 when doing an `x' command are all wrong.)
1600 Note that these problems generally don't affect the native SVR4
1601 C compiler because it doesn't allow the use of -O with -g and
1602 because when it is *not* optimizing, it allocates a memory
1603 location for each floating-point variable, and the memory
1604 location is what gets described in the DWARF AT_location
1605 attribute for the variable in question.
1606 Regardless of the severe mental illness of the x86/svr4 SDB, we
1607 do something sensible here and we use the following DWARF
1608 register numbers. Note that these are all stack-top-relative
1610 11 for %st(0) (gcc regno = 8)
1611 12 for %st(1) (gcc regno = 9)
1612 13 for %st(2) (gcc regno = 10)
1613 14 for %st(3) (gcc regno = 11)
1614 15 for %st(4) (gcc regno = 12)
1615 16 for %st(5) (gcc regno = 13)
1616 17 for %st(6) (gcc regno = 14)
1617 18 for %st(7) (gcc regno = 15)
1619 int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER] =
1621 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */
1622 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */
1623 -1, 9, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1624 21, 22, 23, 24, 25, 26, 27, 28, /* SSE registers */
1625 29, 30, 31, 32, 33, 34, 35, 36, /* MMX registers */
1626 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1627 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1630 /* Test and compare insns in i386.md store the information needed to
1631 generate branch and scc insns here. */
1633 rtx ix86_compare_op0 = NULL_RTX;
1634 rtx ix86_compare_op1 = NULL_RTX;
1635 rtx ix86_compare_emitted = NULL_RTX;
1637 /* Define the structure for the machine field in struct function. */
1639 struct stack_local_entry GTY(())
1641 unsigned short mode;
1644 struct stack_local_entry *next;
1647 /* Structure describing stack frame layout.
1648 Stack grows downward:
1654 saved frame pointer if frame_pointer_needed
1655 <- HARD_FRAME_POINTER
1664 [va_arg registers] (
1665 > to_allocate <- FRAME_POINTER
1677 HOST_WIDE_INT frame;
1679 int outgoing_arguments_size;
1682 HOST_WIDE_INT to_allocate;
1683 /* The offsets relative to ARG_POINTER. */
1684 HOST_WIDE_INT frame_pointer_offset;
1685 HOST_WIDE_INT hard_frame_pointer_offset;
1686 HOST_WIDE_INT stack_pointer_offset;
1688 /* When save_regs_using_mov is set, emit prologue using
1689 move instead of push instructions. */
1690 bool save_regs_using_mov;
1693 /* Code model option. */
1694 enum cmodel ix86_cmodel;
1696 enum asm_dialect ix86_asm_dialect = ASM_ATT;
1698 enum tls_dialect ix86_tls_dialect = TLS_DIALECT_GNU;
1700 /* Which unit we are generating floating point math for. */
1701 enum fpmath_unit ix86_fpmath;
1703 /* Which cpu are we scheduling for. */
1704 enum attr_cpu ix86_schedule;
1706 /* Which cpu are we optimizing for. */
1707 enum processor_type ix86_tune;
1709 /* Which instruction set architecture to use. */
1710 enum processor_type ix86_arch;
1712 /* true if sse prefetch instruction is not NOOP. */
1713 int x86_prefetch_sse;
1715 /* ix86_regparm_string as a number */
1716 static int ix86_regparm;
1718 /* -mstackrealign option */
1719 extern int ix86_force_align_arg_pointer;
1720 static const char ix86_force_align_arg_pointer_string[]
1721 = "force_align_arg_pointer";
1723 static rtx (*ix86_gen_leave) (void);
1724 static rtx (*ix86_gen_pop1) (rtx);
1725 static rtx (*ix86_gen_add3) (rtx, rtx, rtx);
1726 static rtx (*ix86_gen_sub3) (rtx, rtx, rtx);
1727 static rtx (*ix86_gen_sub3_carry) (rtx, rtx, rtx, rtx);
1728 static rtx (*ix86_gen_one_cmpl2) (rtx, rtx);
1729 static rtx (*ix86_gen_monitor) (rtx, rtx, rtx);
1730 static rtx (*ix86_gen_andsp) (rtx, rtx, rtx);
1732 /* Preferred alignment for stack boundary in bits. */
1733 unsigned int ix86_preferred_stack_boundary;
1735 /* Alignment for incoming stack boundary in bits specified at
1737 static unsigned int ix86_user_incoming_stack_boundary;
1739 /* Default alignment for incoming stack boundary in bits. */
1740 static unsigned int ix86_default_incoming_stack_boundary;
1742 /* Alignment for incoming stack boundary in bits. */
1743 unsigned int ix86_incoming_stack_boundary;
1745 /* Values 1-5: see jump.c */
1746 int ix86_branch_cost;
1748 /* Calling abi specific va_list type nodes. */
1749 static GTY(()) tree sysv_va_list_type_node;
1750 static GTY(()) tree ms_va_list_type_node;
1752 /* Variables which are this size or smaller are put in the data/bss
1753 or ldata/lbss sections. */
1755 int ix86_section_threshold = 65536;
1757 /* Prefix built by ASM_GENERATE_INTERNAL_LABEL. */
1758 char internal_label_prefix[16];
1759 int internal_label_prefix_len;
1761 /* Fence to use after loop using movnt. */
1764 /* Register class used for passing given 64bit part of the argument.
1765 These represent classes as documented by the PS ABI, with the exception
1766 of SSESF, SSEDF classes, that are basically SSE class, just gcc will
1767 use SF or DFmode move instead of DImode to avoid reformatting penalties.
1769 Similarly we play games with INTEGERSI_CLASS to use cheaper SImode moves
1770 whenever possible (upper half does contain padding). */
1771 enum x86_64_reg_class
1774 X86_64_INTEGER_CLASS,
1775 X86_64_INTEGERSI_CLASS,
1783 X86_64_COMPLEX_X87_CLASS,
1786 static const char * const x86_64_reg_class_name[] =
1788 "no", "integer", "integerSI", "sse", "sseSF", "sseDF",
1789 "sseup", "x87", "x87up", "cplx87", "no"
1792 #define MAX_CLASSES 4
1794 /* Table of constants used by fldpi, fldln2, etc.... */
1795 static REAL_VALUE_TYPE ext_80387_constants_table [5];
1796 static bool ext_80387_constants_init = 0;
1799 static struct machine_function * ix86_init_machine_status (void);
1800 static rtx ix86_function_value (const_tree, const_tree, bool);
1801 static int ix86_function_regparm (const_tree, const_tree);
1802 static void ix86_compute_frame_layout (struct ix86_frame *);
1803 static bool ix86_expand_vector_init_one_nonzero (bool, enum machine_mode,
1805 static void ix86_add_new_builtins (int);
1807 enum ix86_function_specific_strings
1809 IX86_FUNCTION_SPECIFIC_ARCH,
1810 IX86_FUNCTION_SPECIFIC_TUNE,
1811 IX86_FUNCTION_SPECIFIC_FPMATH,
1812 IX86_FUNCTION_SPECIFIC_MAX
1815 static char *ix86_target_string (int, int, const char *, const char *,
1816 const char *, bool);
1817 static void ix86_debug_options (void) ATTRIBUTE_UNUSED;
1818 static void ix86_function_specific_save (struct cl_target_option *);
1819 static void ix86_function_specific_restore (struct cl_target_option *);
1820 static void ix86_function_specific_print (FILE *, int,
1821 struct cl_target_option *);
1822 static bool ix86_valid_target_attribute_p (tree, tree, tree, int);
1823 static bool ix86_valid_target_attribute_inner_p (tree, char *[]);
1824 static bool ix86_can_inline_p (tree, tree);
1825 static void ix86_set_current_function (tree);
1828 /* The svr4 ABI for the i386 says that records and unions are returned
1830 #ifndef DEFAULT_PCC_STRUCT_RETURN
1831 #define DEFAULT_PCC_STRUCT_RETURN 1
1834 /* Whether -mtune= or -march= were specified */
1835 static int ix86_tune_defaulted;
1836 static int ix86_arch_specified;
1838 /* Bit flags that specify the ISA we are compiling for. */
1839 int ix86_isa_flags = TARGET_64BIT_DEFAULT | TARGET_SUBTARGET_ISA_DEFAULT;
1841 /* A mask of ix86_isa_flags that includes bit X if X
1842 was set or cleared on the command line. */
1843 static int ix86_isa_flags_explicit;
1845 /* Define a set of ISAs which are available when a given ISA is
1846 enabled. MMX and SSE ISAs are handled separately. */
1848 #define OPTION_MASK_ISA_MMX_SET OPTION_MASK_ISA_MMX
1849 #define OPTION_MASK_ISA_3DNOW_SET \
1850 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_MMX_SET)
1852 #define OPTION_MASK_ISA_SSE_SET OPTION_MASK_ISA_SSE
1853 #define OPTION_MASK_ISA_SSE2_SET \
1854 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE_SET)
1855 #define OPTION_MASK_ISA_SSE3_SET \
1856 (OPTION_MASK_ISA_SSE3 | OPTION_MASK_ISA_SSE2_SET)
1857 #define OPTION_MASK_ISA_SSSE3_SET \
1858 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE3_SET)
1859 #define OPTION_MASK_ISA_SSE4_1_SET \
1860 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSSE3_SET)
1861 #define OPTION_MASK_ISA_SSE4_2_SET \
1862 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_SSE4_1_SET)
1863 #define OPTION_MASK_ISA_AVX_SET \
1864 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_SSE4_2_SET)
1865 #define OPTION_MASK_ISA_FMA_SET \
1866 (OPTION_MASK_ISA_FMA | OPTION_MASK_ISA_AVX_SET)
1868 /* SSE4 includes both SSE4.1 and SSE4.2. -msse4 should be the same
1870 #define OPTION_MASK_ISA_SSE4_SET OPTION_MASK_ISA_SSE4_2_SET
1872 #define OPTION_MASK_ISA_SSE4A_SET \
1873 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE3_SET)
1874 #define OPTION_MASK_ISA_SSE5_SET \
1875 (OPTION_MASK_ISA_SSE5 | OPTION_MASK_ISA_SSE4A_SET)
1877 /* AES and PCLMUL need SSE2 because they use xmm registers */
1878 #define OPTION_MASK_ISA_AES_SET \
1879 (OPTION_MASK_ISA_AES | OPTION_MASK_ISA_SSE2_SET)
1880 #define OPTION_MASK_ISA_PCLMUL_SET \
1881 (OPTION_MASK_ISA_PCLMUL | OPTION_MASK_ISA_SSE2_SET)
1883 #define OPTION_MASK_ISA_ABM_SET \
1884 (OPTION_MASK_ISA_ABM | OPTION_MASK_ISA_POPCNT)
1885 #define OPTION_MASK_ISA_POPCNT_SET OPTION_MASK_ISA_POPCNT
1886 #define OPTION_MASK_ISA_CX16_SET OPTION_MASK_ISA_CX16
1887 #define OPTION_MASK_ISA_SAHF_SET OPTION_MASK_ISA_SAHF
1889 /* Define a set of ISAs which aren't available when a given ISA is
1890 disabled. MMX and SSE ISAs are handled separately. */
1892 #define OPTION_MASK_ISA_MMX_UNSET \
1893 (OPTION_MASK_ISA_MMX | OPTION_MASK_ISA_3DNOW_UNSET)
1894 #define OPTION_MASK_ISA_3DNOW_UNSET \
1895 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_3DNOW_A_UNSET)
1896 #define OPTION_MASK_ISA_3DNOW_A_UNSET OPTION_MASK_ISA_3DNOW_A
1898 #define OPTION_MASK_ISA_SSE_UNSET \
1899 (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_SSE2_UNSET)
1900 #define OPTION_MASK_ISA_SSE2_UNSET \
1901 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE3_UNSET)
1902 #define OPTION_MASK_ISA_SSE3_UNSET \
1903 (OPTION_MASK_ISA_SSE3 \
1904 | OPTION_MASK_ISA_SSSE3_UNSET \
1905 | OPTION_MASK_ISA_SSE4A_UNSET )
1906 #define OPTION_MASK_ISA_SSSE3_UNSET \
1907 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE4_1_UNSET)
1908 #define OPTION_MASK_ISA_SSE4_1_UNSET \
1909 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSE4_2_UNSET)
1910 #define OPTION_MASK_ISA_SSE4_2_UNSET \
1911 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_AVX_UNSET )
1912 #define OPTION_MASK_ISA_AVX_UNSET \
1913 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_FMA_UNSET)
1914 #define OPTION_MASK_ISA_FMA_UNSET OPTION_MASK_ISA_FMA
1916 /* SSE4 includes both SSE4.1 and SSE4.2. -mno-sse4 should the same
1918 #define OPTION_MASK_ISA_SSE4_UNSET OPTION_MASK_ISA_SSE4_1_UNSET
1920 #define OPTION_MASK_ISA_SSE4A_UNSET \
1921 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE5_UNSET)
1922 #define OPTION_MASK_ISA_SSE5_UNSET OPTION_MASK_ISA_SSE5
1923 #define OPTION_MASK_ISA_AES_UNSET OPTION_MASK_ISA_AES
1924 #define OPTION_MASK_ISA_PCLMUL_UNSET OPTION_MASK_ISA_PCLMUL
1925 #define OPTION_MASK_ISA_ABM_UNSET OPTION_MASK_ISA_ABM
1926 #define OPTION_MASK_ISA_POPCNT_UNSET OPTION_MASK_ISA_POPCNT
1927 #define OPTION_MASK_ISA_CX16_UNSET OPTION_MASK_ISA_CX16
1928 #define OPTION_MASK_ISA_SAHF_UNSET OPTION_MASK_ISA_SAHF
1930 /* Vectorization library interface and handlers. */
1931 tree (*ix86_veclib_handler)(enum built_in_function, tree, tree) = NULL;
1932 static tree ix86_veclibabi_svml (enum built_in_function, tree, tree);
1933 static tree ix86_veclibabi_acml (enum built_in_function, tree, tree);
1935 /* Processor target table, indexed by processor number */
1938 const struct processor_costs *cost; /* Processor costs */
1939 const int align_loop; /* Default alignments. */
1940 const int align_loop_max_skip;
1941 const int align_jump;
1942 const int align_jump_max_skip;
1943 const int align_func;
1946 static const struct ptt processor_target_table[PROCESSOR_max] =
1948 {&i386_cost, 4, 3, 4, 3, 4},
1949 {&i486_cost, 16, 15, 16, 15, 16},
1950 {&pentium_cost, 16, 7, 16, 7, 16},
1951 {&pentiumpro_cost, 16, 15, 16, 10, 16},
1952 {&geode_cost, 0, 0, 0, 0, 0},
1953 {&k6_cost, 32, 7, 32, 7, 32},
1954 {&athlon_cost, 16, 7, 16, 7, 16},
1955 {&pentium4_cost, 0, 0, 0, 0, 0},
1956 {&k8_cost, 16, 7, 16, 7, 16},
1957 {&nocona_cost, 0, 0, 0, 0, 0},
1958 {&core2_cost, 16, 10, 16, 10, 16},
1959 {&generic32_cost, 16, 7, 16, 7, 16},
1960 {&generic64_cost, 16, 10, 16, 10, 16},
1961 {&amdfam10_cost, 32, 24, 32, 7, 32}
1964 static const char *const cpu_names[TARGET_CPU_DEFAULT_max] =
1989 /* Implement TARGET_HANDLE_OPTION. */
1992 ix86_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, int value)
1999 ix86_isa_flags |= OPTION_MASK_ISA_MMX_SET;
2000 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_SET;
2004 ix86_isa_flags &= ~OPTION_MASK_ISA_MMX_UNSET;
2005 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_UNSET;
2012 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_SET;
2013 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_SET;
2017 ix86_isa_flags &= ~OPTION_MASK_ISA_3DNOW_UNSET;
2018 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_UNSET;
2028 ix86_isa_flags |= OPTION_MASK_ISA_SSE_SET;
2029 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_SET;
2033 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE_UNSET;
2034 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_UNSET;
2041 ix86_isa_flags |= OPTION_MASK_ISA_SSE2_SET;
2042 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_SET;
2046 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE2_UNSET;
2047 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_UNSET;
2054 ix86_isa_flags |= OPTION_MASK_ISA_SSE3_SET;
2055 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_SET;
2059 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE3_UNSET;
2060 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_UNSET;
2067 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3_SET;
2068 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_SET;
2072 ix86_isa_flags &= ~OPTION_MASK_ISA_SSSE3_UNSET;
2073 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_UNSET;
2080 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1_SET;
2081 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_SET;
2085 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_1_UNSET;
2086 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_UNSET;
2093 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2_SET;
2094 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_SET;
2098 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_2_UNSET;
2099 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_UNSET;
2106 ix86_isa_flags |= OPTION_MASK_ISA_AVX_SET;
2107 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_SET;
2111 ix86_isa_flags &= ~OPTION_MASK_ISA_AVX_UNSET;
2112 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_UNSET;
2119 ix86_isa_flags |= OPTION_MASK_ISA_FMA_SET;
2120 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_SET;
2124 ix86_isa_flags &= ~OPTION_MASK_ISA_FMA_UNSET;
2125 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_UNSET;
2130 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
2131 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
2135 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
2136 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
2142 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A_SET;
2143 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_SET;
2147 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4A_UNSET;
2148 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_UNSET;
2155 ix86_isa_flags |= OPTION_MASK_ISA_SSE5_SET;
2156 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_SET;
2160 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE5_UNSET;
2161 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_UNSET;
2168 ix86_isa_flags |= OPTION_MASK_ISA_ABM_SET;
2169 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_SET;
2173 ix86_isa_flags &= ~OPTION_MASK_ISA_ABM_UNSET;
2174 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_UNSET;
2181 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT_SET;
2182 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_SET;
2186 ix86_isa_flags &= ~OPTION_MASK_ISA_POPCNT_UNSET;
2187 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_UNSET;
2194 ix86_isa_flags |= OPTION_MASK_ISA_SAHF_SET;
2195 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_SET;
2199 ix86_isa_flags &= ~OPTION_MASK_ISA_SAHF_UNSET;
2200 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_UNSET;
2207 ix86_isa_flags |= OPTION_MASK_ISA_CX16_SET;
2208 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_SET;
2212 ix86_isa_flags &= ~OPTION_MASK_ISA_CX16_UNSET;
2213 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_UNSET;
2220 ix86_isa_flags |= OPTION_MASK_ISA_AES_SET;
2221 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_SET;
2225 ix86_isa_flags &= ~OPTION_MASK_ISA_AES_UNSET;
2226 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_UNSET;
2233 ix86_isa_flags |= OPTION_MASK_ISA_PCLMUL_SET;
2234 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_SET;
2238 ix86_isa_flags &= ~OPTION_MASK_ISA_PCLMUL_UNSET;
2239 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_UNSET;
2248 /* Return a string the documents the current -m options. The caller is
2249 responsible for freeing the string. */
2252 ix86_target_string (int isa, int flags, const char *arch, const char *tune,
2253 const char *fpmath, bool add_nl_p)
2255 struct ix86_target_opts
2257 const char *option; /* option string */
2258 int mask; /* isa mask options */
2261 /* This table is ordered so that options like -msse5 or -msse4.2 that imply
2262 preceding options while match those first. */
2263 static struct ix86_target_opts isa_opts[] =
2265 { "-m64", OPTION_MASK_ISA_64BIT },
2266 { "-msse5", OPTION_MASK_ISA_SSE5 },
2267 { "-msse4a", OPTION_MASK_ISA_SSE4A },
2268 { "-msse4.2", OPTION_MASK_ISA_SSE4_2 },
2269 { "-msse4.1", OPTION_MASK_ISA_SSE4_1 },
2270 { "-mssse3", OPTION_MASK_ISA_SSSE3 },
2271 { "-msse3", OPTION_MASK_ISA_SSE3 },
2272 { "-msse2", OPTION_MASK_ISA_SSE2 },
2273 { "-msse", OPTION_MASK_ISA_SSE },
2274 { "-m3dnow", OPTION_MASK_ISA_3DNOW },
2275 { "-m3dnowa", OPTION_MASK_ISA_3DNOW_A },
2276 { "-mmmx", OPTION_MASK_ISA_MMX },
2277 { "-mabm", OPTION_MASK_ISA_ABM },
2278 { "-mpopcnt", OPTION_MASK_ISA_POPCNT },
2279 { "-maes", OPTION_MASK_ISA_AES },
2280 { "-mpclmul", OPTION_MASK_ISA_PCLMUL },
2284 static struct ix86_target_opts flag_opts[] =
2286 { "-m128bit-long-double", MASK_128BIT_LONG_DOUBLE },
2287 { "-m80387", MASK_80387 },
2288 { "-maccumulate-outgoing-args", MASK_ACCUMULATE_OUTGOING_ARGS },
2289 { "-malign-double", MASK_ALIGN_DOUBLE },
2290 { "-mcld", MASK_CLD },
2291 { "-mfp-ret-in-387", MASK_FLOAT_RETURNS },
2292 { "-mieee-fp", MASK_IEEE_FP },
2293 { "-minline-all-stringops", MASK_INLINE_ALL_STRINGOPS },
2294 { "-minline-stringops-dynamically", MASK_INLINE_STRINGOPS_DYNAMICALLY },
2295 { "-mms-bitfields", MASK_MS_BITFIELD_LAYOUT },
2296 { "-mno-align-stringops", MASK_NO_ALIGN_STRINGOPS },
2297 { "-mno-fancy-math-387", MASK_NO_FANCY_MATH_387 },
2298 { "-mno-fused-madd", MASK_NO_FUSED_MADD },
2299 { "-mno-push-args", MASK_NO_PUSH_ARGS },
2300 { "-mno-red-zone", MASK_NO_RED_ZONE },
2301 { "-momit-leaf-frame-pointer", MASK_OMIT_LEAF_FRAME_POINTER },
2302 { "-mrecip", MASK_RECIP },
2303 { "-mrtd", MASK_RTD },
2304 { "-msseregparm", MASK_SSEREGPARM },
2305 { "-mstack-arg-probe", MASK_STACK_PROBE },
2306 { "-mtls-direct-seg-refs", MASK_TLS_DIRECT_SEG_REFS },
2309 const char *opts[ (sizeof (isa_opts) / sizeof (isa_opts[0])
2310 + sizeof (flag_opts) / sizeof (flag_opts[0])
2314 char target_other[40];
2323 memset (opts, '\0', sizeof (opts));
2325 /* Add -march= option. */
2328 opts[num][0] = "-march=";
2329 opts[num++][1] = arch;
2332 /* Add -mtune= option. */
2335 opts[num][0] = "-mtune=";
2336 opts[num++][1] = tune;
2339 /* Pick out the options in isa options. */
2340 for (i = 0; i < sizeof (isa_opts) / sizeof (isa_opts[0]); i++)
2342 if ((isa & isa_opts[i].mask) != 0)
2344 opts[num++][0] = isa_opts[i].option;
2345 isa &= ~ isa_opts[i].mask;
2349 if (isa && add_nl_p)
2351 opts[num++][0] = isa_other;
2352 sprintf (isa_other, "(other isa: 0x%x)", isa);
2355 /* Add flag options. */
2356 for (i = 0; i < sizeof (flag_opts) / sizeof (flag_opts[0]); i++)
2358 if ((flags & flag_opts[i].mask) != 0)
2360 opts[num++][0] = flag_opts[i].option;
2361 flags &= ~ flag_opts[i].mask;
2365 if (flags && add_nl_p)
2367 opts[num++][0] = target_other;
2368 sprintf (target_other, "(other flags: 0x%x)", isa);
2371 /* Add -fpmath= option. */
2374 opts[num][0] = "-mfpmath=";
2375 opts[num++][1] = fpmath;
2382 gcc_assert (num < sizeof (opts) / sizeof (opts[0]));
2384 /* Size the string. */
2386 sep_len = (add_nl_p) ? 3 : 1;
2387 for (i = 0; i < num; i++)
2390 for (j = 0; j < 2; j++)
2392 len += strlen (opts[i][j]);
2395 /* Build the string. */
2396 ret = ptr = (char *) xmalloc (len);
2399 for (i = 0; i < num; i++)
2403 for (j = 0; j < 2; j++)
2404 len2[j] = (opts[i][j]) ? strlen (opts[i][j]) : 0;
2411 if (add_nl_p && line_len + len2[0] + len2[1] > 70)
2419 for (j = 0; j < 2; j++)
2422 memcpy (ptr, opts[i][j], len2[j]);
2424 line_len += len2[j];
2429 gcc_assert (ret + len >= ptr);
2434 /* Function that is callable from the debugger to print the current
2437 ix86_debug_options (void)
2439 char *opts = ix86_target_string (ix86_isa_flags, target_flags,
2440 ix86_arch_string, ix86_tune_string,
2441 ix86_fpmath_string, true);
2445 fprintf (stderr, "%s\n\n", opts);
2449 fprintf (stderr, "<no options>\n\n");
2454 /* Sometimes certain combinations of command options do not make
2455 sense on a particular target machine. You can define a macro
2456 `OVERRIDE_OPTIONS' to take account of this. This macro, if
2457 defined, is executed once just after all the command options have
2460 Don't use this macro to turn on various extra optimizations for
2461 `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */
2464 override_options (bool main_args_p)
2467 unsigned int ix86_arch_mask, ix86_tune_mask;
2472 /* Comes from final.c -- no real reason to change it. */
2473 #define MAX_CODE_ALIGN 16
2481 PTA_PREFETCH_SSE = 1 << 4,
2483 PTA_3DNOW_A = 1 << 6,
2487 PTA_POPCNT = 1 << 10,
2489 PTA_SSE4A = 1 << 12,
2490 PTA_NO_SAHF = 1 << 13,
2491 PTA_SSE4_1 = 1 << 14,
2492 PTA_SSE4_2 = 1 << 15,
2495 PTA_PCLMUL = 1 << 18,
2502 const char *const name; /* processor name or nickname. */
2503 const enum processor_type processor;
2504 const enum attr_cpu schedule;
2505 const unsigned /*enum pta_flags*/ flags;
2507 const processor_alias_table[] =
2509 {"i386", PROCESSOR_I386, CPU_NONE, 0},
2510 {"i486", PROCESSOR_I486, CPU_NONE, 0},
2511 {"i586", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2512 {"pentium", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2513 {"pentium-mmx", PROCESSOR_PENTIUM, CPU_PENTIUM, PTA_MMX},
2514 {"winchip-c6", PROCESSOR_I486, CPU_NONE, PTA_MMX},
2515 {"winchip2", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2516 {"c3", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2517 {"c3-2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX | PTA_SSE},
2518 {"i686", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2519 {"pentiumpro", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2520 {"pentium2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX},
2521 {"pentium3", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2523 {"pentium3m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2525 {"pentium-m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2526 PTA_MMX | PTA_SSE | PTA_SSE2},
2527 {"pentium4", PROCESSOR_PENTIUM4, CPU_NONE,
2528 PTA_MMX |PTA_SSE | PTA_SSE2},
2529 {"pentium4m", PROCESSOR_PENTIUM4, CPU_NONE,
2530 PTA_MMX | PTA_SSE | PTA_SSE2},
2531 {"prescott", PROCESSOR_NOCONA, CPU_NONE,
2532 PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3},
2533 {"nocona", PROCESSOR_NOCONA, CPU_NONE,
2534 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2535 | PTA_CX16 | PTA_NO_SAHF},
2536 {"core2", PROCESSOR_CORE2, CPU_CORE2,
2537 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2538 | PTA_SSSE3 | PTA_CX16},
2539 {"geode", PROCESSOR_GEODE, CPU_GEODE,
2540 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A |PTA_PREFETCH_SSE},
2541 {"k6", PROCESSOR_K6, CPU_K6, PTA_MMX},
2542 {"k6-2", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2543 {"k6-3", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2544 {"athlon", PROCESSOR_ATHLON, CPU_ATHLON,
2545 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2546 {"athlon-tbird", PROCESSOR_ATHLON, CPU_ATHLON,
2547 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2548 {"athlon-4", PROCESSOR_ATHLON, CPU_ATHLON,
2549 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2550 {"athlon-xp", PROCESSOR_ATHLON, CPU_ATHLON,
2551 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2552 {"athlon-mp", PROCESSOR_ATHLON, CPU_ATHLON,
2553 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2554 {"x86-64", PROCESSOR_K8, CPU_K8,
2555 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_NO_SAHF},
2556 {"k8", PROCESSOR_K8, CPU_K8,
2557 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2558 | PTA_SSE2 | PTA_NO_SAHF},
2559 {"k8-sse3", PROCESSOR_K8, CPU_K8,
2560 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2561 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2562 {"opteron", PROCESSOR_K8, CPU_K8,
2563 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2564 | PTA_SSE2 | PTA_NO_SAHF},
2565 {"opteron-sse3", PROCESSOR_K8, CPU_K8,
2566 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2567 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2568 {"athlon64", PROCESSOR_K8, CPU_K8,
2569 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2570 | PTA_SSE2 | PTA_NO_SAHF},
2571 {"athlon64-sse3", PROCESSOR_K8, CPU_K8,
2572 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2573 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2574 {"athlon-fx", PROCESSOR_K8, CPU_K8,
2575 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2576 | PTA_SSE2 | PTA_NO_SAHF},
2577 {"amdfam10", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2578 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2579 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2580 {"barcelona", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2581 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2582 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2583 {"generic32", PROCESSOR_GENERIC32, CPU_PENTIUMPRO,
2584 0 /* flags are only used for -march switch. */ },
2585 {"generic64", PROCESSOR_GENERIC64, CPU_GENERIC64,
2586 PTA_64BIT /* flags are only used for -march switch. */ },
2589 int const pta_size = ARRAY_SIZE (processor_alias_table);
2591 /* Set up prefix/suffix so the error messages refer to either the command
2592 line argument, or the attribute(target). */
2601 prefix = "option(\"";
2606 #ifdef SUBTARGET_OVERRIDE_OPTIONS
2607 SUBTARGET_OVERRIDE_OPTIONS;
2610 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
2611 SUBSUBTARGET_OVERRIDE_OPTIONS;
2614 /* -fPIC is the default for x86_64. */
2615 if (TARGET_MACHO && TARGET_64BIT)
2618 /* Set the default values for switches whose default depends on TARGET_64BIT
2619 in case they weren't overwritten by command line options. */
2622 /* Mach-O doesn't support omitting the frame pointer for now. */
2623 if (flag_omit_frame_pointer == 2)
2624 flag_omit_frame_pointer = (TARGET_MACHO ? 0 : 1);
2625 if (flag_asynchronous_unwind_tables == 2)
2626 flag_asynchronous_unwind_tables = 1;
2627 if (flag_pcc_struct_return == 2)
2628 flag_pcc_struct_return = 0;
2632 if (flag_omit_frame_pointer == 2)
2633 flag_omit_frame_pointer = 0;
2634 if (flag_asynchronous_unwind_tables == 2)
2635 flag_asynchronous_unwind_tables = 0;
2636 if (flag_pcc_struct_return == 2)
2637 flag_pcc_struct_return = DEFAULT_PCC_STRUCT_RETURN;
2640 /* Need to check -mtune=generic first. */
2641 if (ix86_tune_string)
2643 if (!strcmp (ix86_tune_string, "generic")
2644 || !strcmp (ix86_tune_string, "i686")
2645 /* As special support for cross compilers we read -mtune=native
2646 as -mtune=generic. With native compilers we won't see the
2647 -mtune=native, as it was changed by the driver. */
2648 || !strcmp (ix86_tune_string, "native"))
2651 ix86_tune_string = "generic64";
2653 ix86_tune_string = "generic32";
2655 /* If this call is for setting the option attribute, allow the
2656 generic32/generic64 that was previously set. */
2657 else if (!main_args_p
2658 && (!strcmp (ix86_tune_string, "generic32")
2659 || !strcmp (ix86_tune_string, "generic64")))
2661 else if (!strncmp (ix86_tune_string, "generic", 7))
2662 error ("bad value (%s) for %stune=%s %s",
2663 ix86_tune_string, prefix, suffix, sw);
2667 if (ix86_arch_string)
2668 ix86_tune_string = ix86_arch_string;
2669 if (!ix86_tune_string)
2671 ix86_tune_string = cpu_names[TARGET_CPU_DEFAULT];
2672 ix86_tune_defaulted = 1;
2675 /* ix86_tune_string is set to ix86_arch_string or defaulted. We
2676 need to use a sensible tune option. */
2677 if (!strcmp (ix86_tune_string, "generic")
2678 || !strcmp (ix86_tune_string, "x86-64")
2679 || !strcmp (ix86_tune_string, "i686"))
2682 ix86_tune_string = "generic64";
2684 ix86_tune_string = "generic32";
2687 if (ix86_stringop_string)
2689 if (!strcmp (ix86_stringop_string, "rep_byte"))
2690 stringop_alg = rep_prefix_1_byte;
2691 else if (!strcmp (ix86_stringop_string, "libcall"))
2692 stringop_alg = libcall;
2693 else if (!strcmp (ix86_stringop_string, "rep_4byte"))
2694 stringop_alg = rep_prefix_4_byte;
2695 else if (!strcmp (ix86_stringop_string, "rep_8byte"))
2696 stringop_alg = rep_prefix_8_byte;
2697 else if (!strcmp (ix86_stringop_string, "byte_loop"))
2698 stringop_alg = loop_1_byte;
2699 else if (!strcmp (ix86_stringop_string, "loop"))
2700 stringop_alg = loop;
2701 else if (!strcmp (ix86_stringop_string, "unrolled_loop"))
2702 stringop_alg = unrolled_loop;
2704 error ("bad value (%s) for %sstringop-strategy=%s %s",
2705 ix86_stringop_string, prefix, suffix, sw);
2707 if (!strcmp (ix86_tune_string, "x86-64"))
2708 warning (OPT_Wdeprecated, "%stune=x86-64%s is deprecated. Use "
2709 "%stune=k8%s or %stune=generic%s instead as appropriate.",
2710 prefix, suffix, prefix, suffix, prefix, suffix);
2712 if (!ix86_arch_string)
2713 ix86_arch_string = TARGET_64BIT ? "x86-64" : "i386";
2715 ix86_arch_specified = 1;
2717 if (!strcmp (ix86_arch_string, "generic"))
2718 error ("generic CPU can be used only for %stune=%s %s",
2719 prefix, suffix, sw);
2720 if (!strncmp (ix86_arch_string, "generic", 7))
2721 error ("bad value (%s) for %sarch=%s %s",
2722 ix86_arch_string, prefix, suffix, sw);
2724 if (ix86_cmodel_string != 0)
2726 if (!strcmp (ix86_cmodel_string, "small"))
2727 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2728 else if (!strcmp (ix86_cmodel_string, "medium"))
2729 ix86_cmodel = flag_pic ? CM_MEDIUM_PIC : CM_MEDIUM;
2730 else if (!strcmp (ix86_cmodel_string, "large"))
2731 ix86_cmodel = flag_pic ? CM_LARGE_PIC : CM_LARGE;
2733 error ("code model %s does not support PIC mode", ix86_cmodel_string);
2734 else if (!strcmp (ix86_cmodel_string, "32"))
2735 ix86_cmodel = CM_32;
2736 else if (!strcmp (ix86_cmodel_string, "kernel") && !flag_pic)
2737 ix86_cmodel = CM_KERNEL;
2739 error ("bad value (%s) for %scmodel=%s %s",
2740 ix86_cmodel_string, prefix, suffix, sw);
2744 /* For TARGET_64BIT and MS_ABI, force pic on, in order to enable the
2745 use of rip-relative addressing. This eliminates fixups that
2746 would otherwise be needed if this object is to be placed in a
2747 DLL, and is essentially just as efficient as direct addressing. */
2748 if (TARGET_64BIT && DEFAULT_ABI == MS_ABI)
2749 ix86_cmodel = CM_SMALL_PIC, flag_pic = 1;
2750 else if (TARGET_64BIT)
2751 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2753 ix86_cmodel = CM_32;
2755 if (ix86_asm_string != 0)
2758 && !strcmp (ix86_asm_string, "intel"))
2759 ix86_asm_dialect = ASM_INTEL;
2760 else if (!strcmp (ix86_asm_string, "att"))
2761 ix86_asm_dialect = ASM_ATT;
2763 error ("bad value (%s) for %sasm=%s %s",
2764 ix86_asm_string, prefix, suffix, sw);
2766 if ((TARGET_64BIT == 0) != (ix86_cmodel == CM_32))
2767 error ("code model %qs not supported in the %s bit mode",
2768 ix86_cmodel_string, TARGET_64BIT ? "64" : "32");
2769 if ((TARGET_64BIT != 0) != ((ix86_isa_flags & OPTION_MASK_ISA_64BIT) != 0))
2770 sorry ("%i-bit mode not compiled in",
2771 (ix86_isa_flags & OPTION_MASK_ISA_64BIT) ? 64 : 32);
2773 for (i = 0; i < pta_size; i++)
2774 if (! strcmp (ix86_arch_string, processor_alias_table[i].name))
2776 ix86_schedule = processor_alias_table[i].schedule;
2777 ix86_arch = processor_alias_table[i].processor;
2778 /* Default cpu tuning to the architecture. */
2779 ix86_tune = ix86_arch;
2781 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
2782 error ("CPU you selected does not support x86-64 "
2785 if (processor_alias_table[i].flags & PTA_MMX
2786 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_MMX))
2787 ix86_isa_flags |= OPTION_MASK_ISA_MMX;
2788 if (processor_alias_table[i].flags & PTA_3DNOW
2789 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW))
2790 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW;
2791 if (processor_alias_table[i].flags & PTA_3DNOW_A
2792 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW_A))
2793 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_A;
2794 if (processor_alias_table[i].flags & PTA_SSE
2795 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE))
2796 ix86_isa_flags |= OPTION_MASK_ISA_SSE;
2797 if (processor_alias_table[i].flags & PTA_SSE2
2798 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE2))
2799 ix86_isa_flags |= OPTION_MASK_ISA_SSE2;
2800 if (processor_alias_table[i].flags & PTA_SSE3
2801 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE3))
2802 ix86_isa_flags |= OPTION_MASK_ISA_SSE3;
2803 if (processor_alias_table[i].flags & PTA_SSSE3
2804 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSSE3))
2805 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3;
2806 if (processor_alias_table[i].flags & PTA_SSE4_1
2807 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_1))
2808 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1;
2809 if (processor_alias_table[i].flags & PTA_SSE4_2
2810 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_2))
2811 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2;
2812 if (processor_alias_table[i].flags & PTA_AVX
2813 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_AVX))
2814 ix86_isa_flags |= OPTION_MASK_ISA_AVX;
2815 if (processor_alias_table[i].flags & PTA_FMA
2816 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_FMA))
2817 ix86_isa_flags |= OPTION_MASK_ISA_FMA;
2818 if (processor_alias_table[i].flags & PTA_SSE4A
2819 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4A))
2820 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A;
2821 if (processor_alias_table[i].flags & PTA_SSE5
2822 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE5))
2823 ix86_isa_flags |= OPTION_MASK_ISA_SSE5;
2824 if (processor_alias_table[i].flags & PTA_ABM
2825 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_ABM))
2826 ix86_isa_flags |= OPTION_MASK_ISA_ABM;
2827 if (processor_alias_table[i].flags & PTA_CX16
2828 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_CX16))
2829 ix86_isa_flags |= OPTION_MASK_ISA_CX16;
2830 if (processor_alias_table[i].flags & (PTA_POPCNT | PTA_ABM)
2831 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_POPCNT))
2832 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT;
2833 if (!(TARGET_64BIT && (processor_alias_table[i].flags & PTA_NO_SAHF))
2834 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SAHF))
2835 ix86_isa_flags |= OPTION_MASK_ISA_SAHF;
2836 if (processor_alias_table[i].flags & PTA_AES
2837 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_AES))
2838 ix86_isa_flags |= OPTION_MASK_ISA_AES;
2839 if (processor_alias_table[i].flags & PTA_PCLMUL
2840 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_PCLMUL))
2841 ix86_isa_flags |= OPTION_MASK_ISA_PCLMUL;
2842 if (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE))
2843 x86_prefetch_sse = true;
2849 error ("bad value (%s) for %sarch=%s %s",
2850 ix86_arch_string, prefix, suffix, sw);
2852 ix86_arch_mask = 1u << ix86_arch;
2853 for (i = 0; i < X86_ARCH_LAST; ++i)
2854 ix86_arch_features[i] = !!(initial_ix86_arch_features[i] & ix86_arch_mask);
2856 for (i = 0; i < pta_size; i++)
2857 if (! strcmp (ix86_tune_string, processor_alias_table[i].name))
2859 ix86_schedule = processor_alias_table[i].schedule;
2860 ix86_tune = processor_alias_table[i].processor;
2861 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
2863 if (ix86_tune_defaulted)
2865 ix86_tune_string = "x86-64";
2866 for (i = 0; i < pta_size; i++)
2867 if (! strcmp (ix86_tune_string,
2868 processor_alias_table[i].name))
2870 ix86_schedule = processor_alias_table[i].schedule;
2871 ix86_tune = processor_alias_table[i].processor;
2874 error ("CPU you selected does not support x86-64 "
2877 /* Intel CPUs have always interpreted SSE prefetch instructions as
2878 NOPs; so, we can enable SSE prefetch instructions even when
2879 -mtune (rather than -march) points us to a processor that has them.
2880 However, the VIA C3 gives a SIGILL, so we only do that for i686 and
2881 higher processors. */
2883 && (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE)))
2884 x86_prefetch_sse = true;
2888 error ("bad value (%s) for %stune=%s %s",
2889 ix86_tune_string, prefix, suffix, sw);
2891 ix86_tune_mask = 1u << ix86_tune;
2892 for (i = 0; i < X86_TUNE_LAST; ++i)
2893 ix86_tune_features[i] = !!(initial_ix86_tune_features[i] & ix86_tune_mask);
2896 ix86_cost = &ix86_size_cost;
2898 ix86_cost = processor_target_table[ix86_tune].cost;
2900 /* Arrange to set up i386_stack_locals for all functions. */
2901 init_machine_status = ix86_init_machine_status;
2903 /* Validate -mregparm= value. */
2904 if (ix86_regparm_string)
2907 warning (0, "%sregparm%s is ignored in 64-bit mode", prefix, suffix);
2908 i = atoi (ix86_regparm_string);
2909 if (i < 0 || i > REGPARM_MAX)
2910 error ("%sregparm=%d%s is not between 0 and %d",
2911 prefix, i, suffix, REGPARM_MAX);
2916 ix86_regparm = REGPARM_MAX;
2918 /* If the user has provided any of the -malign-* options,
2919 warn and use that value only if -falign-* is not set.
2920 Remove this code in GCC 3.2 or later. */
2921 if (ix86_align_loops_string)
2923 warning (0, "%salign-loops%s is obsolete, use %salign-loops%s",
2924 prefix, suffix, prefix, suffix);
2925 if (align_loops == 0)
2927 i = atoi (ix86_align_loops_string);
2928 if (i < 0 || i > MAX_CODE_ALIGN)
2929 error ("%salign-loops=%d%s is not between 0 and %d",
2930 prefix, i, suffix, MAX_CODE_ALIGN);
2932 align_loops = 1 << i;
2936 if (ix86_align_jumps_string)
2938 warning (0, "%salign-jumps%s is obsolete, use %salign-jumps%s",
2939 prefix, suffix, prefix, suffix);
2940 if (align_jumps == 0)
2942 i = atoi (ix86_align_jumps_string);
2943 if (i < 0 || i > MAX_CODE_ALIGN)
2944 error ("%salign-loops=%d%s is not between 0 and %d",
2945 prefix, i, suffix, MAX_CODE_ALIGN);
2947 align_jumps = 1 << i;
2951 if (ix86_align_funcs_string)
2953 warning (0, "%salign-functions%s is obsolete, use %salign-functions%s",
2954 prefix, suffix, prefix, suffix);
2955 if (align_functions == 0)
2957 i = atoi (ix86_align_funcs_string);
2958 if (i < 0 || i > MAX_CODE_ALIGN)
2959 error ("%salign-loops=%d%s is not between 0 and %d",
2960 prefix, i, suffix, MAX_CODE_ALIGN);
2962 align_functions = 1 << i;
2966 /* Default align_* from the processor table. */
2967 if (align_loops == 0)
2969 align_loops = processor_target_table[ix86_tune].align_loop;
2970 align_loops_max_skip = processor_target_table[ix86_tune].align_loop_max_skip;
2972 if (align_jumps == 0)
2974 align_jumps = processor_target_table[ix86_tune].align_jump;
2975 align_jumps_max_skip = processor_target_table[ix86_tune].align_jump_max_skip;
2977 if (align_functions == 0)
2979 align_functions = processor_target_table[ix86_tune].align_func;
2982 /* Validate -mbranch-cost= value, or provide default. */
2983 ix86_branch_cost = ix86_cost->branch_cost;
2984 if (ix86_branch_cost_string)
2986 i = atoi (ix86_branch_cost_string);
2988 error ("%sbranch-cost=%d%s is not between 0 and 5", prefix, i, suffix);
2990 ix86_branch_cost = i;
2992 if (ix86_section_threshold_string)
2994 i = atoi (ix86_section_threshold_string);
2996 error ("%slarge-data-threshold=%d%s is negative", prefix, i, suffix);
2998 ix86_section_threshold = i;
3001 if (ix86_tls_dialect_string)
3003 if (strcmp (ix86_tls_dialect_string, "gnu") == 0)
3004 ix86_tls_dialect = TLS_DIALECT_GNU;
3005 else if (strcmp (ix86_tls_dialect_string, "gnu2") == 0)
3006 ix86_tls_dialect = TLS_DIALECT_GNU2;
3007 else if (strcmp (ix86_tls_dialect_string, "sun") == 0)
3008 ix86_tls_dialect = TLS_DIALECT_SUN;
3010 error ("bad value (%s) for %stls-dialect=%s %s",
3011 ix86_tls_dialect_string, prefix, suffix, sw);
3014 if (ix87_precision_string)
3016 i = atoi (ix87_precision_string);
3017 if (i != 32 && i != 64 && i != 80)
3018 error ("pc%d is not valid precision setting (32, 64 or 80)", i);
3023 target_flags |= TARGET_SUBTARGET64_DEFAULT & ~target_flags_explicit;
3025 /* Enable by default the SSE and MMX builtins. Do allow the user to
3026 explicitly disable any of these. In particular, disabling SSE and
3027 MMX for kernel code is extremely useful. */
3028 if (!ix86_arch_specified)
3030 |= ((OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_MMX
3031 | TARGET_SUBTARGET64_ISA_DEFAULT) & ~ix86_isa_flags_explicit);
3034 warning (0, "%srtd%s is ignored in 64bit mode", prefix, suffix);
3038 target_flags |= TARGET_SUBTARGET32_DEFAULT & ~target_flags_explicit;
3040 if (!ix86_arch_specified)
3042 |= TARGET_SUBTARGET32_ISA_DEFAULT & ~ix86_isa_flags_explicit;
3044 /* i386 ABI does not specify red zone. It still makes sense to use it
3045 when programmer takes care to stack from being destroyed. */
3046 if (!(target_flags_explicit & MASK_NO_RED_ZONE))
3047 target_flags |= MASK_NO_RED_ZONE;
3050 /* Keep nonleaf frame pointers. */
3051 if (flag_omit_frame_pointer)
3052 target_flags &= ~MASK_OMIT_LEAF_FRAME_POINTER;
3053 else if (TARGET_OMIT_LEAF_FRAME_POINTER)
3054 flag_omit_frame_pointer = 1;
3056 /* If we're doing fast math, we don't care about comparison order
3057 wrt NaNs. This lets us use a shorter comparison sequence. */
3058 if (flag_finite_math_only)
3059 target_flags &= ~MASK_IEEE_FP;
3061 /* If the architecture always has an FPU, turn off NO_FANCY_MATH_387,
3062 since the insns won't need emulation. */
3063 if (x86_arch_always_fancy_math_387 & ix86_arch_mask)
3064 target_flags &= ~MASK_NO_FANCY_MATH_387;
3066 /* Likewise, if the target doesn't have a 387, or we've specified
3067 software floating point, don't use 387 inline intrinsics. */
3069 target_flags |= MASK_NO_FANCY_MATH_387;
3071 /* Turn on MMX builtins for -msse. */
3074 ix86_isa_flags |= OPTION_MASK_ISA_MMX & ~ix86_isa_flags_explicit;
3075 x86_prefetch_sse = true;
3078 /* Turn on popcnt instruction for -msse4.2 or -mabm. */
3079 if (TARGET_SSE4_2 || TARGET_ABM)
3080 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT & ~ix86_isa_flags_explicit;
3082 /* Validate -mpreferred-stack-boundary= value or default it to
3083 PREFERRED_STACK_BOUNDARY_DEFAULT. */
3084 ix86_preferred_stack_boundary = PREFERRED_STACK_BOUNDARY_DEFAULT;
3085 if (ix86_preferred_stack_boundary_string)
3087 i = atoi (ix86_preferred_stack_boundary_string);
3088 if (i < (TARGET_64BIT ? 4 : 2) || i > 12)
3089 error ("%spreferred-stack-boundary=%d%s is not between %d and 12",
3090 prefix, i, suffix, TARGET_64BIT ? 4 : 2);
3092 ix86_preferred_stack_boundary = (1 << i) * BITS_PER_UNIT;
3095 /* Set the default value for -mstackrealign. */
3096 if (ix86_force_align_arg_pointer == -1)
3097 ix86_force_align_arg_pointer = STACK_REALIGN_DEFAULT;
3099 /* Validate -mincoming-stack-boundary= value or default it to
3100 MIN_STACK_BOUNDARY/PREFERRED_STACK_BOUNDARY. */
3101 if (ix86_force_align_arg_pointer)
3102 ix86_default_incoming_stack_boundary = MIN_STACK_BOUNDARY;
3104 ix86_default_incoming_stack_boundary = PREFERRED_STACK_BOUNDARY;
3105 ix86_incoming_stack_boundary = ix86_default_incoming_stack_boundary;
3106 if (ix86_incoming_stack_boundary_string)
3108 i = atoi (ix86_incoming_stack_boundary_string);
3109 if (i < (TARGET_64BIT ? 4 : 2) || i > 12)
3110 error ("-mincoming-stack-boundary=%d is not between %d and 12",
3111 i, TARGET_64BIT ? 4 : 2);
3114 ix86_user_incoming_stack_boundary = (1 << i) * BITS_PER_UNIT;
3115 ix86_incoming_stack_boundary
3116 = ix86_user_incoming_stack_boundary;
3120 /* Accept -msseregparm only if at least SSE support is enabled. */
3121 if (TARGET_SSEREGPARM
3123 error ("%ssseregparm%s used without SSE enabled", prefix, suffix);
3125 ix86_fpmath = TARGET_FPMATH_DEFAULT;
3126 if (ix86_fpmath_string != 0)
3128 if (! strcmp (ix86_fpmath_string, "387"))
3129 ix86_fpmath = FPMATH_387;
3130 else if (! strcmp (ix86_fpmath_string, "sse"))
3134 warning (0, "SSE instruction set disabled, using 387 arithmetics");
3135 ix86_fpmath = FPMATH_387;
3138 ix86_fpmath = FPMATH_SSE;
3140 else if (! strcmp (ix86_fpmath_string, "387,sse")
3141 || ! strcmp (ix86_fpmath_string, "387+sse")
3142 || ! strcmp (ix86_fpmath_string, "sse,387")
3143 || ! strcmp (ix86_fpmath_string, "sse+387")
3144 || ! strcmp (ix86_fpmath_string, "both"))
3148 warning (0, "SSE instruction set disabled, using 387 arithmetics");
3149 ix86_fpmath = FPMATH_387;
3151 else if (!TARGET_80387)
3153 warning (0, "387 instruction set disabled, using SSE arithmetics");
3154 ix86_fpmath = FPMATH_SSE;
3157 ix86_fpmath = (enum fpmath_unit) (FPMATH_SSE | FPMATH_387);
3160 error ("bad value (%s) for %sfpmath=%s %s",
3161 ix86_fpmath_string, prefix, suffix, sw);
3164 /* If the i387 is disabled, then do not return values in it. */
3166 target_flags &= ~MASK_FLOAT_RETURNS;
3168 /* Use external vectorized library in vectorizing intrinsics. */
3169 if (ix86_veclibabi_string)
3171 if (strcmp (ix86_veclibabi_string, "svml") == 0)
3172 ix86_veclib_handler = ix86_veclibabi_svml;
3173 else if (strcmp (ix86_veclibabi_string, "acml") == 0)
3174 ix86_veclib_handler = ix86_veclibabi_acml;
3176 error ("unknown vectorization library ABI type (%s) for "
3177 "%sveclibabi=%s %s", ix86_veclibabi_string,
3178 prefix, suffix, sw);
3181 if ((x86_accumulate_outgoing_args & ix86_tune_mask)
3182 && !(target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3184 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3186 /* ??? Unwind info is not correct around the CFG unless either a frame
3187 pointer is present or M_A_O_A is set. Fixing this requires rewriting
3188 unwind info generation to be aware of the CFG and propagating states
3190 if ((flag_unwind_tables || flag_asynchronous_unwind_tables
3191 || flag_exceptions || flag_non_call_exceptions)
3192 && flag_omit_frame_pointer
3193 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
3195 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3196 warning (0, "unwind tables currently require either a frame pointer "
3197 "or %saccumulate-outgoing-args%s for correctness",
3199 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3202 /* If stack probes are required, the space used for large function
3203 arguments on the stack must also be probed, so enable
3204 -maccumulate-outgoing-args so this happens in the prologue. */
3205 if (TARGET_STACK_PROBE
3206 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
3208 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
3209 warning (0, "stack probing requires %saccumulate-outgoing-args%s "
3210 "for correctness", prefix, suffix);
3211 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
3214 /* For sane SSE instruction set generation we need fcomi instruction.
3215 It is safe to enable all CMOVE instructions. */
3219 /* Figure out what ASM_GENERATE_INTERNAL_LABEL builds as a prefix. */
3222 ASM_GENERATE_INTERNAL_LABEL (internal_label_prefix, "LX", 0);
3223 p = strchr (internal_label_prefix, 'X');
3224 internal_label_prefix_len = p - internal_label_prefix;
3228 /* When scheduling description is not available, disable scheduler pass
3229 so it won't slow down the compilation and make x87 code slower. */
3230 if (!TARGET_SCHEDULE)
3231 flag_schedule_insns_after_reload = flag_schedule_insns = 0;
3233 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES))
3234 set_param_value ("simultaneous-prefetches",
3235 ix86_cost->simultaneous_prefetches);
3236 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE))
3237 set_param_value ("l1-cache-line-size", ix86_cost->prefetch_block);
3238 if (!PARAM_SET_P (PARAM_L1_CACHE_SIZE))
3239 set_param_value ("l1-cache-size", ix86_cost->l1_cache_size);
3240 if (!PARAM_SET_P (PARAM_L2_CACHE_SIZE))
3241 set_param_value ("l2-cache-size", ix86_cost->l2_cache_size);
3243 /* If using typedef char *va_list, signal that __builtin_va_start (&ap, 0)
3244 can be optimized to ap = __builtin_next_arg (0). */
3246 targetm.expand_builtin_va_start = NULL;
3250 ix86_gen_leave = gen_leave_rex64;
3251 ix86_gen_pop1 = gen_popdi1;
3252 ix86_gen_add3 = gen_adddi3;
3253 ix86_gen_sub3 = gen_subdi3;
3254 ix86_gen_sub3_carry = gen_subdi3_carry_rex64;
3255 ix86_gen_one_cmpl2 = gen_one_cmpldi2;
3256 ix86_gen_monitor = gen_sse3_monitor64;
3257 ix86_gen_andsp = gen_anddi3;
3261 ix86_gen_leave = gen_leave;
3262 ix86_gen_pop1 = gen_popsi1;
3263 ix86_gen_add3 = gen_addsi3;
3264 ix86_gen_sub3 = gen_subsi3;
3265 ix86_gen_sub3_carry = gen_subsi3_carry;
3266 ix86_gen_one_cmpl2 = gen_one_cmplsi2;
3267 ix86_gen_monitor = gen_sse3_monitor;
3268 ix86_gen_andsp = gen_andsi3;
3272 /* Use -mcld by default for 32-bit code if configured with --enable-cld. */
3274 target_flags |= MASK_CLD & ~target_flags_explicit;
3277 /* Save the initial options in case the user does function specific options */
3279 target_option_default_node = target_option_current_node
3280 = build_target_option_node ();
3283 /* Save the current options */
3286 ix86_function_specific_save (struct cl_target_option *ptr)
3288 gcc_assert (IN_RANGE (ix86_arch, 0, 255));
3289 gcc_assert (IN_RANGE (ix86_schedule, 0, 255));
3290 gcc_assert (IN_RANGE (ix86_tune, 0, 255));
3291 gcc_assert (IN_RANGE (ix86_fpmath, 0, 255));
3292 gcc_assert (IN_RANGE (ix86_branch_cost, 0, 255));
3294 ptr->arch = ix86_arch;
3295 ptr->schedule = ix86_schedule;
3296 ptr->tune = ix86_tune;
3297 ptr->fpmath = ix86_fpmath;
3298 ptr->branch_cost = ix86_branch_cost;
3299 ptr->tune_defaulted = ix86_tune_defaulted;
3300 ptr->arch_specified = ix86_arch_specified;
3301 ptr->ix86_isa_flags_explicit = ix86_isa_flags_explicit;
3302 ptr->target_flags_explicit = target_flags_explicit;
3305 /* Restore the current options */
3308 ix86_function_specific_restore (struct cl_target_option *ptr)
3310 enum processor_type old_tune = ix86_tune;
3311 enum processor_type old_arch = ix86_arch;
3312 unsigned int ix86_arch_mask, ix86_tune_mask;
3315 ix86_arch = ptr->arch;
3316 ix86_schedule = ptr->schedule;
3317 ix86_tune = ptr->tune;
3318 ix86_fpmath = ptr->fpmath;
3319 ix86_branch_cost = ptr->branch_cost;
3320 ix86_tune_defaulted = ptr->tune_defaulted;
3321 ix86_arch_specified = ptr->arch_specified;
3322 ix86_isa_flags_explicit = ptr->ix86_isa_flags_explicit;
3323 target_flags_explicit = ptr->target_flags_explicit;
3325 /* Recreate the arch feature tests if the arch changed */
3326 if (old_arch != ix86_arch)
3328 ix86_arch_mask = 1u << ix86_arch;
3329 for (i = 0; i < X86_ARCH_LAST; ++i)
3330 ix86_arch_features[i]
3331 = !!(initial_ix86_arch_features[i] & ix86_arch_mask);
3334 /* Recreate the tune optimization tests */
3335 if (old_tune != ix86_tune)
3337 ix86_tune_mask = 1u << ix86_tune;
3338 for (i = 0; i < X86_TUNE_LAST; ++i)
3339 ix86_tune_features[i]
3340 = !!(initial_ix86_tune_features[i] & ix86_tune_mask);
3344 /* Print the current options */
3347 ix86_function_specific_print (FILE *file, int indent,
3348 struct cl_target_option *ptr)
3351 = ix86_target_string (ptr->ix86_isa_flags, ptr->target_flags,
3352 NULL, NULL, NULL, false);
3354 fprintf (file, "%*sarch = %d (%s)\n",
3357 ((ptr->arch < TARGET_CPU_DEFAULT_max)
3358 ? cpu_names[ptr->arch]
3361 fprintf (file, "%*stune = %d (%s)\n",
3364 ((ptr->tune < TARGET_CPU_DEFAULT_max)
3365 ? cpu_names[ptr->tune]
3368 fprintf (file, "%*sfpmath = %d%s%s\n", indent, "", ptr->fpmath,
3369 (ptr->fpmath & FPMATH_387) ? ", 387" : "",
3370 (ptr->fpmath & FPMATH_SSE) ? ", sse" : "");
3371 fprintf (file, "%*sbranch_cost = %d\n", indent, "", ptr->branch_cost);
3375 fprintf (file, "%*s%s\n", indent, "", target_string);
3376 free (target_string);
3381 /* Inner function to process the attribute((target(...))), take an argument and
3382 set the current options from the argument. If we have a list, recursively go
3386 ix86_valid_target_attribute_inner_p (tree args, char *p_strings[])
3391 #define IX86_ATTR_ISA(S,O) { S, sizeof (S)-1, ix86_opt_isa, O, 0 }
3392 #define IX86_ATTR_STR(S,O) { S, sizeof (S)-1, ix86_opt_str, O, 0 }
3393 #define IX86_ATTR_YES(S,O,M) { S, sizeof (S)-1, ix86_opt_yes, O, M }
3394 #define IX86_ATTR_NO(S,O,M) { S, sizeof (S)-1, ix86_opt_no, O, M }
3409 enum ix86_opt_type type;
3414 IX86_ATTR_ISA ("3dnow", OPT_m3dnow),
3415 IX86_ATTR_ISA ("abm", OPT_mabm),
3416 IX86_ATTR_ISA ("aes", OPT_maes),
3417 IX86_ATTR_ISA ("avx", OPT_mavx),
3418 IX86_ATTR_ISA ("mmx", OPT_mmmx),
3419 IX86_ATTR_ISA ("pclmul", OPT_mpclmul),
3420 IX86_ATTR_ISA ("popcnt", OPT_mpopcnt),
3421 IX86_ATTR_ISA ("sse", OPT_msse),
3422 IX86_ATTR_ISA ("sse2", OPT_msse2),
3423 IX86_ATTR_ISA ("sse3", OPT_msse3),
3424 IX86_ATTR_ISA ("sse4", OPT_msse4),
3425 IX86_ATTR_ISA ("sse4.1", OPT_msse4_1),
3426 IX86_ATTR_ISA ("sse4.2", OPT_msse4_2),
3427 IX86_ATTR_ISA ("sse4a", OPT_msse4a),
3428 IX86_ATTR_ISA ("sse5", OPT_msse5),
3429 IX86_ATTR_ISA ("ssse3", OPT_mssse3),
3431 /* string options */
3432 IX86_ATTR_STR ("arch=", IX86_FUNCTION_SPECIFIC_ARCH),
3433 IX86_ATTR_STR ("fpmath=", IX86_FUNCTION_SPECIFIC_FPMATH),
3434 IX86_ATTR_STR ("tune=", IX86_FUNCTION_SPECIFIC_TUNE),
3437 IX86_ATTR_YES ("cld",
3441 IX86_ATTR_NO ("fancy-math-387",
3442 OPT_mfancy_math_387,
3443 MASK_NO_FANCY_MATH_387),
3445 IX86_ATTR_NO ("fused-madd",
3447 MASK_NO_FUSED_MADD),
3449 IX86_ATTR_YES ("ieee-fp",
3453 IX86_ATTR_YES ("inline-all-stringops",
3454 OPT_minline_all_stringops,
3455 MASK_INLINE_ALL_STRINGOPS),
3457 IX86_ATTR_YES ("inline-stringops-dynamically",
3458 OPT_minline_stringops_dynamically,
3459 MASK_INLINE_STRINGOPS_DYNAMICALLY),
3461 IX86_ATTR_NO ("align-stringops",
3462 OPT_mno_align_stringops,
3463 MASK_NO_ALIGN_STRINGOPS),
3465 IX86_ATTR_YES ("recip",
3471 /* If this is a list, recurse to get the options. */
3472 if (TREE_CODE (args) == TREE_LIST)
3476 for (; args; args = TREE_CHAIN (args))
3477 if (TREE_VALUE (args)
3478 && !ix86_valid_target_attribute_inner_p (TREE_VALUE (args), p_strings))
3484 else if (TREE_CODE (args) != STRING_CST)
3487 /* Handle multiple arguments separated by commas. */
3488 next_optstr = ASTRDUP (TREE_STRING_POINTER (args));
3490 while (next_optstr && *next_optstr != '\0')
3492 char *p = next_optstr;
3494 char *comma = strchr (next_optstr, ',');
3495 const char *opt_string;
3496 size_t len, opt_len;
3501 enum ix86_opt_type type = ix86_opt_unknown;
3507 len = comma - next_optstr;
3508 next_optstr = comma + 1;
3516 /* Recognize no-xxx. */
3517 if (len > 3 && p[0] == 'n' && p[1] == 'o' && p[2] == '-')
3526 /* Find the option. */
3529 for (i = 0; i < sizeof (attrs) / sizeof (attrs[0]); i++)
3531 type = attrs[i].type;
3532 opt_len = attrs[i].len;
3533 if (ch == attrs[i].string[0]
3534 && ((type != ix86_opt_str) ? len == opt_len : len > opt_len)
3535 && memcmp (p, attrs[i].string, opt_len) == 0)
3538 mask = attrs[i].mask;
3539 opt_string = attrs[i].string;
3544 /* Process the option. */
3547 error ("attribute(target(\"%s\")) is unknown", orig_p);
3551 else if (type == ix86_opt_isa)
3552 ix86_handle_option (opt, p, opt_set_p);
3554 else if (type == ix86_opt_yes || type == ix86_opt_no)
3556 if (type == ix86_opt_no)
3557 opt_set_p = !opt_set_p;
3560 target_flags |= mask;
3562 target_flags &= ~mask;
3565 else if (type == ix86_opt_str)
3569 error ("option(\"%s\") was already specified", opt_string);
3573 p_strings[opt] = xstrdup (p + opt_len);
3583 /* Return a TARGET_OPTION_NODE tree of the target options listed or NULL. */
3586 ix86_valid_target_attribute_tree (tree args)
3588 const char *orig_arch_string = ix86_arch_string;
3589 const char *orig_tune_string = ix86_tune_string;
3590 const char *orig_fpmath_string = ix86_fpmath_string;
3591 int orig_tune_defaulted = ix86_tune_defaulted;
3592 int orig_arch_specified = ix86_arch_specified;
3593 char *option_strings[IX86_FUNCTION_SPECIFIC_MAX] = { NULL, NULL, NULL };
3596 struct cl_target_option *def
3597 = TREE_TARGET_OPTION (target_option_default_node);
3599 /* Process each of the options on the chain. */
3600 if (! ix86_valid_target_attribute_inner_p (args, option_strings))
3603 /* If the changed options are different from the default, rerun override_options,
3604 and then save the options away. The string options are are attribute options,
3605 and will be undone when we copy the save structure. */
3606 if (ix86_isa_flags != def->ix86_isa_flags
3607 || target_flags != def->target_flags
3608 || option_strings[IX86_FUNCTION_SPECIFIC_ARCH]
3609 || option_strings[IX86_FUNCTION_SPECIFIC_TUNE]
3610 || option_strings[IX86_FUNCTION_SPECIFIC_FPMATH])
3612 /* If we are using the default tune= or arch=, undo the string assigned,
3613 and use the default. */
3614 if (option_strings[IX86_FUNCTION_SPECIFIC_ARCH])
3615 ix86_arch_string = option_strings[IX86_FUNCTION_SPECIFIC_ARCH];
3616 else if (!orig_arch_specified)
3617 ix86_arch_string = NULL;
3619 if (option_strings[IX86_FUNCTION_SPECIFIC_TUNE])
3620 ix86_tune_string = option_strings[IX86_FUNCTION_SPECIFIC_TUNE];
3621 else if (orig_tune_defaulted)
3622 ix86_tune_string = NULL;
3624 /* If fpmath= is not set, and we now have sse2 on 32-bit, use it. */
3625 if (option_strings[IX86_FUNCTION_SPECIFIC_FPMATH])
3626 ix86_fpmath_string = option_strings[IX86_FUNCTION_SPECIFIC_FPMATH];
3627 else if (!TARGET_64BIT && TARGET_SSE)
3628 ix86_fpmath_string = "sse,387";
3630 /* Do any overrides, such as arch=xxx, or tune=xxx support. */
3631 override_options (false);
3633 /* Add any builtin functions with the new isa if any. */
3634 ix86_add_new_builtins (ix86_isa_flags);
3636 /* Save the current options unless we are validating options for
3638 t = build_target_option_node ();
3640 ix86_arch_string = orig_arch_string;
3641 ix86_tune_string = orig_tune_string;
3642 ix86_fpmath_string = orig_fpmath_string;
3644 /* Free up memory allocated to hold the strings */
3645 for (i = 0; i < IX86_FUNCTION_SPECIFIC_MAX; i++)
3646 if (option_strings[i])
3647 free (option_strings[i]);
3653 /* Hook to validate attribute((target("string"))). */
3656 ix86_valid_target_attribute_p (tree fndecl,
3657 tree ARG_UNUSED (name),
3659 int ARG_UNUSED (flags))
3661 struct cl_target_option cur_target;
3663 tree old_optimize = build_optimization_node ();
3664 tree new_target, new_optimize;
3665 tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl);
3667 /* If the function changed the optimization levels as well as setting target
3668 options, start with the optimizations specified. */
3669 if (func_optimize && func_optimize != old_optimize)
3670 cl_optimization_restore (TREE_OPTIMIZATION (func_optimize));
3672 /* The target attributes may also change some optimization flags, so update
3673 the optimization options if necessary. */
3674 cl_target_option_save (&cur_target);
3675 new_target = ix86_valid_target_attribute_tree (args);
3676 new_optimize = build_optimization_node ();
3683 DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = new_target;
3685 if (old_optimize != new_optimize)
3686 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize;
3689 cl_target_option_restore (&cur_target);
3691 if (old_optimize != new_optimize)
3692 cl_optimization_restore (TREE_OPTIMIZATION (old_optimize));
3698 /* Hook to determine if one function can safely inline another. */
3701 ix86_can_inline_p (tree caller, tree callee)
3704 tree caller_tree = DECL_FUNCTION_SPECIFIC_TARGET (caller);
3705 tree callee_tree = DECL_FUNCTION_SPECIFIC_TARGET (callee);
3707 /* If callee has no option attributes, then it is ok to inline. */
3711 /* If caller has no option attributes, but callee does then it is not ok to
3713 else if (!caller_tree)
3718 struct cl_target_option *caller_opts = TREE_TARGET_OPTION (caller_tree);
3719 struct cl_target_option *callee_opts = TREE_TARGET_OPTION (callee_tree);
3721 /* Callee's isa options should a subset of the caller's, i.e. a SSE5 function
3722 can inline a SSE2 function but a SSE2 function can't inline a SSE5
3724 if ((caller_opts->ix86_isa_flags & callee_opts->ix86_isa_flags)
3725 != callee_opts->ix86_isa_flags)
3728 /* See if we have the same non-isa options. */
3729 else if (caller_opts->target_flags != callee_opts->target_flags)
3732 /* See if arch, tune, etc. are the same. */
3733 else if (caller_opts->arch != callee_opts->arch)
3736 else if (caller_opts->tune != callee_opts->tune)
3739 else if (caller_opts->fpmath != callee_opts->fpmath)
3742 else if (caller_opts->branch_cost != callee_opts->branch_cost)
3753 /* Remember the last target of ix86_set_current_function. */
3754 static GTY(()) tree ix86_previous_fndecl;
3756 /* Establish appropriate back-end context for processing the function
3757 FNDECL. The argument might be NULL to indicate processing at top
3758 level, outside of any function scope. */
3760 ix86_set_current_function (tree fndecl)
3762 /* Only change the context if the function changes. This hook is called
3763 several times in the course of compiling a function, and we don't want to
3764 slow things down too much or call target_reinit when it isn't safe. */
3765 if (fndecl && fndecl != ix86_previous_fndecl)
3767 tree old_tree = (ix86_previous_fndecl
3768 ? DECL_FUNCTION_SPECIFIC_TARGET (ix86_previous_fndecl)
3771 tree new_tree = (fndecl
3772 ? DECL_FUNCTION_SPECIFIC_TARGET (fndecl)
3775 ix86_previous_fndecl = fndecl;
3776 if (old_tree == new_tree)
3781 cl_target_option_restore (TREE_TARGET_OPTION (new_tree));
3787 struct cl_target_option *def
3788 = TREE_TARGET_OPTION (target_option_current_node);
3790 cl_target_option_restore (def);
3797 /* Return true if this goes in large data/bss. */
3800 ix86_in_large_data_p (tree exp)
3802 if (ix86_cmodel != CM_MEDIUM && ix86_cmodel != CM_MEDIUM_PIC)
3805 /* Functions are never large data. */
3806 if (TREE_CODE (exp) == FUNCTION_DECL)
3809 if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
3811 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));
3812 if (strcmp (section, ".ldata") == 0
3813 || strcmp (section, ".lbss") == 0)
3819 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
3821 /* If this is an incomplete type with size 0, then we can't put it
3822 in data because it might be too big when completed. */
3823 if (!size || size > ix86_section_threshold)
3830 /* Switch to the appropriate section for output of DECL.
3831 DECL is either a `VAR_DECL' node or a constant of some sort.
3832 RELOC indicates whether forming the initial value of DECL requires
3833 link-time relocations. */
3835 static section * x86_64_elf_select_section (tree, int, unsigned HOST_WIDE_INT)
3839 x86_64_elf_select_section (tree decl, int reloc,
3840 unsigned HOST_WIDE_INT align)
3842 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
3843 && ix86_in_large_data_p (decl))
3845 const char *sname = NULL;
3846 unsigned int flags = SECTION_WRITE;
3847 switch (categorize_decl_for_section (decl, reloc))
3852 case SECCAT_DATA_REL:
3853 sname = ".ldata.rel";
3855 case SECCAT_DATA_REL_LOCAL:
3856 sname = ".ldata.rel.local";
3858 case SECCAT_DATA_REL_RO:
3859 sname = ".ldata.rel.ro";
3861 case SECCAT_DATA_REL_RO_LOCAL:
3862 sname = ".ldata.rel.ro.local";
3866 flags |= SECTION_BSS;
3869 case SECCAT_RODATA_MERGE_STR:
3870 case SECCAT_RODATA_MERGE_STR_INIT:
3871 case SECCAT_RODATA_MERGE_CONST:
3875 case SECCAT_SRODATA:
3882 /* We don't split these for medium model. Place them into
3883 default sections and hope for best. */
3885 case SECCAT_EMUTLS_VAR:
3886 case SECCAT_EMUTLS_TMPL:
3891 /* We might get called with string constants, but get_named_section
3892 doesn't like them as they are not DECLs. Also, we need to set
3893 flags in that case. */
3895 return get_section (sname, flags, NULL);
3896 return get_named_section (decl, sname, reloc);
3899 return default_elf_select_section (decl, reloc, align);
3902 /* Build up a unique section name, expressed as a
3903 STRING_CST node, and assign it to DECL_SECTION_NAME (decl).
3904 RELOC indicates whether the initial value of EXP requires
3905 link-time relocations. */
3907 static void ATTRIBUTE_UNUSED
3908 x86_64_elf_unique_section (tree decl, int reloc)
3910 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
3911 && ix86_in_large_data_p (decl))
3913 const char *prefix = NULL;
3914 /* We only need to use .gnu.linkonce if we don't have COMDAT groups. */
3915 bool one_only = DECL_ONE_ONLY (decl) && !HAVE_COMDAT_GROUP;
3917 switch (categorize_decl_for_section (decl, reloc))
3920 case SECCAT_DATA_REL:
3921 case SECCAT_DATA_REL_LOCAL:
3922 case SECCAT_DATA_REL_RO:
3923 case SECCAT_DATA_REL_RO_LOCAL:
3924 prefix = one_only ? ".ld" : ".ldata";
3927 prefix = one_only ? ".lb" : ".lbss";
3930 case SECCAT_RODATA_MERGE_STR:
3931 case SECCAT_RODATA_MERGE_STR_INIT:
3932 case SECCAT_RODATA_MERGE_CONST:
3933 prefix = one_only ? ".lr" : ".lrodata";
3935 case SECCAT_SRODATA:
3942 /* We don't split these for medium model. Place them into
3943 default sections and hope for best. */
3945 case SECCAT_EMUTLS_VAR:
3946 prefix = targetm.emutls.var_section;
3948 case SECCAT_EMUTLS_TMPL:
3949 prefix = targetm.emutls.tmpl_section;
3954 const char *name, *linkonce;
3957 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
3958 name = targetm.strip_name_encoding (name);
3960 /* If we're using one_only, then there needs to be a .gnu.linkonce
3961 prefix to the section name. */
3962 linkonce = one_only ? ".gnu.linkonce" : "";
3964 string = ACONCAT ((linkonce, prefix, ".", name, NULL));
3966 DECL_SECTION_NAME (decl) = build_string (strlen (string), string);
3970 default_unique_section (decl, reloc);
3973 #ifdef COMMON_ASM_OP
3974 /* This says how to output assembler code to declare an
3975 uninitialized external linkage data object.
3977 For medium model x86-64 we need to use .largecomm opcode for
3980 x86_elf_aligned_common (FILE *file,
3981 const char *name, unsigned HOST_WIDE_INT size,
3984 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
3985 && size > (unsigned int)ix86_section_threshold)
3986 fprintf (file, ".largecomm\t");
3988 fprintf (file, "%s", COMMON_ASM_OP);
3989 assemble_name (file, name);
3990 fprintf (file, ","HOST_WIDE_INT_PRINT_UNSIGNED",%u\n",
3991 size, align / BITS_PER_UNIT);
3995 /* Utility function for targets to use in implementing
3996 ASM_OUTPUT_ALIGNED_BSS. */
3999 x86_output_aligned_bss (FILE *file, tree decl ATTRIBUTE_UNUSED,
4000 const char *name, unsigned HOST_WIDE_INT size,
4003 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
4004 && size > (unsigned int)ix86_section_threshold)
4005 switch_to_section (get_named_section (decl, ".lbss", 0));
4007 switch_to_section (bss_section);
4008 ASM_OUTPUT_ALIGN (file, floor_log2 (align / BITS_PER_UNIT));
4009 #ifdef ASM_DECLARE_OBJECT_NAME
4010 last_assemble_variable_decl = decl;
4011 ASM_DECLARE_OBJECT_NAME (file, name, decl);
4013 /* Standard thing is just output label for the object. */
4014 ASM_OUTPUT_LABEL (file, name);
4015 #endif /* ASM_DECLARE_OBJECT_NAME */
4016 ASM_OUTPUT_SKIP (file, size ? size : 1);
4020 optimization_options (int level, int size ATTRIBUTE_UNUSED)
4022 /* For -O2 and beyond, turn off -fschedule-insns by default. It tends to
4023 make the problem with not enough registers even worse. */
4024 #ifdef INSN_SCHEDULING
4026 flag_schedule_insns = 0;
4030 /* The Darwin libraries never set errno, so we might as well
4031 avoid calling them when that's the only reason we would. */
4032 flag_errno_math = 0;
4034 /* The default values of these switches depend on the TARGET_64BIT
4035 that is not known at this moment. Mark these values with 2 and
4036 let user the to override these. In case there is no command line option
4037 specifying them, we will set the defaults in override_options. */
4039 flag_omit_frame_pointer = 2;
4040 flag_pcc_struct_return = 2;
4041 flag_asynchronous_unwind_tables = 2;
4042 flag_vect_cost_model = 1;
4043 #ifdef SUBTARGET_OPTIMIZATION_OPTIONS
4044 SUBTARGET_OPTIMIZATION_OPTIONS;
4048 /* Decide whether we can make a sibling call to a function. DECL is the
4049 declaration of the function being targeted by the call and EXP is the
4050 CALL_EXPR representing the call. */
4053 ix86_function_ok_for_sibcall (tree decl, tree exp)
4058 /* If we are generating position-independent code, we cannot sibcall
4059 optimize any indirect call, or a direct call to a global function,
4060 as the PLT requires %ebx be live. */
4061 if (!TARGET_64BIT && flag_pic && (!decl || !targetm.binds_local_p (decl)))
4068 func = TREE_TYPE (CALL_EXPR_FN (exp));
4069 if (POINTER_TYPE_P (func))
4070 func = TREE_TYPE (func);
4073 /* Check that the return value locations are the same. Like
4074 if we are returning floats on the 80387 register stack, we cannot
4075 make a sibcall from a function that doesn't return a float to a
4076 function that does or, conversely, from a function that does return
4077 a float to a function that doesn't; the necessary stack adjustment
4078 would not be executed. This is also the place we notice
4079 differences in the return value ABI. Note that it is ok for one
4080 of the functions to have void return type as long as the return
4081 value of the other is passed in a register. */
4082 a = ix86_function_value (TREE_TYPE (exp), func, false);
4083 b = ix86_function_value (TREE_TYPE (DECL_RESULT (cfun->decl)),
4085 if (STACK_REG_P (a) || STACK_REG_P (b))
4087 if (!rtx_equal_p (a, b))
4090 else if (VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun->decl))))
4092 else if (!rtx_equal_p (a, b))
4095 /* If this call is indirect, we'll need to be able to use a call-clobbered
4096 register for the address of the target function. Make sure that all
4097 such registers are not used for passing parameters. */
4098 if (!decl && !TARGET_64BIT)
4102 /* We're looking at the CALL_EXPR, we need the type of the function. */
4103 type = CALL_EXPR_FN (exp); /* pointer expression */
4104 type = TREE_TYPE (type); /* pointer type */
4105 type = TREE_TYPE (type); /* function type */
4107 if (ix86_function_regparm (type, NULL) >= 3)
4109 /* ??? Need to count the actual number of registers to be used,
4110 not the possible number of registers. Fix later. */
4115 /* Dllimport'd functions are also called indirectly. */
4116 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
4118 && decl && DECL_DLLIMPORT_P (decl)
4119 && ix86_function_regparm (TREE_TYPE (decl), NULL) >= 3)
4122 /* If we need to align the outgoing stack, then sibcalling would
4123 unalign the stack, which may break the called function. */
4124 if (ix86_incoming_stack_boundary < PREFERRED_STACK_BOUNDARY)
4127 /* Otherwise okay. That also includes certain types of indirect calls. */
4131 /* Handle "cdecl", "stdcall", "fastcall", "regparm" and "sseregparm"
4132 calling convention attributes;
4133 arguments as in struct attribute_spec.handler. */
4136 ix86_handle_cconv_attribute (tree *node, tree name,
4138 int flags ATTRIBUTE_UNUSED,
4141 if (TREE_CODE (*node) != FUNCTION_TYPE
4142 && TREE_CODE (*node) != METHOD_TYPE
4143 && TREE_CODE (*node) != FIELD_DECL
4144 && TREE_CODE (*node) != TYPE_DECL)
4146 warning (OPT_Wattributes, "%qs attribute only applies to functions",
4147 IDENTIFIER_POINTER (name));
4148 *no_add_attrs = true;
4152 /* Can combine regparm with all attributes but fastcall. */
4153 if (is_attribute_p ("regparm", name))
4157 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4159 error ("fastcall and regparm attributes are not compatible");
4162 cst = TREE_VALUE (args);
4163 if (TREE_CODE (cst) != INTEGER_CST)
4165 warning (OPT_Wattributes,
4166 "%qs attribute requires an integer constant argument",
4167 IDENTIFIER_POINTER (name));
4168 *no_add_attrs = true;
4170 else if (compare_tree_int (cst, REGPARM_MAX) > 0)
4172 warning (OPT_Wattributes, "argument to %qs attribute larger than %d",
4173 IDENTIFIER_POINTER (name), REGPARM_MAX);
4174 *no_add_attrs = true;
4182 /* Do not warn when emulating the MS ABI. */
4183 if (TREE_CODE (*node) != FUNCTION_TYPE || ix86_function_type_abi (*node)!=MS_ABI)
4184 warning (OPT_Wattributes, "%qs attribute ignored",
4185 IDENTIFIER_POINTER (name));
4186 *no_add_attrs = true;
4190 /* Can combine fastcall with stdcall (redundant) and sseregparm. */
4191 if (is_attribute_p ("fastcall", name))
4193 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
4195 error ("fastcall and cdecl attributes are not compatible");
4197 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
4199 error ("fastcall and stdcall attributes are not compatible");
4201 if (lookup_attribute ("regparm", TYPE_ATTRIBUTES (*node)))
4203 error ("fastcall and regparm attributes are not compatible");
4207 /* Can combine stdcall with fastcall (redundant), regparm and
4209 else if (is_attribute_p ("stdcall", name))
4211 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
4213 error ("stdcall and cdecl attributes are not compatible");
4215 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4217 error ("stdcall and fastcall attributes are not compatible");
4221 /* Can combine cdecl with regparm and sseregparm. */
4222 else if (is_attribute_p ("cdecl", name))
4224 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
4226 error ("stdcall and cdecl attributes are not compatible");
4228 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
4230 error ("fastcall and cdecl attributes are not compatible");
4234 /* Can combine sseregparm with all attributes. */
4239 /* Return 0 if the attributes for two types are incompatible, 1 if they
4240 are compatible, and 2 if they are nearly compatible (which causes a
4241 warning to be generated). */
4244 ix86_comp_type_attributes (const_tree type1, const_tree type2)
4246 /* Check for mismatch of non-default calling convention. */
4247 const char *const rtdstr = TARGET_RTD ? "cdecl" : "stdcall";
4249 if (TREE_CODE (type1) != FUNCTION_TYPE
4250 && TREE_CODE (type1) != METHOD_TYPE)
4253 /* Check for mismatched fastcall/regparm types. */
4254 if ((!lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type1))
4255 != !lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type2)))
4256 || (ix86_function_regparm (type1, NULL)
4257 != ix86_function_regparm (type2, NULL)))
4260 /* Check for mismatched sseregparm types. */
4261 if (!lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type1))
4262 != !lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type2)))
4265 /* Check for mismatched return types (cdecl vs stdcall). */
4266 if (!lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type1))
4267 != !lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type2)))
4273 /* Return the regparm value for a function with the indicated TYPE and DECL.
4274 DECL may be NULL when calling function indirectly
4275 or considering a libcall. */
4278 ix86_function_regparm (const_tree type, const_tree decl)
4281 int regparm = ix86_regparm;
4283 static bool error_issued;
4287 if (ix86_function_type_abi (type) == DEFAULT_ABI)
4289 return DEFAULT_ABI != SYSV_ABI ? X86_64_REGPARM_MAX : X64_REGPARM_MAX;
4292 attr = lookup_attribute ("regparm", TYPE_ATTRIBUTES (type));
4296 = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attr)));
4298 if (decl && TREE_CODE (decl) == FUNCTION_DECL)
4300 /* We can't use regparm(3) for nested functions because
4301 these pass static chain pointer in %ecx register. */
4302 if (!error_issued && regparm == 3
4303 && decl_function_context (decl)
4304 && !DECL_NO_STATIC_CHAIN (decl))
4306 error ("nested functions are limited to 2 register parameters");
4307 error_issued = true;
4315 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type)))
4318 /* Use register calling convention for local functions when possible. */
4319 if (decl && TREE_CODE (decl) == FUNCTION_DECL
4322 /* FIXME: remove this CONST_CAST when cgraph.[ch] is constified. */
4323 struct cgraph_local_info *i = cgraph_local_info (CONST_CAST_TREE(decl));
4326 int local_regparm, globals = 0, regno;
4329 /* Make sure no regparm register is taken by a
4330 fixed register variable. */
4331 for (local_regparm = 0; local_regparm < REGPARM_MAX; local_regparm++)
4332 if (fixed_regs[local_regparm])
4335 /* We can't use regparm(3) for nested functions as these use
4336 static chain pointer in third argument. */
4337 if (local_regparm == 3
4338 && decl_function_context (decl)
4339 && !DECL_NO_STATIC_CHAIN (decl))
4342 /* If the function realigns its stackpointer, the prologue will
4343 clobber %ecx. If we've already generated code for the callee,
4344 the callee DECL_STRUCT_FUNCTION is gone, so we fall back to
4345 scanning the attributes for the self-realigning property. */
4346 f = DECL_STRUCT_FUNCTION (decl);
4347 /* Since current internal arg pointer won't conflict with
4348 parameter passing regs, so no need to change stack
4349 realignment and adjust regparm number.
4351 Each fixed register usage increases register pressure,
4352 so less registers should be used for argument passing.
4353 This functionality can be overriden by an explicit
4355 for (regno = 0; regno <= DI_REG; regno++)
4356 if (fixed_regs[regno])
4360 = globals < local_regparm ? local_regparm - globals : 0;
4362 if (local_regparm > regparm)
4363 regparm = local_regparm;
4370 /* Return 1 or 2, if we can pass up to SSE_REGPARM_MAX SFmode (1) and
4371 DFmode (2) arguments in SSE registers for a function with the
4372 indicated TYPE and DECL. DECL may be NULL when calling function
4373 indirectly or considering a libcall. Otherwise return 0. */
4376 ix86_function_sseregparm (const_tree type, const_tree decl, bool warn)
4378 gcc_assert (!TARGET_64BIT);
4380 /* Use SSE registers to pass SFmode and DFmode arguments if requested
4381 by the sseregparm attribute. */
4382 if (TARGET_SSEREGPARM
4383 || (type && lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type))))
4390 error ("Calling %qD with attribute sseregparm without "
4391 "SSE/SSE2 enabled", decl);
4393 error ("Calling %qT with attribute sseregparm without "
4394 "SSE/SSE2 enabled", type);
4402 /* For local functions, pass up to SSE_REGPARM_MAX SFmode
4403 (and DFmode for SSE2) arguments in SSE registers. */
4404 if (decl && TARGET_SSE_MATH && !profile_flag)
4406 /* FIXME: remove this CONST_CAST when cgraph.[ch] is constified. */
4407 struct cgraph_local_info *i = cgraph_local_info (CONST_CAST_TREE(decl));
4409 return TARGET_SSE2 ? 2 : 1;
4415 /* Return true if EAX is live at the start of the function. Used by
4416 ix86_expand_prologue to determine if we need special help before
4417 calling allocate_stack_worker. */
4420 ix86_eax_live_at_start_p (void)
4422 /* Cheat. Don't bother working forward from ix86_function_regparm
4423 to the function type to whether an actual argument is located in
4424 eax. Instead just look at cfg info, which is still close enough
4425 to correct at this point. This gives false positives for broken
4426 functions that might use uninitialized data that happens to be
4427 allocated in eax, but who cares? */
4428 return REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR), 0);
4431 /* Value is the number of bytes of arguments automatically
4432 popped when returning from a subroutine call.
4433 FUNDECL is the declaration node of the function (as a tree),
4434 FUNTYPE is the data type of the function (as a tree),
4435 or for a library call it is an identifier node for the subroutine name.
4436 SIZE is the number of bytes of arguments passed on the stack.
4438 On the 80386, the RTD insn may be used to pop them if the number
4439 of args is fixed, but if the number is variable then the caller
4440 must pop them all. RTD can't be used for library calls now
4441 because the library is compiled with the Unix compiler.
4442 Use of RTD is a selectable option, since it is incompatible with
4443 standard Unix calling sequences. If the option is not selected,
4444 the caller must always pop the args.
4446 The attribute stdcall is equivalent to RTD on a per module basis. */
4449 ix86_return_pops_args (tree fundecl, tree funtype, int size)
4453 /* None of the 64-bit ABIs pop arguments. */
4457 rtd = TARGET_RTD && (!fundecl || TREE_CODE (fundecl) != IDENTIFIER_NODE);
4459 /* Cdecl functions override -mrtd, and never pop the stack. */
4460 if (! lookup_attribute ("cdecl", TYPE_ATTRIBUTES (funtype)))
4462 /* Stdcall and fastcall functions will pop the stack if not
4464 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (funtype))
4465 || lookup_attribute ("fastcall", TYPE_ATTRIBUTES (funtype)))
4468 if (rtd && ! stdarg_p (funtype))
4472 /* Lose any fake structure return argument if it is passed on the stack. */
4473 if (aggregate_value_p (TREE_TYPE (funtype), fundecl)
4474 && !KEEP_AGGREGATE_RETURN_POINTER)
4476 int nregs = ix86_function_regparm (funtype, fundecl);
4478 return GET_MODE_SIZE (Pmode);
4484 /* Argument support functions. */
4486 /* Return true when register may be used to pass function parameters. */
4488 ix86_function_arg_regno_p (int regno)
4491 const int *parm_regs;
4496 return (regno < REGPARM_MAX
4497 || (TARGET_SSE && SSE_REGNO_P (regno) && !fixed_regs[regno]));
4499 return (regno < REGPARM_MAX
4500 || (TARGET_MMX && MMX_REGNO_P (regno)
4501 && (regno < FIRST_MMX_REG + MMX_REGPARM_MAX))
4502 || (TARGET_SSE && SSE_REGNO_P (regno)
4503 && (regno < FIRST_SSE_REG + SSE_REGPARM_MAX)));
4508 if (SSE_REGNO_P (regno) && TARGET_SSE)
4513 if (TARGET_SSE && SSE_REGNO_P (regno)
4514 && (regno < FIRST_SSE_REG + SSE_REGPARM_MAX))
4518 /* TODO: The function should depend on current function ABI but
4519 builtins.c would need updating then. Therefore we use the
4522 /* RAX is used as hidden argument to va_arg functions. */
4523 if (DEFAULT_ABI == SYSV_ABI && regno == AX_REG)
4526 if (DEFAULT_ABI == MS_ABI)
4527 parm_regs = x86_64_ms_abi_int_parameter_registers;
4529 parm_regs = x86_64_int_parameter_registers;
4530 for (i = 0; i < (DEFAULT_ABI == MS_ABI ? X64_REGPARM_MAX
4531 : X86_64_REGPARM_MAX); i++)
4532 if (regno == parm_regs[i])
4537 /* Return if we do not know how to pass TYPE solely in registers. */
4540 ix86_must_pass_in_stack (enum machine_mode mode, const_tree type)
4542 if (must_pass_in_stack_var_size_or_pad (mode, type))
4545 /* For 32-bit, we want TImode aggregates to go on the stack. But watch out!
4546 The layout_type routine is crafty and tries to trick us into passing
4547 currently unsupported vector types on the stack by using TImode. */
4548 return (!TARGET_64BIT && mode == TImode
4549 && type && TREE_CODE (type) != VECTOR_TYPE);
4552 /* It returns the size, in bytes, of the area reserved for arguments passed
4553 in registers for the function represented by fndecl dependent to the used
4556 ix86_reg_parm_stack_space (const_tree fndecl)
4558 int call_abi = SYSV_ABI;
4559 if (fndecl != NULL_TREE && TREE_CODE (fndecl) == FUNCTION_DECL)
4560 call_abi = ix86_function_abi (fndecl);
4562 call_abi = ix86_function_type_abi (fndecl);
4563 if (call_abi == MS_ABI)
4568 /* Returns value SYSV_ABI, MS_ABI dependent on fntype, specifying the
4571 ix86_function_type_abi (const_tree fntype)
4573 if (TARGET_64BIT && fntype != NULL)
4576 if (DEFAULT_ABI == SYSV_ABI)
4577 abi = lookup_attribute ("ms_abi", TYPE_ATTRIBUTES (fntype)) ? MS_ABI : SYSV_ABI;
4579 abi = lookup_attribute ("sysv_abi", TYPE_ATTRIBUTES (fntype)) ? SYSV_ABI : MS_ABI;
4587 ix86_function_abi (const_tree fndecl)
4591 return ix86_function_type_abi (TREE_TYPE (fndecl));
4594 /* Returns value SYSV_ABI, MS_ABI dependent on cfun, specifying the
4597 ix86_cfun_abi (void)
4599 if (! cfun || ! TARGET_64BIT)
4601 return cfun->machine->call_abi;
4605 extern void init_regs (void);
4607 /* Implementation of call abi switching target hook. Specific to FNDECL
4608 the specific call register sets are set. See also CONDITIONAL_REGISTER_USAGE
4609 for more details. */
4611 ix86_call_abi_override (const_tree fndecl)
4613 if (fndecl == NULL_TREE)
4614 cfun->machine->call_abi = DEFAULT_ABI;
4616 cfun->machine->call_abi = ix86_function_type_abi (TREE_TYPE (fndecl));
4619 /* MS and SYSV ABI have different set of call used registers. Avoid expensive
4620 re-initialization of init_regs each time we switch function context since
4621 this is needed only during RTL expansion. */
4623 ix86_maybe_switch_abi (void)
4626 call_used_regs[4 /*RSI*/] == (cfun->machine->call_abi == MS_ABI))
4630 /* Initialize a variable CUM of type CUMULATIVE_ARGS
4631 for a call to a function whose data type is FNTYPE.
4632 For a library call, FNTYPE is 0. */
4635 init_cumulative_args (CUMULATIVE_ARGS *cum, /* Argument info to initialize */
4636 tree fntype, /* tree ptr for function decl */
4637 rtx libname, /* SYMBOL_REF of library name or 0 */
4640 struct cgraph_local_info *i = fndecl ? cgraph_local_info (fndecl) : NULL;
4641 memset (cum, 0, sizeof (*cum));
4644 cum->call_abi = ix86_function_abi (fndecl);
4646 cum->call_abi = ix86_function_type_abi (fntype);
4647 /* Set up the number of registers to use for passing arguments. */
4649 if (cum->call_abi == MS_ABI && !ACCUMULATE_OUTGOING_ARGS)
4650 sorry ("ms_abi attribute require -maccumulate-outgoing-args or subtarget optimization implying it");
4651 cum->nregs = ix86_regparm;
4654 if (cum->call_abi != DEFAULT_ABI)
4655 cum->nregs = DEFAULT_ABI != SYSV_ABI ? X86_64_REGPARM_MAX
4660 cum->sse_nregs = SSE_REGPARM_MAX;
4663 if (cum->call_abi != DEFAULT_ABI)
4664 cum->sse_nregs = DEFAULT_ABI != SYSV_ABI ? X86_64_SSE_REGPARM_MAX
4665 : X64_SSE_REGPARM_MAX;
4669 cum->mmx_nregs = MMX_REGPARM_MAX;
4670 cum->warn_avx = true;
4671 cum->warn_sse = true;
4672 cum->warn_mmx = true;
4674 /* Because type might mismatch in between caller and callee, we need to
4675 use actual type of function for local calls.
4676 FIXME: cgraph_analyze can be told to actually record if function uses
4677 va_start so for local functions maybe_vaarg can be made aggressive
4679 FIXME: once typesytem is fixed, we won't need this code anymore. */
4681 fntype = TREE_TYPE (fndecl);
4682 cum->maybe_vaarg = (fntype
4683 ? (!prototype_p (fntype) || stdarg_p (fntype))
4688 /* If there are variable arguments, then we won't pass anything
4689 in registers in 32-bit mode. */
4690 if (stdarg_p (fntype))
4701 /* Use ecx and edx registers if function has fastcall attribute,
4702 else look for regparm information. */
4705 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (fntype)))
4711 cum->nregs = ix86_function_regparm (fntype, fndecl);
4714 /* Set up the number of SSE registers used for passing SFmode
4715 and DFmode arguments. Warn for mismatching ABI. */
4716 cum->float_in_sse = ix86_function_sseregparm (fntype, fndecl, true);
4720 /* Return the "natural" mode for TYPE. In most cases, this is just TYPE_MODE.
4721 But in the case of vector types, it is some vector mode.
4723 When we have only some of our vector isa extensions enabled, then there
4724 are some modes for which vector_mode_supported_p is false. For these
4725 modes, the generic vector support in gcc will choose some non-vector mode
4726 in order to implement the type. By computing the natural mode, we'll
4727 select the proper ABI location for the operand and not depend on whatever
4728 the middle-end decides to do with these vector types. */
4730 static enum machine_mode
4731 type_natural_mode (const_tree type)
4733 enum machine_mode mode = TYPE_MODE (type);
4735 if (TREE_CODE (type) == VECTOR_TYPE && !VECTOR_MODE_P (mode))
4737 HOST_WIDE_INT size = int_size_in_bytes (type);
4738 if ((size == 8 || size == 16)
4739 /* ??? Generic code allows us to create width 1 vectors. Ignore. */
4740 && TYPE_VECTOR_SUBPARTS (type) > 1)
4742 enum machine_mode innermode = TYPE_MODE (TREE_TYPE (type));
4744 if (TREE_CODE (TREE_TYPE (type)) == REAL_TYPE)
4745 mode = MIN_MODE_VECTOR_FLOAT;
4747 mode = MIN_MODE_VECTOR_INT;
4749 /* Get the mode which has this inner mode and number of units. */
4750 for (; mode != VOIDmode; mode = GET_MODE_WIDER_MODE (mode))
4751 if (GET_MODE_NUNITS (mode) == TYPE_VECTOR_SUBPARTS (type)
4752 && GET_MODE_INNER (mode) == innermode)
4762 /* We want to pass a value in REGNO whose "natural" mode is MODE. However,
4763 this may not agree with the mode that the type system has chosen for the
4764 register, which is ORIG_MODE. If ORIG_MODE is not BLKmode, then we can
4765 go ahead and use it. Otherwise we have to build a PARALLEL instead. */
4768 gen_reg_or_parallel (enum machine_mode mode, enum machine_mode orig_mode,
4773 if (orig_mode != BLKmode)
4774 tmp = gen_rtx_REG (orig_mode, regno);
4777 tmp = gen_rtx_REG (mode, regno);
4778 tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, const0_rtx);
4779 tmp = gen_rtx_PARALLEL (orig_mode, gen_rtvec (1, tmp));
4785 /* x86-64 register passing implementation. See x86-64 ABI for details. Goal
4786 of this code is to classify each 8bytes of incoming argument by the register
4787 class and assign registers accordingly. */
4789 /* Return the union class of CLASS1 and CLASS2.
4790 See the x86-64 PS ABI for details. */
4792 static enum x86_64_reg_class
4793 merge_classes (enum x86_64_reg_class class1, enum x86_64_reg_class class2)
4795 /* Rule #1: If both classes are equal, this is the resulting class. */
4796 if (class1 == class2)
4799 /* Rule #2: If one of the classes is NO_CLASS, the resulting class is
4801 if (class1 == X86_64_NO_CLASS)
4803 if (class2 == X86_64_NO_CLASS)
4806 /* Rule #3: If one of the classes is MEMORY, the result is MEMORY. */
4807 if (class1 == X86_64_MEMORY_CLASS || class2 == X86_64_MEMORY_CLASS)
4808 return X86_64_MEMORY_CLASS;
4810 /* Rule #4: If one of the classes is INTEGER, the result is INTEGER. */
4811 if ((class1 == X86_64_INTEGERSI_CLASS && class2 == X86_64_SSESF_CLASS)
4812 || (class2 == X86_64_INTEGERSI_CLASS && class1 == X86_64_SSESF_CLASS))
4813 return X86_64_INTEGERSI_CLASS;
4814 if (class1 == X86_64_INTEGER_CLASS || class1 == X86_64_INTEGERSI_CLASS
4815 || class2 == X86_64_INTEGER_CLASS || class2 == X86_64_INTEGERSI_CLASS)
4816 return X86_64_INTEGER_CLASS;
4818 /* Rule #5: If one of the classes is X87, X87UP, or COMPLEX_X87 class,
4820 if (class1 == X86_64_X87_CLASS
4821 || class1 == X86_64_X87UP_CLASS
4822 || class1 == X86_64_COMPLEX_X87_CLASS
4823 || class2 == X86_64_X87_CLASS
4824 || class2 == X86_64_X87UP_CLASS
4825 || class2 == X86_64_COMPLEX_X87_CLASS)
4826 return X86_64_MEMORY_CLASS;
4828 /* Rule #6: Otherwise class SSE is used. */
4829 return X86_64_SSE_CLASS;
4832 /* Classify the argument of type TYPE and mode MODE.
4833 CLASSES will be filled by the register class used to pass each word
4834 of the operand. The number of words is returned. In case the parameter
4835 should be passed in memory, 0 is returned. As a special case for zero
4836 sized containers, classes[0] will be NO_CLASS and 1 is returned.
4838 BIT_OFFSET is used internally for handling records and specifies offset
4839 of the offset in bits modulo 256 to avoid overflow cases.
4841 See the x86-64 PS ABI for details.
4845 classify_argument (enum machine_mode mode, const_tree type,
4846 enum x86_64_reg_class classes[MAX_CLASSES], int bit_offset)
4848 HOST_WIDE_INT bytes =
4849 (mode == BLKmode) ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode);
4850 int words = (bytes + (bit_offset % 64) / 8 + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
4852 /* Variable sized entities are always passed/returned in memory. */
4856 if (mode != VOIDmode
4857 && targetm.calls.must_pass_in_stack (mode, type))
4860 if (type && AGGREGATE_TYPE_P (type))
4864 enum x86_64_reg_class subclasses[MAX_CLASSES];
4866 /* On x86-64 we pass structures larger than 16 bytes on the stack. */
4870 for (i = 0; i < words; i++)
4871 classes[i] = X86_64_NO_CLASS;
4873 /* Zero sized arrays or structures are NO_CLASS. We return 0 to
4874 signalize memory class, so handle it as special case. */
4877 classes[0] = X86_64_NO_CLASS;
4881 /* Classify each field of record and merge classes. */
4882 switch (TREE_CODE (type))
4885 /* And now merge the fields of structure. */
4886 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4888 if (TREE_CODE (field) == FIELD_DECL)
4892 if (TREE_TYPE (field) == error_mark_node)
4895 /* Bitfields are always classified as integer. Handle them
4896 early, since later code would consider them to be
4897 misaligned integers. */
4898 if (DECL_BIT_FIELD (field))
4900 for (i = (int_bit_position (field) + (bit_offset % 64)) / 8 / 8;
4901 i < ((int_bit_position (field) + (bit_offset % 64))
4902 + tree_low_cst (DECL_SIZE (field), 0)
4905 merge_classes (X86_64_INTEGER_CLASS,
4910 num = classify_argument (TYPE_MODE (TREE_TYPE (field)),
4911 TREE_TYPE (field), subclasses,
4912 (int_bit_position (field)
4913 + bit_offset) % 256);
4916 for (i = 0; i < num; i++)
4919 (int_bit_position (field) + (bit_offset % 64)) / 8 / 8;
4921 merge_classes (subclasses[i], classes[i + pos]);
4929 /* Arrays are handled as small records. */
4932 num = classify_argument (TYPE_MODE (TREE_TYPE (type)),
4933 TREE_TYPE (type), subclasses, bit_offset);
4937 /* The partial classes are now full classes. */
4938 if (subclasses[0] == X86_64_SSESF_CLASS && bytes != 4)
4939 subclasses[0] = X86_64_SSE_CLASS;
4940 if (subclasses[0] == X86_64_INTEGERSI_CLASS
4941 && !((bit_offset % 64) == 0 && bytes == 4))
4942 subclasses[0] = X86_64_INTEGER_CLASS;
4944 for (i = 0; i < words; i++)
4945 classes[i] = subclasses[i % num];
4950 case QUAL_UNION_TYPE:
4951 /* Unions are similar to RECORD_TYPE but offset is always 0.
4953 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
4955 if (TREE_CODE (field) == FIELD_DECL)
4959 if (TREE_TYPE (field) == error_mark_node)
4962 num = classify_argument (TYPE_MODE (TREE_TYPE (field)),
4963 TREE_TYPE (field), subclasses,
4967 for (i = 0; i < num; i++)
4968 classes[i] = merge_classes (subclasses[i], classes[i]);
4977 /* Final merger cleanup. */
4978 for (i = 0; i < words; i++)
4980 /* If one class is MEMORY, everything should be passed in
4982 if (classes[i] == X86_64_MEMORY_CLASS)
4985 /* The X86_64_SSEUP_CLASS should be always preceded by
4986 X86_64_SSE_CLASS. */
4987 if (classes[i] == X86_64_SSEUP_CLASS
4988 && (i == 0 || classes[i - 1] != X86_64_SSE_CLASS))
4989 classes[i] = X86_64_SSE_CLASS;
4991 /* X86_64_X87UP_CLASS should be preceded by X86_64_X87_CLASS. */
4992 if (classes[i] == X86_64_X87UP_CLASS
4993 && (i == 0 || classes[i - 1] != X86_64_X87_CLASS))
4994 classes[i] = X86_64_SSE_CLASS;
4999 /* Compute alignment needed. We align all types to natural boundaries with
5000 exception of XFmode that is aligned to 64bits. */
5001 if (mode != VOIDmode && mode != BLKmode)
5003 int mode_alignment = GET_MODE_BITSIZE (mode);
5006 mode_alignment = 128;
5007 else if (mode == XCmode)
5008 mode_alignment = 256;
5009 if (COMPLEX_MODE_P (mode))
5010 mode_alignment /= 2;
5011 /* Misaligned fields are always returned in memory. */
5012 if (bit_offset % mode_alignment)
5016 /* for V1xx modes, just use the base mode */
5017 if (VECTOR_MODE_P (mode) && mode != V1DImode
5018 && GET_MODE_SIZE (GET_MODE_INNER (mode)) == bytes)
5019 mode = GET_MODE_INNER (mode);
5021 /* Classification of atomic types. */
5026 classes[0] = X86_64_SSE_CLASS;
5029 classes[0] = X86_64_SSE_CLASS;
5030 classes[1] = X86_64_SSEUP_CLASS;
5040 int size = (bit_offset % 64)+ (int) GET_MODE_BITSIZE (mode);
5044 classes[0] = X86_64_INTEGERSI_CLASS;
5047 else if (size <= 64)
5049 classes[0] = X86_64_INTEGER_CLASS;
5052 else if (size <= 64+32)
5054 classes[0] = X86_64_INTEGER_CLASS;
5055 classes[1] = X86_64_INTEGERSI_CLASS;
5058 else if (size <= 64+64)
5060 classes[0] = classes[1] = X86_64_INTEGER_CLASS;
5068 classes[0] = classes[1] = X86_64_INTEGER_CLASS;
5075 if (!(bit_offset % 64))
5076 classes[0] = X86_64_SSESF_CLASS;
5078 classes[0] = X86_64_SSE_CLASS;
5081 classes[0] = X86_64_SSEDF_CLASS;
5084 classes[0] = X86_64_X87_CLASS;
5085 classes[1] = X86_64_X87UP_CLASS;
5088 classes[0] = X86_64_SSE_CLASS;
5089 classes[1] = X86_64_SSEUP_CLASS;
5092 classes[0] = X86_64_SSE_CLASS;
5095 classes[0] = X86_64_SSEDF_CLASS;
5096 classes[1] = X86_64_SSEDF_CLASS;
5099 classes[0] = X86_64_COMPLEX_X87_CLASS;
5102 /* This modes is larger than 16 bytes. */
5110 classes[0] = X86_64_AVX_CLASS;
5118 classes[0] = X86_64_SSE_CLASS;
5119 classes[1] = X86_64_SSEUP_CLASS;
5126 classes[0] = X86_64_SSE_CLASS;
5132 gcc_assert (VECTOR_MODE_P (mode));
5137 gcc_assert (GET_MODE_CLASS (GET_MODE_INNER (mode)) == MODE_INT);
5139 if (bit_offset + GET_MODE_BITSIZE (mode) <= 32)
5140 classes[0] = X86_64_INTEGERSI_CLASS;
5142 classes[0] = X86_64_INTEGER_CLASS;
5143 classes[1] = X86_64_INTEGER_CLASS;
5144 return 1 + (bytes > 8);
5148 /* Examine the argument and return set number of register required in each
5149 class. Return 0 iff parameter should be passed in memory. */
5151 examine_argument (enum machine_mode mode, const_tree type, int in_return,
5152 int *int_nregs, int *sse_nregs)
5154 enum x86_64_reg_class regclass[MAX_CLASSES];
5155 int n = classify_argument (mode, type, regclass, 0);
5161 for (n--; n >= 0; n--)
5162 switch (regclass[n])
5164 case X86_64_INTEGER_CLASS:
5165 case X86_64_INTEGERSI_CLASS:
5168 case X86_64_AVX_CLASS:
5169 case X86_64_SSE_CLASS:
5170 case X86_64_SSESF_CLASS:
5171 case X86_64_SSEDF_CLASS:
5174 case X86_64_NO_CLASS:
5175 case X86_64_SSEUP_CLASS:
5177 case X86_64_X87_CLASS:
5178 case X86_64_X87UP_CLASS:
5182 case X86_64_COMPLEX_X87_CLASS:
5183 return in_return ? 2 : 0;
5184 case X86_64_MEMORY_CLASS:
5190 /* Construct container for the argument used by GCC interface. See
5191 FUNCTION_ARG for the detailed description. */
5194 construct_container (enum machine_mode mode, enum machine_mode orig_mode,
5195 const_tree type, int in_return, int nintregs, int nsseregs,
5196 const int *intreg, int sse_regno)
5198 /* The following variables hold the static issued_error state. */
5199 static bool issued_sse_arg_error;
5200 static bool issued_sse_ret_error;
5201 static bool issued_x87_ret_error;
5203 enum machine_mode tmpmode;
5205 (mode == BLKmode) ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode);
5206 enum x86_64_reg_class regclass[MAX_CLASSES];
5210 int needed_sseregs, needed_intregs;
5211 rtx exp[MAX_CLASSES];
5214 n = classify_argument (mode, type, regclass, 0);
5217 if (!examine_argument (mode, type, in_return, &needed_intregs,
5220 if (needed_intregs > nintregs || needed_sseregs > nsseregs)
5223 /* We allowed the user to turn off SSE for kernel mode. Don't crash if
5224 some less clueful developer tries to use floating-point anyway. */
5225 if (needed_sseregs && !TARGET_SSE)
5229 if (!issued_sse_ret_error)
5231 error ("SSE register return with SSE disabled");
5232 issued_sse_ret_error = true;
5235 else if (!issued_sse_arg_error)
5237 error ("SSE register argument with SSE disabled");
5238 issued_sse_arg_error = true;
5243 /* Likewise, error if the ABI requires us to return values in the
5244 x87 registers and the user specified -mno-80387. */
5245 if (!TARGET_80387 && in_return)
5246 for (i = 0; i < n; i++)
5247 if (regclass[i] == X86_64_X87_CLASS
5248 || regclass[i] == X86_64_X87UP_CLASS
5249 || regclass[i] == X86_64_COMPLEX_X87_CLASS)
5251 if (!issued_x87_ret_error)
5253 error ("x87 register return with x87 disabled");
5254 issued_x87_ret_error = true;
5259 /* First construct simple cases. Avoid SCmode, since we want to use
5260 single register to pass this type. */
5261 if (n == 1 && mode != SCmode)
5262 switch (regclass[0])
5264 case X86_64_INTEGER_CLASS:
5265 case X86_64_INTEGERSI_CLASS:
5266 return gen_rtx_REG (mode, intreg[0]);
5267 case X86_64_AVX_CLASS:
5268 case X86_64_SSE_CLASS:
5269 case X86_64_SSESF_CLASS:
5270 case X86_64_SSEDF_CLASS:
5271 return gen_reg_or_parallel (mode, orig_mode, SSE_REGNO (sse_regno));
5272 case X86_64_X87_CLASS:
5273 case X86_64_COMPLEX_X87_CLASS:
5274 return gen_rtx_REG (mode, FIRST_STACK_REG);
5275 case X86_64_NO_CLASS:
5276 /* Zero sized array, struct or class. */
5281 if (n == 2 && regclass[0] == X86_64_SSE_CLASS
5282 && regclass[1] == X86_64_SSEUP_CLASS && mode != BLKmode)
5283 return gen_rtx_REG (mode, SSE_REGNO (sse_regno));
5286 && regclass[0] == X86_64_X87_CLASS && regclass[1] == X86_64_X87UP_CLASS)
5287 return gen_rtx_REG (XFmode, FIRST_STACK_REG);
5288 if (n == 2 && regclass[0] == X86_64_INTEGER_CLASS
5289 && regclass[1] == X86_64_INTEGER_CLASS
5290 && (mode == CDImode || mode == TImode || mode == TFmode)
5291 && intreg[0] + 1 == intreg[1])
5292 return gen_rtx_REG (mode, intreg[0]);
5294 /* Otherwise figure out the entries of the PARALLEL. */
5295 for (i = 0; i < n; i++)
5297 switch (regclass[i])
5299 case X86_64_NO_CLASS:
5301 case X86_64_INTEGER_CLASS:
5302 case X86_64_INTEGERSI_CLASS:
5303 /* Merge TImodes on aligned occasions here too. */
5304 if (i * 8 + 8 > bytes)
5305 tmpmode = mode_for_size ((bytes - i * 8) * BITS_PER_UNIT, MODE_INT, 0);
5306 else if (regclass[i] == X86_64_INTEGERSI_CLASS)
5310 /* We've requested 24 bytes we don't have mode for. Use DImode. */
5311 if (tmpmode == BLKmode)
5313 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5314 gen_rtx_REG (tmpmode, *intreg),
5318 case X86_64_SSESF_CLASS:
5319 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5320 gen_rtx_REG (SFmode,
5321 SSE_REGNO (sse_regno)),
5325 case X86_64_SSEDF_CLASS:
5326 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5327 gen_rtx_REG (DFmode,
5328 SSE_REGNO (sse_regno)),
5332 case X86_64_SSE_CLASS:
5333 if (i < n - 1 && regclass[i + 1] == X86_64_SSEUP_CLASS)
5337 exp [nexps++] = gen_rtx_EXPR_LIST (VOIDmode,
5338 gen_rtx_REG (tmpmode,
5339 SSE_REGNO (sse_regno)),
5341 if (tmpmode == TImode)
5350 /* Empty aligned struct, union or class. */
5354 ret = gen_rtx_PARALLEL (mode, rtvec_alloc (nexps));
5355 for (i = 0; i < nexps; i++)
5356 XVECEXP (ret, 0, i) = exp [i];
5360 /* Update the data in CUM to advance over an argument of mode MODE
5361 and data type TYPE. (TYPE is null for libcalls where that information
5362 may not be available.) */
5365 function_arg_advance_32 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5366 tree type, HOST_WIDE_INT bytes, HOST_WIDE_INT words)
5382 cum->words += words;
5383 cum->nregs -= words;
5384 cum->regno += words;
5386 if (cum->nregs <= 0)
5394 if (cum->float_in_sse < 2)
5397 if (cum->float_in_sse < 1)
5415 if (!type || !AGGREGATE_TYPE_P (type))
5417 cum->sse_words += words;
5418 cum->sse_nregs -= 1;
5419 cum->sse_regno += 1;
5420 if (cum->sse_nregs <= 0)
5433 if (!type || !AGGREGATE_TYPE_P (type))
5435 cum->mmx_words += words;
5436 cum->mmx_nregs -= 1;
5437 cum->mmx_regno += 1;
5438 if (cum->mmx_nregs <= 0)
5449 function_arg_advance_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5450 tree type, HOST_WIDE_INT words, int named)
5452 int int_nregs, sse_nregs;
5454 /* Unnamed 256bit vector mode parameters are passed on stack. */
5455 if (!named && VALID_AVX256_REG_MODE (mode))
5458 if (!examine_argument (mode, type, 0, &int_nregs, &sse_nregs))
5459 cum->words += words;
5460 else if (sse_nregs <= cum->sse_nregs && int_nregs <= cum->nregs)
5462 cum->nregs -= int_nregs;
5463 cum->sse_nregs -= sse_nregs;
5464 cum->regno += int_nregs;
5465 cum->sse_regno += sse_nregs;
5468 cum->words += words;
5472 function_arg_advance_ms_64 (CUMULATIVE_ARGS *cum, HOST_WIDE_INT bytes,
5473 HOST_WIDE_INT words)
5475 /* Otherwise, this should be passed indirect. */
5476 gcc_assert (bytes == 1 || bytes == 2 || bytes == 4 || bytes == 8);
5478 cum->words += words;
5487 function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5488 tree type, int named)
5490 HOST_WIDE_INT bytes, words;
5492 if (mode == BLKmode)
5493 bytes = int_size_in_bytes (type);
5495 bytes = GET_MODE_SIZE (mode);
5496 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5499 mode = type_natural_mode (type);
5501 if (TARGET_64BIT && (cum ? cum->call_abi : DEFAULT_ABI) == MS_ABI)
5502 function_arg_advance_ms_64 (cum, bytes, words);
5503 else if (TARGET_64BIT)
5504 function_arg_advance_64 (cum, mode, type, words, named);
5506 function_arg_advance_32 (cum, mode, type, bytes, words);
5509 /* Define where to put the arguments to a function.
5510 Value is zero to push the argument on the stack,
5511 or a hard register in which to store the argument.
5513 MODE is the argument's machine mode.
5514 TYPE is the data type of the argument (as a tree).
5515 This is null for libcalls where that information may
5517 CUM is a variable of type CUMULATIVE_ARGS which gives info about
5518 the preceding args and about the function being called.
5519 NAMED is nonzero if this argument is a named parameter
5520 (otherwise it is an extra parameter matching an ellipsis). */
5523 function_arg_32 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5524 enum machine_mode orig_mode, tree type,
5525 HOST_WIDE_INT bytes, HOST_WIDE_INT words)
5527 static bool warnedavx, warnedsse, warnedmmx;
5529 /* Avoid the AL settings for the Unix64 ABI. */
5530 if (mode == VOIDmode)
5546 if (words <= cum->nregs)
5548 int regno = cum->regno;
5550 /* Fastcall allocates the first two DWORD (SImode) or
5551 smaller arguments to ECX and EDX if it isn't an
5557 || (type && AGGREGATE_TYPE_P (type)))
5560 /* ECX not EAX is the first allocated register. */
5561 if (regno == AX_REG)
5564 return gen_rtx_REG (mode, regno);
5569 if (cum->float_in_sse < 2)
5572 if (cum->float_in_sse < 1)
5576 /* In 32bit, we pass TImode in xmm registers. */
5583 if (!type || !AGGREGATE_TYPE_P (type))
5585 if (!TARGET_SSE && !warnedsse && cum->warn_sse)
5588 warning (0, "SSE vector argument without SSE enabled "
5592 return gen_reg_or_parallel (mode, orig_mode,
5593 cum->sse_regno + FIRST_SSE_REG);
5598 /* In 32bit, we pass OImode in ymm registers. */
5605 if (!type || !AGGREGATE_TYPE_P (type))
5607 if (!TARGET_AVX && !warnedavx && cum->warn_avx)
5610 warning (0, "AVX vector argument without AVX enabled "
5614 return gen_reg_or_parallel (mode, orig_mode,
5615 cum->sse_regno + FIRST_SSE_REG);
5624 if (!type || !AGGREGATE_TYPE_P (type))
5626 if (!TARGET_MMX && !warnedmmx && cum->warn_mmx)
5629 warning (0, "MMX vector argument without MMX enabled "
5633 return gen_reg_or_parallel (mode, orig_mode,
5634 cum->mmx_regno + FIRST_MMX_REG);
5643 function_arg_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5644 enum machine_mode orig_mode, tree type, int named)
5646 static bool warnedavx;
5648 /* Handle a hidden AL argument containing number of registers
5649 for varargs x86-64 functions. */
5650 if (mode == VOIDmode)
5651 return GEN_INT (cum->maybe_vaarg
5652 ? (cum->sse_nregs < 0
5653 ? (cum->call_abi == DEFAULT_ABI
5655 : (DEFAULT_ABI != SYSV_ABI ? X86_64_SSE_REGPARM_MAX
5656 : X64_SSE_REGPARM_MAX))
5671 /* In 64bit, we pass TImode in interger registers and OImode on
5673 if (!type || !AGGREGATE_TYPE_P (type))
5675 if (!TARGET_AVX && !warnedavx && cum->warn_avx)
5678 warning (0, "AVX vector argument without AVX enabled "
5683 /* Unnamed 256bit vector mode parameters are passed on stack. */
5689 return construct_container (mode, orig_mode, type, 0, cum->nregs,
5691 &x86_64_int_parameter_registers [cum->regno],
5696 function_arg_ms_64 (CUMULATIVE_ARGS *cum, enum machine_mode mode,
5697 enum machine_mode orig_mode, int named,
5698 HOST_WIDE_INT bytes)
5702 /* We need to add clobber for MS_ABI->SYSV ABI calls in expand_call.
5703 We use value of -2 to specify that current function call is MSABI. */
5704 if (mode == VOIDmode)
5705 return GEN_INT (-2);
5707 /* If we've run out of registers, it goes on the stack. */
5708 if (cum->nregs == 0)
5711 regno = x86_64_ms_abi_int_parameter_registers[cum->regno];
5713 /* Only floating point modes are passed in anything but integer regs. */
5714 if (TARGET_SSE && (mode == SFmode || mode == DFmode))
5717 regno = cum->regno + FIRST_SSE_REG;
5722 /* Unnamed floating parameters are passed in both the
5723 SSE and integer registers. */
5724 t1 = gen_rtx_REG (mode, cum->regno + FIRST_SSE_REG);
5725 t2 = gen_rtx_REG (mode, regno);
5726 t1 = gen_rtx_EXPR_LIST (VOIDmode, t1, const0_rtx);
5727 t2 = gen_rtx_EXPR_LIST (VOIDmode, t2, const0_rtx);
5728 return gen_rtx_PARALLEL (mode, gen_rtvec (2, t1, t2));
5731 /* Handle aggregated types passed in register. */
5732 if (orig_mode == BLKmode)
5734 if (bytes > 0 && bytes <= 8)
5735 mode = (bytes > 4 ? DImode : SImode);
5736 if (mode == BLKmode)
5740 return gen_reg_or_parallel (mode, orig_mode, regno);
5744 function_arg (CUMULATIVE_ARGS *cum, enum machine_mode omode,
5745 tree type, int named)
5747 enum machine_mode mode = omode;
5748 HOST_WIDE_INT bytes, words;
5750 if (mode == BLKmode)
5751 bytes = int_size_in_bytes (type);
5753 bytes = GET_MODE_SIZE (mode);
5754 words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
5756 /* To simplify the code below, represent vector types with a vector mode
5757 even if MMX/SSE are not active. */
5758 if (type && TREE_CODE (type) == VECTOR_TYPE)
5759 mode = type_natural_mode (type);
5761 if (TARGET_64BIT && (cum ? cum->call_abi : DEFAULT_ABI) == MS_ABI)
5762 return function_arg_ms_64 (cum, mode, omode, named, bytes);
5763 else if (TARGET_64BIT)
5764 return function_arg_64 (cum, mode, omode, type, named);
5766 return function_arg_32 (cum, mode, omode, type, bytes, words);
5769 /* A C expression that indicates when an argument must be passed by
5770 reference. If nonzero for an argument, a copy of that argument is
5771 made in memory and a pointer to the argument is passed instead of
5772 the argument itself. The pointer is passed in whatever way is
5773 appropriate for passing a pointer to that type. */
5776 ix86_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
5777 enum machine_mode mode ATTRIBUTE_UNUSED,
5778 const_tree type, bool named ATTRIBUTE_UNUSED)
5780 /* See Windows x64 Software Convention. */
5781 if (TARGET_64BIT && (cum ? cum->call_abi : DEFAULT_ABI) == MS_ABI)
5783 int msize = (int) GET_MODE_SIZE (mode);
5786 /* Arrays are passed by reference. */
5787 if (TREE_CODE (type) == ARRAY_TYPE)
5790 if (AGGREGATE_TYPE_P (type))
5792 /* Structs/unions of sizes other than 8, 16, 32, or 64 bits
5793 are passed by reference. */
5794 msize = int_size_in_bytes (type);
5798 /* __m128 is passed by reference. */
5800 case 1: case 2: case 4: case 8:
5806 else if (TARGET_64BIT && type && int_size_in_bytes (type) == -1)
5812 /* Return true when TYPE should be 128bit aligned for 32bit argument passing
5815 contains_aligned_value_p (tree type)
5817 enum machine_mode mode = TYPE_MODE (type);
5818 if (((TARGET_SSE && SSE_REG_MODE_P (mode))
5822 && (!TYPE_USER_ALIGN (type) || TYPE_ALIGN (type) > 128))
5824 if (TYPE_ALIGN (type) < 128)
5827 if (AGGREGATE_TYPE_P (type))
5829 /* Walk the aggregates recursively. */
5830 switch (TREE_CODE (type))
5834 case QUAL_UNION_TYPE:
5838 /* Walk all the structure fields. */
5839 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
5841 if (TREE_CODE (field) == FIELD_DECL
5842 && contains_aligned_value_p (TREE_TYPE (field)))
5849 /* Just for use if some languages passes arrays by value. */
5850 if (contains_aligned_value_p (TREE_TYPE (type)))
5861 /* Gives the alignment boundary, in bits, of an argument with the
5862 specified mode and type. */
5865 ix86_function_arg_boundary (enum machine_mode mode, tree type)
5870 /* Since canonical type is used for call, we convert it to
5871 canonical type if needed. */
5872 if (!TYPE_STRUCTURAL_EQUALITY_P (type))
5873 type = TYPE_CANONICAL (type);
5874 align = TYPE_ALIGN (type);
5877 align = GET_MODE_ALIGNMENT (mode);
5878 if (align < PARM_BOUNDARY)
5879 align = PARM_BOUNDARY;
5880 /* In 32bit, only _Decimal128 and __float128 are aligned to their
5881 natural boundaries. */
5882 if (!TARGET_64BIT && mode != TDmode && mode != TFmode)
5884 /* i386 ABI defines all arguments to be 4 byte aligned. We have to
5885 make an exception for SSE modes since these require 128bit
5888 The handling here differs from field_alignment. ICC aligns MMX
5889 arguments to 4 byte boundaries, while structure fields are aligned
5890 to 8 byte boundaries. */
5893 if (!(TARGET_SSE && SSE_REG_MODE_P (mode)))
5894 align = PARM_BOUNDARY;
5898 if (!contains_aligned_value_p (type))
5899 align = PARM_BOUNDARY;
5902 if (align > BIGGEST_ALIGNMENT)
5903 align = BIGGEST_ALIGNMENT;
5907 /* Return true if N is a possible register number of function value. */
5910 ix86_function_value_regno_p (int regno)
5917 case FIRST_FLOAT_REG:
5918 /* TODO: The function should depend on current function ABI but
5919 builtins.c would need updating then. Therefore we use the
5921 if (TARGET_64BIT && DEFAULT_ABI == MS_ABI)
5923 return TARGET_FLOAT_RETURNS_IN_80387;
5929 if (TARGET_MACHO || TARGET_64BIT)
5937 /* Define how to find the value returned by a function.
5938 VALTYPE is the data type of the value (as a tree).
5939 If the precise function being called is known, FUNC is its FUNCTION_DECL;
5940 otherwise, FUNC is 0. */
5943 function_value_32 (enum machine_mode orig_mode, enum machine_mode mode,
5944 const_tree fntype, const_tree fn)
5948 /* 8-byte vector modes in %mm0. See ix86_return_in_memory for where
5949 we normally prevent this case when mmx is not available. However
5950 some ABIs may require the result to be returned like DImode. */
5951 if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 8)
5952 regno = TARGET_MMX ? FIRST_MMX_REG : 0;
5954 /* 16-byte vector modes in %xmm0. See ix86_return_in_memory for where
5955 we prevent this case when sse is not available. However some ABIs
5956 may require the result to be returned like integer TImode. */
5957 else if (mode == TImode
5958 || (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 16))
5959 regno = TARGET_SSE ? FIRST_SSE_REG : 0;
5961 /* Floating point return values in %st(0) (unless -mno-fp-ret-in-387). */
5962 else if (X87_FLOAT_MODE_P (mode) && TARGET_FLOAT_RETURNS_IN_80387)
5963 regno = FIRST_FLOAT_REG;
5965 /* Most things go in %eax. */
5968 /* Override FP return register with %xmm0 for local functions when
5969 SSE math is enabled or for functions with sseregparm attribute. */
5970 if ((fn || fntype) && (mode == SFmode || mode == DFmode))
5972 int sse_level = ix86_function_sseregparm (fntype, fn, false);
5973 if ((sse_level >= 1 && mode == SFmode)
5974 || (sse_level == 2 && mode == DFmode))
5975 regno = FIRST_SSE_REG;
5978 return gen_rtx_REG (orig_mode, regno);
5982 function_value_64 (enum machine_mode orig_mode, enum machine_mode mode,
5987 /* Handle libcalls, which don't provide a type node. */
5988 if (valtype == NULL)
6000 return gen_rtx_REG (mode, FIRST_SSE_REG);
6003 return gen_rtx_REG (mode, FIRST_FLOAT_REG);
6007 return gen_rtx_REG (mode, AX_REG);
6011 ret = construct_container (mode, orig_mode, valtype, 1,
6012 X86_64_REGPARM_MAX, X86_64_SSE_REGPARM_MAX,
6013 x86_64_int_return_registers, 0);
6015 /* For zero sized structures, construct_container returns NULL, but we
6016 need to keep rest of compiler happy by returning meaningful value. */
6018 ret = gen_rtx_REG (orig_mode, AX_REG);
6024 function_value_ms_64 (enum machine_mode orig_mode, enum machine_mode mode)
6026 unsigned int regno = AX_REG;
6030 switch (GET_MODE_SIZE (mode))
6033 if((SCALAR_INT_MODE_P (mode) || VECTOR_MODE_P (mode))
6034 && !COMPLEX_MODE_P (mode))
6035 regno = FIRST_SSE_REG;
6039 if (mode == SFmode || mode == DFmode)
6040 regno = FIRST_SSE_REG;
6046 return gen_rtx_REG (orig_mode, regno);
6050 ix86_function_value_1 (const_tree valtype, const_tree fntype_or_decl,
6051 enum machine_mode orig_mode, enum machine_mode mode)
6053 const_tree fn, fntype;
6056 if (fntype_or_decl && DECL_P (fntype_or_decl))
6057 fn = fntype_or_decl;
6058 fntype = fn ? TREE_TYPE (fn) : fntype_or_decl;
6060 if (TARGET_64BIT && ix86_function_type_abi (fntype) == MS_ABI)
6061 return function_value_ms_64 (orig_mode, mode);
6062 else if (TARGET_64BIT)
6063 return function_value_64 (orig_mode, mode, valtype);
6065 return function_value_32 (orig_mode, mode, fntype, fn);
6069 ix86_function_value (const_tree valtype, const_tree fntype_or_decl,
6070 bool outgoing ATTRIBUTE_UNUSED)
6072 enum machine_mode mode, orig_mode;
6074 orig_mode = TYPE_MODE (valtype);
6075 mode = type_natural_mode (valtype);
6076 return ix86_function_value_1 (valtype, fntype_or_decl, orig_mode, mode);
6080 ix86_libcall_value (enum machine_mode mode)
6082 return ix86_function_value_1 (NULL, NULL, mode, mode);
6085 /* Return true iff type is returned in memory. */
6087 static int ATTRIBUTE_UNUSED
6088 return_in_memory_32 (const_tree type, enum machine_mode mode)
6092 if (mode == BLKmode)
6095 size = int_size_in_bytes (type);
6097 if (MS_AGGREGATE_RETURN && AGGREGATE_TYPE_P (type) && size <= 8)
6100 if (VECTOR_MODE_P (mode) || mode == TImode)
6102 /* User-created vectors small enough to fit in EAX. */
6106 /* MMX/3dNow values are returned in MM0,
6107 except when it doesn't exits. */
6109 return (TARGET_MMX ? 0 : 1);
6111 /* SSE values are returned in XMM0, except when it doesn't exist. */
6113 return (TARGET_SSE ? 0 : 1);
6124 static int ATTRIBUTE_UNUSED
6125 return_in_memory_64 (const_tree type, enum machine_mode mode)
6127 int needed_intregs, needed_sseregs;
6128 return !examine_argument (mode, type, 1, &needed_intregs, &needed_sseregs);
6131 static int ATTRIBUTE_UNUSED
6132 return_in_memory_ms_64 (const_tree type, enum machine_mode mode)
6134 HOST_WIDE_INT size = int_size_in_bytes (type);
6136 /* __m128 is returned in xmm0. */
6137 if ((SCALAR_INT_MODE_P (mode) || VECTOR_MODE_P (mode))
6138 && !COMPLEX_MODE_P (mode) && (GET_MODE_SIZE (mode) == 16 || size == 16))
6141 /* Otherwise, the size must be exactly in [1248]. */
6142 return (size != 1 && size != 2 && size != 4 && size != 8);
6146 ix86_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6148 #ifdef SUBTARGET_RETURN_IN_MEMORY
6149 return SUBTARGET_RETURN_IN_MEMORY (type, fntype);
6151 const enum machine_mode mode = type_natural_mode (type);
6153 if (TARGET_64BIT_MS_ABI)
6154 return return_in_memory_ms_64 (type, mode);
6155 else if (TARGET_64BIT)
6156 return return_in_memory_64 (type, mode);
6158 return return_in_memory_32 (type, mode);
6162 /* Return false iff TYPE is returned in memory. This version is used
6163 on Solaris 10. It is similar to the generic ix86_return_in_memory,
6164 but differs notably in that when MMX is available, 8-byte vectors
6165 are returned in memory, rather than in MMX registers. */
6168 ix86_sol10_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
6171 enum machine_mode mode = type_natural_mode (type);
6174 return return_in_memory_64 (type, mode);
6176 if (mode == BLKmode)
6179 size = int_size_in_bytes (type);
6181 if (VECTOR_MODE_P (mode))
6183 /* Return in memory only if MMX registers *are* available. This
6184 seems backwards, but it is consistent with the existing
6191 else if (mode == TImode)
6193 else if (mode == XFmode)
6199 /* When returning SSE vector types, we have a choice of either
6200 (1) being abi incompatible with a -march switch, or
6201 (2) generating an error.
6202 Given no good solution, I think the safest thing is one warning.
6203 The user won't be able to use -Werror, but....
6205 Choose the STRUCT_VALUE_RTX hook because that's (at present) only
6206 called in response to actually generating a caller or callee that
6207 uses such a type. As opposed to TARGET_RETURN_IN_MEMORY, which is called
6208 via aggregate_value_p for general type probing from tree-ssa. */
6211 ix86_struct_value_rtx (tree type, int incoming ATTRIBUTE_UNUSED)
6213 static bool warnedsse, warnedmmx;
6215 if (!TARGET_64BIT && type)
6217 /* Look at the return type of the function, not the function type. */
6218 enum machine_mode mode = TYPE_MODE (TREE_TYPE (type));
6220 if (!TARGET_SSE && !warnedsse)
6223 || (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 16))
6226 warning (0, "SSE vector return without SSE enabled "
6231 if (!TARGET_MMX && !warnedmmx)
6233 if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 8)
6236 warning (0, "MMX vector return without MMX enabled "
6246 /* Create the va_list data type. */
6248 /* Returns the calling convention specific va_list date type.
6249 The argument ABI can be DEFAULT_ABI, MS_ABI, or SYSV_ABI. */
6252 ix86_build_builtin_va_list_abi (enum calling_abi abi)
6254 tree f_gpr, f_fpr, f_ovf, f_sav, record, type_decl;
6256 /* For i386 we use plain pointer to argument area. */
6257 if (!TARGET_64BIT || abi == MS_ABI)
6258 return build_pointer_type (char_type_node);
6260 record = (*lang_hooks.types.make_type) (RECORD_TYPE);
6261 type_decl = build_decl (TYPE_DECL, get_identifier ("__va_list_tag"), record);
6263 f_gpr = build_decl (FIELD_DECL, get_identifier ("gp_offset"),
6264 unsigned_type_node);
6265 f_fpr = build_decl (FIELD_DECL, get_identifier ("fp_offset"),
6266 unsigned_type_node);
6267 f_ovf = build_decl (FIELD_DECL, get_identifier ("overflow_arg_area"),
6269 f_sav = build_decl (FIELD_DECL, get_identifier ("reg_save_area"),
6272 va_list_gpr_counter_field = f_gpr;
6273 va_list_fpr_counter_field = f_fpr;
6275 DECL_FIELD_CONTEXT (f_gpr) = record;
6276 DECL_FIELD_CONTEXT (f_fpr) = record;
6277 DECL_FIELD_CONTEXT (f_ovf) = record;
6278 DECL_FIELD_CONTEXT (f_sav) = record;
6280 TREE_CHAIN (record) = type_decl;
6281 TYPE_NAME (record) = type_decl;
6282 TYPE_FIELDS (record) = f_gpr;
6283 TREE_CHAIN (f_gpr) = f_fpr;
6284 TREE_CHAIN (f_fpr) = f_ovf;
6285 TREE_CHAIN (f_ovf) = f_sav;
6287 layout_type (record);
6289 /* The correct type is an array type of one element. */
6290 return build_array_type (record, build_index_type (size_zero_node));
6293 /* Setup the builtin va_list data type and for 64-bit the additional
6294 calling convention specific va_list data types. */
6297 ix86_build_builtin_va_list (void)
6299 tree ret = ix86_build_builtin_va_list_abi (DEFAULT_ABI);
6301 /* Initialize abi specific va_list builtin types. */
6305 if (DEFAULT_ABI == MS_ABI)
6307 t = ix86_build_builtin_va_list_abi (SYSV_ABI);
6308 if (TREE_CODE (t) != RECORD_TYPE)
6309 t = build_variant_type_copy (t);
6310 sysv_va_list_type_node = t;
6315 if (TREE_CODE (t) != RECORD_TYPE)
6316 t = build_variant_type_copy (t);
6317 sysv_va_list_type_node = t;
6319 if (DEFAULT_ABI != MS_ABI)
6321 t = ix86_build_builtin_va_list_abi (MS_ABI);
6322 if (TREE_CODE (t) != RECORD_TYPE)
6323 t = build_variant_type_copy (t);
6324 ms_va_list_type_node = t;
6329 if (TREE_CODE (t) != RECORD_TYPE)
6330 t = build_variant_type_copy (t);
6331 ms_va_list_type_node = t;
6338 /* Worker function for TARGET_SETUP_INCOMING_VARARGS. */
6341 setup_incoming_varargs_64 (CUMULATIVE_ARGS *cum)
6350 int regparm = ix86_regparm;
6352 if (cum->call_abi != DEFAULT_ABI)
6353 regparm = DEFAULT_ABI != SYSV_ABI ? X86_64_REGPARM_MAX : X64_REGPARM_MAX;
6355 /* GPR size of varargs save area. */
6356 if (cfun->va_list_gpr_size)
6357 ix86_varargs_gpr_size = X86_64_REGPARM_MAX * UNITS_PER_WORD;
6359 ix86_varargs_gpr_size = 0;
6361 /* FPR size of varargs save area. We don't need it if we don't pass
6362 anything in SSE registers. */
6363 if (cum->sse_nregs && cfun->va_list_fpr_size)
6364 ix86_varargs_fpr_size = X86_64_SSE_REGPARM_MAX * 16;
6366 ix86_varargs_fpr_size = 0;
6368 if (! ix86_varargs_gpr_size && ! ix86_varargs_fpr_size)
6371 save_area = frame_pointer_rtx;
6372 set = get_varargs_alias_set ();
6374 for (i = cum->regno;
6376 && i < cum->regno + cfun->va_list_gpr_size / UNITS_PER_WORD;
6379 mem = gen_rtx_MEM (Pmode,
6380 plus_constant (save_area, i * UNITS_PER_WORD));
6381 MEM_NOTRAP_P (mem) = 1;
6382 set_mem_alias_set (mem, set);
6383 emit_move_insn (mem, gen_rtx_REG (Pmode,
6384 x86_64_int_parameter_registers[i]));
6387 if (ix86_varargs_fpr_size)
6389 /* Now emit code to save SSE registers. The AX parameter contains number
6390 of SSE parameter registers used to call this function. We use
6391 sse_prologue_save insn template that produces computed jump across
6392 SSE saves. We need some preparation work to get this working. */
6394 label = gen_label_rtx ();
6395 label_ref = gen_rtx_LABEL_REF (Pmode, label);
6397 /* Compute address to jump to :
6398 label - eax*4 + nnamed_sse_arguments*4 Or
6399 label - eax*5 + nnamed_sse_arguments*5 for AVX. */
6400 tmp_reg = gen_reg_rtx (Pmode);
6401 nsse_reg = gen_reg_rtx (Pmode);
6402 emit_insn (gen_zero_extendqidi2 (nsse_reg, gen_rtx_REG (QImode, AX_REG)));
6403 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6404 gen_rtx_MULT (Pmode, nsse_reg,
6407 /* vmovaps is one byte longer than movaps. */
6409 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6410 gen_rtx_PLUS (Pmode, tmp_reg,
6416 gen_rtx_CONST (DImode,
6417 gen_rtx_PLUS (DImode,
6419 GEN_INT (cum->sse_regno
6420 * (TARGET_AVX ? 5 : 4)))));
6422 emit_move_insn (nsse_reg, label_ref);
6423 emit_insn (gen_subdi3 (nsse_reg, nsse_reg, tmp_reg));
6425 /* Compute address of memory block we save into. We always use pointer
6426 pointing 127 bytes after first byte to store - this is needed to keep
6427 instruction size limited by 4 bytes (5 bytes for AVX) with one
6428 byte displacement. */
6429 tmp_reg = gen_reg_rtx (Pmode);
6430 emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
6431 plus_constant (save_area,
6432 ix86_varargs_gpr_size + 127)));
6433 mem = gen_rtx_MEM (BLKmode, plus_constant (tmp_reg, -127));
6434 MEM_NOTRAP_P (mem) = 1;
6435 set_mem_alias_set (mem, set);
6436 set_mem_align (mem, BITS_PER_WORD);
6438 /* And finally do the dirty job! */
6439 emit_insn (gen_sse_prologue_save (mem, nsse_reg,
6440 GEN_INT (cum->sse_regno), label));
6445 setup_incoming_varargs_ms_64 (CUMULATIVE_ARGS *cum)
6447 alias_set_type set = get_varargs_alias_set ();
6450 for (i = cum->regno; i < X64_REGPARM_MAX; i++)
6454 mem = gen_rtx_MEM (Pmode,
6455 plus_constant (virtual_incoming_args_rtx,
6456 i * UNITS_PER_WORD));
6457 MEM_NOTRAP_P (mem) = 1;
6458 set_mem_alias_set (mem, set);
6460 reg = gen_rtx_REG (Pmode, x86_64_ms_abi_int_parameter_registers[i]);
6461 emit_move_insn (mem, reg);
6466 ix86_setup_incoming_varargs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
6467 tree type, int *pretend_size ATTRIBUTE_UNUSED,
6470 CUMULATIVE_ARGS next_cum;
6473 /* This argument doesn't appear to be used anymore. Which is good,
6474 because the old code here didn't suppress rtl generation. */
6475 gcc_assert (!no_rtl);
6480 fntype = TREE_TYPE (current_function_decl);
6482 /* For varargs, we do not want to skip the dummy va_dcl argument.
6483 For stdargs, we do want to skip the last named argument. */
6485 if (stdarg_p (fntype))
6486 function_arg_advance (&next_cum, mode, type, 1);
6488 if (cum->call_abi == MS_ABI)
6489 setup_incoming_varargs_ms_64 (&next_cum);
6491 setup_incoming_varargs_64 (&next_cum);
6494 /* Checks if TYPE is of kind va_list char *. */
6497 is_va_list_char_pointer (tree type)
6501 /* For 32-bit it is always true. */
6504 canonic = ix86_canonical_va_list_type (type);
6505 return (canonic == ms_va_list_type_node
6506 || (DEFAULT_ABI == MS_ABI && canonic == va_list_type_node));
6509 /* Implement va_start. */
6512 ix86_va_start (tree valist, rtx nextarg)
6514 HOST_WIDE_INT words, n_gpr, n_fpr;
6515 tree f_gpr, f_fpr, f_ovf, f_sav;
6516 tree gpr, fpr, ovf, sav, t;
6519 /* Only 64bit target needs something special. */
6520 if (!TARGET_64BIT || is_va_list_char_pointer (TREE_TYPE (valist)))
6522 std_expand_builtin_va_start (valist, nextarg);
6526 f_gpr = TYPE_FIELDS (TREE_TYPE (sysv_va_list_type_node));
6527 f_fpr = TREE_CHAIN (f_gpr);
6528 f_ovf = TREE_CHAIN (f_fpr);
6529 f_sav = TREE_CHAIN (f_ovf);
6531 valist = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (valist)), valist);
6532 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr), valist, f_gpr, NULL_TREE);
6533 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
6534 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
6535 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
6537 /* Count number of gp and fp argument registers used. */
6538 words = crtl->args.info.words;
6539 n_gpr = crtl->args.info.regno;
6540 n_fpr = crtl->args.info.sse_regno;
6542 if (cfun->va_list_gpr_size)
6544 type = TREE_TYPE (gpr);
6545 t = build2 (MODIFY_EXPR, type,
6546 gpr, build_int_cst (type, n_gpr * 8));
6547 TREE_SIDE_EFFECTS (t) = 1;
6548 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6551 if (TARGET_SSE && cfun->va_list_fpr_size)
6553 type = TREE_TYPE (fpr);
6554 t = build2 (MODIFY_EXPR, type, fpr,
6555 build_int_cst (type, n_fpr * 16 + 8*X86_64_REGPARM_MAX));
6556 TREE_SIDE_EFFECTS (t) = 1;
6557 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6560 /* Find the overflow area. */
6561 type = TREE_TYPE (ovf);
6562 t = make_tree (type, crtl->args.internal_arg_pointer);
6564 t = build2 (POINTER_PLUS_EXPR, type, t,
6565 size_int (words * UNITS_PER_WORD));
6566 t = build2 (MODIFY_EXPR, type, ovf, t);
6567 TREE_SIDE_EFFECTS (t) = 1;
6568 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6570 if (ix86_varargs_gpr_size || ix86_varargs_fpr_size)
6572 /* Find the register save area.
6573 Prologue of the function save it right above stack frame. */
6574 type = TREE_TYPE (sav);
6575 t = make_tree (type, frame_pointer_rtx);
6576 if (!ix86_varargs_gpr_size)
6577 t = build2 (POINTER_PLUS_EXPR, type, t,
6578 size_int (-8 * X86_64_REGPARM_MAX));
6579 t = build2 (MODIFY_EXPR, type, sav, t);
6580 TREE_SIDE_EFFECTS (t) = 1;
6581 expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
6585 /* Implement va_arg. */
6588 ix86_gimplify_va_arg (tree valist, tree type, gimple_seq *pre_p,
6591 static const int intreg[6] = { 0, 1, 2, 3, 4, 5 };
6592 tree f_gpr, f_fpr, f_ovf, f_sav;
6593 tree gpr, fpr, ovf, sav, t;
6595 tree lab_false, lab_over = NULL_TREE;
6600 enum machine_mode nat_mode;
6603 /* Only 64bit target needs something special. */
6604 if (!TARGET_64BIT || is_va_list_char_pointer (TREE_TYPE (valist)))
6605 return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
6607 f_gpr = TYPE_FIELDS (TREE_TYPE (sysv_va_list_type_node));
6608 f_fpr = TREE_CHAIN (f_gpr);
6609 f_ovf = TREE_CHAIN (f_fpr);
6610 f_sav = TREE_CHAIN (f_ovf);
6612 gpr = build3 (COMPONENT_REF, TREE_TYPE (f_gpr),
6613 build_va_arg_indirect_ref (valist), f_gpr, NULL_TREE);
6614 valist = build_va_arg_indirect_ref (valist);
6615 fpr = build3 (COMPONENT_REF, TREE_TYPE (f_fpr), valist, f_fpr, NULL_TREE);
6616 ovf = build3 (COMPONENT_REF, TREE_TYPE (f_ovf), valist, f_ovf, NULL_TREE);
6617 sav = build3 (COMPONENT_REF, TREE_TYPE (f_sav), valist, f_sav, NULL_TREE);
6619 indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false);
6621 type = build_pointer_type (type);
6622 size = int_size_in_bytes (type);
6623 rsize = (size + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
6625 nat_mode = type_natural_mode (type);
6634 /* Unnamed 256bit vector mode parameters are passed on stack. */
6635 if (ix86_cfun_abi () == SYSV_ABI)
6642 container = construct_container (nat_mode, TYPE_MODE (type),
6643 type, 0, X86_64_REGPARM_MAX,
6644 X86_64_SSE_REGPARM_MAX, intreg,
6649 /* Pull the value out of the saved registers. */
6651 addr = create_tmp_var (ptr_type_node, "addr");
6652 DECL_POINTER_ALIAS_SET (addr) = get_varargs_alias_set ();
6656 int needed_intregs, needed_sseregs;
6658 tree int_addr, sse_addr;
6660 lab_false = create_artificial_label ();
6661 lab_over = create_artificial_label ();
6663 examine_argument (nat_mode, type, 0, &needed_intregs, &needed_sseregs);
6665 need_temp = (!REG_P (container)
6666 && ((needed_intregs && TYPE_ALIGN (type) > 64)
6667 || TYPE_ALIGN (type) > 128));
6669 /* In case we are passing structure, verify that it is consecutive block
6670 on the register save area. If not we need to do moves. */
6671 if (!need_temp && !REG_P (container))
6673 /* Verify that all registers are strictly consecutive */
6674 if (SSE_REGNO_P (REGNO (XEXP (XVECEXP (container, 0, 0), 0))))
6678 for (i = 0; i < XVECLEN (container, 0) && !need_temp; i++)
6680 rtx slot = XVECEXP (container, 0, i);
6681 if (REGNO (XEXP (slot, 0)) != FIRST_SSE_REG + (unsigned int) i
6682 || INTVAL (XEXP (slot, 1)) != i * 16)
6690 for (i = 0; i < XVECLEN (container, 0) && !need_temp; i++)
6692 rtx slot = XVECEXP (container, 0, i);
6693 if (REGNO (XEXP (slot, 0)) != (unsigned int) i
6694 || INTVAL (XEXP (slot, 1)) != i * 8)
6706 int_addr = create_tmp_var (ptr_type_node, "int_addr");
6707 DECL_POINTER_ALIAS_SET (int_addr) = get_varargs_alias_set ();
6708 sse_addr = create_tmp_var (ptr_type_node, "sse_addr");
6709 DECL_POINTER_ALIAS_SET (sse_addr) = get_varargs_alias_set ();
6712 /* First ensure that we fit completely in registers. */
6715 t = build_int_cst (TREE_TYPE (gpr),
6716 (X86_64_REGPARM_MAX - needed_intregs + 1) * 8);
6717 t = build2 (GE_EXPR, boolean_type_node, gpr, t);
6718 t2 = build1 (GOTO_EXPR, void_type_node, lab_false);
6719 t = build3 (COND_EXPR, void_type_node, t, t2, NULL_TREE);
6720 gimplify_and_add (t, pre_p);
6724 t = build_int_cst (TREE_TYPE (fpr),
6725 (X86_64_SSE_REGPARM_MAX - needed_sseregs + 1) * 16
6726 + X86_64_REGPARM_MAX * 8);
6727 t = build2 (GE_EXPR, boolean_type_node, fpr, t);
6728 t2 = build1 (GOTO_EXPR, void_type_node, lab_false);
6729 t = build3 (COND_EXPR, void_type_node, t, t2, NULL_TREE);
6730 gimplify_and_add (t, pre_p);
6733 /* Compute index to start of area used for integer regs. */
6736 /* int_addr = gpr + sav; */
6737 t = fold_convert (sizetype, gpr);
6738 t = build2 (POINTER_PLUS_EXPR, ptr_type_node, sav, t);
6739 gimplify_assign (int_addr, t, pre_p);
6743 /* sse_addr = fpr + sav; */
6744 t = fold_convert (sizetype, fpr);
6745 t = build2 (POINTER_PLUS_EXPR, ptr_type_node, sav, t);
6746 gimplify_assign (sse_addr, t, pre_p);
6751 tree temp = create_tmp_var (type, "va_arg_tmp");
6754 t = build1 (ADDR_EXPR, build_pointer_type (type), temp);
6755 gimplify_assign (addr, t, pre_p);
6757 for (i = 0; i < XVECLEN (container, 0); i++)
6759 rtx slot = XVECEXP (container, 0, i);
6760 rtx reg = XEXP (slot, 0);
6761 enum machine_mode mode = GET_MODE (reg);
6762 tree piece_type = lang_hooks.types.type_for_mode (mode, 1);
6763 tree addr_type = build_pointer_type (piece_type);
6764 tree daddr_type = build_pointer_type_for_mode (piece_type,
6768 tree dest_addr, dest;
6770 if (SSE_REGNO_P (REGNO (reg)))
6772 src_addr = sse_addr;
6773 src_offset = (REGNO (reg) - FIRST_SSE_REG) * 16;
6777 src_addr = int_addr;
6778 src_offset = REGNO (reg) * 8;
6780 src_addr = fold_convert (addr_type, src_addr);
6781 src_addr = fold_build2 (POINTER_PLUS_EXPR, addr_type, src_addr,
6782 size_int (src_offset));
6783 src = build_va_arg_indirect_ref (src_addr);
6785 dest_addr = fold_convert (daddr_type, addr);
6786 dest_addr = fold_build2 (POINTER_PLUS_EXPR, daddr_type, dest_addr,
6787 size_int (INTVAL (XEXP (slot, 1))));
6788 dest = build_va_arg_indirect_ref (dest_addr);
6790 gimplify_assign (dest, src, pre_p);
6796 t = build2 (PLUS_EXPR, TREE_TYPE (gpr), gpr,
6797 build_int_cst (TREE_TYPE (gpr), needed_intregs * 8));
6798 gimplify_assign (gpr, t, pre_p);
6803 t = build2 (PLUS_EXPR, TREE_TYPE (fpr), fpr,
6804 build_int_cst (TREE_TYPE (fpr), needed_sseregs * 16));
6805 gimplify_assign (fpr, t, pre_p);
6808 gimple_seq_add_stmt (pre_p, gimple_build_goto (lab_over));
6810 gimple_seq_add_stmt (pre_p, gimple_build_label (lab_false));
6813 /* ... otherwise out of the overflow area. */
6815 /* When we align parameter on stack for caller, if the parameter
6816 alignment is beyond MAX_SUPPORTED_STACK_ALIGNMENT, it will be
6817 aligned at MAX_SUPPORTED_STACK_ALIGNMENT. We will match callee
6818 here with caller. */
6819 arg_boundary = FUNCTION_ARG_BOUNDARY (VOIDmode, type);
6820 if ((unsigned int) arg_boundary > MAX_SUPPORTED_STACK_ALIGNMENT)
6821 arg_boundary = MAX_SUPPORTED_STACK_ALIGNMENT;
6823 /* Care for on-stack alignment if needed. */
6824 if (arg_boundary <= 64
6825 || integer_zerop (TYPE_SIZE (type)))
6829 HOST_WIDE_INT align = arg_boundary / 8;
6830 t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (ovf), ovf,
6831 size_int (align - 1));
6832 t = fold_convert (sizetype, t);
6833 t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t,
6835 t = fold_convert (TREE_TYPE (ovf), t);
6837 gimplify_expr (&t, pre_p, NULL, is_gimple_val, fb_rvalue);
6838 gimplify_assign (addr, t, pre_p);
6840 t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (t), t,
6841 size_int (rsize * UNITS_PER_WORD));
6842 gimplify_assign (unshare_expr (ovf), t, pre_p);
6845 gimple_seq_add_stmt (pre_p, gimple_build_label (lab_over));
6847 ptrtype = build_pointer_type (type);
6848 addr = fold_convert (ptrtype, addr);
6851 addr = build_va_arg_indirect_ref (addr);
6852 return build_va_arg_indirect_ref (addr);
6855 /* Return nonzero if OPNUM's MEM should be matched
6856 in movabs* patterns. */
6859 ix86_check_movabs (rtx insn, int opnum)
6863 set = PATTERN (insn);
6864 if (GET_CODE (set) == PARALLEL)
6865 set = XVECEXP (set, 0, 0);
6866 gcc_assert (GET_CODE (set) == SET);
6867 mem = XEXP (set, opnum);
6868 while (GET_CODE (mem) == SUBREG)
6869 mem = SUBREG_REG (mem);
6870 gcc_assert (MEM_P (mem));
6871 return (volatile_ok || !MEM_VOLATILE_P (mem));
6874 /* Initialize the table of extra 80387 mathematical constants. */
6877 init_ext_80387_constants (void)
6879 static const char * cst[5] =
6881 "0.3010299956639811952256464283594894482", /* 0: fldlg2 */
6882 "0.6931471805599453094286904741849753009", /* 1: fldln2 */
6883 "1.4426950408889634073876517827983434472", /* 2: fldl2e */
6884 "3.3219280948873623478083405569094566090", /* 3: fldl2t */
6885 "3.1415926535897932385128089594061862044", /* 4: fldpi */
6889 for (i = 0; i < 5; i++)
6891 real_from_string (&ext_80387_constants_table[i], cst[i]);
6892 /* Ensure each constant is rounded to XFmode precision. */
6893 real_convert (&ext_80387_constants_table[i],
6894 XFmode, &ext_80387_constants_table[i]);
6897 ext_80387_constants_init = 1;
6900 /* Return true if the constant is something that can be loaded with
6901 a special instruction. */
6904 standard_80387_constant_p (rtx x)
6906 enum machine_mode mode = GET_MODE (x);
6910 if (!(X87_FLOAT_MODE_P (mode) && (GET_CODE (x) == CONST_DOUBLE)))
6913 if (x == CONST0_RTX (mode))
6915 if (x == CONST1_RTX (mode))
6918 REAL_VALUE_FROM_CONST_DOUBLE (r, x);
6920 /* For XFmode constants, try to find a special 80387 instruction when
6921 optimizing for size or on those CPUs that benefit from them. */
6923 && (optimize_function_for_size_p (cfun) || TARGET_EXT_80387_CONSTANTS))
6927 if (! ext_80387_constants_init)
6928 init_ext_80387_constants ();
6930 for (i = 0; i < 5; i++)
6931 if (real_identical (&r, &ext_80387_constants_table[i]))
6935 /* Load of the constant -0.0 or -1.0 will be split as
6936 fldz;fchs or fld1;fchs sequence. */
6937 if (real_isnegzero (&r))
6939 if (real_identical (&r, &dconstm1))
6945 /* Return the opcode of the special instruction to be used to load
6949 standard_80387_constant_opcode (rtx x)
6951 switch (standard_80387_constant_p (x))
6975 /* Return the CONST_DOUBLE representing the 80387 constant that is
6976 loaded by the specified special instruction. The argument IDX
6977 matches the return value from standard_80387_constant_p. */
6980 standard_80387_constant_rtx (int idx)
6984 if (! ext_80387_constants_init)
6985 init_ext_80387_constants ();
7001 return CONST_DOUBLE_FROM_REAL_VALUE (ext_80387_constants_table[i],
7005 /* Return 1 if mode is a valid mode for sse. */
7007 standard_sse_mode_p (enum machine_mode mode)
7024 /* Return 1 if X is all 0s. For all 1s, return 2 if X is in 128bit
7025 SSE modes and SSE2 is enabled, return 3 if X is in 256bit AVX
7026 modes and AVX is enabled. */
7029 standard_sse_constant_p (rtx x)
7031 enum machine_mode mode = GET_MODE (x);
7033 if (x == const0_rtx || x == CONST0_RTX (GET_MODE (x)))
7035 if (vector_all_ones_operand (x, mode))
7037 if (standard_sse_mode_p (mode))
7038 return TARGET_SSE2 ? 2 : -2;
7039 else if (VALID_AVX256_REG_MODE (mode))
7040 return TARGET_AVX ? 3 : -3;
7046 /* Return the opcode of the special instruction to be used to load
7050 standard_sse_constant_opcode (rtx insn, rtx x)
7052 switch (standard_sse_constant_p (x))
7055 switch (get_attr_mode (insn))
7058 return TARGET_AVX ? "vxorps\t%0, %0, %0" : "xorps\t%0, %0";
7060 return TARGET_AVX ? "vxorpd\t%0, %0, %0" : "xorpd\t%0, %0";
7062 return TARGET_AVX ? "vpxor\t%0, %0, %0" : "pxor\t%0, %0";
7064 return "vxorps\t%x0, %x0, %x0";
7066 return "vxorpd\t%x0, %x0, %x0";
7068 return "vpxor\t%x0, %x0, %x0";
7074 switch (get_attr_mode (insn))
7079 return "vpcmpeqd\t%0, %0, %0";
7085 return "pcmpeqd\t%0, %0";
7090 /* Returns 1 if OP contains a symbol reference */
7093 symbolic_reference_mentioned_p (rtx op)
7098 if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
7101 fmt = GET_RTX_FORMAT (GET_CODE (op));
7102 for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--)
7108 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
7109 if (symbolic_reference_mentioned_p (XVECEXP (op, i, j)))
7113 else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i)))
7120 /* Return 1 if it is appropriate to emit `ret' instructions in the
7121 body of a function. Do this only if the epilogue is simple, needing a
7122 couple of insns. Prior to reloading, we can't tell how many registers
7123 must be saved, so return 0 then. Return 0 if there is no frame
7124 marker to de-allocate. */
7127 ix86_can_use_return_insn_p (void)
7129 struct ix86_frame frame;
7131 if (! reload_completed || frame_pointer_needed)
7134 /* Don't allow more than 32 pop, since that's all we can do
7135 with one instruction. */
7136 if (crtl->args.pops_args
7137 && crtl->args.size >= 32768)
7140 ix86_compute_frame_layout (&frame);
7141 return frame.to_allocate == 0 && frame.nregs == 0;
7144 /* Value should be nonzero if functions must have frame pointers.
7145 Zero means the frame pointer need not be set up (and parms may
7146 be accessed via the stack pointer) in functions that seem suitable. */
7149 ix86_frame_pointer_required (void)
7151 /* If we accessed previous frames, then the generated code expects
7152 to be able to access the saved ebp value in our frame. */
7153 if (cfun->machine->accesses_prev_frame)
7156 /* Several x86 os'es need a frame pointer for other reasons,
7157 usually pertaining to setjmp. */
7158 if (SUBTARGET_FRAME_POINTER_REQUIRED)
7161 /* In override_options, TARGET_OMIT_LEAF_FRAME_POINTER turns off
7162 the frame pointer by default. Turn it back on now if we've not
7163 got a leaf function. */
7164 if (TARGET_OMIT_LEAF_FRAME_POINTER
7165 && (!current_function_is_leaf
7166 || ix86_current_function_calls_tls_descriptor))
7175 /* Record that the current function accesses previous call frames. */
7178 ix86_setup_frame_addresses (void)
7180 cfun->machine->accesses_prev_frame = 1;
7183 #if (defined(HAVE_GAS_HIDDEN) && (SUPPORTS_ONE_ONLY - 0)) || TARGET_MACHO
7184 # define USE_HIDDEN_LINKONCE 1
7186 # define USE_HIDDEN_LINKONCE 0
7189 static int pic_labels_used;
7191 /* Fills in the label name that should be used for a pc thunk for
7192 the given register. */
7195 get_pc_thunk_name (char name[32], unsigned int regno)
7197 gcc_assert (!TARGET_64BIT);
7199 if (USE_HIDDEN_LINKONCE)
7200 sprintf (name, "__i686.get_pc_thunk.%s", reg_names[regno]);
7202 ASM_GENERATE_INTERNAL_LABEL (name, "LPR", regno);
7206 /* This function generates code for -fpic that loads %ebx with
7207 the return address of the caller and then returns. */
7210 ix86_file_end (void)
7215 for (regno = 0; regno < 8; ++regno)
7219 if (! ((pic_labels_used >> regno) & 1))
7222 get_pc_thunk_name (name, regno);
7227 switch_to_section (darwin_sections[text_coal_section]);
7228 fputs ("\t.weak_definition\t", asm_out_file);
7229 assemble_name (asm_out_file, name);
7230 fputs ("\n\t.private_extern\t", asm_out_file);
7231 assemble_name (asm_out_file, name);
7232 fputs ("\n", asm_out_file);
7233 ASM_OUTPUT_LABEL (asm_out_file, name);
7237 if (USE_HIDDEN_LINKONCE)
7241 decl = build_decl (FUNCTION_DECL, get_identifier (name),
7243 TREE_PUBLIC (decl) = 1;
7244 TREE_STATIC (decl) = 1;
7245 DECL_ONE_ONLY (decl) = 1;
7247 (*targetm.asm_out.unique_section) (decl, 0);
7248 switch_to_section (get_named_section (decl, NULL, 0));
7250 (*targetm.asm_out.globalize_label) (asm_out_file, name);
7251 fputs ("\t.hidden\t", asm_out_file);
7252 assemble_name (asm_out_file, name);
7253 fputc ('\n', asm_out_file);
7254 ASM_DECLARE_FUNCTION_NAME (asm_out_file, name, decl);
7258 switch_to_section (text_section);
7259 ASM_OUTPUT_LABEL (asm_out_file, name);
7262 xops[0] = gen_rtx_REG (Pmode, regno);
7263 xops[1] = gen_rtx_MEM (Pmode, stack_pointer_rtx);
7264 output_asm_insn ("mov%z0\t{%1, %0|%0, %1}", xops);
7265 output_asm_insn ("ret", xops);
7268 if (NEED_INDICATE_EXEC_STACK)
7269 file_end_indicate_exec_stack ();
7272 /* Emit code for the SET_GOT patterns. */
7275 output_set_got (rtx dest, rtx label ATTRIBUTE_UNUSED)
7281 if (TARGET_VXWORKS_RTP && flag_pic)
7283 /* Load (*VXWORKS_GOTT_BASE) into the PIC register. */
7284 xops[2] = gen_rtx_MEM (Pmode,
7285 gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE));
7286 output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops);
7288 /* Load (*VXWORKS_GOTT_BASE)[VXWORKS_GOTT_INDEX] into the PIC register.
7289 Use %P and a local symbol in order to print VXWORKS_GOTT_INDEX as
7290 an unadorned address. */
7291 xops[2] = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_INDEX);
7292 SYMBOL_REF_FLAGS (xops[2]) |= SYMBOL_FLAG_LOCAL;
7293 output_asm_insn ("mov{l}\t{%P2(%0), %0|%0, DWORD PTR %P2[%0]}", xops);
7297 xops[1] = gen_rtx_SYMBOL_REF (Pmode, GOT_SYMBOL_NAME);
7299 if (! TARGET_DEEP_BRANCH_PREDICTION || !flag_pic)
7301 xops[2] = gen_rtx_LABEL_REF (Pmode, label ? label : gen_label_rtx ());
7304 output_asm_insn ("mov%z0\t{%2, %0|%0, %2}", xops);
7306 output_asm_insn ("call\t%a2", xops);
7309 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
7310 is what will be referenced by the Mach-O PIC subsystem. */
7312 ASM_OUTPUT_LABEL (asm_out_file, MACHOPIC_FUNCTION_BASE_NAME);
7315 (*targetm.asm_out.internal_label) (asm_out_file, "L",
7316 CODE_LABEL_NUMBER (XEXP (xops[2], 0)));
7319 output_asm_insn ("pop%z0\t%0", xops);
7324 get_pc_thunk_name (name, REGNO (dest));
7325 pic_labels_used |= 1 << REGNO (dest);
7327 xops[2] = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (name));
7328 xops[2] = gen_rtx_MEM (QImode, xops[2]);
7329 output_asm_insn ("call\t%X2", xops);
7330 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
7331 is what will be referenced by the Mach-O PIC subsystem. */
7334 ASM_OUTPUT_LABEL (asm_out_file, MACHOPIC_FUNCTION_BASE_NAME);
7336 targetm.asm_out.internal_label (asm_out_file, "L",
7337 CODE_LABEL_NUMBER (label));
7344 if (!flag_pic || TARGET_DEEP_BRANCH_PREDICTION)
7345 output_asm_insn ("add%z0\t{%1, %0|%0, %1}", xops);
7347 output_asm_insn ("add%z0\t{%1+[.-%a2], %0|%0, %1+(.-%a2)}", xops);
7352 /* Generate an "push" pattern for input ARG. */
7357 return gen_rtx_SET (VOIDmode,
7359 gen_rtx_PRE_DEC (Pmode,
7360 stack_pointer_rtx)),
7364 /* Return >= 0 if there is an unused call-clobbered register available
7365 for the entire function. */
7368 ix86_select_alt_pic_regnum (void)
7370 if (current_function_is_leaf && !crtl->profile
7371 && !ix86_current_function_calls_tls_descriptor)
7374 /* Can't use the same register for both PIC and DRAP. */
7376 drap = REGNO (crtl->drap_reg);
7379 for (i = 2; i >= 0; --i)
7380 if (i != drap && !df_regs_ever_live_p (i))
7384 return INVALID_REGNUM;
7387 /* Return 1 if we need to save REGNO. */
7389 ix86_save_reg (unsigned int regno, int maybe_eh_return)
7391 if (pic_offset_table_rtx
7392 && regno == REAL_PIC_OFFSET_TABLE_REGNUM
7393 && (df_regs_ever_live_p (REAL_PIC_OFFSET_TABLE_REGNUM)
7395 || crtl->calls_eh_return
7396 || crtl->uses_const_pool))
7398 if (ix86_select_alt_pic_regnum () != INVALID_REGNUM)
7403 if (crtl->calls_eh_return && maybe_eh_return)
7408 unsigned test = EH_RETURN_DATA_REGNO (i);
7409 if (test == INVALID_REGNUM)
7417 && regno == REGNO (crtl->drap_reg))
7420 return (df_regs_ever_live_p (regno)
7421 && !call_used_regs[regno]
7422 && !fixed_regs[regno]
7423 && (regno != HARD_FRAME_POINTER_REGNUM || !frame_pointer_needed));
7426 /* Return number of saved general prupose registers. */
7429 ix86_nsaved_regs (void)
7434 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7435 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7440 /* Return number of saved SSE registrers. */
7443 ix86_nsaved_sseregs (void)
7448 if (ix86_cfun_abi () != MS_ABI)
7450 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7451 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7456 /* Given FROM and TO register numbers, say whether this elimination is
7457 allowed. If stack alignment is needed, we can only replace argument
7458 pointer with hard frame pointer, or replace frame pointer with stack
7459 pointer. Otherwise, frame pointer elimination is automatically
7460 handled and all other eliminations are valid. */
7463 ix86_can_eliminate (int from, int to)
7465 if (stack_realign_fp)
7466 return ((from == ARG_POINTER_REGNUM
7467 && to == HARD_FRAME_POINTER_REGNUM)
7468 || (from == FRAME_POINTER_REGNUM
7469 && to == STACK_POINTER_REGNUM));
7471 return to == STACK_POINTER_REGNUM ? !frame_pointer_needed : 1;
7474 /* Return the offset between two registers, one to be eliminated, and the other
7475 its replacement, at the start of a routine. */
7478 ix86_initial_elimination_offset (int from, int to)
7480 struct ix86_frame frame;
7481 ix86_compute_frame_layout (&frame);
7483 if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
7484 return frame.hard_frame_pointer_offset;
7485 else if (from == FRAME_POINTER_REGNUM
7486 && to == HARD_FRAME_POINTER_REGNUM)
7487 return frame.hard_frame_pointer_offset - frame.frame_pointer_offset;
7490 gcc_assert (to == STACK_POINTER_REGNUM);
7492 if (from == ARG_POINTER_REGNUM)
7493 return frame.stack_pointer_offset;
7495 gcc_assert (from == FRAME_POINTER_REGNUM);
7496 return frame.stack_pointer_offset - frame.frame_pointer_offset;
7500 /* Fill structure ix86_frame about frame of currently computed function. */
7503 ix86_compute_frame_layout (struct ix86_frame *frame)
7505 HOST_WIDE_INT total_size;
7506 unsigned int stack_alignment_needed;
7507 HOST_WIDE_INT offset;
7508 unsigned int preferred_alignment;
7509 HOST_WIDE_INT size = get_frame_size ();
7511 frame->nregs = ix86_nsaved_regs ();
7512 frame->nsseregs = ix86_nsaved_sseregs ();
7515 stack_alignment_needed = crtl->stack_alignment_needed / BITS_PER_UNIT;
7516 preferred_alignment = crtl->preferred_stack_boundary / BITS_PER_UNIT;
7518 /* MS ABI seem to require stack alignment to be always 16 except for function
7520 if (ix86_cfun_abi () == MS_ABI && preferred_alignment < 16)
7522 preferred_alignment = 16;
7523 stack_alignment_needed = 16;
7524 crtl->preferred_stack_boundary = 128;
7525 crtl->stack_alignment_needed = 128;
7528 gcc_assert (!size || stack_alignment_needed);
7529 gcc_assert (preferred_alignment >= STACK_BOUNDARY / BITS_PER_UNIT);
7530 gcc_assert (preferred_alignment <= stack_alignment_needed);
7532 /* During reload iteration the amount of registers saved can change.
7533 Recompute the value as needed. Do not recompute when amount of registers
7534 didn't change as reload does multiple calls to the function and does not
7535 expect the decision to change within single iteration. */
7536 if (!optimize_function_for_size_p (cfun)
7537 && cfun->machine->use_fast_prologue_epilogue_nregs != frame->nregs)
7539 int count = frame->nregs;
7541 cfun->machine->use_fast_prologue_epilogue_nregs = count;
7542 /* The fast prologue uses move instead of push to save registers. This
7543 is significantly longer, but also executes faster as modern hardware
7544 can execute the moves in parallel, but can't do that for push/pop.
7546 Be careful about choosing what prologue to emit: When function takes
7547 many instructions to execute we may use slow version as well as in
7548 case function is known to be outside hot spot (this is known with
7549 feedback only). Weight the size of function by number of registers
7550 to save as it is cheap to use one or two push instructions but very
7551 slow to use many of them. */
7553 count = (count - 1) * FAST_PROLOGUE_INSN_COUNT;
7554 if (cfun->function_frequency < FUNCTION_FREQUENCY_NORMAL
7555 || (flag_branch_probabilities
7556 && cfun->function_frequency < FUNCTION_FREQUENCY_HOT))
7557 cfun->machine->use_fast_prologue_epilogue = false;
7559 cfun->machine->use_fast_prologue_epilogue
7560 = !expensive_function_p (count);
7562 if (TARGET_PROLOGUE_USING_MOVE
7563 && cfun->machine->use_fast_prologue_epilogue)
7564 frame->save_regs_using_mov = true;
7566 frame->save_regs_using_mov = false;
7569 /* Skip return address and saved base pointer. */
7570 offset = frame_pointer_needed ? UNITS_PER_WORD * 2 : UNITS_PER_WORD;
7572 frame->hard_frame_pointer_offset = offset;
7574 /* Set offset to aligned because the realigned frame starts from
7576 if (stack_realign_fp)
7577 offset = (offset + stack_alignment_needed -1) & -stack_alignment_needed;
7579 /* Register save area */
7580 offset += frame->nregs * UNITS_PER_WORD;
7582 /* Align SSE reg save area. */
7583 if (frame->nsseregs)
7584 frame->padding0 = ((offset + 16 - 1) & -16) - offset;
7586 frame->padding0 = 0;
7588 /* SSE register save area. */
7589 offset += frame->padding0 + frame->nsseregs * 16;
7592 frame->va_arg_size = ix86_varargs_gpr_size + ix86_varargs_fpr_size;
7593 offset += frame->va_arg_size;
7595 /* Align start of frame for local function. */
7596 frame->padding1 = ((offset + stack_alignment_needed - 1)
7597 & -stack_alignment_needed) - offset;
7599 offset += frame->padding1;
7601 /* Frame pointer points here. */
7602 frame->frame_pointer_offset = offset;
7606 /* Add outgoing arguments area. Can be skipped if we eliminated
7607 all the function calls as dead code.
7608 Skipping is however impossible when function calls alloca. Alloca
7609 expander assumes that last crtl->outgoing_args_size
7610 of stack frame are unused. */
7611 if (ACCUMULATE_OUTGOING_ARGS
7612 && (!current_function_is_leaf || cfun->calls_alloca
7613 || ix86_current_function_calls_tls_descriptor))
7615 offset += crtl->outgoing_args_size;
7616 frame->outgoing_arguments_size = crtl->outgoing_args_size;
7619 frame->outgoing_arguments_size = 0;
7621 /* Align stack boundary. Only needed if we're calling another function
7623 if (!current_function_is_leaf || cfun->calls_alloca
7624 || ix86_current_function_calls_tls_descriptor)
7625 frame->padding2 = ((offset + preferred_alignment - 1)
7626 & -preferred_alignment) - offset;
7628 frame->padding2 = 0;
7630 offset += frame->padding2;
7632 /* We've reached end of stack frame. */
7633 frame->stack_pointer_offset = offset;
7635 /* Size prologue needs to allocate. */
7636 frame->to_allocate =
7637 (size + frame->padding1 + frame->padding2
7638 + frame->outgoing_arguments_size + frame->va_arg_size);
7640 if ((!frame->to_allocate && frame->nregs <= 1)
7641 || (TARGET_64BIT && frame->to_allocate >= (HOST_WIDE_INT) 0x80000000))
7642 frame->save_regs_using_mov = false;
7644 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE && current_function_sp_is_unchanging
7645 && current_function_is_leaf
7646 && !ix86_current_function_calls_tls_descriptor)
7648 frame->red_zone_size = frame->to_allocate;
7649 if (frame->save_regs_using_mov)
7650 frame->red_zone_size += frame->nregs * UNITS_PER_WORD;
7651 if (frame->red_zone_size > RED_ZONE_SIZE - RED_ZONE_RESERVE)
7652 frame->red_zone_size = RED_ZONE_SIZE - RED_ZONE_RESERVE;
7655 frame->red_zone_size = 0;
7656 frame->to_allocate -= frame->red_zone_size;
7657 frame->stack_pointer_offset -= frame->red_zone_size;
7659 fprintf (stderr, "\n");
7660 fprintf (stderr, "size: %ld\n", (long)size);
7661 fprintf (stderr, "nregs: %ld\n", (long)frame->nregs);
7662 fprintf (stderr, "nsseregs: %ld\n", (long)frame->nsseregs);
7663 fprintf (stderr, "padding0: %ld\n", (long)frame->padding0);
7664 fprintf (stderr, "alignment1: %ld\n", (long)stack_alignment_needed);
7665 fprintf (stderr, "padding1: %ld\n", (long)frame->padding1);
7666 fprintf (stderr, "va_arg: %ld\n", (long)frame->va_arg_size);
7667 fprintf (stderr, "padding2: %ld\n", (long)frame->padding2);
7668 fprintf (stderr, "to_allocate: %ld\n", (long)frame->to_allocate);
7669 fprintf (stderr, "red_zone_size: %ld\n", (long)frame->red_zone_size);
7670 fprintf (stderr, "frame_pointer_offset: %ld\n", (long)frame->frame_pointer_offset);
7671 fprintf (stderr, "hard_frame_pointer_offset: %ld\n",
7672 (long)frame->hard_frame_pointer_offset);
7673 fprintf (stderr, "stack_pointer_offset: %ld\n", (long)frame->stack_pointer_offset);
7674 fprintf (stderr, "current_function_is_leaf: %ld\n", (long)current_function_is_leaf);
7675 fprintf (stderr, "cfun->calls_alloca: %ld\n", (long)cfun->calls_alloca);
7676 fprintf (stderr, "x86_current_function_calls_tls_descriptor: %ld\n", (long)ix86_current_function_calls_tls_descriptor);
7680 /* Emit code to save registers in the prologue. */
7683 ix86_emit_save_regs (void)
7688 for (regno = FIRST_PSEUDO_REGISTER - 1; regno-- > 0; )
7689 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7691 insn = emit_insn (gen_push (gen_rtx_REG (Pmode, regno)));
7692 RTX_FRAME_RELATED_P (insn) = 1;
7696 /* Emit code to save registers using MOV insns. First register
7697 is restored from POINTER + OFFSET. */
7699 ix86_emit_save_regs_using_mov (rtx pointer, HOST_WIDE_INT offset)
7704 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7705 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7707 insn = emit_move_insn (adjust_address (gen_rtx_MEM (Pmode, pointer),
7709 gen_rtx_REG (Pmode, regno));
7710 RTX_FRAME_RELATED_P (insn) = 1;
7711 offset += UNITS_PER_WORD;
7715 /* Emit code to save registers using MOV insns. First register
7716 is restored from POINTER + OFFSET. */
7718 ix86_emit_save_sse_regs_using_mov (rtx pointer, HOST_WIDE_INT offset)
7724 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
7725 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, true))
7727 mem = adjust_address (gen_rtx_MEM (TImode, pointer), TImode, offset);
7728 set_mem_align (mem, 128);
7729 insn = emit_move_insn (mem, gen_rtx_REG (TImode, regno));
7730 RTX_FRAME_RELATED_P (insn) = 1;
7735 /* Expand prologue or epilogue stack adjustment.
7736 The pattern exist to put a dependency on all ebp-based memory accesses.
7737 STYLE should be negative if instructions should be marked as frame related,
7738 zero if %r11 register is live and cannot be freely used and positive
7742 pro_epilogue_adjust_stack (rtx dest, rtx src, rtx offset, int style)
7747 insn = emit_insn (gen_pro_epilogue_adjust_stack_1 (dest, src, offset));
7748 else if (x86_64_immediate_operand (offset, DImode))
7749 insn = emit_insn (gen_pro_epilogue_adjust_stack_rex64 (dest, src, offset));
7753 /* r11 is used by indirect sibcall return as well, set before the
7754 epilogue and used after the epilogue. ATM indirect sibcall
7755 shouldn't be used together with huge frame sizes in one
7756 function because of the frame_size check in sibcall.c. */
7758 r11 = gen_rtx_REG (DImode, R11_REG);
7759 insn = emit_insn (gen_rtx_SET (DImode, r11, offset));
7761 RTX_FRAME_RELATED_P (insn) = 1;
7762 insn = emit_insn (gen_pro_epilogue_adjust_stack_rex64_2 (dest, src, r11,
7766 RTX_FRAME_RELATED_P (insn) = 1;
7769 /* Find an available register to be used as dynamic realign argument
7770 pointer regsiter. Such a register will be written in prologue and
7771 used in begin of body, so it must not be
7772 1. parameter passing register.
7774 We reuse static-chain register if it is available. Otherwise, we
7775 use DI for i386 and R13 for x86-64. We chose R13 since it has
7778 Return: the regno of chosen register. */
7781 find_drap_reg (void)
7783 tree decl = cfun->decl;
7787 /* Use R13 for nested function or function need static chain.
7788 Since function with tail call may use any caller-saved
7789 registers in epilogue, DRAP must not use caller-saved
7790 register in such case. */
7791 if ((decl_function_context (decl)
7792 && !DECL_NO_STATIC_CHAIN (decl))
7793 || crtl->tail_call_emit)
7800 /* Use DI for nested function or function need static chain.
7801 Since function with tail call may use any caller-saved
7802 registers in epilogue, DRAP must not use caller-saved
7803 register in such case. */
7804 if ((decl_function_context (decl)
7805 && !DECL_NO_STATIC_CHAIN (decl))
7806 || crtl->tail_call_emit)
7809 /* Reuse static chain register if it isn't used for parameter
7811 if (ix86_function_regparm (TREE_TYPE (decl), decl) <= 2
7812 && !lookup_attribute ("fastcall",
7813 TYPE_ATTRIBUTES (TREE_TYPE (decl))))
7820 /* Update incoming stack boundary and estimated stack alignment. */
7823 ix86_update_stack_boundary (void)
7825 /* Prefer the one specified at command line. */
7826 ix86_incoming_stack_boundary
7827 = (ix86_user_incoming_stack_boundary
7828 ? ix86_user_incoming_stack_boundary
7829 : ix86_default_incoming_stack_boundary);
7831 /* Incoming stack alignment can be changed on individual functions
7832 via force_align_arg_pointer attribute. We use the smallest
7833 incoming stack boundary. */
7834 if (ix86_incoming_stack_boundary > MIN_STACK_BOUNDARY
7835 && lookup_attribute (ix86_force_align_arg_pointer_string,
7836 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
7837 ix86_incoming_stack_boundary = MIN_STACK_BOUNDARY;
7839 /* The incoming stack frame has to be aligned at least at
7840 parm_stack_boundary. */
7841 if (ix86_incoming_stack_boundary < crtl->parm_stack_boundary)
7842 ix86_incoming_stack_boundary = crtl->parm_stack_boundary;
7844 /* Stack at entrance of main is aligned by runtime. We use the
7845 smallest incoming stack boundary. */
7846 if (ix86_incoming_stack_boundary > MAIN_STACK_BOUNDARY
7847 && DECL_NAME (current_function_decl)
7848 && MAIN_NAME_P (DECL_NAME (current_function_decl))
7849 && DECL_FILE_SCOPE_P (current_function_decl))
7850 ix86_incoming_stack_boundary = MAIN_STACK_BOUNDARY;
7852 /* x86_64 vararg needs 16byte stack alignment for register save
7856 && crtl->stack_alignment_estimated < 128)
7857 crtl->stack_alignment_estimated = 128;
7860 /* Handle the TARGET_GET_DRAP_RTX hook. Return NULL if no DRAP is
7861 needed or an rtx for DRAP otherwise. */
7864 ix86_get_drap_rtx (void)
7866 if (ix86_force_drap || !ACCUMULATE_OUTGOING_ARGS)
7867 crtl->need_drap = true;
7869 if (stack_realign_drap)
7871 /* Assign DRAP to vDRAP and returns vDRAP */
7872 unsigned int regno = find_drap_reg ();
7877 arg_ptr = gen_rtx_REG (Pmode, regno);
7878 crtl->drap_reg = arg_ptr;
7881 drap_vreg = copy_to_reg (arg_ptr);
7885 insn = emit_insn_before (seq, NEXT_INSN (entry_of_function ()));
7886 RTX_FRAME_RELATED_P (insn) = 1;
7893 /* Handle the TARGET_INTERNAL_ARG_POINTER hook. */
7896 ix86_internal_arg_pointer (void)
7898 return virtual_incoming_args_rtx;
7901 /* Handle the TARGET_DWARF_HANDLE_FRAME_UNSPEC hook.
7902 This is called from dwarf2out.c to emit call frame instructions
7903 for frame-related insns containing UNSPECs and UNSPEC_VOLATILEs. */
7905 ix86_dwarf_handle_frame_unspec (const char *label, rtx pattern, int index)
7907 rtx unspec = SET_SRC (pattern);
7908 gcc_assert (GET_CODE (unspec) == UNSPEC);
7912 case UNSPEC_REG_SAVE:
7913 dwarf2out_reg_save_reg (label, XVECEXP (unspec, 0, 0),
7914 SET_DEST (pattern));
7916 case UNSPEC_DEF_CFA:
7917 dwarf2out_def_cfa (label, REGNO (SET_DEST (pattern)),
7918 INTVAL (XVECEXP (unspec, 0, 0)));
7925 /* Finalize stack_realign_needed flag, which will guide prologue/epilogue
7926 to be generated in correct form. */
7928 ix86_finalize_stack_realign_flags (void)
7930 /* Check if stack realign is really needed after reload, and
7931 stores result in cfun */
7932 unsigned int incoming_stack_boundary
7933 = (crtl->parm_stack_boundary > ix86_incoming_stack_boundary
7934 ? crtl->parm_stack_boundary : ix86_incoming_stack_boundary);
7935 unsigned int stack_realign = (incoming_stack_boundary
7936 < (current_function_is_leaf
7937 ? crtl->max_used_stack_slot_alignment
7938 : crtl->stack_alignment_needed));
7940 if (crtl->stack_realign_finalized)
7942 /* After stack_realign_needed is finalized, we can't no longer
7944 gcc_assert (crtl->stack_realign_needed == stack_realign);
7948 crtl->stack_realign_needed = stack_realign;
7949 crtl->stack_realign_finalized = true;
7953 /* Expand the prologue into a bunch of separate insns. */
7956 ix86_expand_prologue (void)
7960 struct ix86_frame frame;
7961 HOST_WIDE_INT allocate;
7963 ix86_finalize_stack_realign_flags ();
7965 /* DRAP should not coexist with stack_realign_fp */
7966 gcc_assert (!(crtl->drap_reg && stack_realign_fp));
7968 ix86_compute_frame_layout (&frame);
7970 /* Emit prologue code to adjust stack alignment and setup DRAP, in case
7971 of DRAP is needed and stack realignment is really needed after reload */
7972 if (crtl->drap_reg && crtl->stack_realign_needed)
7975 int align_bytes = crtl->stack_alignment_needed / BITS_PER_UNIT;
7976 int param_ptr_offset = (call_used_regs[REGNO (crtl->drap_reg)]
7977 ? 0 : UNITS_PER_WORD);
7979 gcc_assert (stack_realign_drap);
7981 /* Grab the argument pointer. */
7982 x = plus_constant (stack_pointer_rtx,
7983 (UNITS_PER_WORD + param_ptr_offset));
7986 /* Only need to push parameter pointer reg if it is caller
7988 if (!call_used_regs[REGNO (crtl->drap_reg)])
7990 /* Push arg pointer reg */
7991 insn = emit_insn (gen_push (y));
7992 RTX_FRAME_RELATED_P (insn) = 1;
7995 insn = emit_insn (gen_rtx_SET (VOIDmode, y, x));
7996 RTX_FRAME_RELATED_P (insn) = 1;
7998 /* Align the stack. */
7999 insn = emit_insn ((*ix86_gen_andsp) (stack_pointer_rtx,
8001 GEN_INT (-align_bytes)));
8002 RTX_FRAME_RELATED_P (insn) = 1;
8004 /* Replicate the return address on the stack so that return
8005 address can be reached via (argp - 1) slot. This is needed
8006 to implement macro RETURN_ADDR_RTX and intrinsic function
8007 expand_builtin_return_addr etc. */
8009 x = gen_frame_mem (Pmode,
8010 plus_constant (x, -UNITS_PER_WORD));
8011 insn = emit_insn (gen_push (x));
8012 RTX_FRAME_RELATED_P (insn) = 1;
8015 /* Note: AT&T enter does NOT have reversed args. Enter is probably
8016 slower on all targets. Also sdb doesn't like it. */
8018 if (frame_pointer_needed)
8020 insn = emit_insn (gen_push (hard_frame_pointer_rtx));
8021 RTX_FRAME_RELATED_P (insn) = 1;
8023 insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);
8024 RTX_FRAME_RELATED_P (insn) = 1;
8027 if (stack_realign_fp)
8029 int align_bytes = crtl->stack_alignment_needed / BITS_PER_UNIT;
8030 gcc_assert (align_bytes > MIN_STACK_BOUNDARY / BITS_PER_UNIT);
8032 /* Align the stack. */
8033 insn = emit_insn ((*ix86_gen_andsp) (stack_pointer_rtx,
8035 GEN_INT (-align_bytes)));
8036 RTX_FRAME_RELATED_P (insn) = 1;
8039 allocate = frame.to_allocate + frame.nsseregs * 16 + frame.padding0;
8041 if (!frame.save_regs_using_mov)
8042 ix86_emit_save_regs ();
8044 allocate += frame.nregs * UNITS_PER_WORD;
8046 /* When using red zone we may start register saving before allocating
8047 the stack frame saving one cycle of the prologue. However I will
8048 avoid doing this if I am going to have to probe the stack since
8049 at least on x86_64 the stack probe can turn into a call that clobbers
8050 a red zone location */
8051 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE && frame.save_regs_using_mov
8052 && (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT))
8053 ix86_emit_save_regs_using_mov ((frame_pointer_needed
8054 && !crtl->stack_realign_needed)
8055 ? hard_frame_pointer_rtx
8056 : stack_pointer_rtx,
8057 -frame.nregs * UNITS_PER_WORD);
8061 else if (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT)
8062 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
8063 GEN_INT (-allocate), -1);
8066 /* Only valid for Win32. */
8067 rtx eax = gen_rtx_REG (Pmode, AX_REG);
8071 gcc_assert (!TARGET_64BIT || cfun->machine->call_abi == MS_ABI);
8073 if (cfun->machine->call_abi == MS_ABI)
8076 eax_live = ix86_eax_live_at_start_p ();
8080 emit_insn (gen_push (eax));
8081 allocate -= UNITS_PER_WORD;
8084 emit_move_insn (eax, GEN_INT (allocate));
8087 insn = gen_allocate_stack_worker_64 (eax, eax);
8089 insn = gen_allocate_stack_worker_32 (eax, eax);
8090 insn = emit_insn (insn);
8091 RTX_FRAME_RELATED_P (insn) = 1;
8092 t = gen_rtx_PLUS (Pmode, stack_pointer_rtx, GEN_INT (-allocate));
8093 t = gen_rtx_SET (VOIDmode, stack_pointer_rtx, t);
8094 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
8095 t, REG_NOTES (insn));
8099 if (frame_pointer_needed)
8100 t = plus_constant (hard_frame_pointer_rtx,
8103 - frame.nregs * UNITS_PER_WORD);
8105 t = plus_constant (stack_pointer_rtx, allocate);
8106 emit_move_insn (eax, gen_rtx_MEM (Pmode, t));
8110 if (frame.save_regs_using_mov
8111 && !(!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE
8112 && (! TARGET_STACK_PROBE || allocate < CHECK_STACK_LIMIT)))
8114 if (!frame_pointer_needed
8115 || !frame.to_allocate
8116 || crtl->stack_realign_needed)
8117 ix86_emit_save_regs_using_mov (stack_pointer_rtx,
8119 + frame.nsseregs * 16 + frame.padding0);
8121 ix86_emit_save_regs_using_mov (hard_frame_pointer_rtx,
8122 -frame.nregs * UNITS_PER_WORD);
8124 if (!frame_pointer_needed
8125 || !frame.to_allocate
8126 || crtl->stack_realign_needed)
8127 ix86_emit_save_sse_regs_using_mov (stack_pointer_rtx,
8130 ix86_emit_save_sse_regs_using_mov (hard_frame_pointer_rtx,
8131 - frame.nregs * UNITS_PER_WORD
8132 - frame.nsseregs * 16
8135 pic_reg_used = false;
8136 if (pic_offset_table_rtx
8137 && (df_regs_ever_live_p (REAL_PIC_OFFSET_TABLE_REGNUM)
8140 unsigned int alt_pic_reg_used = ix86_select_alt_pic_regnum ();
8142 if (alt_pic_reg_used != INVALID_REGNUM)
8143 SET_REGNO (pic_offset_table_rtx, alt_pic_reg_used);
8145 pic_reg_used = true;
8152 if (ix86_cmodel == CM_LARGE_PIC)
8154 rtx tmp_reg = gen_rtx_REG (DImode, R11_REG);
8155 rtx label = gen_label_rtx ();
8157 LABEL_PRESERVE_P (label) = 1;
8158 gcc_assert (REGNO (pic_offset_table_rtx) != REGNO (tmp_reg));
8159 insn = emit_insn (gen_set_rip_rex64 (pic_offset_table_rtx, label));
8160 insn = emit_insn (gen_set_got_offset_rex64 (tmp_reg, label));
8161 insn = emit_insn (gen_adddi3 (pic_offset_table_rtx,
8162 pic_offset_table_rtx, tmp_reg));
8165 insn = emit_insn (gen_set_got_rex64 (pic_offset_table_rtx));
8168 insn = emit_insn (gen_set_got (pic_offset_table_rtx));
8171 /* Prevent function calls from being scheduled before the call to mcount.
8172 In the pic_reg_used case, make sure that the got load isn't deleted. */
8176 emit_insn (gen_prologue_use (pic_offset_table_rtx));
8177 emit_insn (gen_blockage ());
8180 if (crtl->drap_reg && !crtl->stack_realign_needed)
8182 /* vDRAP is setup but after reload it turns out stack realign
8183 isn't necessary, here we will emit prologue to setup DRAP
8184 without stack realign adjustment */
8185 int drap_bp_offset = UNITS_PER_WORD * 2;
8186 rtx x = plus_constant (hard_frame_pointer_rtx, drap_bp_offset);
8187 insn = emit_insn (gen_rtx_SET (VOIDmode, crtl->drap_reg, x));
8190 /* Emit cld instruction if stringops are used in the function. */
8191 if (TARGET_CLD && ix86_current_function_needs_cld)
8192 emit_insn (gen_cld ());
8195 /* Emit code to restore saved registers using MOV insns. First register
8196 is restored from POINTER + OFFSET. */
8198 ix86_emit_restore_regs_using_mov (rtx pointer, HOST_WIDE_INT offset,
8199 int maybe_eh_return)
8202 rtx base_address = gen_rtx_MEM (Pmode, pointer);
8204 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8205 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, maybe_eh_return))
8207 /* Ensure that adjust_address won't be forced to produce pointer
8208 out of range allowed by x86-64 instruction set. */
8209 if (TARGET_64BIT && offset != trunc_int_for_mode (offset, SImode))
8213 r11 = gen_rtx_REG (DImode, R11_REG);
8214 emit_move_insn (r11, GEN_INT (offset));
8215 emit_insn (gen_adddi3 (r11, r11, pointer));
8216 base_address = gen_rtx_MEM (Pmode, r11);
8219 emit_move_insn (gen_rtx_REG (Pmode, regno),
8220 adjust_address (base_address, Pmode, offset));
8221 offset += UNITS_PER_WORD;
8225 /* Emit code to restore saved registers using MOV insns. First register
8226 is restored from POINTER + OFFSET. */
8228 ix86_emit_restore_sse_regs_using_mov (rtx pointer, HOST_WIDE_INT offset,
8229 int maybe_eh_return)
8232 rtx base_address = gen_rtx_MEM (TImode, pointer);
8235 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8236 if (SSE_REGNO_P (regno) && ix86_save_reg (regno, maybe_eh_return))
8238 /* Ensure that adjust_address won't be forced to produce pointer
8239 out of range allowed by x86-64 instruction set. */
8240 if (TARGET_64BIT && offset != trunc_int_for_mode (offset, SImode))
8244 r11 = gen_rtx_REG (DImode, R11_REG);
8245 emit_move_insn (r11, GEN_INT (offset));
8246 emit_insn (gen_adddi3 (r11, r11, pointer));
8247 base_address = gen_rtx_MEM (TImode, r11);
8250 mem = adjust_address (base_address, TImode, offset);
8251 set_mem_align (mem, 128);
8252 emit_move_insn (gen_rtx_REG (TImode, regno), mem);
8257 /* Restore function stack, frame, and registers. */
8260 ix86_expand_epilogue (int style)
8264 struct ix86_frame frame;
8265 HOST_WIDE_INT offset;
8267 ix86_finalize_stack_realign_flags ();
8269 /* When stack is realigned, SP must be valid. */
8270 sp_valid = (!frame_pointer_needed
8271 || current_function_sp_is_unchanging
8272 || stack_realign_fp);
8274 ix86_compute_frame_layout (&frame);
8276 /* Calculate start of saved registers relative to ebp. Special care
8277 must be taken for the normal return case of a function using
8278 eh_return: the eax and edx registers are marked as saved, but not
8279 restored along this path. */
8280 offset = frame.nregs;
8281 if (crtl->calls_eh_return && style != 2)
8283 offset *= -UNITS_PER_WORD;
8284 offset -= frame.nsseregs * 16 + frame.padding0;
8286 /* If we're only restoring one register and sp is not valid then
8287 using a move instruction to restore the register since it's
8288 less work than reloading sp and popping the register.
8290 The default code result in stack adjustment using add/lea instruction,
8291 while this code results in LEAVE instruction (or discrete equivalent),
8292 so it is profitable in some other cases as well. Especially when there
8293 are no registers to restore. We also use this code when TARGET_USE_LEAVE
8294 and there is exactly one register to pop. This heuristic may need some
8295 tuning in future. */
8296 if ((!sp_valid && frame.nregs <= 1)
8297 || (TARGET_EPILOGUE_USING_MOVE
8298 && cfun->machine->use_fast_prologue_epilogue
8299 && (frame.nregs > 1 || frame.to_allocate))
8300 || (frame_pointer_needed && !frame.nregs && frame.to_allocate)
8301 || (frame_pointer_needed && TARGET_USE_LEAVE
8302 && cfun->machine->use_fast_prologue_epilogue
8303 && frame.nregs == 1)
8304 || crtl->calls_eh_return)
8306 /* Restore registers. We can use ebp or esp to address the memory
8307 locations. If both are available, default to ebp, since offsets
8308 are known to be small. Only exception is esp pointing directly
8309 to the end of block of saved registers, where we may simplify
8312 If we are realigning stack with bp and sp, regs restore can't
8313 be addressed by bp. sp must be used instead. */
8315 if (!frame_pointer_needed
8316 || (sp_valid && !frame.to_allocate)
8317 || stack_realign_fp)
8319 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
8320 frame.to_allocate, style == 2);
8321 ix86_emit_restore_regs_using_mov (stack_pointer_rtx,
8323 + frame.nsseregs * 16
8324 + frame.padding0, style == 2);
8328 ix86_emit_restore_sse_regs_using_mov (hard_frame_pointer_rtx,
8329 offset, style == 2);
8330 ix86_emit_restore_regs_using_mov (hard_frame_pointer_rtx,
8332 + frame.nsseregs * 16
8333 + frame.padding0, style == 2);
8336 /* eh_return epilogues need %ecx added to the stack pointer. */
8339 rtx tmp, sa = EH_RETURN_STACKADJ_RTX;
8341 /* Stack align doesn't work with eh_return. */
8342 gcc_assert (!crtl->stack_realign_needed);
8344 if (frame_pointer_needed)
8346 tmp = gen_rtx_PLUS (Pmode, hard_frame_pointer_rtx, sa);
8347 tmp = plus_constant (tmp, UNITS_PER_WORD);
8348 emit_insn (gen_rtx_SET (VOIDmode, sa, tmp));
8350 tmp = gen_rtx_MEM (Pmode, hard_frame_pointer_rtx);
8351 emit_move_insn (hard_frame_pointer_rtx, tmp);
8353 pro_epilogue_adjust_stack (stack_pointer_rtx, sa,
8358 tmp = gen_rtx_PLUS (Pmode, stack_pointer_rtx, sa);
8359 tmp = plus_constant (tmp, (frame.to_allocate
8360 + frame.nregs * UNITS_PER_WORD
8361 + frame.nsseregs * 16
8363 emit_insn (gen_rtx_SET (VOIDmode, stack_pointer_rtx, tmp));
8366 else if (!frame_pointer_needed)
8367 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
8368 GEN_INT (frame.to_allocate
8369 + frame.nregs * UNITS_PER_WORD
8370 + frame.nsseregs * 16
8373 /* If not an i386, mov & pop is faster than "leave". */
8374 else if (TARGET_USE_LEAVE || optimize_function_for_size_p (cfun)
8375 || !cfun->machine->use_fast_prologue_epilogue)
8376 emit_insn ((*ix86_gen_leave) ());
8379 pro_epilogue_adjust_stack (stack_pointer_rtx,
8380 hard_frame_pointer_rtx,
8383 emit_insn ((*ix86_gen_pop1) (hard_frame_pointer_rtx));
8388 /* First step is to deallocate the stack frame so that we can
8391 If we realign stack with frame pointer, then stack pointer
8392 won't be able to recover via lea $offset(%bp), %sp, because
8393 there is a padding area between bp and sp for realign.
8394 "add $to_allocate, %sp" must be used instead. */
8397 gcc_assert (frame_pointer_needed);
8398 gcc_assert (!stack_realign_fp);
8399 pro_epilogue_adjust_stack (stack_pointer_rtx,
8400 hard_frame_pointer_rtx,
8401 GEN_INT (offset), style);
8402 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
8403 frame.to_allocate, style == 2);
8404 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
8405 GEN_INT (frame.nsseregs * 16), style);
8407 else if (frame.to_allocate || frame.nsseregs)
8409 ix86_emit_restore_sse_regs_using_mov (stack_pointer_rtx,
8412 pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
8413 GEN_INT (frame.to_allocate
8414 + frame.nsseregs * 16
8415 + frame.padding0), style);
8418 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
8419 if (!SSE_REGNO_P (regno) && ix86_save_reg (regno, false))
8420 emit_insn ((*ix86_gen_pop1) (gen_rtx_REG (Pmode, regno)));
8421 if (frame_pointer_needed)
8423 /* Leave results in shorter dependency chains on CPUs that are
8424 able to grok it fast. */
8425 if (TARGET_USE_LEAVE)
8426 emit_insn ((*ix86_gen_leave) ());
8429 /* For stack realigned really happens, recover stack
8430 pointer to hard frame pointer is a must, if not using
8432 if (stack_realign_fp)
8433 pro_epilogue_adjust_stack (stack_pointer_rtx,
8434 hard_frame_pointer_rtx,
8436 emit_insn ((*ix86_gen_pop1) (hard_frame_pointer_rtx));
8441 if (crtl->drap_reg && crtl->stack_realign_needed)
8443 int param_ptr_offset = (call_used_regs[REGNO (crtl->drap_reg)]
8444 ? 0 : UNITS_PER_WORD);
8445 gcc_assert (stack_realign_drap);
8446 emit_insn ((*ix86_gen_add3) (stack_pointer_rtx,
8448 GEN_INT (-(UNITS_PER_WORD
8449 + param_ptr_offset))));
8450 if (!call_used_regs[REGNO (crtl->drap_reg)])
8451 emit_insn ((*ix86_gen_pop1) (crtl->drap_reg));
8455 /* Sibcall epilogues don't want a return instruction. */
8459 if (crtl->args.pops_args && crtl->args.size)
8461 rtx popc = GEN_INT (crtl->args.pops_args);
8463 /* i386 can only pop 64K bytes. If asked to pop more, pop
8464 return address, do explicit add, and jump indirectly to the
8467 if (crtl->args.pops_args >= 65536)
8469 rtx ecx = gen_rtx_REG (SImode, CX_REG);
8471 /* There is no "pascal" calling convention in any 64bit ABI. */
8472 gcc_assert (!TARGET_64BIT);
8474 emit_insn (gen_popsi1 (ecx));
8475 emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, popc));
8476 emit_jump_insn (gen_return_indirect_internal (ecx));
8479 emit_jump_insn (gen_return_pop_internal (popc));
8482 emit_jump_insn (gen_return_internal ());
8485 /* Reset from the function's potential modifications. */
8488 ix86_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
8489 HOST_WIDE_INT size ATTRIBUTE_UNUSED)
8491 if (pic_offset_table_rtx)
8492 SET_REGNO (pic_offset_table_rtx, REAL_PIC_OFFSET_TABLE_REGNUM);
8494 /* Mach-O doesn't support labels at the end of objects, so if
8495 it looks like we might want one, insert a NOP. */
8497 rtx insn = get_last_insn ();
8500 && NOTE_KIND (insn) != NOTE_INSN_DELETED_LABEL)
8501 insn = PREV_INSN (insn);
8505 && NOTE_KIND (insn) == NOTE_INSN_DELETED_LABEL)))
8506 fputs ("\tnop\n", file);
8512 /* Extract the parts of an RTL expression that is a valid memory address
8513 for an instruction. Return 0 if the structure of the address is
8514 grossly off. Return -1 if the address contains ASHIFT, so it is not
8515 strictly valid, but still used for computing length of lea instruction. */
8518 ix86_decompose_address (rtx addr, struct ix86_address *out)
8520 rtx base = NULL_RTX, index = NULL_RTX, disp = NULL_RTX;
8521 rtx base_reg, index_reg;
8522 HOST_WIDE_INT scale = 1;
8523 rtx scale_rtx = NULL_RTX;
8525 enum ix86_address_seg seg = SEG_DEFAULT;
8527 if (REG_P (addr) || GET_CODE (addr) == SUBREG)
8529 else if (GET_CODE (addr) == PLUS)
8539 addends[n++] = XEXP (op, 1);
8542 while (GET_CODE (op) == PLUS);
8547 for (i = n; i >= 0; --i)
8550 switch (GET_CODE (op))
8555 index = XEXP (op, 0);
8556 scale_rtx = XEXP (op, 1);
8560 if (XINT (op, 1) == UNSPEC_TP
8561 && TARGET_TLS_DIRECT_SEG_REFS
8562 && seg == SEG_DEFAULT)
8563 seg = TARGET_64BIT ? SEG_FS : SEG_GS;
8592 else if (GET_CODE (addr) == MULT)
8594 index = XEXP (addr, 0); /* index*scale */
8595 scale_rtx = XEXP (addr, 1);
8597 else if (GET_CODE (addr) == ASHIFT)
8601 /* We're called for lea too, which implements ashift on occasion. */
8602 index = XEXP (addr, 0);
8603 tmp = XEXP (addr, 1);
8604 if (!CONST_INT_P (tmp))
8606 scale = INTVAL (tmp);
8607 if ((unsigned HOST_WIDE_INT) scale > 3)
8613 disp = addr; /* displacement */
8615 /* Extract the integral value of scale. */
8618 if (!CONST_INT_P (scale_rtx))
8620 scale = INTVAL (scale_rtx);
8623 base_reg = base && GET_CODE (base) == SUBREG ? SUBREG_REG (base) : base;
8624 index_reg = index && GET_CODE (index) == SUBREG ? SUBREG_REG (index) : index;
8626 /* Allow arg pointer and stack pointer as index if there is not scaling. */
8627 if (base_reg && index_reg && scale == 1
8628 && (index_reg == arg_pointer_rtx
8629 || index_reg == frame_pointer_rtx
8630 || (REG_P (index_reg) && REGNO (index_reg) == STACK_POINTER_REGNUM)))
8633 tmp = base, base = index, index = tmp;
8634 tmp = base_reg, base_reg = index_reg, index_reg = tmp;
8637 /* Special case: %ebp cannot be encoded as a base without a displacement. */
8638 if ((base_reg == hard_frame_pointer_rtx
8639 || base_reg == frame_pointer_rtx
8640 || base_reg == arg_pointer_rtx) && !disp)
8643 /* Special case: on K6, [%esi] makes the instruction vector decoded.
8644 Avoid this by transforming to [%esi+0].
8645 Reload calls address legitimization without cfun defined, so we need
8646 to test cfun for being non-NULL. */
8647 if (TARGET_K6 && cfun && optimize_function_for_speed_p (cfun)
8648 && base_reg && !index_reg && !disp
8650 && REGNO_REG_CLASS (REGNO (base_reg)) == SIREG)
8653 /* Special case: encode reg+reg instead of reg*2. */
8654 if (!base && index && scale && scale == 2)
8655 base = index, base_reg = index_reg, scale = 1;
8657 /* Special case: scaling cannot be encoded without base or displacement. */
8658 if (!base && !disp && index && scale != 1)
8670 /* Return cost of the memory address x.
8671 For i386, it is better to use a complex address than let gcc copy
8672 the address into a reg and make a new pseudo. But not if the address
8673 requires to two regs - that would mean more pseudos with longer
8676 ix86_address_cost (rtx x, bool speed ATTRIBUTE_UNUSED)
8678 struct ix86_address parts;
8680 int ok = ix86_decompose_address (x, &parts);
8684 if (parts.base && GET_CODE (parts.base) == SUBREG)
8685 parts.base = SUBREG_REG (parts.base);
8686 if (parts.index && GET_CODE (parts.index) == SUBREG)
8687 parts.index = SUBREG_REG (parts.index);
8689 /* Attempt to minimize number of registers in the address. */
8691 && (!REG_P (parts.base) || REGNO (parts.base) >= FIRST_PSEUDO_REGISTER))
8693 && (!REG_P (parts.index)
8694 || REGNO (parts.index) >= FIRST_PSEUDO_REGISTER)))
8698 && (!REG_P (parts.base) || REGNO (parts.base) >= FIRST_PSEUDO_REGISTER)
8700 && (!REG_P (parts.index) || REGNO (parts.index) >= FIRST_PSEUDO_REGISTER)
8701 && parts.base != parts.index)
8704 /* AMD-K6 don't like addresses with ModR/M set to 00_xxx_100b,
8705 since it's predecode logic can't detect the length of instructions
8706 and it degenerates to vector decoded. Increase cost of such
8707 addresses here. The penalty is minimally 2 cycles. It may be worthwhile
8708 to split such addresses or even refuse such addresses at all.
8710 Following addressing modes are affected:
8715 The first and last case may be avoidable by explicitly coding the zero in
8716 memory address, but I don't have AMD-K6 machine handy to check this
8720 && ((!parts.disp && parts.base && parts.index && parts.scale != 1)
8721 || (parts.disp && !parts.base && parts.index && parts.scale != 1)
8722 || (!parts.disp && parts.base && parts.index && parts.scale == 1)))
8728 /* Allow {LABEL | SYMBOL}_REF - SYMBOL_REF-FOR-PICBASE for Mach-O as
8729 this is used for to form addresses to local data when -fPIC is in
8733 darwin_local_data_pic (rtx disp)
8735 return (GET_CODE (disp) == UNSPEC
8736 && XINT (disp, 1) == UNSPEC_MACHOPIC_OFFSET);
8739 /* Determine if a given RTX is a valid constant. We already know this
8740 satisfies CONSTANT_P. */
8743 legitimate_constant_p (rtx x)
8745 switch (GET_CODE (x))
8750 if (GET_CODE (x) == PLUS)
8752 if (!CONST_INT_P (XEXP (x, 1)))
8757 if (TARGET_MACHO && darwin_local_data_pic (x))
8760 /* Only some unspecs are valid as "constants". */
8761 if (GET_CODE (x) == UNSPEC)
8762 switch (XINT (x, 1))
8767 return TARGET_64BIT;
8770 x = XVECEXP (x, 0, 0);
8771 return (GET_CODE (x) == SYMBOL_REF
8772 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_EXEC);
8774 x = XVECEXP (x, 0, 0);
8775 return (GET_CODE (x) == SYMBOL_REF
8776 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC);
8781 /* We must have drilled down to a symbol. */
8782 if (GET_CODE (x) == LABEL_REF)
8784 if (GET_CODE (x) != SYMBOL_REF)
8789 /* TLS symbols are never valid. */
8790 if (SYMBOL_REF_TLS_MODEL (x))
8793 /* DLLIMPORT symbols are never valid. */
8794 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
8795 && SYMBOL_REF_DLLIMPORT_P (x))
8800 if (GET_MODE (x) == TImode
8801 && x != CONST0_RTX (TImode)
8807 if (x == CONST0_RTX (GET_MODE (x)))
8815 /* Otherwise we handle everything else in the move patterns. */
8819 /* Determine if it's legal to put X into the constant pool. This
8820 is not possible for the address of thread-local symbols, which
8821 is checked above. */
8824 ix86_cannot_force_const_mem (rtx x)
8826 /* We can always put integral constants and vectors in memory. */
8827 switch (GET_CODE (x))
8837 return !legitimate_constant_p (x);
8840 /* Determine if a given RTX is a valid constant address. */
8843 constant_address_p (rtx x)
8845 return CONSTANT_P (x) && legitimate_address_p (Pmode, x, 1);
8848 /* Nonzero if the constant value X is a legitimate general operand
8849 when generating PIC code. It is given that flag_pic is on and
8850 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
8853 legitimate_pic_operand_p (rtx x)
8857 switch (GET_CODE (x))
8860 inner = XEXP (x, 0);
8861 if (GET_CODE (inner) == PLUS
8862 && CONST_INT_P (XEXP (inner, 1)))
8863 inner = XEXP (inner, 0);
8865 /* Only some unspecs are valid as "constants". */
8866 if (GET_CODE (inner) == UNSPEC)
8867 switch (XINT (inner, 1))
8872 return TARGET_64BIT;
8874 x = XVECEXP (inner, 0, 0);
8875 return (GET_CODE (x) == SYMBOL_REF
8876 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_EXEC);
8877 case UNSPEC_MACHOPIC_OFFSET:
8878 return legitimate_pic_address_disp_p (x);
8886 return legitimate_pic_address_disp_p (x);
8893 /* Determine if a given CONST RTX is a valid memory displacement
8897 legitimate_pic_address_disp_p (rtx disp)
8901 /* In 64bit mode we can allow direct addresses of symbols and labels
8902 when they are not dynamic symbols. */
8905 rtx op0 = disp, op1;
8907 switch (GET_CODE (disp))
8913 if (GET_CODE (XEXP (disp, 0)) != PLUS)
8915 op0 = XEXP (XEXP (disp, 0), 0);
8916 op1 = XEXP (XEXP (disp, 0), 1);
8917 if (!CONST_INT_P (op1)
8918 || INTVAL (op1) >= 16*1024*1024
8919 || INTVAL (op1) < -16*1024*1024)
8921 if (GET_CODE (op0) == LABEL_REF)
8923 if (GET_CODE (op0) != SYMBOL_REF)
8928 /* TLS references should always be enclosed in UNSPEC. */
8929 if (SYMBOL_REF_TLS_MODEL (op0))
8931 if (!SYMBOL_REF_FAR_ADDR_P (op0) && SYMBOL_REF_LOCAL_P (op0)
8932 && ix86_cmodel != CM_LARGE_PIC)
8940 if (GET_CODE (disp) != CONST)
8942 disp = XEXP (disp, 0);
8946 /* We are unsafe to allow PLUS expressions. This limit allowed distance
8947 of GOT tables. We should not need these anyway. */
8948 if (GET_CODE (disp) != UNSPEC
8949 || (XINT (disp, 1) != UNSPEC_GOTPCREL
8950 && XINT (disp, 1) != UNSPEC_GOTOFF
8951 && XINT (disp, 1) != UNSPEC_PLTOFF))
8954 if (GET_CODE (XVECEXP (disp, 0, 0)) != SYMBOL_REF
8955 && GET_CODE (XVECEXP (disp, 0, 0)) != LABEL_REF)
8961 if (GET_CODE (disp) == PLUS)
8963 if (!CONST_INT_P (XEXP (disp, 1)))
8965 disp = XEXP (disp, 0);
8969 if (TARGET_MACHO && darwin_local_data_pic (disp))
8972 if (GET_CODE (disp) != UNSPEC)
8975 switch (XINT (disp, 1))
8980 /* We need to check for both symbols and labels because VxWorks loads
8981 text labels with @GOT rather than @GOTOFF. See gotoff_operand for
8983 return (GET_CODE (XVECEXP (disp, 0, 0)) == SYMBOL_REF
8984 || GET_CODE (XVECEXP (disp, 0, 0)) == LABEL_REF);
8986 /* Refuse GOTOFF in 64bit mode since it is always 64bit when used.
8987 While ABI specify also 32bit relocation but we don't produce it in
8988 small PIC model at all. */
8989 if ((GET_CODE (XVECEXP (disp, 0, 0)) == SYMBOL_REF
8990 || GET_CODE (XVECEXP (disp, 0, 0)) == LABEL_REF)
8992 return gotoff_operand (XVECEXP (disp, 0, 0), Pmode);
8994 case UNSPEC_GOTTPOFF:
8995 case UNSPEC_GOTNTPOFF:
8996 case UNSPEC_INDNTPOFF:
8999 disp = XVECEXP (disp, 0, 0);
9000 return (GET_CODE (disp) == SYMBOL_REF
9001 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_INITIAL_EXEC);
9003 disp = XVECEXP (disp, 0, 0);
9004 return (GET_CODE (disp) == SYMBOL_REF
9005 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_LOCAL_EXEC);
9007 disp = XVECEXP (disp, 0, 0);
9008 return (GET_CODE (disp) == SYMBOL_REF
9009 && SYMBOL_REF_TLS_MODEL (disp) == TLS_MODEL_LOCAL_DYNAMIC);
9015 /* GO_IF_LEGITIMATE_ADDRESS recognizes an RTL expression that is a valid
9016 memory address for an instruction. The MODE argument is the machine mode
9017 for the MEM expression that wants to use this address.
9019 It only recognizes address in canonical form. LEGITIMIZE_ADDRESS should
9020 convert common non-canonical forms to canonical form so that they will
9024 legitimate_address_p (enum machine_mode mode ATTRIBUTE_UNUSED,
9025 rtx addr, int strict)
9027 struct ix86_address parts;
9028 rtx base, index, disp;
9029 HOST_WIDE_INT scale;
9030 const char *reason = NULL;
9031 rtx reason_rtx = NULL_RTX;
9033 if (ix86_decompose_address (addr, &parts) <= 0)
9035 reason = "decomposition failed";
9040 index = parts.index;
9042 scale = parts.scale;
9044 /* Validate base register.
9046 Don't allow SUBREG's that span more than a word here. It can lead to spill
9047 failures when the base is one word out of a two word structure, which is
9048 represented internally as a DImode int. */
9057 else if (GET_CODE (base) == SUBREG
9058 && REG_P (SUBREG_REG (base))
9059 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (base)))
9061 reg = SUBREG_REG (base);
9064 reason = "base is not a register";
9068 if (GET_MODE (base) != Pmode)
9070 reason = "base is not in Pmode";
9074 if ((strict && ! REG_OK_FOR_BASE_STRICT_P (reg))
9075 || (! strict && ! REG_OK_FOR_BASE_NONSTRICT_P (reg)))
9077 reason = "base is not valid";
9082 /* Validate index register.
9084 Don't allow SUBREG's that span more than a word here -- same as above. */
9093 else if (GET_CODE (index) == SUBREG
9094 && REG_P (SUBREG_REG (index))
9095 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (index)))
9097 reg = SUBREG_REG (index);
9100 reason = "index is not a register";
9104 if (GET_MODE (index) != Pmode)
9106 reason = "index is not in Pmode";
9110 if ((strict && ! REG_OK_FOR_INDEX_STRICT_P (reg))
9111 || (! strict && ! REG_OK_FOR_INDEX_NONSTRICT_P (reg)))
9113 reason = "index is not valid";
9118 /* Validate scale factor. */
9121 reason_rtx = GEN_INT (scale);
9124 reason = "scale without index";
9128 if (scale != 2 && scale != 4 && scale != 8)
9130 reason = "scale is not a valid multiplier";
9135 /* Validate displacement. */
9140 if (GET_CODE (disp) == CONST
9141 && GET_CODE (XEXP (disp, 0)) == UNSPEC
9142 && XINT (XEXP (disp, 0), 1) != UNSPEC_MACHOPIC_OFFSET)
9143 switch (XINT (XEXP (disp, 0), 1))
9145 /* Refuse GOTOFF and GOT in 64bit mode since it is always 64bit when
9146 used. While ABI specify also 32bit relocations, we don't produce
9147 them at all and use IP relative instead. */
9150 gcc_assert (flag_pic);
9152 goto is_legitimate_pic;
9153 reason = "64bit address unspec";
9156 case UNSPEC_GOTPCREL:
9157 gcc_assert (flag_pic);
9158 goto is_legitimate_pic;
9160 case UNSPEC_GOTTPOFF:
9161 case UNSPEC_GOTNTPOFF:
9162 case UNSPEC_INDNTPOFF:
9168 reason = "invalid address unspec";
9172 else if (SYMBOLIC_CONST (disp)
9176 && MACHOPIC_INDIRECT
9177 && !machopic_operand_p (disp)
9183 if (TARGET_64BIT && (index || base))
9185 /* foo@dtpoff(%rX) is ok. */
9186 if (GET_CODE (disp) != CONST
9187 || GET_CODE (XEXP (disp, 0)) != PLUS
9188 || GET_CODE (XEXP (XEXP (disp, 0), 0)) != UNSPEC
9189 || !CONST_INT_P (XEXP (XEXP (disp, 0), 1))
9190 || (XINT (XEXP (XEXP (disp, 0), 0), 1) != UNSPEC_DTPOFF
9191 && XINT (XEXP (XEXP (disp, 0), 0), 1) != UNSPEC_NTPOFF))
9193 reason = "non-constant pic memory reference";
9197 else if (! legitimate_pic_address_disp_p (disp))
9199 reason = "displacement is an invalid pic construct";
9203 /* This code used to verify that a symbolic pic displacement
9204 includes the pic_offset_table_rtx register.
9206 While this is good idea, unfortunately these constructs may
9207 be created by "adds using lea" optimization for incorrect
9216 This code is nonsensical, but results in addressing
9217 GOT table with pic_offset_table_rtx base. We can't
9218 just refuse it easily, since it gets matched by
9219 "addsi3" pattern, that later gets split to lea in the
9220 case output register differs from input. While this
9221 can be handled by separate addsi pattern for this case
9222 that never results in lea, this seems to be easier and
9223 correct fix for crash to disable this test. */
9225 else if (GET_CODE (disp) != LABEL_REF
9226 && !CONST_INT_P (disp)
9227 && (GET_CODE (disp) != CONST
9228 || !legitimate_constant_p (disp))
9229 && (GET_CODE (disp) != SYMBOL_REF
9230 || !legitimate_constant_p (disp)))
9232 reason = "displacement is not constant";
9235 else if (TARGET_64BIT
9236 && !x86_64_immediate_operand (disp, VOIDmode))
9238 reason = "displacement is out of range";
9243 /* Everything looks valid. */
9250 /* Return a unique alias set for the GOT. */
9252 static alias_set_type
9253 ix86_GOT_alias_set (void)
9255 static alias_set_type set = -1;
9257 set = new_alias_set ();
9261 /* Return a legitimate reference for ORIG (an address) using the
9262 register REG. If REG is 0, a new pseudo is generated.
9264 There are two types of references that must be handled:
9266 1. Global data references must load the address from the GOT, via
9267 the PIC reg. An insn is emitted to do this load, and the reg is
9270 2. Static data references, constant pool addresses, and code labels
9271 compute the address as an offset from the GOT, whose base is in
9272 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
9273 differentiate them from global data objects. The returned
9274 address is the PIC reg + an unspec constant.
9276 GO_IF_LEGITIMATE_ADDRESS rejects symbolic references unless the PIC
9277 reg also appears in the address. */
9280 legitimize_pic_address (rtx orig, rtx reg)
9287 if (TARGET_MACHO && !TARGET_64BIT)
9290 reg = gen_reg_rtx (Pmode);
9291 /* Use the generic Mach-O PIC machinery. */
9292 return machopic_legitimize_pic_address (orig, GET_MODE (orig), reg);
9296 if (TARGET_64BIT && legitimate_pic_address_disp_p (addr))
9298 else if (TARGET_64BIT
9299 && ix86_cmodel != CM_SMALL_PIC
9300 && gotoff_operand (addr, Pmode))
9303 /* This symbol may be referenced via a displacement from the PIC
9304 base address (@GOTOFF). */
9306 if (reload_in_progress)
9307 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
9308 if (GET_CODE (addr) == CONST)
9309 addr = XEXP (addr, 0);
9310 if (GET_CODE (addr) == PLUS)
9312 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, XEXP (addr, 0)),
9314 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, XEXP (addr, 1));
9317 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTOFF);
9318 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
9320 tmpreg = gen_reg_rtx (Pmode);
9323 emit_move_insn (tmpreg, new_rtx);
9327 new_rtx = expand_simple_binop (Pmode, PLUS, reg, pic_offset_table_rtx,
9328 tmpreg, 1, OPTAB_DIRECT);
9331 else new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, tmpreg);
9333 else if (!TARGET_64BIT && gotoff_operand (addr, Pmode))
9335 /* This symbol may be referenced via a displacement from the PIC
9336 base address (@GOTOFF). */
9338 if (reload_in_progress)
9339 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
9340 if (GET_CODE (addr) == CONST)
9341 addr = XEXP (addr, 0);
9342 if (GET_CODE (addr) == PLUS)
9344 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, XEXP (addr, 0)),
9346 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, XEXP (addr, 1));
9349 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTOFF);
9350 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
9351 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
9355 emit_move_insn (reg, new_rtx);
9359 else if ((GET_CODE (addr) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (addr) == 0)
9360 /* We can't use @GOTOFF for text labels on VxWorks;
9361 see gotoff_operand. */
9362 || (TARGET_VXWORKS_RTP && GET_CODE (addr) == LABEL_REF))
9364 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES)
9366 if (GET_CODE (addr) == SYMBOL_REF && SYMBOL_REF_DLLIMPORT_P (addr))
9367 return legitimize_dllimport_symbol (addr, true);
9368 if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS
9369 && GET_CODE (XEXP (XEXP (addr, 0), 0)) == SYMBOL_REF
9370 && SYMBOL_REF_DLLIMPORT_P (XEXP (XEXP (addr, 0), 0)))
9372 rtx t = legitimize_dllimport_symbol (XEXP (XEXP (addr, 0), 0), true);
9373 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (addr, 0), 1));
9377 if (TARGET_64BIT && ix86_cmodel != CM_LARGE_PIC)
9379 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOTPCREL);
9380 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
9381 new_rtx = gen_const_mem (Pmode, new_rtx);
9382 set_mem_alias_set (new_rtx, ix86_GOT_alias_set ());
9385 reg = gen_reg_rtx (Pmode);
9386 /* Use directly gen_movsi, otherwise the address is loaded
9387 into register for CSE. We don't want to CSE this addresses,
9388 instead we CSE addresses from the GOT table, so skip this. */
9389 emit_insn (gen_movsi (reg, new_rtx));
9394 /* This symbol must be referenced via a load from the
9395 Global Offset Table (@GOT). */
9397 if (reload_in_progress)
9398 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
9399 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), UNSPEC_GOT);
9400 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
9402 new_rtx = force_reg (Pmode, new_rtx);
9403 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
9404 new_rtx = gen_const_mem (Pmode, new_rtx);
9405 set_mem_alias_set (new_rtx, ix86_GOT_alias_set ());
9408 reg = gen_reg_rtx (Pmode);
9409 emit_move_insn (reg, new_rtx);
9415 if (CONST_INT_P (addr)
9416 && !x86_64_immediate_operand (addr, VOIDmode))
9420 emit_move_insn (reg, addr);
9424 new_rtx = force_reg (Pmode, addr);
9426 else if (GET_CODE (addr) == CONST)
9428 addr = XEXP (addr, 0);
9430 /* We must match stuff we generate before. Assume the only
9431 unspecs that can get here are ours. Not that we could do
9432 anything with them anyway.... */
9433 if (GET_CODE (addr) == UNSPEC
9434 || (GET_CODE (addr) == PLUS
9435 && GET_CODE (XEXP (addr, 0)) == UNSPEC))
9437 gcc_assert (GET_CODE (addr) == PLUS);
9439 if (GET_CODE (addr) == PLUS)
9441 rtx op0 = XEXP (addr, 0), op1 = XEXP (addr, 1);
9443 /* Check first to see if this is a constant offset from a @GOTOFF
9444 symbol reference. */
9445 if (gotoff_operand (op0, Pmode)
9446 && CONST_INT_P (op1))
9450 if (reload_in_progress)
9451 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
9452 new_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, op0),
9454 new_rtx = gen_rtx_PLUS (Pmode, new_rtx, op1);
9455 new_rtx = gen_rtx_CONST (Pmode, new_rtx);
9456 new_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new_rtx);
9460 emit_move_insn (reg, new_rtx);
9466 if (INTVAL (op1) < -16*1024*1024
9467 || INTVAL (op1) >= 16*1024*1024)
9469 if (!x86_64_immediate_operand (op1, Pmode))
9470 op1 = force_reg (Pmode, op1);
9471 new_rtx = gen_rtx_PLUS (Pmode, force_reg (Pmode, op0), op1);
9477 base = legitimize_pic_address (XEXP (addr, 0), reg);
9478 new_rtx = legitimize_pic_address (XEXP (addr, 1),
9479 base == reg ? NULL_RTX : reg);
9481 if (CONST_INT_P (new_rtx))
9482 new_rtx = plus_constant (base, INTVAL (new_rtx));
9485 if (GET_CODE (new_rtx) == PLUS && CONSTANT_P (XEXP (new_rtx, 1)))
9487 base = gen_rtx_PLUS (Pmode, base, XEXP (new_rtx, 0));
9488 new_rtx = XEXP (new_rtx, 1);
9490 new_rtx = gen_rtx_PLUS (Pmode, base, new_rtx);
9498 /* Load the thread pointer. If TO_REG is true, force it into a register. */
9501 get_thread_pointer (int to_reg)
9505 tp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), UNSPEC_TP);
9509 reg = gen_reg_rtx (Pmode);
9510 insn = gen_rtx_SET (VOIDmode, reg, tp);
9511 insn = emit_insn (insn);
9516 /* A subroutine of legitimize_address and ix86_expand_move. FOR_MOV is
9517 false if we expect this to be used for a memory address and true if
9518 we expect to load the address into a register. */
9521 legitimize_tls_address (rtx x, enum tls_model model, int for_mov)
9523 rtx dest, base, off, pic, tp;
9528 case TLS_MODEL_GLOBAL_DYNAMIC:
9529 dest = gen_reg_rtx (Pmode);
9530 tp = TARGET_GNU2_TLS ? get_thread_pointer (1) : 0;
9532 if (TARGET_64BIT && ! TARGET_GNU2_TLS)
9534 rtx rax = gen_rtx_REG (Pmode, AX_REG), insns;
9537 emit_call_insn (gen_tls_global_dynamic_64 (rax, x));
9538 insns = get_insns ();
9541 RTL_CONST_CALL_P (insns) = 1;
9542 emit_libcall_block (insns, dest, rax, x);
9544 else if (TARGET_64BIT && TARGET_GNU2_TLS)
9545 emit_insn (gen_tls_global_dynamic_64 (dest, x));
9547 emit_insn (gen_tls_global_dynamic_32 (dest, x));
9549 if (TARGET_GNU2_TLS)
9551 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, tp, dest));
9553 set_unique_reg_note (get_last_insn (), REG_EQUIV, x);
9557 case TLS_MODEL_LOCAL_DYNAMIC:
9558 base = gen_reg_rtx (Pmode);
9559 tp = TARGET_GNU2_TLS ? get_thread_pointer (1) : 0;
9561 if (TARGET_64BIT && ! TARGET_GNU2_TLS)
9563 rtx rax = gen_rtx_REG (Pmode, AX_REG), insns, note;
9566 emit_call_insn (gen_tls_local_dynamic_base_64 (rax));
9567 insns = get_insns ();
9570 note = gen_rtx_EXPR_LIST (VOIDmode, const0_rtx, NULL);
9571 note = gen_rtx_EXPR_LIST (VOIDmode, ix86_tls_get_addr (), note);
9572 RTL_CONST_CALL_P (insns) = 1;
9573 emit_libcall_block (insns, base, rax, note);
9575 else if (TARGET_64BIT && TARGET_GNU2_TLS)
9576 emit_insn (gen_tls_local_dynamic_base_64 (base));
9578 emit_insn (gen_tls_local_dynamic_base_32 (base));
9580 if (TARGET_GNU2_TLS)
9582 rtx x = ix86_tls_module_base ();
9584 set_unique_reg_note (get_last_insn (), REG_EQUIV,
9585 gen_rtx_MINUS (Pmode, x, tp));
9588 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), UNSPEC_DTPOFF);
9589 off = gen_rtx_CONST (Pmode, off);
9591 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, base, off));
9593 if (TARGET_GNU2_TLS)
9595 dest = force_reg (Pmode, gen_rtx_PLUS (Pmode, dest, tp));
9597 set_unique_reg_note (get_last_insn (), REG_EQUIV, x);
9602 case TLS_MODEL_INITIAL_EXEC:
9606 type = UNSPEC_GOTNTPOFF;
9610 if (reload_in_progress)
9611 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
9612 pic = pic_offset_table_rtx;
9613 type = TARGET_ANY_GNU_TLS ? UNSPEC_GOTNTPOFF : UNSPEC_GOTTPOFF;
9615 else if (!TARGET_ANY_GNU_TLS)
9617 pic = gen_reg_rtx (Pmode);
9618 emit_insn (gen_set_got (pic));
9619 type = UNSPEC_GOTTPOFF;
9624 type = UNSPEC_INDNTPOFF;
9627 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x), type);
9628 off = gen_rtx_CONST (Pmode, off);
9630 off = gen_rtx_PLUS (Pmode, pic, off);
9631 off = gen_const_mem (Pmode, off);
9632 set_mem_alias_set (off, ix86_GOT_alias_set ());
9634 if (TARGET_64BIT || TARGET_ANY_GNU_TLS)
9636 base = get_thread_pointer (for_mov || !TARGET_TLS_DIRECT_SEG_REFS);
9637 off = force_reg (Pmode, off);
9638 return gen_rtx_PLUS (Pmode, base, off);
9642 base = get_thread_pointer (true);
9643 dest = gen_reg_rtx (Pmode);
9644 emit_insn (gen_subsi3 (dest, base, off));
9648 case TLS_MODEL_LOCAL_EXEC:
9649 off = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, x),
9650 (TARGET_64BIT || TARGET_ANY_GNU_TLS)
9651 ? UNSPEC_NTPOFF : UNSPEC_TPOFF);
9652 off = gen_rtx_CONST (Pmode, off);
9654 if (TARGET_64BIT || TARGET_ANY_GNU_TLS)
9656 base = get_thread_pointer (for_mov || !TARGET_TLS_DIRECT_SEG_REFS);
9657 return gen_rtx_PLUS (Pmode, base, off);
9661 base = get_thread_pointer (true);
9662 dest = gen_reg_rtx (Pmode);
9663 emit_insn (gen_subsi3 (dest, base, off));
9674 /* Create or return the unique __imp_DECL dllimport symbol corresponding
9677 static GTY((if_marked ("tree_map_marked_p"), param_is (struct tree_map)))
9678 htab_t dllimport_map;
9681 get_dllimport_decl (tree decl)
9683 struct tree_map *h, in;
9687 size_t namelen, prefixlen;
9693 dllimport_map = htab_create_ggc (512, tree_map_hash, tree_map_eq, 0);
9695 in.hash = htab_hash_pointer (decl);
9696 in.base.from = decl;
9697 loc = htab_find_slot_with_hash (dllimport_map, &in, in.hash, INSERT);
9698 h = (struct tree_map *) *loc;
9702 *loc = h = GGC_NEW (struct tree_map);
9704 h->base.from = decl;
9705 h->to = to = build_decl (VAR_DECL, NULL, ptr_type_node);
9706 DECL_ARTIFICIAL (to) = 1;
9707 DECL_IGNORED_P (to) = 1;
9708 DECL_EXTERNAL (to) = 1;
9709 TREE_READONLY (to) = 1;
9711 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
9712 name = targetm.strip_name_encoding (name);
9713 prefix = name[0] == FASTCALL_PREFIX || user_label_prefix[0] == 0
9714 ? "*__imp_" : "*__imp__";
9715 namelen = strlen (name);
9716 prefixlen = strlen (prefix);
9717 imp_name = (char *) alloca (namelen + prefixlen + 1);
9718 memcpy (imp_name, prefix, prefixlen);
9719 memcpy (imp_name + prefixlen, name, namelen + 1);
9721 name = ggc_alloc_string (imp_name, namelen + prefixlen);
9722 rtl = gen_rtx_SYMBOL_REF (Pmode, name);
9723 SET_SYMBOL_REF_DECL (rtl, to);
9724 SYMBOL_REF_FLAGS (rtl) = SYMBOL_FLAG_LOCAL;
9726 rtl = gen_const_mem (Pmode, rtl);
9727 set_mem_alias_set (rtl, ix86_GOT_alias_set ());
9729 SET_DECL_RTL (to, rtl);
9730 SET_DECL_ASSEMBLER_NAME (to, get_identifier (name));
9735 /* Expand SYMBOL into its corresponding dllimport symbol. WANT_REG is
9736 true if we require the result be a register. */
9739 legitimize_dllimport_symbol (rtx symbol, bool want_reg)
9744 gcc_assert (SYMBOL_REF_DECL (symbol));
9745 imp_decl = get_dllimport_decl (SYMBOL_REF_DECL (symbol));
9747 x = DECL_RTL (imp_decl);
9749 x = force_reg (Pmode, x);
9753 /* Try machine-dependent ways of modifying an illegitimate address
9754 to be legitimate. If we find one, return the new, valid address.
9755 This macro is used in only one place: `memory_address' in explow.c.
9757 OLDX is the address as it was before break_out_memory_refs was called.
9758 In some cases it is useful to look at this to decide what needs to be done.
9760 MODE and WIN are passed so that this macro can use
9761 GO_IF_LEGITIMATE_ADDRESS.
9763 It is always safe for this macro to do nothing. It exists to recognize
9764 opportunities to optimize the output.
9766 For the 80386, we handle X+REG by loading X into a register R and
9767 using R+REG. R will go in a general reg and indexing will be used.
9768 However, if REG is a broken-out memory address or multiplication,
9769 nothing needs to be done because REG can certainly go in a general reg.
9771 When -fpic is used, special handling is needed for symbolic references.
9772 See comments by legitimize_pic_address in i386.c for details. */
9775 legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, enum machine_mode mode)
9780 log = GET_CODE (x) == SYMBOL_REF ? SYMBOL_REF_TLS_MODEL (x) : 0;
9782 return legitimize_tls_address (x, (enum tls_model) log, false);
9783 if (GET_CODE (x) == CONST
9784 && GET_CODE (XEXP (x, 0)) == PLUS
9785 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
9786 && (log = SYMBOL_REF_TLS_MODEL (XEXP (XEXP (x, 0), 0))))
9788 rtx t = legitimize_tls_address (XEXP (XEXP (x, 0), 0),
9789 (enum tls_model) log, false);
9790 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (x, 0), 1));
9793 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES)
9795 if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_DLLIMPORT_P (x))
9796 return legitimize_dllimport_symbol (x, true);
9797 if (GET_CODE (x) == CONST
9798 && GET_CODE (XEXP (x, 0)) == PLUS
9799 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
9800 && SYMBOL_REF_DLLIMPORT_P (XEXP (XEXP (x, 0), 0)))
9802 rtx t = legitimize_dllimport_symbol (XEXP (XEXP (x, 0), 0), true);
9803 return gen_rtx_PLUS (Pmode, t, XEXP (XEXP (x, 0), 1));
9807 if (flag_pic && SYMBOLIC_CONST (x))
9808 return legitimize_pic_address (x, 0);
9810 /* Canonicalize shifts by 0, 1, 2, 3 into multiply */
9811 if (GET_CODE (x) == ASHIFT
9812 && CONST_INT_P (XEXP (x, 1))
9813 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (x, 1)) < 4)
9816 log = INTVAL (XEXP (x, 1));
9817 x = gen_rtx_MULT (Pmode, force_reg (Pmode, XEXP (x, 0)),
9818 GEN_INT (1 << log));
9821 if (GET_CODE (x) == PLUS)
9823 /* Canonicalize shifts by 0, 1, 2, 3 into multiply. */
9825 if (GET_CODE (XEXP (x, 0)) == ASHIFT
9826 && CONST_INT_P (XEXP (XEXP (x, 0), 1))
9827 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (x, 0), 1)) < 4)
9830 log = INTVAL (XEXP (XEXP (x, 0), 1));
9831 XEXP (x, 0) = gen_rtx_MULT (Pmode,
9832 force_reg (Pmode, XEXP (XEXP (x, 0), 0)),
9833 GEN_INT (1 << log));
9836 if (GET_CODE (XEXP (x, 1)) == ASHIFT
9837 && CONST_INT_P (XEXP (XEXP (x, 1), 1))
9838 && (unsigned HOST_WIDE_INT) INTVAL (XEXP (XEXP (x, 1), 1)) < 4)
9841 log = INTVAL (XEXP (XEXP (x, 1), 1));
9842 XEXP (x, 1) = gen_rtx_MULT (Pmode,
9843 force_reg (Pmode, XEXP (XEXP (x, 1), 0)),
9844 GEN_INT (1 << log));
9847 /* Put multiply first if it isn't already. */
9848 if (GET_CODE (XEXP (x, 1)) == MULT)
9850 rtx tmp = XEXP (x, 0);
9851 XEXP (x, 0) = XEXP (x, 1);
9856 /* Canonicalize (plus (mult (reg) (const)) (plus (reg) (const)))
9857 into (plus (plus (mult (reg) (const)) (reg)) (const)). This can be
9858 created by virtual register instantiation, register elimination, and
9859 similar optimizations. */
9860 if (GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == PLUS)
9863 x = gen_rtx_PLUS (Pmode,
9864 gen_rtx_PLUS (Pmode, XEXP (x, 0),
9865 XEXP (XEXP (x, 1), 0)),
9866 XEXP (XEXP (x, 1), 1));
9870 (plus (plus (mult (reg) (const)) (plus (reg) (const))) const)
9871 into (plus (plus (mult (reg) (const)) (reg)) (const)). */
9872 else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS
9873 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
9874 && GET_CODE (XEXP (XEXP (x, 0), 1)) == PLUS
9875 && CONSTANT_P (XEXP (x, 1)))
9878 rtx other = NULL_RTX;
9880 if (CONST_INT_P (XEXP (x, 1)))
9882 constant = XEXP (x, 1);
9883 other = XEXP (XEXP (XEXP (x, 0), 1), 1);
9885 else if (CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 1), 1)))
9887 constant = XEXP (XEXP (XEXP (x, 0), 1), 1);
9888 other = XEXP (x, 1);
9896 x = gen_rtx_PLUS (Pmode,
9897 gen_rtx_PLUS (Pmode, XEXP (XEXP (x, 0), 0),
9898 XEXP (XEXP (XEXP (x, 0), 1), 0)),
9899 plus_constant (other, INTVAL (constant)));
9903 if (changed && legitimate_address_p (mode, x, FALSE))
9906 if (GET_CODE (XEXP (x, 0)) == MULT)
9909 XEXP (x, 0) = force_operand (XEXP (x, 0), 0);
9912 if (GET_CODE (XEXP (x, 1)) == MULT)
9915 XEXP (x, 1) = force_operand (XEXP (x, 1), 0);
9919 && REG_P (XEXP (x, 1))
9920 && REG_P (XEXP (x, 0)))
9923 if (flag_pic && SYMBOLIC_CONST (XEXP (x, 1)))
9926 x = legitimize_pic_address (x, 0);
9929 if (changed && legitimate_address_p (mode, x, FALSE))
9932 if (REG_P (XEXP (x, 0)))
9934 rtx temp = gen_reg_rtx (Pmode);
9935 rtx val = force_operand (XEXP (x, 1), temp);
9937 emit_move_insn (temp, val);
9943 else if (REG_P (XEXP (x, 1)))
9945 rtx temp = gen_reg_rtx (Pmode);
9946 rtx val = force_operand (XEXP (x, 0), temp);
9948 emit_move_insn (temp, val);
9958 /* Print an integer constant expression in assembler syntax. Addition
9959 and subtraction are the only arithmetic that may appear in these
9960 expressions. FILE is the stdio stream to write to, X is the rtx, and
9961 CODE is the operand print code from the output string. */
9964 output_pic_addr_const (FILE *file, rtx x, int code)
9968 switch (GET_CODE (x))
9971 gcc_assert (flag_pic);
9976 if (! TARGET_MACHO || TARGET_64BIT)
9977 output_addr_const (file, x);
9980 const char *name = XSTR (x, 0);
9982 /* Mark the decl as referenced so that cgraph will
9983 output the function. */
9984 if (SYMBOL_REF_DECL (x))
9985 mark_decl_referenced (SYMBOL_REF_DECL (x));
9988 if (MACHOPIC_INDIRECT
9989 && machopic_classify_symbol (x) == MACHOPIC_UNDEFINED_FUNCTION)
9990 name = machopic_indirection_name (x, /*stub_p=*/true);
9992 assemble_name (file, name);
9994 if (!TARGET_MACHO && !(TARGET_64BIT && DEFAULT_ABI == MS_ABI)
9995 && code == 'P' && ! SYMBOL_REF_LOCAL_P (x))
9996 fputs ("@PLT", file);
10003 ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x));
10004 assemble_name (asm_out_file, buf);
10008 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x));
10012 /* This used to output parentheses around the expression,
10013 but that does not work on the 386 (either ATT or BSD assembler). */
10014 output_pic_addr_const (file, XEXP (x, 0), code);
10018 if (GET_MODE (x) == VOIDmode)
10020 /* We can use %d if the number is <32 bits and positive. */
10021 if (CONST_DOUBLE_HIGH (x) || CONST_DOUBLE_LOW (x) < 0)
10022 fprintf (file, "0x%lx%08lx",
10023 (unsigned long) CONST_DOUBLE_HIGH (x),
10024 (unsigned long) CONST_DOUBLE_LOW (x));
10026 fprintf (file, HOST_WIDE_INT_PRINT_DEC, CONST_DOUBLE_LOW (x));
10029 /* We can't handle floating point constants;
10030 PRINT_OPERAND must handle them. */
10031 output_operand_lossage ("floating constant misused");
10035 /* Some assemblers need integer constants to appear first. */
10036 if (CONST_INT_P (XEXP (x, 0)))
10038 output_pic_addr_const (file, XEXP (x, 0), code);
10040 output_pic_addr_const (file, XEXP (x, 1), code);
10044 gcc_assert (CONST_INT_P (XEXP (x, 1)));
10045 output_pic_addr_const (file, XEXP (x, 1), code);
10047 output_pic_addr_const (file, XEXP (x, 0), code);
10053 putc (ASSEMBLER_DIALECT == ASM_INTEL ? '(' : '[', file);
10054 output_pic_addr_const (file, XEXP (x, 0), code);
10056 output_pic_addr_const (file, XEXP (x, 1), code);
10058 putc (ASSEMBLER_DIALECT == ASM_INTEL ? ')' : ']', file);
10062 gcc_assert (XVECLEN (x, 0) == 1);
10063 output_pic_addr_const (file, XVECEXP (x, 0, 0), code);
10064 switch (XINT (x, 1))
10067 fputs ("@GOT", file);
10069 case UNSPEC_GOTOFF:
10070 fputs ("@GOTOFF", file);
10072 case UNSPEC_PLTOFF:
10073 fputs ("@PLTOFF", file);
10075 case UNSPEC_GOTPCREL:
10076 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
10077 "@GOTPCREL(%rip)" : "@GOTPCREL[rip]", file);
10079 case UNSPEC_GOTTPOFF:
10080 /* FIXME: This might be @TPOFF in Sun ld too. */
10081 fputs ("@GOTTPOFF", file);
10084 fputs ("@TPOFF", file);
10086 case UNSPEC_NTPOFF:
10088 fputs ("@TPOFF", file);
10090 fputs ("@NTPOFF", file);
10092 case UNSPEC_DTPOFF:
10093 fputs ("@DTPOFF", file);
10095 case UNSPEC_GOTNTPOFF:
10097 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
10098 "@GOTTPOFF(%rip)": "@GOTTPOFF[rip]", file);
10100 fputs ("@GOTNTPOFF", file);
10102 case UNSPEC_INDNTPOFF:
10103 fputs ("@INDNTPOFF", file);
10106 case UNSPEC_MACHOPIC_OFFSET:
10108 machopic_output_function_base_name (file);
10112 output_operand_lossage ("invalid UNSPEC as operand");
10118 output_operand_lossage ("invalid expression as operand");
10122 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
10123 We need to emit DTP-relative relocations. */
10125 static void ATTRIBUTE_UNUSED
10126 i386_output_dwarf_dtprel (FILE *file, int size, rtx x)
10128 fputs (ASM_LONG, file);
10129 output_addr_const (file, x);
10130 fputs ("@DTPOFF", file);
10136 fputs (", 0", file);
10139 gcc_unreachable ();
10143 /* Return true if X is a representation of the PIC register. This copes
10144 with calls from ix86_find_base_term, where the register might have
10145 been replaced by a cselib value. */
10148 ix86_pic_register_p (rtx x)
10150 if (GET_CODE (x) == VALUE)
10151 return (pic_offset_table_rtx
10152 && rtx_equal_for_cselib_p (x, pic_offset_table_rtx));
10154 return REG_P (x) && REGNO (x) == PIC_OFFSET_TABLE_REGNUM;
10157 /* In the name of slightly smaller debug output, and to cater to
10158 general assembler lossage, recognize PIC+GOTOFF and turn it back
10159 into a direct symbol reference.
10161 On Darwin, this is necessary to avoid a crash, because Darwin
10162 has a different PIC label for each routine but the DWARF debugging
10163 information is not associated with any particular routine, so it's
10164 necessary to remove references to the PIC label from RTL stored by
10165 the DWARF output code. */
10168 ix86_delegitimize_address (rtx orig_x)
10171 /* reg_addend is NULL or a multiple of some register. */
10172 rtx reg_addend = NULL_RTX;
10173 /* const_addend is NULL or a const_int. */
10174 rtx const_addend = NULL_RTX;
10175 /* This is the result, or NULL. */
10176 rtx result = NULL_RTX;
10183 if (GET_CODE (x) != CONST
10184 || GET_CODE (XEXP (x, 0)) != UNSPEC
10185 || XINT (XEXP (x, 0), 1) != UNSPEC_GOTPCREL
10186 || !MEM_P (orig_x))
10188 return XVECEXP (XEXP (x, 0), 0, 0);
10191 if (GET_CODE (x) != PLUS
10192 || GET_CODE (XEXP (x, 1)) != CONST)
10195 if (ix86_pic_register_p (XEXP (x, 0)))
10196 /* %ebx + GOT/GOTOFF */
10198 else if (GET_CODE (XEXP (x, 0)) == PLUS)
10200 /* %ebx + %reg * scale + GOT/GOTOFF */
10201 reg_addend = XEXP (x, 0);
10202 if (ix86_pic_register_p (XEXP (reg_addend, 0)))
10203 reg_addend = XEXP (reg_addend, 1);
10204 else if (ix86_pic_register_p (XEXP (reg_addend, 1)))
10205 reg_addend = XEXP (reg_addend, 0);
10208 if (!REG_P (reg_addend)
10209 && GET_CODE (reg_addend) != MULT
10210 && GET_CODE (reg_addend) != ASHIFT)
10216 x = XEXP (XEXP (x, 1), 0);
10217 if (GET_CODE (x) == PLUS
10218 && CONST_INT_P (XEXP (x, 1)))
10220 const_addend = XEXP (x, 1);
10224 if (GET_CODE (x) == UNSPEC
10225 && ((XINT (x, 1) == UNSPEC_GOT && MEM_P (orig_x))
10226 || (XINT (x, 1) == UNSPEC_GOTOFF && !MEM_P (orig_x))))
10227 result = XVECEXP (x, 0, 0);
10229 if (TARGET_MACHO && darwin_local_data_pic (x)
10230 && !MEM_P (orig_x))
10231 result = XVECEXP (x, 0, 0);
10237 result = gen_rtx_CONST (Pmode, gen_rtx_PLUS (Pmode, result, const_addend));
10239 result = gen_rtx_PLUS (Pmode, reg_addend, result);
10243 /* If X is a machine specific address (i.e. a symbol or label being
10244 referenced as a displacement from the GOT implemented using an
10245 UNSPEC), then return the base term. Otherwise return X. */
10248 ix86_find_base_term (rtx x)
10254 if (GET_CODE (x) != CONST)
10256 term = XEXP (x, 0);
10257 if (GET_CODE (term) == PLUS
10258 && (CONST_INT_P (XEXP (term, 1))
10259 || GET_CODE (XEXP (term, 1)) == CONST_DOUBLE))
10260 term = XEXP (term, 0);
10261 if (GET_CODE (term) != UNSPEC
10262 || XINT (term, 1) != UNSPEC_GOTPCREL)
10265 return XVECEXP (term, 0, 0);
10268 return ix86_delegitimize_address (x);
10272 put_condition_code (enum rtx_code code, enum machine_mode mode, int reverse,
10273 int fp, FILE *file)
10275 const char *suffix;
10277 if (mode == CCFPmode || mode == CCFPUmode)
10279 enum rtx_code second_code, bypass_code;
10280 ix86_fp_comparison_codes (code, &bypass_code, &code, &second_code);
10281 gcc_assert (bypass_code == UNKNOWN && second_code == UNKNOWN);
10282 code = ix86_fp_compare_code_to_integer (code);
10286 code = reverse_condition (code);
10337 gcc_assert (mode == CCmode || mode == CCNOmode || mode == CCGCmode);
10341 /* ??? Use "nbe" instead of "a" for fcmov lossage on some assemblers.
10342 Those same assemblers have the same but opposite lossage on cmov. */
10343 if (mode == CCmode)
10344 suffix = fp ? "nbe" : "a";
10345 else if (mode == CCCmode)
10348 gcc_unreachable ();
10364 gcc_unreachable ();
10368 gcc_assert (mode == CCmode || mode == CCCmode);
10385 gcc_unreachable ();
10389 /* ??? As above. */
10390 gcc_assert (mode == CCmode || mode == CCCmode);
10391 suffix = fp ? "nb" : "ae";
10394 gcc_assert (mode == CCmode || mode == CCGCmode || mode == CCNOmode);
10398 /* ??? As above. */
10399 if (mode == CCmode)
10401 else if (mode == CCCmode)
10402 suffix = fp ? "nb" : "ae";
10404 gcc_unreachable ();
10407 suffix = fp ? "u" : "p";
10410 suffix = fp ? "nu" : "np";
10413 gcc_unreachable ();
10415 fputs (suffix, file);
10418 /* Print the name of register X to FILE based on its machine mode and number.
10419 If CODE is 'w', pretend the mode is HImode.
10420 If CODE is 'b', pretend the mode is QImode.
10421 If CODE is 'k', pretend the mode is SImode.
10422 If CODE is 'q', pretend the mode is DImode.
10423 If CODE is 'x', pretend the mode is V4SFmode.
10424 If CODE is 't', pretend the mode is V8SFmode.
10425 If CODE is 'h', pretend the reg is the 'high' byte register.
10426 If CODE is 'y', print "st(0)" instead of "st", if the reg is stack op.
10427 If CODE is 'd', duplicate the operand for AVX instruction.
10431 print_reg (rtx x, int code, FILE *file)
10434 bool duplicated = code == 'd' && TARGET_AVX;
10436 gcc_assert (x == pc_rtx
10437 || (REGNO (x) != ARG_POINTER_REGNUM
10438 && REGNO (x) != FRAME_POINTER_REGNUM
10439 && REGNO (x) != FLAGS_REG
10440 && REGNO (x) != FPSR_REG
10441 && REGNO (x) != FPCR_REG));
10443 if (ASSEMBLER_DIALECT == ASM_ATT)
10448 gcc_assert (TARGET_64BIT);
10449 fputs ("rip", file);
10453 if (code == 'w' || MMX_REG_P (x))
10455 else if (code == 'b')
10457 else if (code == 'k')
10459 else if (code == 'q')
10461 else if (code == 'y')
10463 else if (code == 'h')
10465 else if (code == 'x')
10467 else if (code == 't')
10470 code = GET_MODE_SIZE (GET_MODE (x));
10472 /* Irritatingly, AMD extended registers use different naming convention
10473 from the normal registers. */
10474 if (REX_INT_REG_P (x))
10476 gcc_assert (TARGET_64BIT);
10480 error ("extended registers have no high halves");
10483 fprintf (file, "r%ib", REGNO (x) - FIRST_REX_INT_REG + 8);
10486 fprintf (file, "r%iw", REGNO (x) - FIRST_REX_INT_REG + 8);
10489 fprintf (file, "r%id", REGNO (x) - FIRST_REX_INT_REG + 8);
10492 fprintf (file, "r%i", REGNO (x) - FIRST_REX_INT_REG + 8);
10495 error ("unsupported operand size for extended register");
10505 if (STACK_TOP_P (x))
10514 if (! ANY_FP_REG_P (x))
10515 putc (code == 8 && TARGET_64BIT ? 'r' : 'e', file);
10520 reg = hi_reg_name[REGNO (x)];
10523 if (REGNO (x) >= ARRAY_SIZE (qi_reg_name))
10525 reg = qi_reg_name[REGNO (x)];
10528 if (REGNO (x) >= ARRAY_SIZE (qi_high_reg_name))
10530 reg = qi_high_reg_name[REGNO (x)];
10535 gcc_assert (!duplicated);
10537 fputs (hi_reg_name[REGNO (x)] + 1, file);
10542 gcc_unreachable ();
10548 if (ASSEMBLER_DIALECT == ASM_ATT)
10549 fprintf (file, ", %%%s", reg);
10551 fprintf (file, ", %s", reg);
10555 /* Locate some local-dynamic symbol still in use by this function
10556 so that we can print its name in some tls_local_dynamic_base
10560 get_some_local_dynamic_name_1 (rtx *px, void *data ATTRIBUTE_UNUSED)
10564 if (GET_CODE (x) == SYMBOL_REF
10565 && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC)
10567 cfun->machine->some_ld_name = XSTR (x, 0);
10574 static const char *
10575 get_some_local_dynamic_name (void)
10579 if (cfun->machine->some_ld_name)
10580 return cfun->machine->some_ld_name;
10582 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
10584 && for_each_rtx (&PATTERN (insn), get_some_local_dynamic_name_1, 0))
10585 return cfun->machine->some_ld_name;
10587 gcc_unreachable ();
10590 /* Meaning of CODE:
10591 L,W,B,Q,S,T -- print the opcode suffix for specified size of operand.
10592 C -- print opcode suffix for set/cmov insn.
10593 c -- like C, but print reversed condition
10594 E,e -- likewise, but for compare-and-branch fused insn.
10595 F,f -- likewise, but for floating-point.
10596 O -- if HAVE_AS_IX86_CMOV_SUN_SYNTAX, expand to "w.", "l." or "q.",
10598 R -- print the prefix for register names.
10599 z -- print the opcode suffix for the size of the current operand.
10600 * -- print a star (in certain assembler syntax)
10601 A -- print an absolute memory reference.
10602 w -- print the operand as if it's a "word" (HImode) even if it isn't.
10603 s -- print a shift double count, followed by the assemblers argument
10605 b -- print the QImode name of the register for the indicated operand.
10606 %b0 would print %al if operands[0] is reg 0.
10607 w -- likewise, print the HImode name of the register.
10608 k -- likewise, print the SImode name of the register.
10609 q -- likewise, print the DImode name of the register.
10610 x -- likewise, print the V4SFmode name of the register.
10611 t -- likewise, print the V8SFmode name of the register.
10612 h -- print the QImode name for a "high" register, either ah, bh, ch or dh.
10613 y -- print "st(0)" instead of "st" as a register.
10614 d -- print duplicated register operand for AVX instruction.
10615 D -- print condition for SSE cmp instruction.
10616 P -- if PIC, print an @PLT suffix.
10617 X -- don't print any sort of PIC '@' suffix for a symbol.
10618 & -- print some in-use local-dynamic symbol name.
10619 H -- print a memory address offset by 8; used for sse high-parts
10620 Y -- print condition for SSE5 com* instruction.
10621 + -- print a branch hint as 'cs' or 'ds' prefix
10622 ; -- print a semicolon (after prefixes due to bug in older gas).
10626 print_operand (FILE *file, rtx x, int code)
10633 if (ASSEMBLER_DIALECT == ASM_ATT)
10638 assemble_name (file, get_some_local_dynamic_name ());
10642 switch (ASSEMBLER_DIALECT)
10649 /* Intel syntax. For absolute addresses, registers should not
10650 be surrounded by braces. */
10654 PRINT_OPERAND (file, x, 0);
10661 gcc_unreachable ();
10664 PRINT_OPERAND (file, x, 0);
10669 if (ASSEMBLER_DIALECT == ASM_ATT)
10674 if (ASSEMBLER_DIALECT == ASM_ATT)
10679 if (ASSEMBLER_DIALECT == ASM_ATT)
10684 if (ASSEMBLER_DIALECT == ASM_ATT)
10689 if (ASSEMBLER_DIALECT == ASM_ATT)
10694 if (ASSEMBLER_DIALECT == ASM_ATT)
10699 /* 387 opcodes don't get size suffixes if the operands are
10701 if (STACK_REG_P (x))
10704 /* Likewise if using Intel opcodes. */
10705 if (ASSEMBLER_DIALECT == ASM_INTEL)
10708 /* This is the size of op from size of operand. */
10709 switch (GET_MODE_SIZE (GET_MODE (x)))
10718 #ifdef HAVE_GAS_FILDS_FISTS
10728 if (GET_MODE (x) == SFmode)
10743 if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT)
10747 #ifdef GAS_MNEMONICS
10762 gcc_unreachable ();
10779 if (CONST_INT_P (x) || ! SHIFT_DOUBLE_OMITS_COUNT)
10781 PRINT_OPERAND (file, x, 0);
10782 fputs (", ", file);
10787 /* Little bit of braindamage here. The SSE compare instructions
10788 does use completely different names for the comparisons that the
10789 fp conditional moves. */
10792 switch (GET_CODE (x))
10795 fputs ("eq", file);
10798 fputs ("eq_us", file);
10801 fputs ("lt", file);
10804 fputs ("nge", file);
10807 fputs ("le", file);
10810 fputs ("ngt", file);
10813 fputs ("unord", file);
10816 fputs ("neq", file);
10819 fputs ("neq_oq", file);
10822 fputs ("ge", file);
10825 fputs ("nlt", file);
10828 fputs ("gt", file);
10831 fputs ("nle", file);
10834 fputs ("ord", file);
10837 gcc_unreachable ();
10842 switch (GET_CODE (x))
10846 fputs ("eq", file);
10850 fputs ("lt", file);
10854 fputs ("le", file);
10857 fputs ("unord", file);
10861 fputs ("neq", file);
10865 fputs ("nlt", file);
10869 fputs ("nle", file);
10872 fputs ("ord", file);
10875 gcc_unreachable ();
10880 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
10881 if (ASSEMBLER_DIALECT == ASM_ATT)
10883 switch (GET_MODE (x))
10885 case HImode: putc ('w', file); break;
10887 case SFmode: putc ('l', file); break;
10889 case DFmode: putc ('q', file); break;
10890 default: gcc_unreachable ();
10897 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 0, file);
10900 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
10901 if (ASSEMBLER_DIALECT == ASM_ATT)
10904 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 1, file);
10907 /* Like above, but reverse condition */
10909 /* Check to see if argument to %c is really a constant
10910 and not a condition code which needs to be reversed. */
10911 if (!COMPARISON_P (x))
10913 output_operand_lossage ("operand is neither a constant nor a condition code, invalid operand code 'c'");
10916 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 0, file);
10919 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
10920 if (ASSEMBLER_DIALECT == ASM_ATT)
10923 put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 1, file);
10927 put_condition_code (GET_CODE (x), CCmode, 0, 0, file);
10931 put_condition_code (GET_CODE (x), CCmode, 1, 0, file);
10935 /* It doesn't actually matter what mode we use here, as we're
10936 only going to use this for printing. */
10937 x = adjust_address_nv (x, DImode, 8);
10945 || optimize_function_for_size_p (cfun) || !TARGET_BRANCH_PREDICTION_HINTS)
10948 x = find_reg_note (current_output_insn, REG_BR_PROB, 0);
10951 int pred_val = INTVAL (XEXP (x, 0));
10953 if (pred_val < REG_BR_PROB_BASE * 45 / 100
10954 || pred_val > REG_BR_PROB_BASE * 55 / 100)
10956 int taken = pred_val > REG_BR_PROB_BASE / 2;
10957 int cputaken = final_forward_branch_p (current_output_insn) == 0;
10959 /* Emit hints only in the case default branch prediction
10960 heuristics would fail. */
10961 if (taken != cputaken)
10963 /* We use 3e (DS) prefix for taken branches and
10964 2e (CS) prefix for not taken branches. */
10966 fputs ("ds ; ", file);
10968 fputs ("cs ; ", file);
10976 switch (GET_CODE (x))
10979 fputs ("neq", file);
10982 fputs ("eq", file);
10986 fputs (INTEGRAL_MODE_P (GET_MODE (x)) ? "ge" : "unlt", file);
10990 fputs (INTEGRAL_MODE_P (GET_MODE (x)) ? "gt" : "unle", file);
10994 fputs ("le", file);
10998 fputs ("lt", file);
11001 fputs ("unord", file);
11004 fputs ("ord", file);
11007 fputs ("ueq", file);
11010 fputs ("nlt", file);
11013 fputs ("nle", file);
11016 fputs ("ule", file);
11019 fputs ("ult", file);
11022 fputs ("une", file);
11025 gcc_unreachable ();
11031 fputs (" ; ", file);
11038 output_operand_lossage ("invalid operand code '%c'", code);
11043 print_reg (x, code, file);
11045 else if (MEM_P (x))
11047 /* No `byte ptr' prefix for call instructions or BLKmode operands. */
11048 if (ASSEMBLER_DIALECT == ASM_INTEL && code != 'X' && code != 'P'
11049 && GET_MODE (x) != BLKmode)
11052 switch (GET_MODE_SIZE (GET_MODE (x)))
11054 case 1: size = "BYTE"; break;
11055 case 2: size = "WORD"; break;
11056 case 4: size = "DWORD"; break;
11057 case 8: size = "QWORD"; break;
11058 case 12: size = "XWORD"; break;
11060 if (GET_MODE (x) == XFmode)
11066 gcc_unreachable ();
11069 /* Check for explicit size override (codes 'b', 'w' and 'k') */
11072 else if (code == 'w')
11074 else if (code == 'k')
11077 fputs (size, file);
11078 fputs (" PTR ", file);
11082 /* Avoid (%rip) for call operands. */
11083 if (CONSTANT_ADDRESS_P (x) && code == 'P'
11084 && !CONST_INT_P (x))
11085 output_addr_const (file, x);
11086 else if (this_is_asm_operands && ! address_operand (x, VOIDmode))
11087 output_operand_lossage ("invalid constraints for operand");
11089 output_address (x);
11092 else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == SFmode)
11097 REAL_VALUE_FROM_CONST_DOUBLE (r, x);
11098 REAL_VALUE_TO_TARGET_SINGLE (r, l);
11100 if (ASSEMBLER_DIALECT == ASM_ATT)
11102 fprintf (file, "0x%08lx", (long unsigned int) l);
11105 /* These float cases don't actually occur as immediate operands. */
11106 else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == DFmode)
11110 real_to_decimal (dstr, CONST_DOUBLE_REAL_VALUE (x), sizeof (dstr), 0, 1);
11111 fprintf (file, "%s", dstr);
11114 else if (GET_CODE (x) == CONST_DOUBLE
11115 && GET_MODE (x) == XFmode)
11119 real_to_decimal (dstr, CONST_DOUBLE_REAL_VALUE (x), sizeof (dstr), 0, 1);
11120 fprintf (file, "%s", dstr);
11125 /* We have patterns that allow zero sets of memory, for instance.
11126 In 64-bit mode, we should probably support all 8-byte vectors,
11127 since we can in fact encode that into an immediate. */
11128 if (GET_CODE (x) == CONST_VECTOR)
11130 gcc_assert (x == CONST0_RTX (GET_MODE (x)));
11136 if (CONST_INT_P (x) || GET_CODE (x) == CONST_DOUBLE)
11138 if (ASSEMBLER_DIALECT == ASM_ATT)
11141 else if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF
11142 || GET_CODE (x) == LABEL_REF)
11144 if (ASSEMBLER_DIALECT == ASM_ATT)
11147 fputs ("OFFSET FLAT:", file);
11150 if (CONST_INT_P (x))
11151 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x));
11153 output_pic_addr_const (file, x, code);
11155 output_addr_const (file, x);
11159 /* Print a memory operand whose address is ADDR. */
11162 print_operand_address (FILE *file, rtx addr)
11164 struct ix86_address parts;
11165 rtx base, index, disp;
11167 int ok = ix86_decompose_address (addr, &parts);
11172 index = parts.index;
11174 scale = parts.scale;
11182 if (ASSEMBLER_DIALECT == ASM_ATT)
11184 fputs ((parts.seg == SEG_FS ? "fs:" : "gs:"), file);
11187 gcc_unreachable ();
11190 /* Use one byte shorter RIP relative addressing for 64bit mode. */
11191 if (TARGET_64BIT && !base && !index)
11195 if (GET_CODE (disp) == CONST
11196 && GET_CODE (XEXP (disp, 0)) == PLUS
11197 && CONST_INT_P (XEXP (XEXP (disp, 0), 1)))
11198 symbol = XEXP (XEXP (disp, 0), 0);
11200 if (GET_CODE (symbol) == LABEL_REF
11201 || (GET_CODE (symbol) == SYMBOL_REF
11202 && SYMBOL_REF_TLS_MODEL (symbol) == 0))
11205 if (!base && !index)
11207 /* Displacement only requires special attention. */
11209 if (CONST_INT_P (disp))
11211 if (ASSEMBLER_DIALECT == ASM_INTEL && parts.seg == SEG_DEFAULT)
11212 fputs ("ds:", file);
11213 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (disp));
11216 output_pic_addr_const (file, disp, 0);
11218 output_addr_const (file, disp);
11222 if (ASSEMBLER_DIALECT == ASM_ATT)
11227 output_pic_addr_const (file, disp, 0);
11228 else if (GET_CODE (disp) == LABEL_REF)
11229 output_asm_label (disp);
11231 output_addr_const (file, disp);
11236 print_reg (base, 0, file);
11240 print_reg (index, 0, file);
11242 fprintf (file, ",%d", scale);
11248 rtx offset = NULL_RTX;
11252 /* Pull out the offset of a symbol; print any symbol itself. */
11253 if (GET_CODE (disp) == CONST
11254 && GET_CODE (XEXP (disp, 0)) == PLUS
11255 && CONST_INT_P (XEXP (XEXP (disp, 0), 1)))
11257 offset = XEXP (XEXP (disp, 0), 1);
11258 disp = gen_rtx_CONST (VOIDmode,
11259 XEXP (XEXP (disp, 0), 0));
11263 output_pic_addr_const (file, disp, 0);
11264 else if (GET_CODE (disp) == LABEL_REF)
11265 output_asm_label (disp);
11266 else if (CONST_INT_P (disp))
11269 output_addr_const (file, disp);
11275 print_reg (base, 0, file);
11278 if (INTVAL (offset) >= 0)
11280 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset));
11284 fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset));
11291 print_reg (index, 0, file);
11293 fprintf (file, "*%d", scale);
11301 output_addr_const_extra (FILE *file, rtx x)
11305 if (GET_CODE (x) != UNSPEC)
11308 op = XVECEXP (x, 0, 0);
11309 switch (XINT (x, 1))
11311 case UNSPEC_GOTTPOFF:
11312 output_addr_const (file, op);
11313 /* FIXME: This might be @TPOFF in Sun ld. */
11314 fputs ("@GOTTPOFF", file);
11317 output_addr_const (file, op);
11318 fputs ("@TPOFF", file);
11320 case UNSPEC_NTPOFF:
11321 output_addr_const (file, op);
11323 fputs ("@TPOFF", file);
11325 fputs ("@NTPOFF", file);
11327 case UNSPEC_DTPOFF:
11328 output_addr_const (file, op);
11329 fputs ("@DTPOFF", file);
11331 case UNSPEC_GOTNTPOFF:
11332 output_addr_const (file, op);
11334 fputs (ASSEMBLER_DIALECT == ASM_ATT ?
11335 "@GOTTPOFF(%rip)" : "@GOTTPOFF[rip]", file);
11337 fputs ("@GOTNTPOFF", file);
11339 case UNSPEC_INDNTPOFF:
11340 output_addr_const (file, op);
11341 fputs ("@INDNTPOFF", file);
11344 case UNSPEC_MACHOPIC_OFFSET:
11345 output_addr_const (file, op);
11347 machopic_output_function_base_name (file);
11358 /* Split one or more DImode RTL references into pairs of SImode
11359 references. The RTL can be REG, offsettable MEM, integer constant, or
11360 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
11361 split and "num" is its length. lo_half and hi_half are output arrays
11362 that parallel "operands". */
11365 split_di (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
11369 rtx op = operands[num];
11371 /* simplify_subreg refuse to split volatile memory addresses,
11372 but we still have to handle it. */
11375 lo_half[num] = adjust_address (op, SImode, 0);
11376 hi_half[num] = adjust_address (op, SImode, 4);
11380 lo_half[num] = simplify_gen_subreg (SImode, op,
11381 GET_MODE (op) == VOIDmode
11382 ? DImode : GET_MODE (op), 0);
11383 hi_half[num] = simplify_gen_subreg (SImode, op,
11384 GET_MODE (op) == VOIDmode
11385 ? DImode : GET_MODE (op), 4);
11389 /* Split one or more TImode RTL references into pairs of DImode
11390 references. The RTL can be REG, offsettable MEM, integer constant, or
11391 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
11392 split and "num" is its length. lo_half and hi_half are output arrays
11393 that parallel "operands". */
11396 split_ti (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
11400 rtx op = operands[num];
11402 /* simplify_subreg refuse to split volatile memory addresses, but we
11403 still have to handle it. */
11406 lo_half[num] = adjust_address (op, DImode, 0);
11407 hi_half[num] = adjust_address (op, DImode, 8);
11411 lo_half[num] = simplify_gen_subreg (DImode, op, TImode, 0);
11412 hi_half[num] = simplify_gen_subreg (DImode, op, TImode, 8);
11417 /* Output code to perform a 387 binary operation in INSN, one of PLUS,
11418 MINUS, MULT or DIV. OPERANDS are the insn operands, where operands[3]
11419 is the expression of the binary operation. The output may either be
11420 emitted here, or returned to the caller, like all output_* functions.
11422 There is no guarantee that the operands are the same mode, as they
11423 might be within FLOAT or FLOAT_EXTEND expressions. */
11425 #ifndef SYSV386_COMPAT
11426 /* Set to 1 for compatibility with brain-damaged assemblers. No-one
11427 wants to fix the assemblers because that causes incompatibility
11428 with gcc. No-one wants to fix gcc because that causes
11429 incompatibility with assemblers... You can use the option of
11430 -DSYSV386_COMPAT=0 if you recompile both gcc and gas this way. */
11431 #define SYSV386_COMPAT 1
11435 output_387_binary_op (rtx insn, rtx *operands)
11437 static char buf[40];
11440 int is_sse = SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]) || SSE_REG_P (operands[2]);
11442 #ifdef ENABLE_CHECKING
11443 /* Even if we do not want to check the inputs, this documents input
11444 constraints. Which helps in understanding the following code. */
11445 if (STACK_REG_P (operands[0])
11446 && ((REG_P (operands[1])
11447 && REGNO (operands[0]) == REGNO (operands[1])
11448 && (STACK_REG_P (operands[2]) || MEM_P (operands[2])))
11449 || (REG_P (operands[2])
11450 && REGNO (operands[0]) == REGNO (operands[2])
11451 && (STACK_REG_P (operands[1]) || MEM_P (operands[1]))))
11452 && (STACK_TOP_P (operands[1]) || STACK_TOP_P (operands[2])))
11455 gcc_assert (is_sse);
11458 switch (GET_CODE (operands[3]))
11461 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
11462 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
11470 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
11471 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
11479 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
11480 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
11488 if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT
11489 || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT)
11497 gcc_unreachable ();
11504 strcpy (buf, ssep);
11505 if (GET_MODE (operands[0]) == SFmode)
11506 strcat (buf, "ss\t{%2, %1, %0|%0, %1, %2}");
11508 strcat (buf, "sd\t{%2, %1, %0|%0, %1, %2}");
11512 strcpy (buf, ssep + 1);
11513 if (GET_MODE (operands[0]) == SFmode)
11514 strcat (buf, "ss\t{%2, %0|%0, %2}");
11516 strcat (buf, "sd\t{%2, %0|%0, %2}");
11522 switch (GET_CODE (operands[3]))
11526 if (REG_P (operands[2]) && REGNO (operands[0]) == REGNO (operands[2]))
11528 rtx temp = operands[2];
11529 operands[2] = operands[1];
11530 operands[1] = temp;
11533 /* know operands[0] == operands[1]. */
11535 if (MEM_P (operands[2]))
11541 if (find_regno_note (insn, REG_DEAD, REGNO (operands[2])))
11543 if (STACK_TOP_P (operands[0]))
11544 /* How is it that we are storing to a dead operand[2]?
11545 Well, presumably operands[1] is dead too. We can't
11546 store the result to st(0) as st(0) gets popped on this
11547 instruction. Instead store to operands[2] (which I
11548 think has to be st(1)). st(1) will be popped later.
11549 gcc <= 2.8.1 didn't have this check and generated
11550 assembly code that the Unixware assembler rejected. */
11551 p = "p\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
11553 p = "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
11557 if (STACK_TOP_P (operands[0]))
11558 p = "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
11560 p = "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
11565 if (MEM_P (operands[1]))
11571 if (MEM_P (operands[2]))
11577 if (find_regno_note (insn, REG_DEAD, REGNO (operands[2])))
11580 /* The SystemV/386 SVR3.2 assembler, and probably all AT&T
11581 derived assemblers, confusingly reverse the direction of
11582 the operation for fsub{r} and fdiv{r} when the
11583 destination register is not st(0). The Intel assembler
11584 doesn't have this brain damage. Read !SYSV386_COMPAT to
11585 figure out what the hardware really does. */
11586 if (STACK_TOP_P (operands[0]))
11587 p = "{p\t%0, %2|rp\t%2, %0}";
11589 p = "{rp\t%2, %0|p\t%0, %2}";
11591 if (STACK_TOP_P (operands[0]))
11592 /* As above for fmul/fadd, we can't store to st(0). */
11593 p = "rp\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
11595 p = "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
11600 if (find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
11603 if (STACK_TOP_P (operands[0]))
11604 p = "{rp\t%0, %1|p\t%1, %0}";
11606 p = "{p\t%1, %0|rp\t%0, %1}";
11608 if (STACK_TOP_P (operands[0]))
11609 p = "p\t{%0, %1|%1, %0}"; /* st(1) = st(1) op st(0); pop */
11611 p = "rp\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2); pop */
11616 if (STACK_TOP_P (operands[0]))
11618 if (STACK_TOP_P (operands[1]))
11619 p = "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
11621 p = "r\t{%y1, %0|%0, %y1}"; /* st(0) = st(r1) op st(0) */
11624 else if (STACK_TOP_P (operands[1]))
11627 p = "{\t%1, %0|r\t%0, %1}";
11629 p = "r\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2) */
11635 p = "{r\t%2, %0|\t%0, %2}";
11637 p = "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
11643 gcc_unreachable ();
11650 /* Return needed mode for entity in optimize_mode_switching pass. */
11653 ix86_mode_needed (int entity, rtx insn)
11655 enum attr_i387_cw mode;
11657 /* The mode UNINITIALIZED is used to store control word after a
11658 function call or ASM pattern. The mode ANY specify that function
11659 has no requirements on the control word and make no changes in the
11660 bits we are interested in. */
11663 || (NONJUMP_INSN_P (insn)
11664 && (asm_noperands (PATTERN (insn)) >= 0
11665 || GET_CODE (PATTERN (insn)) == ASM_INPUT)))
11666 return I387_CW_UNINITIALIZED;
11668 if (recog_memoized (insn) < 0)
11669 return I387_CW_ANY;
11671 mode = get_attr_i387_cw (insn);
11676 if (mode == I387_CW_TRUNC)
11681 if (mode == I387_CW_FLOOR)
11686 if (mode == I387_CW_CEIL)
11691 if (mode == I387_CW_MASK_PM)
11696 gcc_unreachable ();
11699 return I387_CW_ANY;
11702 /* Output code to initialize control word copies used by trunc?f?i and
11703 rounding patterns. CURRENT_MODE is set to current control word,
11704 while NEW_MODE is set to new control word. */
11707 emit_i387_cw_initialization (int mode)
11709 rtx stored_mode = assign_386_stack_local (HImode, SLOT_CW_STORED);
11712 enum ix86_stack_slot slot;
11714 rtx reg = gen_reg_rtx (HImode);
11716 emit_insn (gen_x86_fnstcw_1 (stored_mode));
11717 emit_move_insn (reg, copy_rtx (stored_mode));
11719 if (TARGET_64BIT || TARGET_PARTIAL_REG_STALL
11720 || optimize_function_for_size_p (cfun))
11724 case I387_CW_TRUNC:
11725 /* round toward zero (truncate) */
11726 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0c00)));
11727 slot = SLOT_CW_TRUNC;
11730 case I387_CW_FLOOR:
11731 /* round down toward -oo */
11732 emit_insn (gen_andhi3 (reg, reg, GEN_INT (~0x0c00)));
11733 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0400)));
11734 slot = SLOT_CW_FLOOR;
11738 /* round up toward +oo */
11739 emit_insn (gen_andhi3 (reg, reg, GEN_INT (~0x0c00)));
11740 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0800)));
11741 slot = SLOT_CW_CEIL;
11744 case I387_CW_MASK_PM:
11745 /* mask precision exception for nearbyint() */
11746 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0020)));
11747 slot = SLOT_CW_MASK_PM;
11751 gcc_unreachable ();
11758 case I387_CW_TRUNC:
11759 /* round toward zero (truncate) */
11760 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0xc)));
11761 slot = SLOT_CW_TRUNC;
11764 case I387_CW_FLOOR:
11765 /* round down toward -oo */
11766 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0x4)));
11767 slot = SLOT_CW_FLOOR;
11771 /* round up toward +oo */
11772 emit_insn (gen_movsi_insv_1 (reg, GEN_INT (0x8)));
11773 slot = SLOT_CW_CEIL;
11776 case I387_CW_MASK_PM:
11777 /* mask precision exception for nearbyint() */
11778 emit_insn (gen_iorhi3 (reg, reg, GEN_INT (0x0020)));
11779 slot = SLOT_CW_MASK_PM;
11783 gcc_unreachable ();
11787 gcc_assert (slot < MAX_386_STACK_LOCALS);
11789 new_mode = assign_386_stack_local (HImode, slot);
11790 emit_move_insn (new_mode, reg);
11793 /* Output code for INSN to convert a float to a signed int. OPERANDS
11794 are the insn operands. The output may be [HSD]Imode and the input
11795 operand may be [SDX]Fmode. */
11798 output_fix_trunc (rtx insn, rtx *operands, int fisttp)
11800 int stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0;
11801 int dimode_p = GET_MODE (operands[0]) == DImode;
11802 int round_mode = get_attr_i387_cw (insn);
11804 /* Jump through a hoop or two for DImode, since the hardware has no
11805 non-popping instruction. We used to do this a different way, but
11806 that was somewhat fragile and broke with post-reload splitters. */
11807 if ((dimode_p || fisttp) && !stack_top_dies)
11808 output_asm_insn ("fld\t%y1", operands);
11810 gcc_assert (STACK_TOP_P (operands[1]));
11811 gcc_assert (MEM_P (operands[0]));
11812 gcc_assert (GET_MODE (operands[1]) != TFmode);
11815 output_asm_insn ("fisttp%z0\t%0", operands);
11818 if (round_mode != I387_CW_ANY)
11819 output_asm_insn ("fldcw\t%3", operands);
11820 if (stack_top_dies || dimode_p)
11821 output_asm_insn ("fistp%z0\t%0", operands);
11823 output_asm_insn ("fist%z0\t%0", operands);
11824 if (round_mode != I387_CW_ANY)
11825 output_asm_insn ("fldcw\t%2", operands);
11831 /* Output code for x87 ffreep insn. The OPNO argument, which may only
11832 have the values zero or one, indicates the ffreep insn's operand
11833 from the OPERANDS array. */
11835 static const char *
11836 output_387_ffreep (rtx *operands ATTRIBUTE_UNUSED, int opno)
11838 if (TARGET_USE_FFREEP)
11839 #if HAVE_AS_IX86_FFREEP
11840 return opno ? "ffreep\t%y1" : "ffreep\t%y0";
11843 static char retval[] = ".word\t0xc_df";
11844 int regno = REGNO (operands[opno]);
11846 gcc_assert (FP_REGNO_P (regno));
11848 retval[9] = '0' + (regno - FIRST_STACK_REG);
11853 return opno ? "fstp\t%y1" : "fstp\t%y0";
11857 /* Output code for INSN to compare OPERANDS. EFLAGS_P is 1 when fcomi
11858 should be used. UNORDERED_P is true when fucom should be used. */
11861 output_fp_compare (rtx insn, rtx *operands, int eflags_p, int unordered_p)
11863 int stack_top_dies;
11864 rtx cmp_op0, cmp_op1;
11865 int is_sse = SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]);
11869 cmp_op0 = operands[0];
11870 cmp_op1 = operands[1];
11874 cmp_op0 = operands[1];
11875 cmp_op1 = operands[2];
11880 static const char ucomiss[] = "vucomiss\t{%1, %0|%0, %1}";
11881 static const char ucomisd[] = "vucomisd\t{%1, %0|%0, %1}";
11882 static const char comiss[] = "vcomiss\t{%1, %0|%0, %1}";
11883 static const char comisd[] = "vcomisd\t{%1, %0|%0, %1}";
11885 if (GET_MODE (operands[0]) == SFmode)
11887 return &ucomiss[TARGET_AVX ? 0 : 1];
11889 return &comiss[TARGET_AVX ? 0 : 1];
11892 return &ucomisd[TARGET_AVX ? 0 : 1];
11894 return &comisd[TARGET_AVX ? 0 : 1];
11897 gcc_assert (STACK_TOP_P (cmp_op0));
11899 stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0;
11901 if (cmp_op1 == CONST0_RTX (GET_MODE (cmp_op1)))
11903 if (stack_top_dies)
11905 output_asm_insn ("ftst\n\tfnstsw\t%0", operands);
11906 return output_387_ffreep (operands, 1);
11909 return "ftst\n\tfnstsw\t%0";
11912 if (STACK_REG_P (cmp_op1)
11914 && find_regno_note (insn, REG_DEAD, REGNO (cmp_op1))
11915 && REGNO (cmp_op1) != FIRST_STACK_REG)
11917 /* If both the top of the 387 stack dies, and the other operand
11918 is also a stack register that dies, then this must be a
11919 `fcompp' float compare */
11923 /* There is no double popping fcomi variant. Fortunately,
11924 eflags is immune from the fstp's cc clobbering. */
11926 output_asm_insn ("fucomip\t{%y1, %0|%0, %y1}", operands);
11928 output_asm_insn ("fcomip\t{%y1, %0|%0, %y1}", operands);
11929 return output_387_ffreep (operands, 0);
11934 return "fucompp\n\tfnstsw\t%0";
11936 return "fcompp\n\tfnstsw\t%0";
11941 /* Encoded here as eflags_p | intmode | unordered_p | stack_top_dies. */
11943 static const char * const alt[16] =
11945 "fcom%z2\t%y2\n\tfnstsw\t%0",
11946 "fcomp%z2\t%y2\n\tfnstsw\t%0",
11947 "fucom%z2\t%y2\n\tfnstsw\t%0",
11948 "fucomp%z2\t%y2\n\tfnstsw\t%0",
11950 "ficom%z2\t%y2\n\tfnstsw\t%0",
11951 "ficomp%z2\t%y2\n\tfnstsw\t%0",
11955 "fcomi\t{%y1, %0|%0, %y1}",
11956 "fcomip\t{%y1, %0|%0, %y1}",
11957 "fucomi\t{%y1, %0|%0, %y1}",
11958 "fucomip\t{%y1, %0|%0, %y1}",
11969 mask = eflags_p << 3;
11970 mask |= (GET_MODE_CLASS (GET_MODE (cmp_op1)) == MODE_INT) << 2;
11971 mask |= unordered_p << 1;
11972 mask |= stack_top_dies;
11974 gcc_assert (mask < 16);
11983 ix86_output_addr_vec_elt (FILE *file, int value)
11985 const char *directive = ASM_LONG;
11989 directive = ASM_QUAD;
11991 gcc_assert (!TARGET_64BIT);
11994 fprintf (file, "%s%s%d\n", directive, LPREFIX, value);
11998 ix86_output_addr_diff_elt (FILE *file, int value, int rel)
12000 const char *directive = ASM_LONG;
12003 if (TARGET_64BIT && CASE_VECTOR_MODE == DImode)
12004 directive = ASM_QUAD;
12006 gcc_assert (!TARGET_64BIT);
12008 /* We can't use @GOTOFF for text labels on VxWorks; see gotoff_operand. */
12009 if (TARGET_64BIT || TARGET_VXWORKS_RTP)
12010 fprintf (file, "%s%s%d-%s%d\n",
12011 directive, LPREFIX, value, LPREFIX, rel);
12012 else if (HAVE_AS_GOTOFF_IN_DATA)
12013 fprintf (file, "%s%s%d@GOTOFF\n", ASM_LONG, LPREFIX, value);
12015 else if (TARGET_MACHO)
12017 fprintf (file, "%s%s%d-", ASM_LONG, LPREFIX, value);
12018 machopic_output_function_base_name (file);
12019 fprintf(file, "\n");
12023 asm_fprintf (file, "%s%U%s+[.-%s%d]\n",
12024 ASM_LONG, GOT_SYMBOL_NAME, LPREFIX, value);
12027 /* Generate either "mov $0, reg" or "xor reg, reg", as appropriate
12031 ix86_expand_clear (rtx dest)
12035 /* We play register width games, which are only valid after reload. */
12036 gcc_assert (reload_completed);
12038 /* Avoid HImode and its attendant prefix byte. */
12039 if (GET_MODE_SIZE (GET_MODE (dest)) < 4)
12040 dest = gen_rtx_REG (SImode, REGNO (dest));
12041 tmp = gen_rtx_SET (VOIDmode, dest, const0_rtx);
12043 /* This predicate should match that for movsi_xor and movdi_xor_rex64. */
12044 if (reload_completed && (!TARGET_USE_MOV0 || optimize_insn_for_speed_p ()))
12046 rtx clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
12047 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob));
12053 /* X is an unchanging MEM. If it is a constant pool reference, return
12054 the constant pool rtx, else NULL. */
12057 maybe_get_pool_constant (rtx x)
12059 x = ix86_delegitimize_address (XEXP (x, 0));
12061 if (GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x))
12062 return get_pool_constant (x);
12068 ix86_expand_move (enum machine_mode mode, rtx operands[])
12071 enum tls_model model;
12076 if (GET_CODE (op1) == SYMBOL_REF)
12078 model = SYMBOL_REF_TLS_MODEL (op1);
12081 op1 = legitimize_tls_address (op1, model, true);
12082 op1 = force_operand (op1, op0);
12086 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
12087 && SYMBOL_REF_DLLIMPORT_P (op1))
12088 op1 = legitimize_dllimport_symbol (op1, false);
12090 else if (GET_CODE (op1) == CONST
12091 && GET_CODE (XEXP (op1, 0)) == PLUS
12092 && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SYMBOL_REF)
12094 rtx addend = XEXP (XEXP (op1, 0), 1);
12095 rtx symbol = XEXP (XEXP (op1, 0), 0);
12098 model = SYMBOL_REF_TLS_MODEL (symbol);
12100 tmp = legitimize_tls_address (symbol, model, true);
12101 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
12102 && SYMBOL_REF_DLLIMPORT_P (symbol))
12103 tmp = legitimize_dllimport_symbol (symbol, true);
12107 tmp = force_operand (tmp, NULL);
12108 tmp = expand_simple_binop (Pmode, PLUS, tmp, addend,
12109 op0, 1, OPTAB_DIRECT);
12115 if (flag_pic && mode == Pmode && symbolic_operand (op1, Pmode))
12117 if (TARGET_MACHO && !TARGET_64BIT)
12122 rtx temp = ((reload_in_progress
12123 || ((op0 && REG_P (op0))
12125 ? op0 : gen_reg_rtx (Pmode));
12126 op1 = machopic_indirect_data_reference (op1, temp);
12127 op1 = machopic_legitimize_pic_address (op1, mode,
12128 temp == op1 ? 0 : temp);
12130 else if (MACHOPIC_INDIRECT)
12131 op1 = machopic_indirect_data_reference (op1, 0);
12139 op1 = force_reg (Pmode, op1);
12140 else if (!TARGET_64BIT || !x86_64_movabs_operand (op1, Pmode))
12142 rtx reg = !can_create_pseudo_p () ? op0 : NULL_RTX;
12143 op1 = legitimize_pic_address (op1, reg);
12152 && (PUSH_ROUNDING (GET_MODE_SIZE (mode)) != GET_MODE_SIZE (mode)
12153 || !push_operand (op0, mode))
12155 op1 = force_reg (mode, op1);
12157 if (push_operand (op0, mode)
12158 && ! general_no_elim_operand (op1, mode))
12159 op1 = copy_to_mode_reg (mode, op1);
12161 /* Force large constants in 64bit compilation into register
12162 to get them CSEed. */
12163 if (can_create_pseudo_p ()
12164 && (mode == DImode) && TARGET_64BIT
12165 && immediate_operand (op1, mode)
12166 && !x86_64_zext_immediate_operand (op1, VOIDmode)
12167 && !register_operand (op0, mode)
12169 op1 = copy_to_mode_reg (mode, op1);
12171 if (can_create_pseudo_p ()
12172 && FLOAT_MODE_P (mode)
12173 && GET_CODE (op1) == CONST_DOUBLE)
12175 /* If we are loading a floating point constant to a register,
12176 force the value to memory now, since we'll get better code
12177 out the back end. */
12179 op1 = validize_mem (force_const_mem (mode, op1));
12180 if (!register_operand (op0, mode))
12182 rtx temp = gen_reg_rtx (mode);
12183 emit_insn (gen_rtx_SET (VOIDmode, temp, op1));
12184 emit_move_insn (op0, temp);
12190 emit_insn (gen_rtx_SET (VOIDmode, op0, op1));
12194 ix86_expand_vector_move (enum machine_mode mode, rtx operands[])
12196 rtx op0 = operands[0], op1 = operands[1];
12197 unsigned int align = GET_MODE_ALIGNMENT (mode);
12199 /* Force constants other than zero into memory. We do not know how
12200 the instructions used to build constants modify the upper 64 bits
12201 of the register, once we have that information we may be able
12202 to handle some of them more efficiently. */
12203 if (can_create_pseudo_p ()
12204 && register_operand (op0, mode)
12205 && (CONSTANT_P (op1)
12206 || (GET_CODE (op1) == SUBREG
12207 && CONSTANT_P (SUBREG_REG (op1))))
12208 && standard_sse_constant_p (op1) <= 0)
12209 op1 = validize_mem (force_const_mem (mode, op1));
12211 /* We need to check memory alignment for SSE mode since attribute
12212 can make operands unaligned. */
12213 if (can_create_pseudo_p ()
12214 && SSE_REG_MODE_P (mode)
12215 && ((MEM_P (op0) && (MEM_ALIGN (op0) < align))
12216 || (MEM_P (op1) && (MEM_ALIGN (op1) < align))))
12220 /* ix86_expand_vector_move_misalign() does not like constants ... */
12221 if (CONSTANT_P (op1)
12222 || (GET_CODE (op1) == SUBREG
12223 && CONSTANT_P (SUBREG_REG (op1))))
12224 op1 = validize_mem (force_const_mem (mode, op1));
12226 /* ... nor both arguments in memory. */
12227 if (!register_operand (op0, mode)
12228 && !register_operand (op1, mode))
12229 op1 = force_reg (mode, op1);
12231 tmp[0] = op0; tmp[1] = op1;
12232 ix86_expand_vector_move_misalign (mode, tmp);
12236 /* Make operand1 a register if it isn't already. */
12237 if (can_create_pseudo_p ()
12238 && !register_operand (op0, mode)
12239 && !register_operand (op1, mode))
12241 emit_move_insn (op0, force_reg (GET_MODE (op0), op1));
12245 emit_insn (gen_rtx_SET (VOIDmode, op0, op1));
12248 /* Implement the movmisalign patterns for SSE. Non-SSE modes go
12249 straight to ix86_expand_vector_move. */
12250 /* Code generation for scalar reg-reg moves of single and double precision data:
12251 if (x86_sse_partial_reg_dependency == true | x86_sse_split_regs == true)
12255 if (x86_sse_partial_reg_dependency == true)
12260 Code generation for scalar loads of double precision data:
12261 if (x86_sse_split_regs == true)
12262 movlpd mem, reg (gas syntax)
12266 Code generation for unaligned packed loads of single precision data
12267 (x86_sse_unaligned_move_optimal overrides x86_sse_partial_reg_dependency):
12268 if (x86_sse_unaligned_move_optimal)
12271 if (x86_sse_partial_reg_dependency == true)
12283 Code generation for unaligned packed loads of double precision data
12284 (x86_sse_unaligned_move_optimal overrides x86_sse_split_regs):
12285 if (x86_sse_unaligned_move_optimal)
12288 if (x86_sse_split_regs == true)
12301 ix86_expand_vector_move_misalign (enum machine_mode mode, rtx operands[])
12310 switch (GET_MODE_CLASS (mode))
12312 case MODE_VECTOR_INT:
12314 switch (GET_MODE_SIZE (mode))
12317 op0 = gen_lowpart (V16QImode, op0);
12318 op1 = gen_lowpart (V16QImode, op1);
12319 emit_insn (gen_avx_movdqu (op0, op1));
12322 op0 = gen_lowpart (V32QImode, op0);
12323 op1 = gen_lowpart (V32QImode, op1);
12324 emit_insn (gen_avx_movdqu256 (op0, op1));
12327 gcc_unreachable ();
12330 case MODE_VECTOR_FLOAT:
12331 op0 = gen_lowpart (mode, op0);
12332 op1 = gen_lowpart (mode, op1);
12337 emit_insn (gen_avx_movups (op0, op1));
12340 emit_insn (gen_avx_movups256 (op0, op1));
12343 emit_insn (gen_avx_movupd (op0, op1));
12346 emit_insn (gen_avx_movupd256 (op0, op1));
12349 gcc_unreachable ();
12354 gcc_unreachable ();
12362 /* If we're optimizing for size, movups is the smallest. */
12363 if (optimize_insn_for_size_p ())
12365 op0 = gen_lowpart (V4SFmode, op0);
12366 op1 = gen_lowpart (V4SFmode, op1);
12367 emit_insn (gen_sse_movups (op0, op1));
12371 /* ??? If we have typed data, then it would appear that using
12372 movdqu is the only way to get unaligned data loaded with
12374 if (TARGET_SSE2 && GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
12376 op0 = gen_lowpart (V16QImode, op0);
12377 op1 = gen_lowpart (V16QImode, op1);
12378 emit_insn (gen_sse2_movdqu (op0, op1));
12382 if (TARGET_SSE2 && mode == V2DFmode)
12386 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL)
12388 op0 = gen_lowpart (V2DFmode, op0);
12389 op1 = gen_lowpart (V2DFmode, op1);
12390 emit_insn (gen_sse2_movupd (op0, op1));
12394 /* When SSE registers are split into halves, we can avoid
12395 writing to the top half twice. */
12396 if (TARGET_SSE_SPLIT_REGS)
12398 emit_clobber (op0);
12403 /* ??? Not sure about the best option for the Intel chips.
12404 The following would seem to satisfy; the register is
12405 entirely cleared, breaking the dependency chain. We
12406 then store to the upper half, with a dependency depth
12407 of one. A rumor has it that Intel recommends two movsd
12408 followed by an unpacklpd, but this is unconfirmed. And
12409 given that the dependency depth of the unpacklpd would
12410 still be one, I'm not sure why this would be better. */
12411 zero = CONST0_RTX (V2DFmode);
12414 m = adjust_address (op1, DFmode, 0);
12415 emit_insn (gen_sse2_loadlpd (op0, zero, m));
12416 m = adjust_address (op1, DFmode, 8);
12417 emit_insn (gen_sse2_loadhpd (op0, op0, m));
12421 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL)
12423 op0 = gen_lowpart (V4SFmode, op0);
12424 op1 = gen_lowpart (V4SFmode, op1);
12425 emit_insn (gen_sse_movups (op0, op1));
12429 if (TARGET_SSE_PARTIAL_REG_DEPENDENCY)
12430 emit_move_insn (op0, CONST0_RTX (mode));
12432 emit_clobber (op0);
12434 if (mode != V4SFmode)
12435 op0 = gen_lowpart (V4SFmode, op0);
12436 m = adjust_address (op1, V2SFmode, 0);
12437 emit_insn (gen_sse_loadlps (op0, op0, m));
12438 m = adjust_address (op1, V2SFmode, 8);
12439 emit_insn (gen_sse_loadhps (op0, op0, m));
12442 else if (MEM_P (op0))
12444 /* If we're optimizing for size, movups is the smallest. */
12445 if (optimize_insn_for_size_p ())
12447 op0 = gen_lowpart (V4SFmode, op0);
12448 op1 = gen_lowpart (V4SFmode, op1);
12449 emit_insn (gen_sse_movups (op0, op1));
12453 /* ??? Similar to above, only less clear because of quote
12454 typeless stores unquote. */
12455 if (TARGET_SSE2 && !TARGET_SSE_TYPELESS_STORES
12456 && GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
12458 op0 = gen_lowpart (V16QImode, op0);
12459 op1 = gen_lowpart (V16QImode, op1);
12460 emit_insn (gen_sse2_movdqu (op0, op1));
12464 if (TARGET_SSE2 && mode == V2DFmode)
12466 m = adjust_address (op0, DFmode, 0);
12467 emit_insn (gen_sse2_storelpd (m, op1));
12468 m = adjust_address (op0, DFmode, 8);
12469 emit_insn (gen_sse2_storehpd (m, op1));
12473 if (mode != V4SFmode)
12474 op1 = gen_lowpart (V4SFmode, op1);
12475 m = adjust_address (op0, V2SFmode, 0);
12476 emit_insn (gen_sse_storelps (m, op1));
12477 m = adjust_address (op0, V2SFmode, 8);
12478 emit_insn (gen_sse_storehps (m, op1));
12482 gcc_unreachable ();
12485 /* Expand a push in MODE. This is some mode for which we do not support
12486 proper push instructions, at least from the registers that we expect
12487 the value to live in. */
12490 ix86_expand_push (enum machine_mode mode, rtx x)
12494 tmp = expand_simple_binop (Pmode, PLUS, stack_pointer_rtx,
12495 GEN_INT (-GET_MODE_SIZE (mode)),
12496 stack_pointer_rtx, 1, OPTAB_DIRECT);
12497 if (tmp != stack_pointer_rtx)
12498 emit_move_insn (stack_pointer_rtx, tmp);
12500 tmp = gen_rtx_MEM (mode, stack_pointer_rtx);
12501 emit_move_insn (tmp, x);
12504 /* Helper function of ix86_fixup_binary_operands to canonicalize
12505 operand order. Returns true if the operands should be swapped. */
12508 ix86_swap_binary_operands_p (enum rtx_code code, enum machine_mode mode,
12511 rtx dst = operands[0];
12512 rtx src1 = operands[1];
12513 rtx src2 = operands[2];
12515 /* If the operation is not commutative, we can't do anything. */
12516 if (GET_RTX_CLASS (code) != RTX_COMM_ARITH)
12519 /* Highest priority is that src1 should match dst. */
12520 if (rtx_equal_p (dst, src1))
12522 if (rtx_equal_p (dst, src2))
12525 /* Next highest priority is that immediate constants come second. */
12526 if (immediate_operand (src2, mode))
12528 if (immediate_operand (src1, mode))
12531 /* Lowest priority is that memory references should come second. */
12541 /* Fix up OPERANDS to satisfy ix86_binary_operator_ok. Return the
12542 destination to use for the operation. If different from the true
12543 destination in operands[0], a copy operation will be required. */
12546 ix86_fixup_binary_operands (enum rtx_code code, enum machine_mode mode,
12549 rtx dst = operands[0];
12550 rtx src1 = operands[1];
12551 rtx src2 = operands[2];
12553 /* Canonicalize operand order. */
12554 if (ix86_swap_binary_operands_p (code, mode, operands))
12558 /* It is invalid to swap operands of different modes. */
12559 gcc_assert (GET_MODE (src1) == GET_MODE (src2));
12566 /* Both source operands cannot be in memory. */
12567 if (MEM_P (src1) && MEM_P (src2))
12569 /* Optimization: Only read from memory once. */
12570 if (rtx_equal_p (src1, src2))
12572 src2 = force_reg (mode, src2);
12576 src2 = force_reg (mode, src2);
12579 /* If the destination is memory, and we do not have matching source
12580 operands, do things in registers. */
12581 if (MEM_P (dst) && !rtx_equal_p (dst, src1))
12582 dst = gen_reg_rtx (mode);
12584 /* Source 1 cannot be a constant. */
12585 if (CONSTANT_P (src1))
12586 src1 = force_reg (mode, src1);
12588 /* Source 1 cannot be a non-matching memory. */
12589 if (MEM_P (src1) && !rtx_equal_p (dst, src1))
12590 src1 = force_reg (mode, src1);
12592 operands[1] = src1;
12593 operands[2] = src2;
12597 /* Similarly, but assume that the destination has already been
12598 set up properly. */
12601 ix86_fixup_binary_operands_no_copy (enum rtx_code code,
12602 enum machine_mode mode, rtx operands[])
12604 rtx dst = ix86_fixup_binary_operands (code, mode, operands);
12605 gcc_assert (dst == operands[0]);
12608 /* Attempt to expand a binary operator. Make the expansion closer to the
12609 actual machine, then just general_operand, which will allow 3 separate
12610 memory references (one output, two input) in a single insn. */
12613 ix86_expand_binary_operator (enum rtx_code code, enum machine_mode mode,
12616 rtx src1, src2, dst, op, clob;
12618 dst = ix86_fixup_binary_operands (code, mode, operands);
12619 src1 = operands[1];
12620 src2 = operands[2];
12622 /* Emit the instruction. */
12624 op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_ee (code, mode, src1, src2));
12625 if (reload_in_progress)
12627 /* Reload doesn't know about the flags register, and doesn't know that
12628 it doesn't want to clobber it. We can only do this with PLUS. */
12629 gcc_assert (code == PLUS);
12634 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
12635 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob)));
12638 /* Fix up the destination if needed. */
12639 if (dst != operands[0])
12640 emit_move_insn (operands[0], dst);
12643 /* Return TRUE or FALSE depending on whether the binary operator meets the
12644 appropriate constraints. */
12647 ix86_binary_operator_ok (enum rtx_code code, enum machine_mode mode,
12650 rtx dst = operands[0];
12651 rtx src1 = operands[1];
12652 rtx src2 = operands[2];
12654 /* Both source operands cannot be in memory. */
12655 if (MEM_P (src1) && MEM_P (src2))
12658 /* Canonicalize operand order for commutative operators. */
12659 if (ix86_swap_binary_operands_p (code, mode, operands))
12666 /* If the destination is memory, we must have a matching source operand. */
12667 if (MEM_P (dst) && !rtx_equal_p (dst, src1))
12670 /* Source 1 cannot be a constant. */
12671 if (CONSTANT_P (src1))
12674 /* Source 1 cannot be a non-matching memory. */
12675 if (MEM_P (src1) && !rtx_equal_p (dst, src1))
12681 /* Attempt to expand a unary operator. Make the expansion closer to the
12682 actual machine, then just general_operand, which will allow 2 separate
12683 memory references (one output, one input) in a single insn. */
12686 ix86_expand_unary_operator (enum rtx_code code, enum machine_mode mode,
12689 int matching_memory;
12690 rtx src, dst, op, clob;
12695 /* If the destination is memory, and we do not have matching source
12696 operands, do things in registers. */
12697 matching_memory = 0;
12700 if (rtx_equal_p (dst, src))
12701 matching_memory = 1;
12703 dst = gen_reg_rtx (mode);
12706 /* When source operand is memory, destination must match. */
12707 if (MEM_P (src) && !matching_memory)
12708 src = force_reg (mode, src);
12710 /* Emit the instruction. */
12712 op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_e (code, mode, src));
12713 if (reload_in_progress || code == NOT)
12715 /* Reload doesn't know about the flags register, and doesn't know that
12716 it doesn't want to clobber it. */
12717 gcc_assert (code == NOT);
12722 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
12723 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob)));
12726 /* Fix up the destination if needed. */
12727 if (dst != operands[0])
12728 emit_move_insn (operands[0], dst);
12731 /* Return TRUE or FALSE depending on whether the unary operator meets the
12732 appropriate constraints. */
12735 ix86_unary_operator_ok (enum rtx_code code ATTRIBUTE_UNUSED,
12736 enum machine_mode mode ATTRIBUTE_UNUSED,
12737 rtx operands[2] ATTRIBUTE_UNUSED)
12739 /* If one of operands is memory, source and destination must match. */
12740 if ((MEM_P (operands[0])
12741 || MEM_P (operands[1]))
12742 && ! rtx_equal_p (operands[0], operands[1]))
12747 /* Post-reload splitter for converting an SF or DFmode value in an
12748 SSE register into an unsigned SImode. */
12751 ix86_split_convert_uns_si_sse (rtx operands[])
12753 enum machine_mode vecmode;
12754 rtx value, large, zero_or_two31, input, two31, x;
12756 large = operands[1];
12757 zero_or_two31 = operands[2];
12758 input = operands[3];
12759 two31 = operands[4];
12760 vecmode = GET_MODE (large);
12761 value = gen_rtx_REG (vecmode, REGNO (operands[0]));
12763 /* Load up the value into the low element. We must ensure that the other
12764 elements are valid floats -- zero is the easiest such value. */
12767 if (vecmode == V4SFmode)
12768 emit_insn (gen_vec_setv4sf_0 (value, CONST0_RTX (V4SFmode), input));
12770 emit_insn (gen_sse2_loadlpd (value, CONST0_RTX (V2DFmode), input));
12774 input = gen_rtx_REG (vecmode, REGNO (input));
12775 emit_move_insn (value, CONST0_RTX (vecmode));
12776 if (vecmode == V4SFmode)
12777 emit_insn (gen_sse_movss (value, value, input));
12779 emit_insn (gen_sse2_movsd (value, value, input));
12782 emit_move_insn (large, two31);
12783 emit_move_insn (zero_or_two31, MEM_P (two31) ? large : two31);
12785 x = gen_rtx_fmt_ee (LE, vecmode, large, value);
12786 emit_insn (gen_rtx_SET (VOIDmode, large, x));
12788 x = gen_rtx_AND (vecmode, zero_or_two31, large);
12789 emit_insn (gen_rtx_SET (VOIDmode, zero_or_two31, x));
12791 x = gen_rtx_MINUS (vecmode, value, zero_or_two31);
12792 emit_insn (gen_rtx_SET (VOIDmode, value, x));
12794 large = gen_rtx_REG (V4SImode, REGNO (large));
12795 emit_insn (gen_ashlv4si3 (large, large, GEN_INT (31)));
12797 x = gen_rtx_REG (V4SImode, REGNO (value));
12798 if (vecmode == V4SFmode)
12799 emit_insn (gen_sse2_cvttps2dq (x, value));
12801 emit_insn (gen_sse2_cvttpd2dq (x, value));
12804 emit_insn (gen_xorv4si3 (value, value, large));
12807 /* Convert an unsigned DImode value into a DFmode, using only SSE.
12808 Expects the 64-bit DImode to be supplied in a pair of integral
12809 registers. Requires SSE2; will use SSE3 if available. For x86_32,
12810 -mfpmath=sse, !optimize_size only. */
12813 ix86_expand_convert_uns_didf_sse (rtx target, rtx input)
12815 REAL_VALUE_TYPE bias_lo_rvt, bias_hi_rvt;
12816 rtx int_xmm, fp_xmm;
12817 rtx biases, exponents;
12820 int_xmm = gen_reg_rtx (V4SImode);
12821 if (TARGET_INTER_UNIT_MOVES)
12822 emit_insn (gen_movdi_to_sse (int_xmm, input));
12823 else if (TARGET_SSE_SPLIT_REGS)
12825 emit_clobber (int_xmm);
12826 emit_move_insn (gen_lowpart (DImode, int_xmm), input);
12830 x = gen_reg_rtx (V2DImode);
12831 ix86_expand_vector_init_one_nonzero (false, V2DImode, x, input, 0);
12832 emit_move_insn (int_xmm, gen_lowpart (V4SImode, x));
12835 x = gen_rtx_CONST_VECTOR (V4SImode,
12836 gen_rtvec (4, GEN_INT (0x43300000UL),
12837 GEN_INT (0x45300000UL),
12838 const0_rtx, const0_rtx));
12839 exponents = validize_mem (force_const_mem (V4SImode, x));
12841 /* int_xmm = {0x45300000UL, fp_xmm/hi, 0x43300000, fp_xmm/lo } */
12842 emit_insn (gen_sse2_punpckldq (int_xmm, int_xmm, exponents));
12844 /* Concatenating (juxtaposing) (0x43300000UL ## fp_value_low_xmm)
12845 yields a valid DF value equal to (0x1.0p52 + double(fp_value_lo_xmm)).
12846 Similarly (0x45300000UL ## fp_value_hi_xmm) yields
12847 (0x1.0p84 + double(fp_value_hi_xmm)).
12848 Note these exponents differ by 32. */
12850 fp_xmm = copy_to_mode_reg (V2DFmode, gen_lowpart (V2DFmode, int_xmm));
12852 /* Subtract off those 0x1.0p52 and 0x1.0p84 biases, to produce values
12853 in [0,2**32-1] and [0]+[2**32,2**64-1] respectively. */
12854 real_ldexp (&bias_lo_rvt, &dconst1, 52);
12855 real_ldexp (&bias_hi_rvt, &dconst1, 84);
12856 biases = const_double_from_real_value (bias_lo_rvt, DFmode);
12857 x = const_double_from_real_value (bias_hi_rvt, DFmode);
12858 biases = gen_rtx_CONST_VECTOR (V2DFmode, gen_rtvec (2, biases, x));
12859 biases = validize_mem (force_const_mem (V2DFmode, biases));
12860 emit_insn (gen_subv2df3 (fp_xmm, fp_xmm, biases));
12862 /* Add the upper and lower DFmode values together. */
12864 emit_insn (gen_sse3_haddv2df3 (fp_xmm, fp_xmm, fp_xmm));
12867 x = copy_to_mode_reg (V2DFmode, fp_xmm);
12868 emit_insn (gen_sse2_unpckhpd (fp_xmm, fp_xmm, fp_xmm));
12869 emit_insn (gen_addv2df3 (fp_xmm, fp_xmm, x));
12872 ix86_expand_vector_extract (false, target, fp_xmm, 0);
12875 /* Not used, but eases macroization of patterns. */
12877 ix86_expand_convert_uns_sixf_sse (rtx target ATTRIBUTE_UNUSED,
12878 rtx input ATTRIBUTE_UNUSED)
12880 gcc_unreachable ();
12883 /* Convert an unsigned SImode value into a DFmode. Only currently used
12884 for SSE, but applicable anywhere. */
12887 ix86_expand_convert_uns_sidf_sse (rtx target, rtx input)
12889 REAL_VALUE_TYPE TWO31r;
12892 x = expand_simple_binop (SImode, PLUS, input, GEN_INT (-2147483647 - 1),
12893 NULL, 1, OPTAB_DIRECT);
12895 fp = gen_reg_rtx (DFmode);
12896 emit_insn (gen_floatsidf2 (fp, x));
12898 real_ldexp (&TWO31r, &dconst1, 31);
12899 x = const_double_from_real_value (TWO31r, DFmode);
12901 x = expand_simple_binop (DFmode, PLUS, fp, x, target, 0, OPTAB_DIRECT);
12903 emit_move_insn (target, x);
12906 /* Convert a signed DImode value into a DFmode. Only used for SSE in
12907 32-bit mode; otherwise we have a direct convert instruction. */
12910 ix86_expand_convert_sign_didf_sse (rtx target, rtx input)
12912 REAL_VALUE_TYPE TWO32r;
12913 rtx fp_lo, fp_hi, x;
12915 fp_lo = gen_reg_rtx (DFmode);
12916 fp_hi = gen_reg_rtx (DFmode);
12918 emit_insn (gen_floatsidf2 (fp_hi, gen_highpart (SImode, input)));
12920 real_ldexp (&TWO32r, &dconst1, 32);
12921 x = const_double_from_real_value (TWO32r, DFmode);
12922 fp_hi = expand_simple_binop (DFmode, MULT, fp_hi, x, fp_hi, 0, OPTAB_DIRECT);
12924 ix86_expand_convert_uns_sidf_sse (fp_lo, gen_lowpart (SImode, input));
12926 x = expand_simple_binop (DFmode, PLUS, fp_hi, fp_lo, target,
12929 emit_move_insn (target, x);
12932 /* Convert an unsigned SImode value into a SFmode, using only SSE.
12933 For x86_32, -mfpmath=sse, !optimize_size only. */
12935 ix86_expand_convert_uns_sisf_sse (rtx target, rtx input)
12937 REAL_VALUE_TYPE ONE16r;
12938 rtx fp_hi, fp_lo, int_hi, int_lo, x;
12940 real_ldexp (&ONE16r, &dconst1, 16);
12941 x = const_double_from_real_value (ONE16r, SFmode);
12942 int_lo = expand_simple_binop (SImode, AND, input, GEN_INT(0xffff),
12943 NULL, 0, OPTAB_DIRECT);
12944 int_hi = expand_simple_binop (SImode, LSHIFTRT, input, GEN_INT(16),
12945 NULL, 0, OPTAB_DIRECT);
12946 fp_hi = gen_reg_rtx (SFmode);
12947 fp_lo = gen_reg_rtx (SFmode);
12948 emit_insn (gen_floatsisf2 (fp_hi, int_hi));
12949 emit_insn (gen_floatsisf2 (fp_lo, int_lo));
12950 fp_hi = expand_simple_binop (SFmode, MULT, fp_hi, x, fp_hi,
12952 fp_hi = expand_simple_binop (SFmode, PLUS, fp_hi, fp_lo, target,
12954 if (!rtx_equal_p (target, fp_hi))
12955 emit_move_insn (target, fp_hi);
12958 /* A subroutine of ix86_build_signbit_mask_vector. If VECT is true,
12959 then replicate the value for all elements of the vector
12963 ix86_build_const_vector (enum machine_mode mode, bool vect, rtx value)
12970 v = gen_rtvec (4, value, value, value, value);
12971 return gen_rtx_CONST_VECTOR (V4SImode, v);
12975 v = gen_rtvec (2, value, value);
12976 return gen_rtx_CONST_VECTOR (V2DImode, v);
12980 v = gen_rtvec (4, value, value, value, value);
12982 v = gen_rtvec (4, value, CONST0_RTX (SFmode),
12983 CONST0_RTX (SFmode), CONST0_RTX (SFmode));
12984 return gen_rtx_CONST_VECTOR (V4SFmode, v);
12988 v = gen_rtvec (2, value, value);
12990 v = gen_rtvec (2, value, CONST0_RTX (DFmode));
12991 return gen_rtx_CONST_VECTOR (V2DFmode, v);
12994 gcc_unreachable ();
12998 /* A subroutine of ix86_expand_fp_absneg_operator, copysign expanders
12999 and ix86_expand_int_vcond. Create a mask for the sign bit in MODE
13000 for an SSE register. If VECT is true, then replicate the mask for
13001 all elements of the vector register. If INVERT is true, then create
13002 a mask excluding the sign bit. */
13005 ix86_build_signbit_mask (enum machine_mode mode, bool vect, bool invert)
13007 enum machine_mode vec_mode, imode;
13008 HOST_WIDE_INT hi, lo;
13013 /* Find the sign bit, sign extended to 2*HWI. */
13019 vec_mode = (mode == SImode) ? V4SImode : V4SFmode;
13020 lo = 0x80000000, hi = lo < 0;
13026 vec_mode = (mode == DImode) ? V2DImode : V2DFmode;
13027 if (HOST_BITS_PER_WIDE_INT >= 64)
13028 lo = (HOST_WIDE_INT)1 << shift, hi = -1;
13030 lo = 0, hi = (HOST_WIDE_INT)1 << (shift - HOST_BITS_PER_WIDE_INT);
13035 vec_mode = VOIDmode;
13036 if (HOST_BITS_PER_WIDE_INT >= 64)
13039 lo = 0, hi = (HOST_WIDE_INT)1 << shift;
13046 lo = 0, hi = (HOST_WIDE_INT)1 << (shift - HOST_BITS_PER_WIDE_INT);
13050 lo = ~lo, hi = ~hi;
13056 mask = immed_double_const (lo, hi, imode);
13058 vec = gen_rtvec (2, v, mask);
13059 v = gen_rtx_CONST_VECTOR (V2DImode, vec);
13060 v = copy_to_mode_reg (mode, gen_lowpart (mode, v));
13067 gcc_unreachable ();
13071 lo = ~lo, hi = ~hi;
13073 /* Force this value into the low part of a fp vector constant. */
13074 mask = immed_double_const (lo, hi, imode);
13075 mask = gen_lowpart (mode, mask);
13077 if (vec_mode == VOIDmode)
13078 return force_reg (mode, mask);
13080 v = ix86_build_const_vector (mode, vect, mask);
13081 return force_reg (vec_mode, v);
13084 /* Generate code for floating point ABS or NEG. */
13087 ix86_expand_fp_absneg_operator (enum rtx_code code, enum machine_mode mode,
13090 rtx mask, set, use, clob, dst, src;
13091 bool use_sse = false;
13092 bool vector_mode = VECTOR_MODE_P (mode);
13093 enum machine_mode elt_mode = mode;
13097 elt_mode = GET_MODE_INNER (mode);
13100 else if (mode == TFmode)
13102 else if (TARGET_SSE_MATH)
13103 use_sse = SSE_FLOAT_MODE_P (mode);
13105 /* NEG and ABS performed with SSE use bitwise mask operations.
13106 Create the appropriate mask now. */
13108 mask = ix86_build_signbit_mask (elt_mode, vector_mode, code == ABS);
13117 set = gen_rtx_fmt_ee (code == NEG ? XOR : AND, mode, src, mask);
13118 set = gen_rtx_SET (VOIDmode, dst, set);
13123 set = gen_rtx_fmt_e (code, mode, src);
13124 set = gen_rtx_SET (VOIDmode, dst, set);
13127 use = gen_rtx_USE (VOIDmode, mask);
13128 clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
13129 emit_insn (gen_rtx_PARALLEL (VOIDmode,
13130 gen_rtvec (3, set, use, clob)));
13137 /* Expand a copysign operation. Special case operand 0 being a constant. */
13140 ix86_expand_copysign (rtx operands[])
13142 enum machine_mode mode;
13143 rtx dest, op0, op1, mask, nmask;
13145 dest = operands[0];
13149 mode = GET_MODE (dest);
13151 if (GET_CODE (op0) == CONST_DOUBLE)
13153 rtx (*copysign_insn)(rtx, rtx, rtx, rtx);
13155 if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0)))
13156 op0 = simplify_unary_operation (ABS, mode, op0, mode);
13158 if (mode == SFmode || mode == DFmode)
13160 enum machine_mode vmode;
13162 vmode = mode == SFmode ? V4SFmode : V2DFmode;
13164 if (op0 == CONST0_RTX (mode))
13165 op0 = CONST0_RTX (vmode);
13170 if (mode == SFmode)
13171 v = gen_rtvec (4, op0, CONST0_RTX (SFmode),
13172 CONST0_RTX (SFmode), CONST0_RTX (SFmode));
13174 v = gen_rtvec (2, op0, CONST0_RTX (DFmode));
13176 op0 = force_reg (vmode, gen_rtx_CONST_VECTOR (vmode, v));
13179 else if (op0 != CONST0_RTX (mode))
13180 op0 = force_reg (mode, op0);
13182 mask = ix86_build_signbit_mask (mode, 0, 0);
13184 if (mode == SFmode)
13185 copysign_insn = gen_copysignsf3_const;
13186 else if (mode == DFmode)
13187 copysign_insn = gen_copysigndf3_const;
13189 copysign_insn = gen_copysigntf3_const;
13191 emit_insn (copysign_insn (dest, op0, op1, mask));
13195 rtx (*copysign_insn)(rtx, rtx, rtx, rtx, rtx, rtx);
13197 nmask = ix86_build_signbit_mask (mode, 0, 1);
13198 mask = ix86_build_signbit_mask (mode, 0, 0);
13200 if (mode == SFmode)
13201 copysign_insn = gen_copysignsf3_var;
13202 else if (mode == DFmode)
13203 copysign_insn = gen_copysigndf3_var;
13205 copysign_insn = gen_copysigntf3_var;
13207 emit_insn (copysign_insn (dest, NULL_RTX, op0, op1, nmask, mask));
13211 /* Deconstruct a copysign operation into bit masks. Operand 0 is known to
13212 be a constant, and so has already been expanded into a vector constant. */
13215 ix86_split_copysign_const (rtx operands[])
13217 enum machine_mode mode, vmode;
13218 rtx dest, op0, op1, mask, x;
13220 dest = operands[0];
13223 mask = operands[3];
13225 mode = GET_MODE (dest);
13226 vmode = GET_MODE (mask);
13228 dest = simplify_gen_subreg (vmode, dest, mode, 0);
13229 x = gen_rtx_AND (vmode, dest, mask);
13230 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
13232 if (op0 != CONST0_RTX (vmode))
13234 x = gen_rtx_IOR (vmode, dest, op0);
13235 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
13239 /* Deconstruct a copysign operation into bit masks. Operand 0 is variable,
13240 so we have to do two masks. */
13243 ix86_split_copysign_var (rtx operands[])
13245 enum machine_mode mode, vmode;
13246 rtx dest, scratch, op0, op1, mask, nmask, x;
13248 dest = operands[0];
13249 scratch = operands[1];
13252 nmask = operands[4];
13253 mask = operands[5];
13255 mode = GET_MODE (dest);
13256 vmode = GET_MODE (mask);
13258 if (rtx_equal_p (op0, op1))
13260 /* Shouldn't happen often (it's useless, obviously), but when it does
13261 we'd generate incorrect code if we continue below. */
13262 emit_move_insn (dest, op0);
13266 if (REG_P (mask) && REGNO (dest) == REGNO (mask)) /* alternative 0 */
13268 gcc_assert (REGNO (op1) == REGNO (scratch));
13270 x = gen_rtx_AND (vmode, scratch, mask);
13271 emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
13274 op0 = simplify_gen_subreg (vmode, op0, mode, 0);
13275 x = gen_rtx_NOT (vmode, dest);
13276 x = gen_rtx_AND (vmode, x, op0);
13277 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
13281 if (REGNO (op1) == REGNO (scratch)) /* alternative 1,3 */
13283 x = gen_rtx_AND (vmode, scratch, mask);
13285 else /* alternative 2,4 */
13287 gcc_assert (REGNO (mask) == REGNO (scratch));
13288 op1 = simplify_gen_subreg (vmode, op1, mode, 0);
13289 x = gen_rtx_AND (vmode, scratch, op1);
13291 emit_insn (gen_rtx_SET (VOIDmode, scratch, x));
13293 if (REGNO (op0) == REGNO (dest)) /* alternative 1,2 */
13295 dest = simplify_gen_subreg (vmode, op0, mode, 0);
13296 x = gen_rtx_AND (vmode, dest, nmask);
13298 else /* alternative 3,4 */
13300 gcc_assert (REGNO (nmask) == REGNO (dest));
13302 op0 = simplify_gen_subreg (vmode, op0, mode, 0);
13303 x = gen_rtx_AND (vmode, dest, op0);
13305 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
13308 x = gen_rtx_IOR (vmode, dest, scratch);
13309 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
13312 /* Return TRUE or FALSE depending on whether the first SET in INSN
13313 has source and destination with matching CC modes, and that the
13314 CC mode is at least as constrained as REQ_MODE. */
13317 ix86_match_ccmode (rtx insn, enum machine_mode req_mode)
13320 enum machine_mode set_mode;
13322 set = PATTERN (insn);
13323 if (GET_CODE (set) == PARALLEL)
13324 set = XVECEXP (set, 0, 0);
13325 gcc_assert (GET_CODE (set) == SET);
13326 gcc_assert (GET_CODE (SET_SRC (set)) == COMPARE);
13328 set_mode = GET_MODE (SET_DEST (set));
13332 if (req_mode != CCNOmode
13333 && (req_mode != CCmode
13334 || XEXP (SET_SRC (set), 1) != const0_rtx))
13338 if (req_mode == CCGCmode)
13342 if (req_mode == CCGOCmode || req_mode == CCNOmode)
13346 if (req_mode == CCZmode)
13357 gcc_unreachable ();
13360 return (GET_MODE (SET_SRC (set)) == set_mode);
13363 /* Generate insn patterns to do an integer compare of OPERANDS. */
13366 ix86_expand_int_compare (enum rtx_code code, rtx op0, rtx op1)
13368 enum machine_mode cmpmode;
13371 cmpmode = SELECT_CC_MODE (code, op0, op1);
13372 flags = gen_rtx_REG (cmpmode, FLAGS_REG);
13374 /* This is very simple, but making the interface the same as in the
13375 FP case makes the rest of the code easier. */
13376 tmp = gen_rtx_COMPARE (cmpmode, op0, op1);
13377 emit_insn (gen_rtx_SET (VOIDmode, flags, tmp));
13379 /* Return the test that should be put into the flags user, i.e.
13380 the bcc, scc, or cmov instruction. */
13381 return gen_rtx_fmt_ee (code, VOIDmode, flags, const0_rtx);
13384 /* Figure out whether to use ordered or unordered fp comparisons.
13385 Return the appropriate mode to use. */
13388 ix86_fp_compare_mode (enum rtx_code code ATTRIBUTE_UNUSED)
13390 /* ??? In order to make all comparisons reversible, we do all comparisons
13391 non-trapping when compiling for IEEE. Once gcc is able to distinguish
13392 all forms trapping and nontrapping comparisons, we can make inequality
13393 comparisons trapping again, since it results in better code when using
13394 FCOM based compares. */
13395 return TARGET_IEEE_FP ? CCFPUmode : CCFPmode;
13399 ix86_cc_mode (enum rtx_code code, rtx op0, rtx op1)
13401 enum machine_mode mode = GET_MODE (op0);
13403 if (SCALAR_FLOAT_MODE_P (mode))
13405 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode));
13406 return ix86_fp_compare_mode (code);
13411 /* Only zero flag is needed. */
13412 case EQ: /* ZF=0 */
13413 case NE: /* ZF!=0 */
13415 /* Codes needing carry flag. */
13416 case GEU: /* CF=0 */
13417 case LTU: /* CF=1 */
13418 /* Detect overflow checks. They need just the carry flag. */
13419 if (GET_CODE (op0) == PLUS
13420 && rtx_equal_p (op1, XEXP (op0, 0)))
13424 case GTU: /* CF=0 & ZF=0 */
13425 case LEU: /* CF=1 | ZF=1 */
13426 /* Detect overflow checks. They need just the carry flag. */
13427 if (GET_CODE (op0) == MINUS
13428 && rtx_equal_p (op1, XEXP (op0, 0)))
13432 /* Codes possibly doable only with sign flag when
13433 comparing against zero. */
13434 case GE: /* SF=OF or SF=0 */
13435 case LT: /* SF<>OF or SF=1 */
13436 if (op1 == const0_rtx)
13439 /* For other cases Carry flag is not required. */
13441 /* Codes doable only with sign flag when comparing
13442 against zero, but we miss jump instruction for it
13443 so we need to use relational tests against overflow
13444 that thus needs to be zero. */
13445 case GT: /* ZF=0 & SF=OF */
13446 case LE: /* ZF=1 | SF<>OF */
13447 if (op1 == const0_rtx)
13451 /* strcmp pattern do (use flags) and combine may ask us for proper
13456 gcc_unreachable ();
13460 /* Return the fixed registers used for condition codes. */
13463 ix86_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2)
13470 /* If two condition code modes are compatible, return a condition code
13471 mode which is compatible with both. Otherwise, return
13474 static enum machine_mode
13475 ix86_cc_modes_compatible (enum machine_mode m1, enum machine_mode m2)
13480 if (GET_MODE_CLASS (m1) != MODE_CC || GET_MODE_CLASS (m2) != MODE_CC)
13483 if ((m1 == CCGCmode && m2 == CCGOCmode)
13484 || (m1 == CCGOCmode && m2 == CCGCmode))
13490 gcc_unreachable ();
13520 /* These are only compatible with themselves, which we already
13526 /* Split comparison code CODE into comparisons we can do using branch
13527 instructions. BYPASS_CODE is comparison code for branch that will
13528 branch around FIRST_CODE and SECOND_CODE. If some of branches
13529 is not required, set value to UNKNOWN.
13530 We never require more than two branches. */
13533 ix86_fp_comparison_codes (enum rtx_code code, enum rtx_code *bypass_code,
13534 enum rtx_code *first_code,
13535 enum rtx_code *second_code)
13537 *first_code = code;
13538 *bypass_code = UNKNOWN;
13539 *second_code = UNKNOWN;
13541 /* The fcomi comparison sets flags as follows:
13551 case GT: /* GTU - CF=0 & ZF=0 */
13552 case GE: /* GEU - CF=0 */
13553 case ORDERED: /* PF=0 */
13554 case UNORDERED: /* PF=1 */
13555 case UNEQ: /* EQ - ZF=1 */
13556 case UNLT: /* LTU - CF=1 */
13557 case UNLE: /* LEU - CF=1 | ZF=1 */
13558 case LTGT: /* EQ - ZF=0 */
13560 case LT: /* LTU - CF=1 - fails on unordered */
13561 *first_code = UNLT;
13562 *bypass_code = UNORDERED;
13564 case LE: /* LEU - CF=1 | ZF=1 - fails on unordered */
13565 *first_code = UNLE;
13566 *bypass_code = UNORDERED;
13568 case EQ: /* EQ - ZF=1 - fails on unordered */
13569 *first_code = UNEQ;
13570 *bypass_code = UNORDERED;
13572 case NE: /* NE - ZF=0 - fails on unordered */
13573 *first_code = LTGT;
13574 *second_code = UNORDERED;
13576 case UNGE: /* GEU - CF=0 - fails on unordered */
13578 *second_code = UNORDERED;
13580 case UNGT: /* GTU - CF=0 & ZF=0 - fails on unordered */
13582 *second_code = UNORDERED;
13585 gcc_unreachable ();
13587 if (!TARGET_IEEE_FP)
13589 *second_code = UNKNOWN;
13590 *bypass_code = UNKNOWN;
13594 /* Return cost of comparison done fcom + arithmetics operations on AX.
13595 All following functions do use number of instructions as a cost metrics.
13596 In future this should be tweaked to compute bytes for optimize_size and
13597 take into account performance of various instructions on various CPUs. */
13599 ix86_fp_comparison_arithmetics_cost (enum rtx_code code)
13601 if (!TARGET_IEEE_FP)
13603 /* The cost of code output by ix86_expand_fp_compare. */
13627 gcc_unreachable ();
13631 /* Return cost of comparison done using fcomi operation.
13632 See ix86_fp_comparison_arithmetics_cost for the metrics. */
13634 ix86_fp_comparison_fcomi_cost (enum rtx_code code)
13636 enum rtx_code bypass_code, first_code, second_code;
13637 /* Return arbitrarily high cost when instruction is not supported - this
13638 prevents gcc from using it. */
13641 ix86_fp_comparison_codes (code, &bypass_code, &first_code, &second_code);
13642 return (bypass_code != UNKNOWN || second_code != UNKNOWN) + 2;
13645 /* Return cost of comparison done using sahf operation.
13646 See ix86_fp_comparison_arithmetics_cost for the metrics. */
13648 ix86_fp_comparison_sahf_cost (enum rtx_code code)
13650 enum rtx_code bypass_code, first_code, second_code;
13651 /* Return arbitrarily high cost when instruction is not preferred - this
13652 avoids gcc from using it. */
13653 if (!(TARGET_SAHF && (TARGET_USE_SAHF || optimize_insn_for_size_p ())))
13655 ix86_fp_comparison_codes (code, &bypass_code, &first_code, &second_code);
13656 return (bypass_code != UNKNOWN || second_code != UNKNOWN) + 3;
13659 /* Compute cost of the comparison done using any method.
13660 See ix86_fp_comparison_arithmetics_cost for the metrics. */
13662 ix86_fp_comparison_cost (enum rtx_code code)
13664 int fcomi_cost, sahf_cost, arithmetics_cost = 1024;
13667 fcomi_cost = ix86_fp_comparison_fcomi_cost (code);
13668 sahf_cost = ix86_fp_comparison_sahf_cost (code);
13670 min = arithmetics_cost = ix86_fp_comparison_arithmetics_cost (code);
13671 if (min > sahf_cost)
13673 if (min > fcomi_cost)
13678 /* Return true if we should use an FCOMI instruction for this
13682 ix86_use_fcomi_compare (enum rtx_code code ATTRIBUTE_UNUSED)
13684 enum rtx_code swapped_code = swap_condition (code);
13686 return ((ix86_fp_comparison_cost (code)
13687 == ix86_fp_comparison_fcomi_cost (code))
13688 || (ix86_fp_comparison_cost (swapped_code)
13689 == ix86_fp_comparison_fcomi_cost (swapped_code)));
13692 /* Swap, force into registers, or otherwise massage the two operands
13693 to a fp comparison. The operands are updated in place; the new
13694 comparison code is returned. */
13696 static enum rtx_code
13697 ix86_prepare_fp_compare_args (enum rtx_code code, rtx *pop0, rtx *pop1)
13699 enum machine_mode fpcmp_mode = ix86_fp_compare_mode (code);
13700 rtx op0 = *pop0, op1 = *pop1;
13701 enum machine_mode op_mode = GET_MODE (op0);
13702 int is_sse = TARGET_SSE_MATH && SSE_FLOAT_MODE_P (op_mode);
13704 /* All of the unordered compare instructions only work on registers.
13705 The same is true of the fcomi compare instructions. The XFmode
13706 compare instructions require registers except when comparing
13707 against zero or when converting operand 1 from fixed point to
13711 && (fpcmp_mode == CCFPUmode
13712 || (op_mode == XFmode
13713 && ! (standard_80387_constant_p (op0) == 1
13714 || standard_80387_constant_p (op1) == 1)
13715 && GET_CODE (op1) != FLOAT)
13716 || ix86_use_fcomi_compare (code)))
13718 op0 = force_reg (op_mode, op0);
13719 op1 = force_reg (op_mode, op1);
13723 /* %%% We only allow op1 in memory; op0 must be st(0). So swap
13724 things around if they appear profitable, otherwise force op0
13725 into a register. */
13727 if (standard_80387_constant_p (op0) == 0
13729 && ! (standard_80387_constant_p (op1) == 0
13733 tmp = op0, op0 = op1, op1 = tmp;
13734 code = swap_condition (code);
13738 op0 = force_reg (op_mode, op0);
13740 if (CONSTANT_P (op1))
13742 int tmp = standard_80387_constant_p (op1);
13744 op1 = validize_mem (force_const_mem (op_mode, op1));
13748 op1 = force_reg (op_mode, op1);
13751 op1 = force_reg (op_mode, op1);
13755 /* Try to rearrange the comparison to make it cheaper. */
13756 if (ix86_fp_comparison_cost (code)
13757 > ix86_fp_comparison_cost (swap_condition (code))
13758 && (REG_P (op1) || can_create_pseudo_p ()))
13761 tmp = op0, op0 = op1, op1 = tmp;
13762 code = swap_condition (code);
13764 op0 = force_reg (op_mode, op0);
13772 /* Convert comparison codes we use to represent FP comparison to integer
13773 code that will result in proper branch. Return UNKNOWN if no such code
13777 ix86_fp_compare_code_to_integer (enum rtx_code code)
13806 /* Generate insn patterns to do a floating point compare of OPERANDS. */
13809 ix86_expand_fp_compare (enum rtx_code code, rtx op0, rtx op1, rtx scratch,
13810 rtx *second_test, rtx *bypass_test)
13812 enum machine_mode fpcmp_mode, intcmp_mode;
13814 int cost = ix86_fp_comparison_cost (code);
13815 enum rtx_code bypass_code, first_code, second_code;
13817 fpcmp_mode = ix86_fp_compare_mode (code);
13818 code = ix86_prepare_fp_compare_args (code, &op0, &op1);
13821 *second_test = NULL_RTX;
13823 *bypass_test = NULL_RTX;
13825 ix86_fp_comparison_codes (code, &bypass_code, &first_code, &second_code);
13827 /* Do fcomi/sahf based test when profitable. */
13828 if (ix86_fp_comparison_arithmetics_cost (code) > cost
13829 && (bypass_code == UNKNOWN || bypass_test)
13830 && (second_code == UNKNOWN || second_test))
13832 tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1);
13833 tmp = gen_rtx_SET (VOIDmode, gen_rtx_REG (fpcmp_mode, FLAGS_REG),
13839 gcc_assert (TARGET_SAHF);
13842 scratch = gen_reg_rtx (HImode);
13843 tmp2 = gen_rtx_CLOBBER (VOIDmode, scratch);
13845 emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, tmp2)));
13848 /* The FP codes work out to act like unsigned. */
13849 intcmp_mode = fpcmp_mode;
13851 if (bypass_code != UNKNOWN)
13852 *bypass_test = gen_rtx_fmt_ee (bypass_code, VOIDmode,
13853 gen_rtx_REG (intcmp_mode, FLAGS_REG),
13855 if (second_code != UNKNOWN)
13856 *second_test = gen_rtx_fmt_ee (second_code, VOIDmode,
13857 gen_rtx_REG (intcmp_mode, FLAGS_REG),
13862 /* Sadness wrt reg-stack pops killing fpsr -- gotta get fnstsw first. */
13863 tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1);
13864 tmp2 = gen_rtx_UNSPEC (HImode, gen_rtvec (1, tmp), UNSPEC_FNSTSW);
13866 scratch = gen_reg_rtx (HImode);
13867 emit_insn (gen_rtx_SET (VOIDmode, scratch, tmp2));
13869 /* In the unordered case, we have to check C2 for NaN's, which
13870 doesn't happen to work out to anything nice combination-wise.
13871 So do some bit twiddling on the value we've got in AH to come
13872 up with an appropriate set of condition codes. */
13874 intcmp_mode = CCNOmode;
13879 if (code == GT || !TARGET_IEEE_FP)
13881 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x45)));
13886 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13887 emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx));
13888 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x44)));
13889 intcmp_mode = CCmode;
13895 if (code == LT && TARGET_IEEE_FP)
13897 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13898 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x01)));
13899 intcmp_mode = CCmode;
13904 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x01)));
13910 if (code == GE || !TARGET_IEEE_FP)
13912 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x05)));
13917 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13918 emit_insn (gen_xorqi_cc_ext_1 (scratch, scratch,
13925 if (code == LE && TARGET_IEEE_FP)
13927 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13928 emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx));
13929 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40)));
13930 intcmp_mode = CCmode;
13935 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x45)));
13941 if (code == EQ && TARGET_IEEE_FP)
13943 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13944 emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40)));
13945 intcmp_mode = CCmode;
13950 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x40)));
13957 if (code == NE && TARGET_IEEE_FP)
13959 emit_insn (gen_andqi_ext_0 (scratch, scratch, GEN_INT (0x45)));
13960 emit_insn (gen_xorqi_cc_ext_1 (scratch, scratch,
13966 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x40)));
13972 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x04)));
13976 emit_insn (gen_testqi_ext_ccno_0 (scratch, GEN_INT (0x04)));
13981 gcc_unreachable ();
13985 /* Return the test that should be put into the flags user, i.e.
13986 the bcc, scc, or cmov instruction. */
13987 return gen_rtx_fmt_ee (code, VOIDmode,
13988 gen_rtx_REG (intcmp_mode, FLAGS_REG),
13993 ix86_expand_compare (enum rtx_code code, rtx *second_test, rtx *bypass_test)
13996 op0 = ix86_compare_op0;
13997 op1 = ix86_compare_op1;
14000 *second_test = NULL_RTX;
14002 *bypass_test = NULL_RTX;
14004 if (ix86_compare_emitted)
14006 ret = gen_rtx_fmt_ee (code, VOIDmode, ix86_compare_emitted, const0_rtx);
14007 ix86_compare_emitted = NULL_RTX;
14009 else if (SCALAR_FLOAT_MODE_P (GET_MODE (op0)))
14011 gcc_assert (!DECIMAL_FLOAT_MODE_P (GET_MODE (op0)));
14012 ret = ix86_expand_fp_compare (code, op0, op1, NULL_RTX,
14013 second_test, bypass_test);
14016 ret = ix86_expand_int_compare (code, op0, op1);
14021 /* Return true if the CODE will result in nontrivial jump sequence. */
14023 ix86_fp_jump_nontrivial_p (enum rtx_code code)
14025 enum rtx_code bypass_code, first_code, second_code;
14028 ix86_fp_comparison_codes (code, &bypass_code, &first_code, &second_code);
14029 return bypass_code != UNKNOWN || second_code != UNKNOWN;
14033 ix86_expand_branch (enum rtx_code code, rtx label)
14037 /* If we have emitted a compare insn, go straight to simple.
14038 ix86_expand_compare won't emit anything if ix86_compare_emitted
14040 if (ix86_compare_emitted)
14043 switch (GET_MODE (ix86_compare_op0))
14049 tmp = ix86_expand_compare (code, NULL, NULL);
14050 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
14051 gen_rtx_LABEL_REF (VOIDmode, label),
14053 emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
14062 enum rtx_code bypass_code, first_code, second_code;
14064 code = ix86_prepare_fp_compare_args (code, &ix86_compare_op0,
14065 &ix86_compare_op1);
14067 ix86_fp_comparison_codes (code, &bypass_code, &first_code, &second_code);
14069 /* Check whether we will use the natural sequence with one jump. If
14070 so, we can expand jump early. Otherwise delay expansion by
14071 creating compound insn to not confuse optimizers. */
14072 if (bypass_code == UNKNOWN && second_code == UNKNOWN)
14074 ix86_split_fp_branch (code, ix86_compare_op0, ix86_compare_op1,
14075 gen_rtx_LABEL_REF (VOIDmode, label),
14076 pc_rtx, NULL_RTX, NULL_RTX);
14080 tmp = gen_rtx_fmt_ee (code, VOIDmode,
14081 ix86_compare_op0, ix86_compare_op1);
14082 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
14083 gen_rtx_LABEL_REF (VOIDmode, label),
14085 tmp = gen_rtx_SET (VOIDmode, pc_rtx, tmp);
14087 use_fcomi = ix86_use_fcomi_compare (code);
14088 vec = rtvec_alloc (3 + !use_fcomi);
14089 RTVEC_ELT (vec, 0) = tmp;
14091 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCFPmode, FPSR_REG));
14093 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCFPmode, FLAGS_REG));
14096 = gen_rtx_CLOBBER (VOIDmode, gen_rtx_SCRATCH (HImode));
14098 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode, vec));
14107 /* Expand DImode branch into multiple compare+branch. */
14109 rtx lo[2], hi[2], label2;
14110 enum rtx_code code1, code2, code3;
14111 enum machine_mode submode;
14113 if (CONSTANT_P (ix86_compare_op0) && ! CONSTANT_P (ix86_compare_op1))
14115 tmp = ix86_compare_op0;
14116 ix86_compare_op0 = ix86_compare_op1;
14117 ix86_compare_op1 = tmp;
14118 code = swap_condition (code);
14120 if (GET_MODE (ix86_compare_op0) == DImode)
14122 split_di (&ix86_compare_op0, 1, lo+0, hi+0);
14123 split_di (&ix86_compare_op1, 1, lo+1, hi+1);
14128 split_ti (&ix86_compare_op0, 1, lo+0, hi+0);
14129 split_ti (&ix86_compare_op1, 1, lo+1, hi+1);
14133 /* When comparing for equality, we can use (hi0^hi1)|(lo0^lo1) to
14134 avoid two branches. This costs one extra insn, so disable when
14135 optimizing for size. */
14137 if ((code == EQ || code == NE)
14138 && (!optimize_insn_for_size_p ()
14139 || hi[1] == const0_rtx || lo[1] == const0_rtx))
14144 if (hi[1] != const0_rtx)
14145 xor1 = expand_binop (submode, xor_optab, xor1, hi[1],
14146 NULL_RTX, 0, OPTAB_WIDEN);
14149 if (lo[1] != const0_rtx)
14150 xor0 = expand_binop (submode, xor_optab, xor0, lo[1],
14151 NULL_RTX, 0, OPTAB_WIDEN);
14153 tmp = expand_binop (submode, ior_optab, xor1, xor0,
14154 NULL_RTX, 0, OPTAB_WIDEN);
14156 ix86_compare_op0 = tmp;
14157 ix86_compare_op1 = const0_rtx;
14158 ix86_expand_branch (code, label);
14162 /* Otherwise, if we are doing less-than or greater-or-equal-than,
14163 op1 is a constant and the low word is zero, then we can just
14164 examine the high word. Similarly for low word -1 and
14165 less-or-equal-than or greater-than. */
14167 if (CONST_INT_P (hi[1]))
14170 case LT: case LTU: case GE: case GEU:
14171 if (lo[1] == const0_rtx)
14173 ix86_compare_op0 = hi[0];
14174 ix86_compare_op1 = hi[1];
14175 ix86_expand_branch (code, label);
14179 case LE: case LEU: case GT: case GTU:
14180 if (lo[1] == constm1_rtx)
14182 ix86_compare_op0 = hi[0];
14183 ix86_compare_op1 = hi[1];
14184 ix86_expand_branch (code, label);
14192 /* Otherwise, we need two or three jumps. */
14194 label2 = gen_label_rtx ();
14197 code2 = swap_condition (code);
14198 code3 = unsigned_condition (code);
14202 case LT: case GT: case LTU: case GTU:
14205 case LE: code1 = LT; code2 = GT; break;
14206 case GE: code1 = GT; code2 = LT; break;
14207 case LEU: code1 = LTU; code2 = GTU; break;
14208 case GEU: code1 = GTU; code2 = LTU; break;
14210 case EQ: code1 = UNKNOWN; code2 = NE; break;
14211 case NE: code2 = UNKNOWN; break;
14214 gcc_unreachable ();
14219 * if (hi(a) < hi(b)) goto true;
14220 * if (hi(a) > hi(b)) goto false;
14221 * if (lo(a) < lo(b)) goto true;
14225 ix86_compare_op0 = hi[0];
14226 ix86_compare_op1 = hi[1];
14228 if (code1 != UNKNOWN)
14229 ix86_expand_branch (code1, label);
14230 if (code2 != UNKNOWN)
14231 ix86_expand_branch (code2, label2);
14233 ix86_compare_op0 = lo[0];
14234 ix86_compare_op1 = lo[1];
14235 ix86_expand_branch (code3, label);
14237 if (code2 != UNKNOWN)
14238 emit_label (label2);
14243 gcc_unreachable ();
14247 /* Split branch based on floating point condition. */
14249 ix86_split_fp_branch (enum rtx_code code, rtx op1, rtx op2,
14250 rtx target1, rtx target2, rtx tmp, rtx pushed)
14252 rtx second, bypass;
14253 rtx label = NULL_RTX;
14255 int bypass_probability = -1, second_probability = -1, probability = -1;
14258 if (target2 != pc_rtx)
14261 code = reverse_condition_maybe_unordered (code);
14266 condition = ix86_expand_fp_compare (code, op1, op2,
14267 tmp, &second, &bypass);
14269 /* Remove pushed operand from stack. */
14271 ix86_free_from_memory (GET_MODE (pushed));
14273 if (split_branch_probability >= 0)
14275 /* Distribute the probabilities across the jumps.
14276 Assume the BYPASS and SECOND to be always test
14278 probability = split_branch_probability;
14280 /* Value of 1 is low enough to make no need for probability
14281 to be updated. Later we may run some experiments and see
14282 if unordered values are more frequent in practice. */
14284 bypass_probability = 1;
14286 second_probability = 1;
14288 if (bypass != NULL_RTX)
14290 label = gen_label_rtx ();
14291 i = emit_jump_insn (gen_rtx_SET
14293 gen_rtx_IF_THEN_ELSE (VOIDmode,
14295 gen_rtx_LABEL_REF (VOIDmode,
14298 if (bypass_probability >= 0)
14300 = gen_rtx_EXPR_LIST (REG_BR_PROB,
14301 GEN_INT (bypass_probability),
14304 i = emit_jump_insn (gen_rtx_SET
14306 gen_rtx_IF_THEN_ELSE (VOIDmode,
14307 condition, target1, target2)));
14308 if (probability >= 0)
14310 = gen_rtx_EXPR_LIST (REG_BR_PROB,
14311 GEN_INT (probability),
14313 if (second != NULL_RTX)
14315 i = emit_jump_insn (gen_rtx_SET
14317 gen_rtx_IF_THEN_ELSE (VOIDmode, second, target1,
14319 if (second_probability >= 0)
14321 = gen_rtx_EXPR_LIST (REG_BR_PROB,
14322 GEN_INT (second_probability),
14325 if (label != NULL_RTX)
14326 emit_label (label);
14330 ix86_expand_setcc (enum rtx_code code, rtx dest)
14332 rtx ret, tmp, tmpreg, equiv;
14333 rtx second_test, bypass_test;
14335 if (GET_MODE (ix86_compare_op0) == (TARGET_64BIT ? TImode : DImode))
14336 return 0; /* FAIL */
14338 gcc_assert (GET_MODE (dest) == QImode);
14340 ret = ix86_expand_compare (code, &second_test, &bypass_test);
14341 PUT_MODE (ret, QImode);
14346 emit_insn (gen_rtx_SET (VOIDmode, tmp, ret));
14347 if (bypass_test || second_test)
14349 rtx test = second_test;
14351 rtx tmp2 = gen_reg_rtx (QImode);
14354 gcc_assert (!second_test);
14355 test = bypass_test;
14357 PUT_CODE (test, reverse_condition_maybe_unordered (GET_CODE (test)));
14359 PUT_MODE (test, QImode);
14360 emit_insn (gen_rtx_SET (VOIDmode, tmp2, test));
14363 emit_insn (gen_andqi3 (tmp, tmpreg, tmp2));
14365 emit_insn (gen_iorqi3 (tmp, tmpreg, tmp2));
14368 /* Attach a REG_EQUAL note describing the comparison result. */
14369 if (ix86_compare_op0 && ix86_compare_op1)
14371 equiv = simplify_gen_relational (code, QImode,
14372 GET_MODE (ix86_compare_op0),
14373 ix86_compare_op0, ix86_compare_op1);
14374 set_unique_reg_note (get_last_insn (), REG_EQUAL, equiv);
14377 return 1; /* DONE */
14380 /* Expand comparison setting or clearing carry flag. Return true when
14381 successful and set pop for the operation. */
14383 ix86_expand_carry_flag_compare (enum rtx_code code, rtx op0, rtx op1, rtx *pop)
14385 enum machine_mode mode =
14386 GET_MODE (op0) != VOIDmode ? GET_MODE (op0) : GET_MODE (op1);
14388 /* Do not handle DImode compares that go through special path. */
14389 if (mode == (TARGET_64BIT ? TImode : DImode))
14392 if (SCALAR_FLOAT_MODE_P (mode))
14394 rtx second_test = NULL, bypass_test = NULL;
14395 rtx compare_op, compare_seq;
14397 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode));
14399 /* Shortcut: following common codes never translate
14400 into carry flag compares. */
14401 if (code == EQ || code == NE || code == UNEQ || code == LTGT
14402 || code == ORDERED || code == UNORDERED)
14405 /* These comparisons require zero flag; swap operands so they won't. */
14406 if ((code == GT || code == UNLE || code == LE || code == UNGT)
14407 && !TARGET_IEEE_FP)
14412 code = swap_condition (code);
14415 /* Try to expand the comparison and verify that we end up with
14416 carry flag based comparison. This fails to be true only when
14417 we decide to expand comparison using arithmetic that is not
14418 too common scenario. */
14420 compare_op = ix86_expand_fp_compare (code, op0, op1, NULL_RTX,
14421 &second_test, &bypass_test);
14422 compare_seq = get_insns ();
14425 if (second_test || bypass_test)
14428 if (GET_MODE (XEXP (compare_op, 0)) == CCFPmode
14429 || GET_MODE (XEXP (compare_op, 0)) == CCFPUmode)
14430 code = ix86_fp_compare_code_to_integer (GET_CODE (compare_op));
14432 code = GET_CODE (compare_op);
14434 if (code != LTU && code != GEU)
14437 emit_insn (compare_seq);
14442 if (!INTEGRAL_MODE_P (mode))
14451 /* Convert a==0 into (unsigned)a<1. */
14454 if (op1 != const0_rtx)
14457 code = (code == EQ ? LTU : GEU);
14460 /* Convert a>b into b<a or a>=b-1. */
14463 if (CONST_INT_P (op1))
14465 op1 = gen_int_mode (INTVAL (op1) + 1, GET_MODE (op0));
14466 /* Bail out on overflow. We still can swap operands but that
14467 would force loading of the constant into register. */
14468 if (op1 == const0_rtx
14469 || !x86_64_immediate_operand (op1, GET_MODE (op1)))
14471 code = (code == GTU ? GEU : LTU);
14478 code = (code == GTU ? LTU : GEU);
14482 /* Convert a>=0 into (unsigned)a<0x80000000. */
14485 if (mode == DImode || op1 != const0_rtx)
14487 op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode);
14488 code = (code == LT ? GEU : LTU);
14492 if (mode == DImode || op1 != constm1_rtx)
14494 op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode);
14495 code = (code == LE ? GEU : LTU);
14501 /* Swapping operands may cause constant to appear as first operand. */
14502 if (!nonimmediate_operand (op0, VOIDmode))
14504 if (!can_create_pseudo_p ())
14506 op0 = force_reg (mode, op0);
14508 ix86_compare_op0 = op0;
14509 ix86_compare_op1 = op1;
14510 *pop = ix86_expand_compare (code, NULL, NULL);
14511 gcc_assert (GET_CODE (*pop) == LTU || GET_CODE (*pop) == GEU);
14516 ix86_expand_int_movcc (rtx operands[])
14518 enum rtx_code code = GET_CODE (operands[1]), compare_code;
14519 rtx compare_seq, compare_op;
14520 rtx second_test, bypass_test;
14521 enum machine_mode mode = GET_MODE (operands[0]);
14522 bool sign_bit_compare_p = false;;
14525 compare_op = ix86_expand_compare (code, &second_test, &bypass_test);
14526 compare_seq = get_insns ();
14529 compare_code = GET_CODE (compare_op);
14531 if ((ix86_compare_op1 == const0_rtx && (code == GE || code == LT))
14532 || (ix86_compare_op1 == constm1_rtx && (code == GT || code == LE)))
14533 sign_bit_compare_p = true;
14535 /* Don't attempt mode expansion here -- if we had to expand 5 or 6
14536 HImode insns, we'd be swallowed in word prefix ops. */
14538 if ((mode != HImode || TARGET_FAST_PREFIX)
14539 && (mode != (TARGET_64BIT ? TImode : DImode))
14540 && CONST_INT_P (operands[2])
14541 && CONST_INT_P (operands[3]))
14543 rtx out = operands[0];
14544 HOST_WIDE_INT ct = INTVAL (operands[2]);
14545 HOST_WIDE_INT cf = INTVAL (operands[3]);
14546 HOST_WIDE_INT diff;
14549 /* Sign bit compares are better done using shifts than we do by using
14551 if (sign_bit_compare_p
14552 || ix86_expand_carry_flag_compare (code, ix86_compare_op0,
14553 ix86_compare_op1, &compare_op))
14555 /* Detect overlap between destination and compare sources. */
14558 if (!sign_bit_compare_p)
14560 bool fpcmp = false;
14562 compare_code = GET_CODE (compare_op);
14564 if (GET_MODE (XEXP (compare_op, 0)) == CCFPmode
14565 || GET_MODE (XEXP (compare_op, 0)) == CCFPUmode)
14568 compare_code = ix86_fp_compare_code_to_integer (compare_code);
14571 /* To simplify rest of code, restrict to the GEU case. */
14572 if (compare_code == LTU)
14574 HOST_WIDE_INT tmp = ct;
14577 compare_code = reverse_condition (compare_code);
14578 code = reverse_condition (code);
14583 PUT_CODE (compare_op,
14584 reverse_condition_maybe_unordered
14585 (GET_CODE (compare_op)));
14587 PUT_CODE (compare_op, reverse_condition (GET_CODE (compare_op)));
14591 if (reg_overlap_mentioned_p (out, ix86_compare_op0)
14592 || reg_overlap_mentioned_p (out, ix86_compare_op1))
14593 tmp = gen_reg_rtx (mode);
14595 if (mode == DImode)
14596 emit_insn (gen_x86_movdicc_0_m1_rex64 (tmp, compare_op));
14598 emit_insn (gen_x86_movsicc_0_m1 (gen_lowpart (SImode, tmp), compare_op));
14602 if (code == GT || code == GE)
14603 code = reverse_condition (code);
14606 HOST_WIDE_INT tmp = ct;
14611 tmp = emit_store_flag (tmp, code, ix86_compare_op0,
14612 ix86_compare_op1, VOIDmode, 0, -1);
14625 tmp = expand_simple_binop (mode, PLUS,
14627 copy_rtx (tmp), 1, OPTAB_DIRECT);
14638 tmp = expand_simple_binop (mode, IOR,
14640 copy_rtx (tmp), 1, OPTAB_DIRECT);
14642 else if (diff == -1 && ct)
14652 tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1);
14654 tmp = expand_simple_binop (mode, PLUS,
14655 copy_rtx (tmp), GEN_INT (cf),
14656 copy_rtx (tmp), 1, OPTAB_DIRECT);
14664 * andl cf - ct, dest
14674 tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1);
14677 tmp = expand_simple_binop (mode, AND,
14679 gen_int_mode (cf - ct, mode),
14680 copy_rtx (tmp), 1, OPTAB_DIRECT);
14682 tmp = expand_simple_binop (mode, PLUS,
14683 copy_rtx (tmp), GEN_INT (ct),
14684 copy_rtx (tmp), 1, OPTAB_DIRECT);
14687 if (!rtx_equal_p (tmp, out))
14688 emit_move_insn (copy_rtx (out), copy_rtx (tmp));
14690 return 1; /* DONE */
14695 enum machine_mode cmp_mode = GET_MODE (ix86_compare_op0);
14698 tmp = ct, ct = cf, cf = tmp;
14701 if (SCALAR_FLOAT_MODE_P (cmp_mode))
14703 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode));
14705 /* We may be reversing unordered compare to normal compare, that
14706 is not valid in general (we may convert non-trapping condition
14707 to trapping one), however on i386 we currently emit all
14708 comparisons unordered. */
14709 compare_code = reverse_condition_maybe_unordered (compare_code);
14710 code = reverse_condition_maybe_unordered (code);
14714 compare_code = reverse_condition (compare_code);
14715 code = reverse_condition (code);
14719 compare_code = UNKNOWN;
14720 if (GET_MODE_CLASS (GET_MODE (ix86_compare_op0)) == MODE_INT
14721 && CONST_INT_P (ix86_compare_op1))
14723 if (ix86_compare_op1 == const0_rtx
14724 && (code == LT || code == GE))
14725 compare_code = code;
14726 else if (ix86_compare_op1 == constm1_rtx)
14730 else if (code == GT)
14735 /* Optimize dest = (op0 < 0) ? -1 : cf. */
14736 if (compare_code != UNKNOWN
14737 && GET_MODE (ix86_compare_op0) == GET_MODE (out)
14738 && (cf == -1 || ct == -1))
14740 /* If lea code below could be used, only optimize
14741 if it results in a 2 insn sequence. */
14743 if (! (diff == 1 || diff == 2 || diff == 4 || diff == 8
14744 || diff == 3 || diff == 5 || diff == 9)
14745 || (compare_code == LT && ct == -1)
14746 || (compare_code == GE && cf == -1))
14749 * notl op1 (if necessary)
14757 code = reverse_condition (code);
14760 out = emit_store_flag (out, code, ix86_compare_op0,
14761 ix86_compare_op1, VOIDmode, 0, -1);
14763 out = expand_simple_binop (mode, IOR,
14765 out, 1, OPTAB_DIRECT);
14766 if (out != operands[0])
14767 emit_move_insn (operands[0], out);
14769 return 1; /* DONE */
14774 if ((diff == 1 || diff == 2 || diff == 4 || diff == 8
14775 || diff == 3 || diff == 5 || diff == 9)
14776 && ((mode != QImode && mode != HImode) || !TARGET_PARTIAL_REG_STALL)
14778 || x86_64_immediate_operand (GEN_INT (cf), VOIDmode)))
14784 * lea cf(dest*(ct-cf)),dest
14788 * This also catches the degenerate setcc-only case.
14794 out = emit_store_flag (out, code, ix86_compare_op0,
14795 ix86_compare_op1, VOIDmode, 0, 1);
14798 /* On x86_64 the lea instruction operates on Pmode, so we need
14799 to get arithmetics done in proper mode to match. */
14801 tmp = copy_rtx (out);
14805 out1 = copy_rtx (out);
14806 tmp = gen_rtx_MULT (mode, out1, GEN_INT (diff & ~1));
14810 tmp = gen_rtx_PLUS (mode, tmp, out1);
14816 tmp = gen_rtx_PLUS (mode, tmp, GEN_INT (cf));
14819 if (!rtx_equal_p (tmp, out))
14822 out = force_operand (tmp, copy_rtx (out));
14824 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (out), copy_rtx (tmp)));
14826 if (!rtx_equal_p (out, operands[0]))
14827 emit_move_insn (operands[0], copy_rtx (out));
14829 return 1; /* DONE */
14833 * General case: Jumpful:
14834 * xorl dest,dest cmpl op1, op2
14835 * cmpl op1, op2 movl ct, dest
14836 * setcc dest jcc 1f
14837 * decl dest movl cf, dest
14838 * andl (cf-ct),dest 1:
14841 * Size 20. Size 14.
14843 * This is reasonably steep, but branch mispredict costs are
14844 * high on modern cpus, so consider failing only if optimizing
14848 if ((!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL))
14849 && BRANCH_COST (optimize_insn_for_speed_p (),
14854 enum machine_mode cmp_mode = GET_MODE (ix86_compare_op0);
14859 if (SCALAR_FLOAT_MODE_P (cmp_mode))
14861 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode));
14863 /* We may be reversing unordered compare to normal compare,
14864 that is not valid in general (we may convert non-trapping
14865 condition to trapping one), however on i386 we currently
14866 emit all comparisons unordered. */
14867 code = reverse_condition_maybe_unordered (code);
14871 code = reverse_condition (code);
14872 if (compare_code != UNKNOWN)
14873 compare_code = reverse_condition (compare_code);
14877 if (compare_code != UNKNOWN)
14879 /* notl op1 (if needed)
14884 For x < 0 (resp. x <= -1) there will be no notl,
14885 so if possible swap the constants to get rid of the
14887 True/false will be -1/0 while code below (store flag
14888 followed by decrement) is 0/-1, so the constants need
14889 to be exchanged once more. */
14891 if (compare_code == GE || !cf)
14893 code = reverse_condition (code);
14898 HOST_WIDE_INT tmp = cf;
14903 out = emit_store_flag (out, code, ix86_compare_op0,
14904 ix86_compare_op1, VOIDmode, 0, -1);
14908 out = emit_store_flag (out, code, ix86_compare_op0,
14909 ix86_compare_op1, VOIDmode, 0, 1);
14911 out = expand_simple_binop (mode, PLUS, copy_rtx (out), constm1_rtx,
14912 copy_rtx (out), 1, OPTAB_DIRECT);
14915 out = expand_simple_binop (mode, AND, copy_rtx (out),
14916 gen_int_mode (cf - ct, mode),
14917 copy_rtx (out), 1, OPTAB_DIRECT);
14919 out = expand_simple_binop (mode, PLUS, copy_rtx (out), GEN_INT (ct),
14920 copy_rtx (out), 1, OPTAB_DIRECT);
14921 if (!rtx_equal_p (out, operands[0]))
14922 emit_move_insn (operands[0], copy_rtx (out));
14924 return 1; /* DONE */
14928 if (!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL))
14930 /* Try a few things more with specific constants and a variable. */
14933 rtx var, orig_out, out, tmp;
14935 if (BRANCH_COST (optimize_insn_for_speed_p (), false) <= 2)
14936 return 0; /* FAIL */
14938 /* If one of the two operands is an interesting constant, load a
14939 constant with the above and mask it in with a logical operation. */
14941 if (CONST_INT_P (operands[2]))
14944 if (INTVAL (operands[2]) == 0 && operands[3] != constm1_rtx)
14945 operands[3] = constm1_rtx, op = and_optab;
14946 else if (INTVAL (operands[2]) == -1 && operands[3] != const0_rtx)
14947 operands[3] = const0_rtx, op = ior_optab;
14949 return 0; /* FAIL */
14951 else if (CONST_INT_P (operands[3]))
14954 if (INTVAL (operands[3]) == 0 && operands[2] != constm1_rtx)
14955 operands[2] = constm1_rtx, op = and_optab;
14956 else if (INTVAL (operands[3]) == -1 && operands[3] != const0_rtx)
14957 operands[2] = const0_rtx, op = ior_optab;
14959 return 0; /* FAIL */
14962 return 0; /* FAIL */
14964 orig_out = operands[0];
14965 tmp = gen_reg_rtx (mode);
14968 /* Recurse to get the constant loaded. */
14969 if (ix86_expand_int_movcc (operands) == 0)
14970 return 0; /* FAIL */
14972 /* Mask in the interesting variable. */
14973 out = expand_binop (mode, op, var, tmp, orig_out, 0,
14975 if (!rtx_equal_p (out, orig_out))
14976 emit_move_insn (copy_rtx (orig_out), copy_rtx (out));
14978 return 1; /* DONE */
14982 * For comparison with above,
14992 if (! nonimmediate_operand (operands[2], mode))
14993 operands[2] = force_reg (mode, operands[2]);
14994 if (! nonimmediate_operand (operands[3], mode))
14995 operands[3] = force_reg (mode, operands[3]);
14997 if (bypass_test && reg_overlap_mentioned_p (operands[0], operands[3]))
14999 rtx tmp = gen_reg_rtx (mode);
15000 emit_move_insn (tmp, operands[3]);
15003 if (second_test && reg_overlap_mentioned_p (operands[0], operands[2]))
15005 rtx tmp = gen_reg_rtx (mode);
15006 emit_move_insn (tmp, operands[2]);
15010 if (! register_operand (operands[2], VOIDmode)
15012 || ! register_operand (operands[3], VOIDmode)))
15013 operands[2] = force_reg (mode, operands[2]);
15016 && ! register_operand (operands[3], VOIDmode))
15017 operands[3] = force_reg (mode, operands[3]);
15019 emit_insn (compare_seq);
15020 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
15021 gen_rtx_IF_THEN_ELSE (mode,
15022 compare_op, operands[2],
15025 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (operands[0]),
15026 gen_rtx_IF_THEN_ELSE (mode,
15028 copy_rtx (operands[3]),
15029 copy_rtx (operands[0]))));
15031 emit_insn (gen_rtx_SET (VOIDmode, copy_rtx (operands[0]),
15032 gen_rtx_IF_THEN_ELSE (mode,
15034 copy_rtx (operands[2]),
15035 copy_rtx (operands[0]))));
15037 return 1; /* DONE */
15040 /* Swap, force into registers, or otherwise massage the two operands
15041 to an sse comparison with a mask result. Thus we differ a bit from
15042 ix86_prepare_fp_compare_args which expects to produce a flags result.
15044 The DEST operand exists to help determine whether to commute commutative
15045 operators. The POP0/POP1 operands are updated in place. The new
15046 comparison code is returned, or UNKNOWN if not implementable. */
15048 static enum rtx_code
15049 ix86_prepare_sse_fp_compare_args (rtx dest, enum rtx_code code,
15050 rtx *pop0, rtx *pop1)
15058 /* We have no LTGT as an operator. We could implement it with
15059 NE & ORDERED, but this requires an extra temporary. It's
15060 not clear that it's worth it. */
15067 /* These are supported directly. */
15074 /* For commutative operators, try to canonicalize the destination
15075 operand to be first in the comparison - this helps reload to
15076 avoid extra moves. */
15077 if (!dest || !rtx_equal_p (dest, *pop1))
15085 /* These are not supported directly. Swap the comparison operands
15086 to transform into something that is supported. */
15090 code = swap_condition (code);
15094 gcc_unreachable ();
15100 /* Detect conditional moves that exactly match min/max operational
15101 semantics. Note that this is IEEE safe, as long as we don't
15102 interchange the operands.
15104 Returns FALSE if this conditional move doesn't match a MIN/MAX,
15105 and TRUE if the operation is successful and instructions are emitted. */
15108 ix86_expand_sse_fp_minmax (rtx dest, enum rtx_code code, rtx cmp_op0,
15109 rtx cmp_op1, rtx if_true, rtx if_false)
15111 enum machine_mode mode;
15117 else if (code == UNGE)
15120 if_true = if_false;
15126 if (rtx_equal_p (cmp_op0, if_true) && rtx_equal_p (cmp_op1, if_false))
15128 else if (rtx_equal_p (cmp_op1, if_true) && rtx_equal_p (cmp_op0, if_false))
15133 mode = GET_MODE (dest);
15135 /* We want to check HONOR_NANS and HONOR_SIGNED_ZEROS here,
15136 but MODE may be a vector mode and thus not appropriate. */
15137 if (!flag_finite_math_only || !flag_unsafe_math_optimizations)
15139 int u = is_min ? UNSPEC_IEEE_MIN : UNSPEC_IEEE_MAX;
15142 if_true = force_reg (mode, if_true);
15143 v = gen_rtvec (2, if_true, if_false);
15144 tmp = gen_rtx_UNSPEC (mode, v, u);
15148 code = is_min ? SMIN : SMAX;
15149 tmp = gen_rtx_fmt_ee (code, mode, if_true, if_false);
15152 emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
15156 /* Expand an sse vector comparison. Return the register with the result. */
15159 ix86_expand_sse_cmp (rtx dest, enum rtx_code code, rtx cmp_op0, rtx cmp_op1,
15160 rtx op_true, rtx op_false)
15162 enum machine_mode mode = GET_MODE (dest);
15165 cmp_op0 = force_reg (mode, cmp_op0);
15166 if (!nonimmediate_operand (cmp_op1, mode))
15167 cmp_op1 = force_reg (mode, cmp_op1);
15170 || reg_overlap_mentioned_p (dest, op_true)
15171 || reg_overlap_mentioned_p (dest, op_false))
15172 dest = gen_reg_rtx (mode);
15174 x = gen_rtx_fmt_ee (code, mode, cmp_op0, cmp_op1);
15175 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
15180 /* Expand DEST = CMP ? OP_TRUE : OP_FALSE into a sequence of logical
15181 operations. This is used for both scalar and vector conditional moves. */
15184 ix86_expand_sse_movcc (rtx dest, rtx cmp, rtx op_true, rtx op_false)
15186 enum machine_mode mode = GET_MODE (dest);
15189 if (op_false == CONST0_RTX (mode))
15191 op_true = force_reg (mode, op_true);
15192 x = gen_rtx_AND (mode, cmp, op_true);
15193 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
15195 else if (op_true == CONST0_RTX (mode))
15197 op_false = force_reg (mode, op_false);
15198 x = gen_rtx_NOT (mode, cmp);
15199 x = gen_rtx_AND (mode, x, op_false);
15200 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
15202 else if (TARGET_SSE5)
15204 rtx pcmov = gen_rtx_SET (mode, dest,
15205 gen_rtx_IF_THEN_ELSE (mode, cmp,
15212 op_true = force_reg (mode, op_true);
15213 op_false = force_reg (mode, op_false);
15215 t2 = gen_reg_rtx (mode);
15217 t3 = gen_reg_rtx (mode);
15221 x = gen_rtx_AND (mode, op_true, cmp);
15222 emit_insn (gen_rtx_SET (VOIDmode, t2, x));
15224 x = gen_rtx_NOT (mode, cmp);
15225 x = gen_rtx_AND (mode, x, op_false);
15226 emit_insn (gen_rtx_SET (VOIDmode, t3, x));
15228 x = gen_rtx_IOR (mode, t3, t2);
15229 emit_insn (gen_rtx_SET (VOIDmode, dest, x));
15233 /* Expand a floating-point conditional move. Return true if successful. */
15236 ix86_expand_fp_movcc (rtx operands[])
15238 enum machine_mode mode = GET_MODE (operands[0]);
15239 enum rtx_code code = GET_CODE (operands[1]);
15240 rtx tmp, compare_op, second_test, bypass_test;
15242 if (TARGET_SSE_MATH && SSE_FLOAT_MODE_P (mode))
15244 enum machine_mode cmode;
15246 /* Since we've no cmove for sse registers, don't force bad register
15247 allocation just to gain access to it. Deny movcc when the
15248 comparison mode doesn't match the move mode. */
15249 cmode = GET_MODE (ix86_compare_op0);
15250 if (cmode == VOIDmode)
15251 cmode = GET_MODE (ix86_compare_op1);
15255 code = ix86_prepare_sse_fp_compare_args (operands[0], code,
15257 &ix86_compare_op1);
15258 if (code == UNKNOWN)
15261 if (ix86_expand_sse_fp_minmax (operands[0], code, ix86_compare_op0,
15262 ix86_compare_op1, operands[2],
15266 tmp = ix86_expand_sse_cmp (operands[0], code, ix86_compare_op0,
15267 ix86_compare_op1, operands[2], operands[3]);
15268 ix86_expand_sse_movcc (operands[0], tmp, operands[2], operands[3]);
15272 /* The floating point conditional move instructions don't directly
15273 support conditions resulting from a signed integer comparison. */
15275 compare_op = ix86_expand_compare (code, &second_test, &bypass_test);
15277 /* The floating point conditional move instructions don't directly
15278 support signed integer comparisons. */
15280 if (!fcmov_comparison_operator (compare_op, VOIDmode))
15282 gcc_assert (!second_test && !bypass_test);
15283 tmp = gen_reg_rtx (QImode);
15284 ix86_expand_setcc (code, tmp);
15286 ix86_compare_op0 = tmp;
15287 ix86_compare_op1 = const0_rtx;
15288 compare_op = ix86_expand_compare (code, &second_test, &bypass_test);
15290 if (bypass_test && reg_overlap_mentioned_p (operands[0], operands[3]))
15292 tmp = gen_reg_rtx (mode);
15293 emit_move_insn (tmp, operands[3]);
15296 if (second_test && reg_overlap_mentioned_p (operands[0], operands[2]))
15298 tmp = gen_reg_rtx (mode);
15299 emit_move_insn (tmp, operands[2]);
15303 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
15304 gen_rtx_IF_THEN_ELSE (mode, compare_op,
15305 operands[2], operands[3])));
15307 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
15308 gen_rtx_IF_THEN_ELSE (mode, bypass_test,
15309 operands[3], operands[0])));
15311 emit_insn (gen_rtx_SET (VOIDmode, operands[0],
15312 gen_rtx_IF_THEN_ELSE (mode, second_test,
15313 operands[2], operands[0])));
15318 /* Expand a floating-point vector conditional move; a vcond operation
15319 rather than a movcc operation. */
15322 ix86_expand_fp_vcond (rtx operands[])
15324 enum rtx_code code = GET_CODE (operands[3]);
15327 code = ix86_prepare_sse_fp_compare_args (operands[0], code,
15328 &operands[4], &operands[5]);
15329 if (code == UNKNOWN)
15332 if (ix86_expand_sse_fp_minmax (operands[0], code, operands[4],
15333 operands[5], operands[1], operands[2]))
15336 cmp = ix86_expand_sse_cmp (operands[0], code, operands[4], operands[5],
15337 operands[1], operands[2]);
15338 ix86_expand_sse_movcc (operands[0], cmp, operands[1], operands[2]);
15342 /* Expand a signed/unsigned integral vector conditional move. */
15345 ix86_expand_int_vcond (rtx operands[])
15347 enum machine_mode mode = GET_MODE (operands[0]);
15348 enum rtx_code code = GET_CODE (operands[3]);
15349 bool negate = false;
15352 cop0 = operands[4];
15353 cop1 = operands[5];
15355 /* SSE5 supports all of the comparisons on all vector int types. */
15358 /* Canonicalize the comparison to EQ, GT, GTU. */
15369 code = reverse_condition (code);
15375 code = reverse_condition (code);
15381 code = swap_condition (code);
15382 x = cop0, cop0 = cop1, cop1 = x;
15386 gcc_unreachable ();
15389 /* Only SSE4.1/SSE4.2 supports V2DImode. */
15390 if (mode == V2DImode)
15395 /* SSE4.1 supports EQ. */
15396 if (!TARGET_SSE4_1)
15402 /* SSE4.2 supports GT/GTU. */
15403 if (!TARGET_SSE4_2)
15408 gcc_unreachable ();
15412 /* Unsigned parallel compare is not supported by the hardware. Play some
15413 tricks to turn this into a signed comparison against 0. */
15416 cop0 = force_reg (mode, cop0);
15425 /* Perform a parallel modulo subtraction. */
15426 t1 = gen_reg_rtx (mode);
15427 emit_insn ((mode == V4SImode
15429 : gen_subv2di3) (t1, cop0, cop1));
15431 /* Extract the original sign bit of op0. */
15432 mask = ix86_build_signbit_mask (GET_MODE_INNER (mode),
15434 t2 = gen_reg_rtx (mode);
15435 emit_insn ((mode == V4SImode
15437 : gen_andv2di3) (t2, cop0, mask));
15439 /* XOR it back into the result of the subtraction. This results
15440 in the sign bit set iff we saw unsigned underflow. */
15441 x = gen_reg_rtx (mode);
15442 emit_insn ((mode == V4SImode
15444 : gen_xorv2di3) (x, t1, t2));
15452 /* Perform a parallel unsigned saturating subtraction. */
15453 x = gen_reg_rtx (mode);
15454 emit_insn (gen_rtx_SET (VOIDmode, x,
15455 gen_rtx_US_MINUS (mode, cop0, cop1)));
15462 gcc_unreachable ();
15466 cop1 = CONST0_RTX (mode);
15470 x = ix86_expand_sse_cmp (operands[0], code, cop0, cop1,
15471 operands[1+negate], operands[2-negate]);
15473 ix86_expand_sse_movcc (operands[0], x, operands[1+negate],
15474 operands[2-negate]);
15478 /* Unpack OP[1] into the next wider integer vector type. UNSIGNED_P is
15479 true if we should do zero extension, else sign extension. HIGH_P is
15480 true if we want the N/2 high elements, else the low elements. */
15483 ix86_expand_sse_unpack (rtx operands[2], bool unsigned_p, bool high_p)
15485 enum machine_mode imode = GET_MODE (operands[1]);
15486 rtx (*unpack)(rtx, rtx, rtx);
15493 unpack = gen_vec_interleave_highv16qi;
15495 unpack = gen_vec_interleave_lowv16qi;
15499 unpack = gen_vec_interleave_highv8hi;
15501 unpack = gen_vec_interleave_lowv8hi;
15505 unpack = gen_vec_interleave_highv4si;
15507 unpack = gen_vec_interleave_lowv4si;
15510 gcc_unreachable ();
15513 dest = gen_lowpart (imode, operands[0]);
15516 se = force_reg (imode, CONST0_RTX (imode));
15518 se = ix86_expand_sse_cmp (gen_reg_rtx (imode), GT, CONST0_RTX (imode),
15519 operands[1], pc_rtx, pc_rtx);
15521 emit_insn (unpack (dest, operands[1], se));
15524 /* This function performs the same task as ix86_expand_sse_unpack,
15525 but with SSE4.1 instructions. */
15528 ix86_expand_sse4_unpack (rtx operands[2], bool unsigned_p, bool high_p)
15530 enum machine_mode imode = GET_MODE (operands[1]);
15531 rtx (*unpack)(rtx, rtx);
15538 unpack = gen_sse4_1_zero_extendv8qiv8hi2;
15540 unpack = gen_sse4_1_extendv8qiv8hi2;
15544 unpack = gen_sse4_1_zero_extendv4hiv4si2;
15546 unpack = gen_sse4_1_extendv4hiv4si2;
15550 unpack = gen_sse4_1_zero_extendv2siv2di2;
15552 unpack = gen_sse4_1_extendv2siv2di2;
15555 gcc_unreachable ();
15558 dest = operands[0];
15561 /* Shift higher 8 bytes to lower 8 bytes. */
15562 src = gen_reg_rtx (imode);
15563 emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode, src),
15564 gen_lowpart (TImode, operands[1]),
15570 emit_insn (unpack (dest, src));
15573 /* This function performs the same task as ix86_expand_sse_unpack,
15574 but with sse5 instructions. */
15577 ix86_expand_sse5_unpack (rtx operands[2], bool unsigned_p, bool high_p)
15579 enum machine_mode imode = GET_MODE (operands[1]);
15580 int pperm_bytes[16];
15582 int h = (high_p) ? 8 : 0;
15585 rtvec v = rtvec_alloc (16);
15588 rtx op0 = operands[0], op1 = operands[1];
15593 vs = rtvec_alloc (8);
15594 h2 = (high_p) ? 8 : 0;
15595 for (i = 0; i < 8; i++)
15597 pperm_bytes[2*i+0] = PPERM_SRC | PPERM_SRC2 | i | h;
15598 pperm_bytes[2*i+1] = ((unsigned_p)
15600 : PPERM_SIGN | PPERM_SRC2 | i | h);
15603 for (i = 0; i < 16; i++)
15604 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15606 for (i = 0; i < 8; i++)
15607 RTVEC_ELT (vs, i) = GEN_INT (i + h2);
15609 p = gen_rtx_PARALLEL (VOIDmode, vs);
15610 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15612 emit_insn (gen_sse5_pperm_zero_v16qi_v8hi (op0, op1, p, x));
15614 emit_insn (gen_sse5_pperm_sign_v16qi_v8hi (op0, op1, p, x));
15618 vs = rtvec_alloc (4);
15619 h2 = (high_p) ? 4 : 0;
15620 for (i = 0; i < 4; i++)
15622 sign_extend = ((unsigned_p)
15624 : PPERM_SIGN | PPERM_SRC2 | ((2*i) + 1 + h));
15625 pperm_bytes[4*i+0] = PPERM_SRC | PPERM_SRC2 | ((2*i) + 0 + h);
15626 pperm_bytes[4*i+1] = PPERM_SRC | PPERM_SRC2 | ((2*i) + 1 + h);
15627 pperm_bytes[4*i+2] = sign_extend;
15628 pperm_bytes[4*i+3] = sign_extend;
15631 for (i = 0; i < 16; i++)
15632 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15634 for (i = 0; i < 4; i++)
15635 RTVEC_ELT (vs, i) = GEN_INT (i + h2);
15637 p = gen_rtx_PARALLEL (VOIDmode, vs);
15638 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15640 emit_insn (gen_sse5_pperm_zero_v8hi_v4si (op0, op1, p, x));
15642 emit_insn (gen_sse5_pperm_sign_v8hi_v4si (op0, op1, p, x));
15646 vs = rtvec_alloc (2);
15647 h2 = (high_p) ? 2 : 0;
15648 for (i = 0; i < 2; i++)
15650 sign_extend = ((unsigned_p)
15652 : PPERM_SIGN | PPERM_SRC2 | ((4*i) + 3 + h));
15653 pperm_bytes[8*i+0] = PPERM_SRC | PPERM_SRC2 | ((4*i) + 0 + h);
15654 pperm_bytes[8*i+1] = PPERM_SRC | PPERM_SRC2 | ((4*i) + 1 + h);
15655 pperm_bytes[8*i+2] = PPERM_SRC | PPERM_SRC2 | ((4*i) + 2 + h);
15656 pperm_bytes[8*i+3] = PPERM_SRC | PPERM_SRC2 | ((4*i) + 3 + h);
15657 pperm_bytes[8*i+4] = sign_extend;
15658 pperm_bytes[8*i+5] = sign_extend;
15659 pperm_bytes[8*i+6] = sign_extend;
15660 pperm_bytes[8*i+7] = sign_extend;
15663 for (i = 0; i < 16; i++)
15664 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15666 for (i = 0; i < 2; i++)
15667 RTVEC_ELT (vs, i) = GEN_INT (i + h2);
15669 p = gen_rtx_PARALLEL (VOIDmode, vs);
15670 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15672 emit_insn (gen_sse5_pperm_zero_v4si_v2di (op0, op1, p, x));
15674 emit_insn (gen_sse5_pperm_sign_v4si_v2di (op0, op1, p, x));
15678 gcc_unreachable ();
15684 /* Pack the high bits from OPERANDS[1] and low bits from OPERANDS[2] into the
15685 next narrower integer vector type */
15687 ix86_expand_sse5_pack (rtx operands[3])
15689 enum machine_mode imode = GET_MODE (operands[0]);
15690 int pperm_bytes[16];
15692 rtvec v = rtvec_alloc (16);
15694 rtx op0 = operands[0];
15695 rtx op1 = operands[1];
15696 rtx op2 = operands[2];
15701 for (i = 0; i < 8; i++)
15703 pperm_bytes[i+0] = PPERM_SRC | PPERM_SRC1 | (i*2);
15704 pperm_bytes[i+8] = PPERM_SRC | PPERM_SRC2 | (i*2);
15707 for (i = 0; i < 16; i++)
15708 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15710 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15711 emit_insn (gen_sse5_pperm_pack_v8hi_v16qi (op0, op1, op2, x));
15715 for (i = 0; i < 4; i++)
15717 pperm_bytes[(2*i)+0] = PPERM_SRC | PPERM_SRC1 | ((i*4) + 0);
15718 pperm_bytes[(2*i)+1] = PPERM_SRC | PPERM_SRC1 | ((i*4) + 1);
15719 pperm_bytes[(2*i)+8] = PPERM_SRC | PPERM_SRC2 | ((i*4) + 0);
15720 pperm_bytes[(2*i)+9] = PPERM_SRC | PPERM_SRC2 | ((i*4) + 1);
15723 for (i = 0; i < 16; i++)
15724 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15726 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15727 emit_insn (gen_sse5_pperm_pack_v4si_v8hi (op0, op1, op2, x));
15731 for (i = 0; i < 2; i++)
15733 pperm_bytes[(4*i)+0] = PPERM_SRC | PPERM_SRC1 | ((i*8) + 0);
15734 pperm_bytes[(4*i)+1] = PPERM_SRC | PPERM_SRC1 | ((i*8) + 1);
15735 pperm_bytes[(4*i)+2] = PPERM_SRC | PPERM_SRC1 | ((i*8) + 2);
15736 pperm_bytes[(4*i)+3] = PPERM_SRC | PPERM_SRC1 | ((i*8) + 3);
15737 pperm_bytes[(4*i)+8] = PPERM_SRC | PPERM_SRC2 | ((i*8) + 0);
15738 pperm_bytes[(4*i)+9] = PPERM_SRC | PPERM_SRC2 | ((i*8) + 1);
15739 pperm_bytes[(4*i)+10] = PPERM_SRC | PPERM_SRC2 | ((i*8) + 2);
15740 pperm_bytes[(4*i)+11] = PPERM_SRC | PPERM_SRC2 | ((i*8) + 3);
15743 for (i = 0; i < 16; i++)
15744 RTVEC_ELT (v, i) = GEN_INT (pperm_bytes[i]);
15746 x = force_reg (V16QImode, gen_rtx_CONST_VECTOR (V16QImode, v));
15747 emit_insn (gen_sse5_pperm_pack_v2di_v4si (op0, op1, op2, x));
15751 gcc_unreachable ();
15757 /* Expand conditional increment or decrement using adb/sbb instructions.
15758 The default case using setcc followed by the conditional move can be
15759 done by generic code. */
15761 ix86_expand_int_addcc (rtx operands[])
15763 enum rtx_code code = GET_CODE (operands[1]);
15765 rtx val = const0_rtx;
15766 bool fpcmp = false;
15767 enum machine_mode mode = GET_MODE (operands[0]);
15769 if (operands[3] != const1_rtx
15770 && operands[3] != constm1_rtx)
15772 if (!ix86_expand_carry_flag_compare (code, ix86_compare_op0,
15773 ix86_compare_op1, &compare_op))
15775 code = GET_CODE (compare_op);
15777 if (GET_MODE (XEXP (compare_op, 0)) == CCFPmode
15778 || GET_MODE (XEXP (compare_op, 0)) == CCFPUmode)
15781 code = ix86_fp_compare_code_to_integer (code);
15788 PUT_CODE (compare_op,
15789 reverse_condition_maybe_unordered
15790 (GET_CODE (compare_op)));
15792 PUT_CODE (compare_op, reverse_condition (GET_CODE (compare_op)));
15794 PUT_MODE (compare_op, mode);
15796 /* Construct either adc or sbb insn. */
15797 if ((code == LTU) == (operands[3] == constm1_rtx))
15799 switch (GET_MODE (operands[0]))
15802 emit_insn (gen_subqi3_carry (operands[0], operands[2], val, compare_op));
15805 emit_insn (gen_subhi3_carry (operands[0], operands[2], val, compare_op));
15808 emit_insn (gen_subsi3_carry (operands[0], operands[2], val, compare_op));
15811 emit_insn (gen_subdi3_carry_rex64 (operands[0], operands[2], val, compare_op));
15814 gcc_unreachable ();
15819 switch (GET_MODE (operands[0]))
15822 emit_insn (gen_addqi3_carry (operands[0], operands[2], val, compare_op));
15825 emit_insn (gen_addhi3_carry (operands[0], operands[2], val, compare_op));
15828 emit_insn (gen_addsi3_carry (operands[0], operands[2], val, compare_op));
15831 emit_insn (gen_adddi3_carry_rex64 (operands[0], operands[2], val, compare_op));
15834 gcc_unreachable ();
15837 return 1; /* DONE */
15841 /* Split operands 0 and 1 into SImode parts. Similar to split_di, but
15842 works for floating pointer parameters and nonoffsetable memories.
15843 For pushes, it returns just stack offsets; the values will be saved
15844 in the right order. Maximally three parts are generated. */
15847 ix86_split_to_parts (rtx operand, rtx *parts, enum machine_mode mode)
15852 size = mode==XFmode ? 3 : GET_MODE_SIZE (mode) / 4;
15854 size = (GET_MODE_SIZE (mode) + 4) / 8;
15856 gcc_assert (!REG_P (operand) || !MMX_REGNO_P (REGNO (operand)));
15857 gcc_assert (size >= 2 && size <= 4);
15859 /* Optimize constant pool reference to immediates. This is used by fp
15860 moves, that force all constants to memory to allow combining. */
15861 if (MEM_P (operand) && MEM_READONLY_P (operand))
15863 rtx tmp = maybe_get_pool_constant (operand);
15868 if (MEM_P (operand) && !offsettable_memref_p (operand))
15870 /* The only non-offsetable memories we handle are pushes. */
15871 int ok = push_operand (operand, VOIDmode);
15875 operand = copy_rtx (operand);
15876 PUT_MODE (operand, Pmode);
15877 parts[0] = parts[1] = parts[2] = parts[3] = operand;
15881 if (GET_CODE (operand) == CONST_VECTOR)
15883 enum machine_mode imode = int_mode_for_mode (mode);
15884 /* Caution: if we looked through a constant pool memory above,
15885 the operand may actually have a different mode now. That's
15886 ok, since we want to pun this all the way back to an integer. */
15887 operand = simplify_subreg (imode, operand, GET_MODE (operand), 0);
15888 gcc_assert (operand != NULL);
15894 if (mode == DImode)
15895 split_di (&operand, 1, &parts[0], &parts[1]);
15900 if (REG_P (operand))
15902 gcc_assert (reload_completed);
15903 for (i = 0; i < size; i++)
15904 parts[i] = gen_rtx_REG (SImode, REGNO (operand) + i);
15906 else if (offsettable_memref_p (operand))
15908 operand = adjust_address (operand, SImode, 0);
15909 parts[0] = operand;
15910 for (i = 1; i < size; i++)
15911 parts[i] = adjust_address (operand, SImode, 4 * i);
15913 else if (GET_CODE (operand) == CONST_DOUBLE)
15918 REAL_VALUE_FROM_CONST_DOUBLE (r, operand);
15922 real_to_target (l, &r, mode);
15923 parts[3] = gen_int_mode (l[3], SImode);
15924 parts[2] = gen_int_mode (l[2], SImode);
15927 REAL_VALUE_TO_TARGET_LONG_DOUBLE (r, l);
15928 parts[2] = gen_int_mode (l[2], SImode);
15931 REAL_VALUE_TO_TARGET_DOUBLE (r, l);
15934 gcc_unreachable ();
15936 parts[1] = gen_int_mode (l[1], SImode);
15937 parts[0] = gen_int_mode (l[0], SImode);
15940 gcc_unreachable ();
15945 if (mode == TImode)
15946 split_ti (&operand, 1, &parts[0], &parts[1]);
15947 if (mode == XFmode || mode == TFmode)
15949 enum machine_mode upper_mode = mode==XFmode ? SImode : DImode;
15950 if (REG_P (operand))
15952 gcc_assert (reload_completed);
15953 parts[0] = gen_rtx_REG (DImode, REGNO (operand) + 0);
15954 parts[1] = gen_rtx_REG (upper_mode, REGNO (operand) + 1);
15956 else if (offsettable_memref_p (operand))
15958 operand = adjust_address (operand, DImode, 0);
15959 parts[0] = operand;
15960 parts[1] = adjust_address (operand, upper_mode, 8);
15962 else if (GET_CODE (operand) == CONST_DOUBLE)
15967 REAL_VALUE_FROM_CONST_DOUBLE (r, operand);
15968 real_to_target (l, &r, mode);
15970 /* Do not use shift by 32 to avoid warning on 32bit systems. */
15971 if (HOST_BITS_PER_WIDE_INT >= 64)
15974 ((l[0] & (((HOST_WIDE_INT) 2 << 31) - 1))
15975 + ((((HOST_WIDE_INT) l[1]) << 31) << 1),
15978 parts[0] = immed_double_const (l[0], l[1], DImode);
15980 if (upper_mode == SImode)
15981 parts[1] = gen_int_mode (l[2], SImode);
15982 else if (HOST_BITS_PER_WIDE_INT >= 64)
15985 ((l[2] & (((HOST_WIDE_INT) 2 << 31) - 1))
15986 + ((((HOST_WIDE_INT) l[3]) << 31) << 1),
15989 parts[1] = immed_double_const (l[2], l[3], DImode);
15992 gcc_unreachable ();
15999 /* Emit insns to perform a move or push of DI, DF, XF, and TF values.
16000 Return false when normal moves are needed; true when all required
16001 insns have been emitted. Operands 2-4 contain the input values
16002 int the correct order; operands 5-7 contain the output values. */
16005 ix86_split_long_move (rtx operands[])
16010 int collisions = 0;
16011 enum machine_mode mode = GET_MODE (operands[0]);
16012 bool collisionparts[4];
16014 /* The DFmode expanders may ask us to move double.
16015 For 64bit target this is single move. By hiding the fact
16016 here we simplify i386.md splitters. */
16017 if (GET_MODE_SIZE (GET_MODE (operands[0])) == 8 && TARGET_64BIT)
16019 /* Optimize constant pool reference to immediates. This is used by
16020 fp moves, that force all constants to memory to allow combining. */
16022 if (MEM_P (operands[1])
16023 && GET_CODE (XEXP (operands[1], 0)) == SYMBOL_REF
16024 && CONSTANT_POOL_ADDRESS_P (XEXP (operands[1], 0)))
16025 operands[1] = get_pool_constant (XEXP (operands[1], 0));
16026 if (push_operand (operands[0], VOIDmode))
16028 operands[0] = copy_rtx (operands[0]);
16029 PUT_MODE (operands[0], Pmode);
16032 operands[0] = gen_lowpart (DImode, operands[0]);
16033 operands[1] = gen_lowpart (DImode, operands[1]);
16034 emit_move_insn (operands[0], operands[1]);
16038 /* The only non-offsettable memory we handle is push. */
16039 if (push_operand (operands[0], VOIDmode))
16042 gcc_assert (!MEM_P (operands[0])
16043 || offsettable_memref_p (operands[0]));
16045 nparts = ix86_split_to_parts (operands[1], part[1], GET_MODE (operands[0]));
16046 ix86_split_to_parts (operands[0], part[0], GET_MODE (operands[0]));
16048 /* When emitting push, take care for source operands on the stack. */
16049 if (push && MEM_P (operands[1])
16050 && reg_overlap_mentioned_p (stack_pointer_rtx, operands[1]))
16051 for (i = 0; i < nparts - 1; i++)
16052 part[1][i] = change_address (part[1][i],
16053 GET_MODE (part[1][i]),
16054 XEXP (part[1][i + 1], 0));
16056 /* We need to do copy in the right order in case an address register
16057 of the source overlaps the destination. */
16058 if (REG_P (part[0][0]) && MEM_P (part[1][0]))
16062 for (i = 0; i < nparts; i++)
16065 = reg_overlap_mentioned_p (part[0][i], XEXP (part[1][0], 0));
16066 if (collisionparts[i])
16070 /* Collision in the middle part can be handled by reordering. */
16071 if (collisions == 1 && nparts == 3 && collisionparts [1])
16073 tmp = part[0][1]; part[0][1] = part[0][2]; part[0][2] = tmp;
16074 tmp = part[1][1]; part[1][1] = part[1][2]; part[1][2] = tmp;
16076 else if (collisions == 1
16078 && (collisionparts [1] || collisionparts [2]))
16080 if (collisionparts [1])
16082 tmp = part[0][1]; part[0][1] = part[0][2]; part[0][2] = tmp;
16083 tmp = part[1][1]; part[1][1] = part[1][2]; part[1][2] = tmp;
16087 tmp = part[0][2]; part[0][2] = part[0][3]; part[0][3] = tmp;
16088 tmp = part[1][2]; part[1][2] = part[1][3]; part[1][3] = tmp;
16092 /* If there are more collisions, we can't handle it by reordering.
16093 Do an lea to the last part and use only one colliding move. */
16094 else if (collisions > 1)
16100 base = part[0][nparts - 1];
16102 /* Handle the case when the last part isn't valid for lea.
16103 Happens in 64-bit mode storing the 12-byte XFmode. */
16104 if (GET_MODE (base) != Pmode)
16105 base = gen_rtx_REG (Pmode, REGNO (base));
16107 emit_insn (gen_rtx_SET (VOIDmode, base, XEXP (part[1][0], 0)));
16108 part[1][0] = replace_equiv_address (part[1][0], base);
16109 for (i = 1; i < nparts; i++)
16111 tmp = plus_constant (base, UNITS_PER_WORD * i);
16112 part[1][i] = replace_equiv_address (part[1][i], tmp);
16123 if (TARGET_128BIT_LONG_DOUBLE && mode == XFmode)
16124 emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, GEN_INT (-4)));
16125 emit_move_insn (part[0][2], part[1][2]);
16127 else if (nparts == 4)
16129 emit_move_insn (part[0][3], part[1][3]);
16130 emit_move_insn (part[0][2], part[1][2]);
16135 /* In 64bit mode we don't have 32bit push available. In case this is
16136 register, it is OK - we will just use larger counterpart. We also
16137 retype memory - these comes from attempt to avoid REX prefix on
16138 moving of second half of TFmode value. */
16139 if (GET_MODE (part[1][1]) == SImode)
16141 switch (GET_CODE (part[1][1]))
16144 part[1][1] = adjust_address (part[1][1], DImode, 0);
16148 part[1][1] = gen_rtx_REG (DImode, REGNO (part[1][1]));
16152 gcc_unreachable ();
16155 if (GET_MODE (part[1][0]) == SImode)
16156 part[1][0] = part[1][1];
16159 emit_move_insn (part[0][1], part[1][1]);
16160 emit_move_insn (part[0][0], part[1][0]);
16164 /* Choose correct order to not overwrite the source before it is copied. */
16165 if ((REG_P (part[0][0])
16166 && REG_P (part[1][1])
16167 && (REGNO (part[0][0]) == REGNO (part[1][1])
16169 && REGNO (part[0][0]) == REGNO (part[1][2]))
16171 && REGNO (part[0][0]) == REGNO (part[1][3]))))
16173 && reg_overlap_mentioned_p (part[0][0], XEXP (part[1][0], 0))))
16175 for (i = 0, j = nparts - 1; i < nparts; i++, j--)
16177 operands[2 + i] = part[0][j];
16178 operands[6 + i] = part[1][j];
16183 for (i = 0; i < nparts; i++)
16185 operands[2 + i] = part[0][i];
16186 operands[6 + i] = part[1][i];
16190 /* If optimizing for size, attempt to locally unCSE nonzero constants. */
16191 if (optimize_insn_for_size_p ())
16193 for (j = 0; j < nparts - 1; j++)
16194 if (CONST_INT_P (operands[6 + j])
16195 && operands[6 + j] != const0_rtx
16196 && REG_P (operands[2 + j]))
16197 for (i = j; i < nparts - 1; i++)
16198 if (CONST_INT_P (operands[7 + i])
16199 && INTVAL (operands[7 + i]) == INTVAL (operands[6 + j]))
16200 operands[7 + i] = operands[2 + j];
16203 for (i = 0; i < nparts; i++)
16204 emit_move_insn (operands[2 + i], operands[6 + i]);
16209 /* Helper function of ix86_split_ashl used to generate an SImode/DImode
16210 left shift by a constant, either using a single shift or
16211 a sequence of add instructions. */
16214 ix86_expand_ashl_const (rtx operand, int count, enum machine_mode mode)
16218 emit_insn ((mode == DImode
16220 : gen_adddi3) (operand, operand, operand));
16222 else if (!optimize_insn_for_size_p ()
16223 && count * ix86_cost->add <= ix86_cost->shift_const)
16226 for (i=0; i<count; i++)
16228 emit_insn ((mode == DImode
16230 : gen_adddi3) (operand, operand, operand));
16234 emit_insn ((mode == DImode
16236 : gen_ashldi3) (operand, operand, GEN_INT (count)));
16240 ix86_split_ashl (rtx *operands, rtx scratch, enum machine_mode mode)
16242 rtx low[2], high[2];
16244 const int single_width = mode == DImode ? 32 : 64;
16246 if (CONST_INT_P (operands[2]))
16248 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
16249 count = INTVAL (operands[2]) & (single_width * 2 - 1);
16251 if (count >= single_width)
16253 emit_move_insn (high[0], low[1]);
16254 emit_move_insn (low[0], const0_rtx);
16256 if (count > single_width)
16257 ix86_expand_ashl_const (high[0], count - single_width, mode);
16261 if (!rtx_equal_p (operands[0], operands[1]))
16262 emit_move_insn (operands[0], operands[1]);
16263 emit_insn ((mode == DImode
16265 : gen_x86_64_shld) (high[0], low[0], GEN_INT (count)));
16266 ix86_expand_ashl_const (low[0], count, mode);
16271 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
16273 if (operands[1] == const1_rtx)
16275 /* Assuming we've chosen a QImode capable registers, then 1 << N
16276 can be done with two 32/64-bit shifts, no branches, no cmoves. */
16277 if (ANY_QI_REG_P (low[0]) && ANY_QI_REG_P (high[0]))
16279 rtx s, d, flags = gen_rtx_REG (CCZmode, FLAGS_REG);
16281 ix86_expand_clear (low[0]);
16282 ix86_expand_clear (high[0]);
16283 emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (single_width)));
16285 d = gen_lowpart (QImode, low[0]);
16286 d = gen_rtx_STRICT_LOW_PART (VOIDmode, d);
16287 s = gen_rtx_EQ (QImode, flags, const0_rtx);
16288 emit_insn (gen_rtx_SET (VOIDmode, d, s));
16290 d = gen_lowpart (QImode, high[0]);
16291 d = gen_rtx_STRICT_LOW_PART (VOIDmode, d);
16292 s = gen_rtx_NE (QImode, flags, const0_rtx);
16293 emit_insn (gen_rtx_SET (VOIDmode, d, s));
16296 /* Otherwise, we can get the same results by manually performing
16297 a bit extract operation on bit 5/6, and then performing the two
16298 shifts. The two methods of getting 0/1 into low/high are exactly
16299 the same size. Avoiding the shift in the bit extract case helps
16300 pentium4 a bit; no one else seems to care much either way. */
16305 if (TARGET_PARTIAL_REG_STALL && !optimize_insn_for_size_p ())
16306 x = gen_rtx_ZERO_EXTEND (mode == DImode ? SImode : DImode, operands[2]);
16308 x = gen_lowpart (mode == DImode ? SImode : DImode, operands[2]);
16309 emit_insn (gen_rtx_SET (VOIDmode, high[0], x));
16311 emit_insn ((mode == DImode
16313 : gen_lshrdi3) (high[0], high[0], GEN_INT (mode == DImode ? 5 : 6)));
16314 emit_insn ((mode == DImode
16316 : gen_anddi3) (high[0], high[0], GEN_INT (1)));
16317 emit_move_insn (low[0], high[0]);
16318 emit_insn ((mode == DImode
16320 : gen_xordi3) (low[0], low[0], GEN_INT (1)));
16323 emit_insn ((mode == DImode
16325 : gen_ashldi3) (low[0], low[0], operands[2]));
16326 emit_insn ((mode == DImode
16328 : gen_ashldi3) (high[0], high[0], operands[2]));
16332 if (operands[1] == constm1_rtx)
16334 /* For -1 << N, we can avoid the shld instruction, because we
16335 know that we're shifting 0...31/63 ones into a -1. */
16336 emit_move_insn (low[0], constm1_rtx);
16337 if (optimize_insn_for_size_p ())
16338 emit_move_insn (high[0], low[0]);
16340 emit_move_insn (high[0], constm1_rtx);
16344 if (!rtx_equal_p (operands[0], operands[1]))
16345 emit_move_insn (operands[0], operands[1]);
16347 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
16348 emit_insn ((mode == DImode
16350 : gen_x86_64_shld) (high[0], low[0], operands[2]));
16353 emit_insn ((mode == DImode ? gen_ashlsi3 : gen_ashldi3) (low[0], low[0], operands[2]));
16355 if (TARGET_CMOVE && scratch)
16357 ix86_expand_clear (scratch);
16358 emit_insn ((mode == DImode
16359 ? gen_x86_shift_adj_1
16360 : gen_x86_64_shift_adj_1) (high[0], low[0], operands[2],
16364 emit_insn ((mode == DImode
16365 ? gen_x86_shift_adj_2
16366 : gen_x86_64_shift_adj_2) (high[0], low[0], operands[2]));
16370 ix86_split_ashr (rtx *operands, rtx scratch, enum machine_mode mode)
16372 rtx low[2], high[2];
16374 const int single_width = mode == DImode ? 32 : 64;
16376 if (CONST_INT_P (operands[2]))
16378 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
16379 count = INTVAL (operands[2]) & (single_width * 2 - 1);
16381 if (count == single_width * 2 - 1)
16383 emit_move_insn (high[0], high[1]);
16384 emit_insn ((mode == DImode
16386 : gen_ashrdi3) (high[0], high[0],
16387 GEN_INT (single_width - 1)));
16388 emit_move_insn (low[0], high[0]);
16391 else if (count >= single_width)
16393 emit_move_insn (low[0], high[1]);
16394 emit_move_insn (high[0], low[0]);
16395 emit_insn ((mode == DImode
16397 : gen_ashrdi3) (high[0], high[0],
16398 GEN_INT (single_width - 1)));
16399 if (count > single_width)
16400 emit_insn ((mode == DImode
16402 : gen_ashrdi3) (low[0], low[0],
16403 GEN_INT (count - single_width)));
16407 if (!rtx_equal_p (operands[0], operands[1]))
16408 emit_move_insn (operands[0], operands[1]);
16409 emit_insn ((mode == DImode
16411 : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
16412 emit_insn ((mode == DImode
16414 : gen_ashrdi3) (high[0], high[0], GEN_INT (count)));
16419 if (!rtx_equal_p (operands[0], operands[1]))
16420 emit_move_insn (operands[0], operands[1]);
16422 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
16424 emit_insn ((mode == DImode
16426 : gen_x86_64_shrd) (low[0], high[0], operands[2]));
16427 emit_insn ((mode == DImode
16429 : gen_ashrdi3) (high[0], high[0], operands[2]));
16431 if (TARGET_CMOVE && scratch)
16433 emit_move_insn (scratch, high[0]);
16434 emit_insn ((mode == DImode
16436 : gen_ashrdi3) (scratch, scratch,
16437 GEN_INT (single_width - 1)));
16438 emit_insn ((mode == DImode
16439 ? gen_x86_shift_adj_1
16440 : gen_x86_64_shift_adj_1) (low[0], high[0], operands[2],
16444 emit_insn ((mode == DImode
16445 ? gen_x86_shift_adj_3
16446 : gen_x86_64_shift_adj_3) (low[0], high[0], operands[2]));
16451 ix86_split_lshr (rtx *operands, rtx scratch, enum machine_mode mode)
16453 rtx low[2], high[2];
16455 const int single_width = mode == DImode ? 32 : 64;
16457 if (CONST_INT_P (operands[2]))
16459 (mode == DImode ? split_di : split_ti) (operands, 2, low, high);
16460 count = INTVAL (operands[2]) & (single_width * 2 - 1);
16462 if (count >= single_width)
16464 emit_move_insn (low[0], high[1]);
16465 ix86_expand_clear (high[0]);
16467 if (count > single_width)
16468 emit_insn ((mode == DImode
16470 : gen_lshrdi3) (low[0], low[0],
16471 GEN_INT (count - single_width)));
16475 if (!rtx_equal_p (operands[0], operands[1]))
16476 emit_move_insn (operands[0], operands[1]);
16477 emit_insn ((mode == DImode
16479 : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
16480 emit_insn ((mode == DImode
16482 : gen_lshrdi3) (high[0], high[0], GEN_INT (count)));
16487 if (!rtx_equal_p (operands[0], operands[1]))
16488 emit_move_insn (operands[0], operands[1]);
16490 (mode == DImode ? split_di : split_ti) (operands, 1, low, high);
16492 emit_insn ((mode == DImode
16494 : gen_x86_64_shrd) (low[0], high[0], operands[2]));
16495 emit_insn ((mode == DImode
16497 : gen_lshrdi3) (high[0], high[0], operands[2]));
16499 /* Heh. By reversing the arguments, we can reuse this pattern. */
16500 if (TARGET_CMOVE && scratch)
16502 ix86_expand_clear (scratch);
16503 emit_insn ((mode == DImode
16504 ? gen_x86_shift_adj_1
16505 : gen_x86_64_shift_adj_1) (low[0], high[0], operands[2],
16509 emit_insn ((mode == DImode
16510 ? gen_x86_shift_adj_2
16511 : gen_x86_64_shift_adj_2) (low[0], high[0], operands[2]));
16515 /* Predict just emitted jump instruction to be taken with probability PROB. */
16517 predict_jump (int prob)
16519 rtx insn = get_last_insn ();
16520 gcc_assert (JUMP_P (insn));
16522 = gen_rtx_EXPR_LIST (REG_BR_PROB,
16527 /* Helper function for the string operations below. Dest VARIABLE whether
16528 it is aligned to VALUE bytes. If true, jump to the label. */
16530 ix86_expand_aligntest (rtx variable, int value, bool epilogue)
16532 rtx label = gen_label_rtx ();
16533 rtx tmpcount = gen_reg_rtx (GET_MODE (variable));
16534 if (GET_MODE (variable) == DImode)
16535 emit_insn (gen_anddi3 (tmpcount, variable, GEN_INT (value)));
16537 emit_insn (gen_andsi3 (tmpcount, variable, GEN_INT (value)));
16538 emit_cmp_and_jump_insns (tmpcount, const0_rtx, EQ, 0, GET_MODE (variable),
16541 predict_jump (REG_BR_PROB_BASE * 50 / 100);
16543 predict_jump (REG_BR_PROB_BASE * 90 / 100);
16547 /* Adjust COUNTER by the VALUE. */
16549 ix86_adjust_counter (rtx countreg, HOST_WIDE_INT value)
16551 if (GET_MODE (countreg) == DImode)
16552 emit_insn (gen_adddi3 (countreg, countreg, GEN_INT (-value)));
16554 emit_insn (gen_addsi3 (countreg, countreg, GEN_INT (-value)));
16557 /* Zero extend possibly SImode EXP to Pmode register. */
16559 ix86_zero_extend_to_Pmode (rtx exp)
16562 if (GET_MODE (exp) == VOIDmode)
16563 return force_reg (Pmode, exp);
16564 if (GET_MODE (exp) == Pmode)
16565 return copy_to_mode_reg (Pmode, exp);
16566 r = gen_reg_rtx (Pmode);
16567 emit_insn (gen_zero_extendsidi2 (r, exp));
16571 /* Divide COUNTREG by SCALE. */
16573 scale_counter (rtx countreg, int scale)
16576 rtx piece_size_mask;
16580 if (CONST_INT_P (countreg))
16581 return GEN_INT (INTVAL (countreg) / scale);
16582 gcc_assert (REG_P (countreg));
16584 piece_size_mask = GEN_INT (scale - 1);
16585 sc = expand_simple_binop (GET_MODE (countreg), LSHIFTRT, countreg,
16586 GEN_INT (exact_log2 (scale)),
16587 NULL, 1, OPTAB_DIRECT);
16591 /* Return mode for the memcpy/memset loop counter. Prefer SImode over
16592 DImode for constant loop counts. */
16594 static enum machine_mode
16595 counter_mode (rtx count_exp)
16597 if (GET_MODE (count_exp) != VOIDmode)
16598 return GET_MODE (count_exp);
16599 if (GET_CODE (count_exp) != CONST_INT)
16601 if (TARGET_64BIT && (INTVAL (count_exp) & ~0xffffffff))
16606 /* When SRCPTR is non-NULL, output simple loop to move memory
16607 pointer to SRCPTR to DESTPTR via chunks of MODE unrolled UNROLL times,
16608 overall size is COUNT specified in bytes. When SRCPTR is NULL, output the
16609 equivalent loop to set memory by VALUE (supposed to be in MODE).
16611 The size is rounded down to whole number of chunk size moved at once.
16612 SRCMEM and DESTMEM provide MEMrtx to feed proper aliasing info. */
16616 expand_set_or_movmem_via_loop (rtx destmem, rtx srcmem,
16617 rtx destptr, rtx srcptr, rtx value,
16618 rtx count, enum machine_mode mode, int unroll,
16621 rtx out_label, top_label, iter, tmp;
16622 enum machine_mode iter_mode = counter_mode (count);
16623 rtx piece_size = GEN_INT (GET_MODE_SIZE (mode) * unroll);
16624 rtx piece_size_mask = GEN_INT (~((GET_MODE_SIZE (mode) * unroll) - 1));
16630 top_label = gen_label_rtx ();
16631 out_label = gen_label_rtx ();
16632 iter = gen_reg_rtx (iter_mode);
16634 size = expand_simple_binop (iter_mode, AND, count, piece_size_mask,
16635 NULL, 1, OPTAB_DIRECT);
16636 /* Those two should combine. */
16637 if (piece_size == const1_rtx)
16639 emit_cmp_and_jump_insns (size, const0_rtx, EQ, NULL_RTX, iter_mode,
16641 predict_jump (REG_BR_PROB_BASE * 10 / 100);
16643 emit_move_insn (iter, const0_rtx);
16645 emit_label (top_label);
16647 tmp = convert_modes (Pmode, iter_mode, iter, true);
16648 x_addr = gen_rtx_PLUS (Pmode, destptr, tmp);
16649 destmem = change_address (destmem, mode, x_addr);
16653 y_addr = gen_rtx_PLUS (Pmode, srcptr, copy_rtx (tmp));
16654 srcmem = change_address (srcmem, mode, y_addr);
16656 /* When unrolling for chips that reorder memory reads and writes,
16657 we can save registers by using single temporary.
16658 Also using 4 temporaries is overkill in 32bit mode. */
16659 if (!TARGET_64BIT && 0)
16661 for (i = 0; i < unroll; i++)
16666 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
16668 adjust_address (copy_rtx (srcmem), mode, GET_MODE_SIZE (mode));
16670 emit_move_insn (destmem, srcmem);
16676 gcc_assert (unroll <= 4);
16677 for (i = 0; i < unroll; i++)
16679 tmpreg[i] = gen_reg_rtx (mode);
16683 adjust_address (copy_rtx (srcmem), mode, GET_MODE_SIZE (mode));
16685 emit_move_insn (tmpreg[i], srcmem);
16687 for (i = 0; i < unroll; i++)
16692 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
16694 emit_move_insn (destmem, tmpreg[i]);
16699 for (i = 0; i < unroll; i++)
16703 adjust_address (copy_rtx (destmem), mode, GET_MODE_SIZE (mode));
16704 emit_move_insn (destmem, value);
16707 tmp = expand_simple_binop (iter_mode, PLUS, iter, piece_size, iter,
16708 true, OPTAB_LIB_WIDEN);
16710 emit_move_insn (iter, tmp);
16712 emit_cmp_and_jump_insns (iter, size, LT, NULL_RTX, iter_mode,
16714 if (expected_size != -1)
16716 expected_size /= GET_MODE_SIZE (mode) * unroll;
16717 if (expected_size == 0)
16719 else if (expected_size > REG_BR_PROB_BASE)
16720 predict_jump (REG_BR_PROB_BASE - 1);
16722 predict_jump (REG_BR_PROB_BASE - (REG_BR_PROB_BASE + expected_size / 2) / expected_size);
16725 predict_jump (REG_BR_PROB_BASE * 80 / 100);
16726 iter = ix86_zero_extend_to_Pmode (iter);
16727 tmp = expand_simple_binop (Pmode, PLUS, destptr, iter, destptr,
16728 true, OPTAB_LIB_WIDEN);
16729 if (tmp != destptr)
16730 emit_move_insn (destptr, tmp);
16733 tmp = expand_simple_binop (Pmode, PLUS, srcptr, iter, srcptr,
16734 true, OPTAB_LIB_WIDEN);
16736 emit_move_insn (srcptr, tmp);
16738 emit_label (out_label);
16741 /* Output "rep; mov" instruction.
16742 Arguments have same meaning as for previous function */
16744 expand_movmem_via_rep_mov (rtx destmem, rtx srcmem,
16745 rtx destptr, rtx srcptr,
16747 enum machine_mode mode)
16753 /* If the size is known, it is shorter to use rep movs. */
16754 if (mode == QImode && CONST_INT_P (count)
16755 && !(INTVAL (count) & 3))
16758 if (destptr != XEXP (destmem, 0) || GET_MODE (destmem) != BLKmode)
16759 destmem = adjust_automodify_address_nv (destmem, BLKmode, destptr, 0);
16760 if (srcptr != XEXP (srcmem, 0) || GET_MODE (srcmem) != BLKmode)
16761 srcmem = adjust_automodify_address_nv (srcmem, BLKmode, srcptr, 0);
16762 countreg = ix86_zero_extend_to_Pmode (scale_counter (count, GET_MODE_SIZE (mode)));
16763 if (mode != QImode)
16765 destexp = gen_rtx_ASHIFT (Pmode, countreg,
16766 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
16767 destexp = gen_rtx_PLUS (Pmode, destexp, destptr);
16768 srcexp = gen_rtx_ASHIFT (Pmode, countreg,
16769 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
16770 srcexp = gen_rtx_PLUS (Pmode, srcexp, srcptr);
16774 destexp = gen_rtx_PLUS (Pmode, destptr, countreg);
16775 srcexp = gen_rtx_PLUS (Pmode, srcptr, countreg);
16777 if (CONST_INT_P (count))
16779 count = GEN_INT (INTVAL (count)
16780 & ~((HOST_WIDE_INT) GET_MODE_SIZE (mode) - 1));
16781 destmem = shallow_copy_rtx (destmem);
16782 srcmem = shallow_copy_rtx (srcmem);
16783 set_mem_size (destmem, count);
16784 set_mem_size (srcmem, count);
16788 if (MEM_SIZE (destmem))
16789 set_mem_size (destmem, NULL_RTX);
16790 if (MEM_SIZE (srcmem))
16791 set_mem_size (srcmem, NULL_RTX);
16793 emit_insn (gen_rep_mov (destptr, destmem, srcptr, srcmem, countreg,
16797 /* Output "rep; stos" instruction.
16798 Arguments have same meaning as for previous function */
16800 expand_setmem_via_rep_stos (rtx destmem, rtx destptr, rtx value,
16801 rtx count, enum machine_mode mode,
16807 if (destptr != XEXP (destmem, 0) || GET_MODE (destmem) != BLKmode)
16808 destmem = adjust_automodify_address_nv (destmem, BLKmode, destptr, 0);
16809 value = force_reg (mode, gen_lowpart (mode, value));
16810 countreg = ix86_zero_extend_to_Pmode (scale_counter (count, GET_MODE_SIZE (mode)));
16811 if (mode != QImode)
16813 destexp = gen_rtx_ASHIFT (Pmode, countreg,
16814 GEN_INT (exact_log2 (GET_MODE_SIZE (mode))));
16815 destexp = gen_rtx_PLUS (Pmode, destexp, destptr);
16818 destexp = gen_rtx_PLUS (Pmode, destptr, countreg);
16819 if (orig_value == const0_rtx && CONST_INT_P (count))
16821 count = GEN_INT (INTVAL (count)
16822 & ~((HOST_WIDE_INT) GET_MODE_SIZE (mode) - 1));
16823 destmem = shallow_copy_rtx (destmem);
16824 set_mem_size (destmem, count);
16826 else if (MEM_SIZE (destmem))
16827 set_mem_size (destmem, NULL_RTX);
16828 emit_insn (gen_rep_stos (destptr, countreg, destmem, value, destexp));
16832 emit_strmov (rtx destmem, rtx srcmem,
16833 rtx destptr, rtx srcptr, enum machine_mode mode, int offset)
16835 rtx src = adjust_automodify_address_nv (srcmem, mode, srcptr, offset);
16836 rtx dest = adjust_automodify_address_nv (destmem, mode, destptr, offset);
16837 emit_insn (gen_strmov (destptr, dest, srcptr, src));
16840 /* Output code to copy at most count & (max_size - 1) bytes from SRC to DEST. */
16842 expand_movmem_epilogue (rtx destmem, rtx srcmem,
16843 rtx destptr, rtx srcptr, rtx count, int max_size)
16846 if (CONST_INT_P (count))
16848 HOST_WIDE_INT countval = INTVAL (count);
16851 if ((countval & 0x10) && max_size > 16)
16855 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset);
16856 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset + 8);
16859 gcc_unreachable ();
16862 if ((countval & 0x08) && max_size > 8)
16865 emit_strmov (destmem, srcmem, destptr, srcptr, DImode, offset);
16868 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset);
16869 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset + 4);
16873 if ((countval & 0x04) && max_size > 4)
16875 emit_strmov (destmem, srcmem, destptr, srcptr, SImode, offset);
16878 if ((countval & 0x02) && max_size > 2)
16880 emit_strmov (destmem, srcmem, destptr, srcptr, HImode, offset);
16883 if ((countval & 0x01) && max_size > 1)
16885 emit_strmov (destmem, srcmem, destptr, srcptr, QImode, offset);
16892 count = expand_simple_binop (GET_MODE (count), AND, count, GEN_INT (max_size - 1),
16893 count, 1, OPTAB_DIRECT);
16894 expand_set_or_movmem_via_loop (destmem, srcmem, destptr, srcptr, NULL,
16895 count, QImode, 1, 4);
16899 /* When there are stringops, we can cheaply increase dest and src pointers.
16900 Otherwise we save code size by maintaining offset (zero is readily
16901 available from preceding rep operation) and using x86 addressing modes.
16903 if (TARGET_SINGLE_STRINGOP)
16907 rtx label = ix86_expand_aligntest (count, 4, true);
16908 src = change_address (srcmem, SImode, srcptr);
16909 dest = change_address (destmem, SImode, destptr);
16910 emit_insn (gen_strmov (destptr, dest, srcptr, src));
16911 emit_label (label);
16912 LABEL_NUSES (label) = 1;
16916 rtx label = ix86_expand_aligntest (count, 2, true);
16917 src = change_address (srcmem, HImode, srcptr);
16918 dest = change_address (destmem, HImode, destptr);
16919 emit_insn (gen_strmov (destptr, dest, srcptr, src));
16920 emit_label (label);
16921 LABEL_NUSES (label) = 1;
16925 rtx label = ix86_expand_aligntest (count, 1, true);
16926 src = change_address (srcmem, QImode, srcptr);
16927 dest = change_address (destmem, QImode, destptr);
16928 emit_insn (gen_strmov (destptr, dest, srcptr, src));
16929 emit_label (label);
16930 LABEL_NUSES (label) = 1;
16935 rtx offset = force_reg (Pmode, const0_rtx);
16940 rtx label = ix86_expand_aligntest (count, 4, true);
16941 src = change_address (srcmem, SImode, srcptr);
16942 dest = change_address (destmem, SImode, destptr);
16943 emit_move_insn (dest, src);
16944 tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (4), NULL,
16945 true, OPTAB_LIB_WIDEN);
16947 emit_move_insn (offset, tmp);
16948 emit_label (label);
16949 LABEL_NUSES (label) = 1;
16953 rtx label = ix86_expand_aligntest (count, 2, true);
16954 tmp = gen_rtx_PLUS (Pmode, srcptr, offset);
16955 src = change_address (srcmem, HImode, tmp);
16956 tmp = gen_rtx_PLUS (Pmode, destptr, offset);
16957 dest = change_address (destmem, HImode, tmp);
16958 emit_move_insn (dest, src);
16959 tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (2), tmp,
16960 true, OPTAB_LIB_WIDEN);
16962 emit_move_insn (offset, tmp);
16963 emit_label (label);
16964 LABEL_NUSES (label) = 1;
16968 rtx label = ix86_expand_aligntest (count, 1, true);
16969 tmp = gen_rtx_PLUS (Pmode, srcptr, offset);
16970 src = change_address (srcmem, QImode, tmp);
16971 tmp = gen_rtx_PLUS (Pmode, destptr, offset);
16972 dest = change_address (destmem, QImode, tmp);
16973 emit_move_insn (dest, src);
16974 emit_label (label);
16975 LABEL_NUSES (label) = 1;
16980 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
16982 expand_setmem_epilogue_via_loop (rtx destmem, rtx destptr, rtx value,
16983 rtx count, int max_size)
16986 expand_simple_binop (counter_mode (count), AND, count,
16987 GEN_INT (max_size - 1), count, 1, OPTAB_DIRECT);
16988 expand_set_or_movmem_via_loop (destmem, NULL, destptr, NULL,
16989 gen_lowpart (QImode, value), count, QImode,
16993 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
16995 expand_setmem_epilogue (rtx destmem, rtx destptr, rtx value, rtx count, int max_size)
16999 if (CONST_INT_P (count))
17001 HOST_WIDE_INT countval = INTVAL (count);
17004 if ((countval & 0x10) && max_size > 16)
17008 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset);
17009 emit_insn (gen_strset (destptr, dest, value));
17010 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset + 8);
17011 emit_insn (gen_strset (destptr, dest, value));
17014 gcc_unreachable ();
17017 if ((countval & 0x08) && max_size > 8)
17021 dest = adjust_automodify_address_nv (destmem, DImode, destptr, offset);
17022 emit_insn (gen_strset (destptr, dest, value));
17026 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset);
17027 emit_insn (gen_strset (destptr, dest, value));
17028 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset + 4);
17029 emit_insn (gen_strset (destptr, dest, value));
17033 if ((countval & 0x04) && max_size > 4)
17035 dest = adjust_automodify_address_nv (destmem, SImode, destptr, offset);
17036 emit_insn (gen_strset (destptr, dest, gen_lowpart (SImode, value)));
17039 if ((countval & 0x02) && max_size > 2)
17041 dest = adjust_automodify_address_nv (destmem, HImode, destptr, offset);
17042 emit_insn (gen_strset (destptr, dest, gen_lowpart (HImode, value)));
17045 if ((countval & 0x01) && max_size > 1)
17047 dest = adjust_automodify_address_nv (destmem, QImode, destptr, offset);
17048 emit_insn (gen_strset (destptr, dest, gen_lowpart (QImode, value)));
17055 expand_setmem_epilogue_via_loop (destmem, destptr, value, count, max_size);
17060 rtx label = ix86_expand_aligntest (count, 16, true);
17063 dest = change_address (destmem, DImode, destptr);
17064 emit_insn (gen_strset (destptr, dest, value));
17065 emit_insn (gen_strset (destptr, dest, value));
17069 dest = change_address (destmem, SImode, destptr);
17070 emit_insn (gen_strset (destptr, dest, value));
17071 emit_insn (gen_strset (destptr, dest, value));
17072 emit_insn (gen_strset (destptr, dest, value));
17073 emit_insn (gen_strset (destptr, dest, value));
17075 emit_label (label);
17076 LABEL_NUSES (label) = 1;
17080 rtx label = ix86_expand_aligntest (count, 8, true);
17083 dest = change_address (destmem, DImode, destptr);
17084 emit_insn (gen_strset (destptr, dest, value));
17088 dest = change_address (destmem, SImode, destptr);
17089 emit_insn (gen_strset (destptr, dest, value));
17090 emit_insn (gen_strset (destptr, dest, value));
17092 emit_label (label);
17093 LABEL_NUSES (label) = 1;
17097 rtx label = ix86_expand_aligntest (count, 4, true);
17098 dest = change_address (destmem, SImode, destptr);
17099 emit_insn (gen_strset (destptr, dest, gen_lowpart (SImode, value)));
17100 emit_label (label);
17101 LABEL_NUSES (label) = 1;
17105 rtx label = ix86_expand_aligntest (count, 2, true);
17106 dest = change_address (destmem, HImode, destptr);
17107 emit_insn (gen_strset (destptr, dest, gen_lowpart (HImode, value)));
17108 emit_label (label);
17109 LABEL_NUSES (label) = 1;
17113 rtx label = ix86_expand_aligntest (count, 1, true);
17114 dest = change_address (destmem, QImode, destptr);
17115 emit_insn (gen_strset (destptr, dest, gen_lowpart (QImode, value)));
17116 emit_label (label);
17117 LABEL_NUSES (label) = 1;
17121 /* Copy enough from DEST to SRC to align DEST known to by aligned by ALIGN to
17122 DESIRED_ALIGNMENT. */
17124 expand_movmem_prologue (rtx destmem, rtx srcmem,
17125 rtx destptr, rtx srcptr, rtx count,
17126 int align, int desired_alignment)
17128 if (align <= 1 && desired_alignment > 1)
17130 rtx label = ix86_expand_aligntest (destptr, 1, false);
17131 srcmem = change_address (srcmem, QImode, srcptr);
17132 destmem = change_address (destmem, QImode, destptr);
17133 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17134 ix86_adjust_counter (count, 1);
17135 emit_label (label);
17136 LABEL_NUSES (label) = 1;
17138 if (align <= 2 && desired_alignment > 2)
17140 rtx label = ix86_expand_aligntest (destptr, 2, false);
17141 srcmem = change_address (srcmem, HImode, srcptr);
17142 destmem = change_address (destmem, HImode, destptr);
17143 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17144 ix86_adjust_counter (count, 2);
17145 emit_label (label);
17146 LABEL_NUSES (label) = 1;
17148 if (align <= 4 && desired_alignment > 4)
17150 rtx label = ix86_expand_aligntest (destptr, 4, false);
17151 srcmem = change_address (srcmem, SImode, srcptr);
17152 destmem = change_address (destmem, SImode, destptr);
17153 emit_insn (gen_strmov (destptr, destmem, srcptr, srcmem));
17154 ix86_adjust_counter (count, 4);
17155 emit_label (label);
17156 LABEL_NUSES (label) = 1;
17158 gcc_assert (desired_alignment <= 8);
17161 /* Copy enough from DST to SRC to align DST known to DESIRED_ALIGN.
17162 ALIGN_BYTES is how many bytes need to be copied. */
17164 expand_constant_movmem_prologue (rtx dst, rtx *srcp, rtx destreg, rtx srcreg,
17165 int desired_align, int align_bytes)
17168 rtx src_size, dst_size;
17170 int src_align_bytes = get_mem_align_offset (src, desired_align * BITS_PER_UNIT);
17171 if (src_align_bytes >= 0)
17172 src_align_bytes = desired_align - src_align_bytes;
17173 src_size = MEM_SIZE (src);
17174 dst_size = MEM_SIZE (dst);
17175 if (align_bytes & 1)
17177 dst = adjust_automodify_address_nv (dst, QImode, destreg, 0);
17178 src = adjust_automodify_address_nv (src, QImode, srcreg, 0);
17180 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17182 if (align_bytes & 2)
17184 dst = adjust_automodify_address_nv (dst, HImode, destreg, off);
17185 src = adjust_automodify_address_nv (src, HImode, srcreg, off);
17186 if (MEM_ALIGN (dst) < 2 * BITS_PER_UNIT)
17187 set_mem_align (dst, 2 * BITS_PER_UNIT);
17188 if (src_align_bytes >= 0
17189 && (src_align_bytes & 1) == (align_bytes & 1)
17190 && MEM_ALIGN (src) < 2 * BITS_PER_UNIT)
17191 set_mem_align (src, 2 * BITS_PER_UNIT);
17193 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17195 if (align_bytes & 4)
17197 dst = adjust_automodify_address_nv (dst, SImode, destreg, off);
17198 src = adjust_automodify_address_nv (src, SImode, srcreg, off);
17199 if (MEM_ALIGN (dst) < 4 * BITS_PER_UNIT)
17200 set_mem_align (dst, 4 * BITS_PER_UNIT);
17201 if (src_align_bytes >= 0)
17203 unsigned int src_align = 0;
17204 if ((src_align_bytes & 3) == (align_bytes & 3))
17206 else if ((src_align_bytes & 1) == (align_bytes & 1))
17208 if (MEM_ALIGN (src) < src_align * BITS_PER_UNIT)
17209 set_mem_align (src, src_align * BITS_PER_UNIT);
17212 emit_insn (gen_strmov (destreg, dst, srcreg, src));
17214 dst = adjust_automodify_address_nv (dst, BLKmode, destreg, off);
17215 src = adjust_automodify_address_nv (src, BLKmode, srcreg, off);
17216 if (MEM_ALIGN (dst) < (unsigned int) desired_align * BITS_PER_UNIT)
17217 set_mem_align (dst, desired_align * BITS_PER_UNIT);
17218 if (src_align_bytes >= 0)
17220 unsigned int src_align = 0;
17221 if ((src_align_bytes & 7) == (align_bytes & 7))
17223 else if ((src_align_bytes & 3) == (align_bytes & 3))
17225 else if ((src_align_bytes & 1) == (align_bytes & 1))
17227 if (src_align > (unsigned int) desired_align)
17228 src_align = desired_align;
17229 if (MEM_ALIGN (src) < src_align * BITS_PER_UNIT)
17230 set_mem_align (src, src_align * BITS_PER_UNIT);
17233 set_mem_size (dst, GEN_INT (INTVAL (dst_size) - align_bytes));
17235 set_mem_size (dst, GEN_INT (INTVAL (src_size) - align_bytes));
17240 /* Set enough from DEST to align DEST known to by aligned by ALIGN to
17241 DESIRED_ALIGNMENT. */
17243 expand_setmem_prologue (rtx destmem, rtx destptr, rtx value, rtx count,
17244 int align, int desired_alignment)
17246 if (align <= 1 && desired_alignment > 1)
17248 rtx label = ix86_expand_aligntest (destptr, 1, false);
17249 destmem = change_address (destmem, QImode, destptr);
17250 emit_insn (gen_strset (destptr, destmem, gen_lowpart (QImode, value)));
17251 ix86_adjust_counter (count, 1);
17252 emit_label (label);
17253 LABEL_NUSES (label) = 1;
17255 if (align <= 2 && desired_alignment > 2)
17257 rtx label = ix86_expand_aligntest (destptr, 2, false);
17258 destmem = change_address (destmem, HImode, destptr);
17259 emit_insn (gen_strset (destptr, destmem, gen_lowpart (HImode, value)));
17260 ix86_adjust_counter (count, 2);
17261 emit_label (label);
17262 LABEL_NUSES (label) = 1;
17264 if (align <= 4 && desired_alignment > 4)
17266 rtx label = ix86_expand_aligntest (destptr, 4, false);
17267 destmem = change_address (destmem, SImode, destptr);
17268 emit_insn (gen_strset (destptr, destmem, gen_lowpart (SImode, value)));
17269 ix86_adjust_counter (count, 4);
17270 emit_label (label);
17271 LABEL_NUSES (label) = 1;
17273 gcc_assert (desired_alignment <= 8);
17276 /* Set enough from DST to align DST known to by aligned by ALIGN to
17277 DESIRED_ALIGN. ALIGN_BYTES is how many bytes need to be stored. */
17279 expand_constant_setmem_prologue (rtx dst, rtx destreg, rtx value,
17280 int desired_align, int align_bytes)
17283 rtx dst_size = MEM_SIZE (dst);
17284 if (align_bytes & 1)
17286 dst = adjust_automodify_address_nv (dst, QImode, destreg, 0);
17288 emit_insn (gen_strset (destreg, dst,
17289 gen_lowpart (QImode, value)));
17291 if (align_bytes & 2)
17293 dst = adjust_automodify_address_nv (dst, HImode, destreg, off);
17294 if (MEM_ALIGN (dst) < 2 * BITS_PER_UNIT)
17295 set_mem_align (dst, 2 * BITS_PER_UNIT);
17297 emit_insn (gen_strset (destreg, dst,
17298 gen_lowpart (HImode, value)));
17300 if (align_bytes & 4)
17302 dst = adjust_automodify_address_nv (dst, SImode, destreg, off);
17303 if (MEM_ALIGN (dst) < 4 * BITS_PER_UNIT)
17304 set_mem_align (dst, 4 * BITS_PER_UNIT);
17306 emit_insn (gen_strset (destreg, dst,
17307 gen_lowpart (SImode, value)));
17309 dst = adjust_automodify_address_nv (dst, BLKmode, destreg, off);
17310 if (MEM_ALIGN (dst) < (unsigned int) desired_align * BITS_PER_UNIT)
17311 set_mem_align (dst, desired_align * BITS_PER_UNIT);
17313 set_mem_size (dst, GEN_INT (INTVAL (dst_size) - align_bytes));
17317 /* Given COUNT and EXPECTED_SIZE, decide on codegen of string operation. */
17318 static enum stringop_alg
17319 decide_alg (HOST_WIDE_INT count, HOST_WIDE_INT expected_size, bool memset,
17320 int *dynamic_check)
17322 const struct stringop_algs * algs;
17323 bool optimize_for_speed;
17324 /* Algorithms using the rep prefix want at least edi and ecx;
17325 additionally, memset wants eax and memcpy wants esi. Don't
17326 consider such algorithms if the user has appropriated those
17327 registers for their own purposes. */
17328 bool rep_prefix_usable = !(fixed_regs[CX_REG] || fixed_regs[DI_REG]
17330 ? fixed_regs[AX_REG] : fixed_regs[SI_REG]));
17332 #define ALG_USABLE_P(alg) (rep_prefix_usable \
17333 || (alg != rep_prefix_1_byte \
17334 && alg != rep_prefix_4_byte \
17335 && alg != rep_prefix_8_byte))
17336 const struct processor_costs *cost;
17338 /* Even if the string operation call is cold, we still might spend a lot
17339 of time processing large blocks. */
17340 if (optimize_function_for_size_p (cfun)
17341 || (optimize_insn_for_size_p ()
17342 && expected_size != -1 && expected_size < 256))
17343 optimize_for_speed = false;
17345 optimize_for_speed = true;
17347 cost = optimize_for_speed ? ix86_cost : &ix86_size_cost;
17349 *dynamic_check = -1;
17351 algs = &cost->memset[TARGET_64BIT != 0];
17353 algs = &cost->memcpy[TARGET_64BIT != 0];
17354 if (stringop_alg != no_stringop && ALG_USABLE_P (stringop_alg))
17355 return stringop_alg;
17356 /* rep; movq or rep; movl is the smallest variant. */
17357 else if (!optimize_for_speed)
17359 if (!count || (count & 3))
17360 return rep_prefix_usable ? rep_prefix_1_byte : loop_1_byte;
17362 return rep_prefix_usable ? rep_prefix_4_byte : loop;
17364 /* Very tiny blocks are best handled via the loop, REP is expensive to setup.
17366 else if (expected_size != -1 && expected_size < 4)
17367 return loop_1_byte;
17368 else if (expected_size != -1)
17371 enum stringop_alg alg = libcall;
17372 for (i = 0; i < NAX_STRINGOP_ALGS; i++)
17374 /* We get here if the algorithms that were not libcall-based
17375 were rep-prefix based and we are unable to use rep prefixes
17376 based on global register usage. Break out of the loop and
17377 use the heuristic below. */
17378 if (algs->size[i].max == 0)
17380 if (algs->size[i].max >= expected_size || algs->size[i].max == -1)
17382 enum stringop_alg candidate = algs->size[i].alg;
17384 if (candidate != libcall && ALG_USABLE_P (candidate))
17386 /* Honor TARGET_INLINE_ALL_STRINGOPS by picking
17387 last non-libcall inline algorithm. */
17388 if (TARGET_INLINE_ALL_STRINGOPS)
17390 /* When the current size is best to be copied by a libcall,
17391 but we are still forced to inline, run the heuristic below
17392 that will pick code for medium sized blocks. */
17393 if (alg != libcall)
17397 else if (ALG_USABLE_P (candidate))
17401 gcc_assert (TARGET_INLINE_ALL_STRINGOPS || !rep_prefix_usable);
17403 /* When asked to inline the call anyway, try to pick meaningful choice.
17404 We look for maximal size of block that is faster to copy by hand and
17405 take blocks of at most of that size guessing that average size will
17406 be roughly half of the block.
17408 If this turns out to be bad, we might simply specify the preferred
17409 choice in ix86_costs. */
17410 if ((TARGET_INLINE_ALL_STRINGOPS || TARGET_INLINE_STRINGOPS_DYNAMICALLY)
17411 && (algs->unknown_size == libcall || !ALG_USABLE_P (algs->unknown_size)))
17414 enum stringop_alg alg;
17416 bool any_alg_usable_p = true;
17418 for (i = 0; i < NAX_STRINGOP_ALGS; i++)
17420 enum stringop_alg candidate = algs->size[i].alg;
17421 any_alg_usable_p = any_alg_usable_p && ALG_USABLE_P (candidate);
17423 if (candidate != libcall && candidate
17424 && ALG_USABLE_P (candidate))
17425 max = algs->size[i].max;
17427 /* If there aren't any usable algorithms, then recursing on
17428 smaller sizes isn't going to find anything. Just return the
17429 simple byte-at-a-time copy loop. */
17430 if (!any_alg_usable_p)
17432 /* Pick something reasonable. */
17433 if (TARGET_INLINE_STRINGOPS_DYNAMICALLY)
17434 *dynamic_check = 128;
17435 return loop_1_byte;
17439 alg = decide_alg (count, max / 2, memset, dynamic_check);
17440 gcc_assert (*dynamic_check == -1);
17441 gcc_assert (alg != libcall);
17442 if (TARGET_INLINE_STRINGOPS_DYNAMICALLY)
17443 *dynamic_check = max;
17446 return ALG_USABLE_P (algs->unknown_size) ? algs->unknown_size : libcall;
17447 #undef ALG_USABLE_P
17450 /* Decide on alignment. We know that the operand is already aligned to ALIGN
17451 (ALIGN can be based on profile feedback and thus it is not 100% guaranteed). */
17453 decide_alignment (int align,
17454 enum stringop_alg alg,
17457 int desired_align = 0;
17461 gcc_unreachable ();
17463 case unrolled_loop:
17464 desired_align = GET_MODE_SIZE (Pmode);
17466 case rep_prefix_8_byte:
17469 case rep_prefix_4_byte:
17470 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
17471 copying whole cacheline at once. */
17472 if (TARGET_PENTIUMPRO)
17477 case rep_prefix_1_byte:
17478 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
17479 copying whole cacheline at once. */
17480 if (TARGET_PENTIUMPRO)
17494 if (desired_align < align)
17495 desired_align = align;
17496 if (expected_size != -1 && expected_size < 4)
17497 desired_align = align;
17498 return desired_align;
17501 /* Return the smallest power of 2 greater than VAL. */
17503 smallest_pow2_greater_than (int val)
17511 /* Expand string move (memcpy) operation. Use i386 string operations when
17512 profitable. expand_setmem contains similar code. The code depends upon
17513 architecture, block size and alignment, but always has the same
17516 1) Prologue guard: Conditional that jumps up to epilogues for small
17517 blocks that can be handled by epilogue alone. This is faster but
17518 also needed for correctness, since prologue assume the block is larger
17519 than the desired alignment.
17521 Optional dynamic check for size and libcall for large
17522 blocks is emitted here too, with -minline-stringops-dynamically.
17524 2) Prologue: copy first few bytes in order to get destination aligned
17525 to DESIRED_ALIGN. It is emitted only when ALIGN is less than
17526 DESIRED_ALIGN and and up to DESIRED_ALIGN - ALIGN bytes can be copied.
17527 We emit either a jump tree on power of two sized blocks, or a byte loop.
17529 3) Main body: the copying loop itself, copying in SIZE_NEEDED chunks
17530 with specified algorithm.
17532 4) Epilogue: code copying tail of the block that is too small to be
17533 handled by main body (or up to size guarded by prologue guard). */
17536 ix86_expand_movmem (rtx dst, rtx src, rtx count_exp, rtx align_exp,
17537 rtx expected_align_exp, rtx expected_size_exp)
17543 rtx jump_around_label = NULL;
17544 HOST_WIDE_INT align = 1;
17545 unsigned HOST_WIDE_INT count = 0;
17546 HOST_WIDE_INT expected_size = -1;
17547 int size_needed = 0, epilogue_size_needed;
17548 int desired_align = 0, align_bytes = 0;
17549 enum stringop_alg alg;
17551 bool need_zero_guard = false;
17553 if (CONST_INT_P (align_exp))
17554 align = INTVAL (align_exp);
17555 /* i386 can do misaligned access on reasonably increased cost. */
17556 if (CONST_INT_P (expected_align_exp)
17557 && INTVAL (expected_align_exp) > align)
17558 align = INTVAL (expected_align_exp);
17559 /* ALIGN is the minimum of destination and source alignment, but we care here
17560 just about destination alignment. */
17561 else if (MEM_ALIGN (dst) > (unsigned HOST_WIDE_INT) align * BITS_PER_UNIT)
17562 align = MEM_ALIGN (dst) / BITS_PER_UNIT;
17564 if (CONST_INT_P (count_exp))
17565 count = expected_size = INTVAL (count_exp);
17566 if (CONST_INT_P (expected_size_exp) && count == 0)
17567 expected_size = INTVAL (expected_size_exp);
17569 /* Make sure we don't need to care about overflow later on. */
17570 if (count > ((unsigned HOST_WIDE_INT) 1 << 30))
17573 /* Step 0: Decide on preferred algorithm, desired alignment and
17574 size of chunks to be copied by main loop. */
17576 alg = decide_alg (count, expected_size, false, &dynamic_check);
17577 desired_align = decide_alignment (align, alg, expected_size);
17579 if (!TARGET_ALIGN_STRINGOPS)
17580 align = desired_align;
17582 if (alg == libcall)
17584 gcc_assert (alg != no_stringop);
17586 count_exp = copy_to_mode_reg (GET_MODE (count_exp), count_exp);
17587 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
17588 srcreg = copy_to_mode_reg (Pmode, XEXP (src, 0));
17593 gcc_unreachable ();
17595 need_zero_guard = true;
17596 size_needed = GET_MODE_SIZE (Pmode);
17598 case unrolled_loop:
17599 need_zero_guard = true;
17600 size_needed = GET_MODE_SIZE (Pmode) * (TARGET_64BIT ? 4 : 2);
17602 case rep_prefix_8_byte:
17605 case rep_prefix_4_byte:
17608 case rep_prefix_1_byte:
17612 need_zero_guard = true;
17617 epilogue_size_needed = size_needed;
17619 /* Step 1: Prologue guard. */
17621 /* Alignment code needs count to be in register. */
17622 if (CONST_INT_P (count_exp) && desired_align > align)
17624 if (INTVAL (count_exp) > desired_align
17625 && INTVAL (count_exp) > size_needed)
17628 = get_mem_align_offset (dst, desired_align * BITS_PER_UNIT);
17629 if (align_bytes <= 0)
17632 align_bytes = desired_align - align_bytes;
17634 if (align_bytes == 0)
17635 count_exp = force_reg (counter_mode (count_exp), count_exp);
17637 gcc_assert (desired_align >= 1 && align >= 1);
17639 /* Ensure that alignment prologue won't copy past end of block. */
17640 if (size_needed > 1 || (desired_align > 1 && desired_align > align))
17642 epilogue_size_needed = MAX (size_needed - 1, desired_align - align);
17643 /* Epilogue always copies COUNT_EXP & EPILOGUE_SIZE_NEEDED bytes.
17644 Make sure it is power of 2. */
17645 epilogue_size_needed = smallest_pow2_greater_than (epilogue_size_needed);
17647 if (CONST_INT_P (count_exp))
17649 if (UINTVAL (count_exp) < (unsigned HOST_WIDE_INT)epilogue_size_needed)
17654 label = gen_label_rtx ();
17655 emit_cmp_and_jump_insns (count_exp,
17656 GEN_INT (epilogue_size_needed),
17657 LTU, 0, counter_mode (count_exp), 1, label);
17658 if (expected_size == -1 || expected_size < epilogue_size_needed)
17659 predict_jump (REG_BR_PROB_BASE * 60 / 100);
17661 predict_jump (REG_BR_PROB_BASE * 20 / 100);
17665 /* Emit code to decide on runtime whether library call or inline should be
17667 if (dynamic_check != -1)
17669 if (CONST_INT_P (count_exp))
17671 if (UINTVAL (count_exp) >= (unsigned HOST_WIDE_INT)dynamic_check)
17673 emit_block_move_via_libcall (dst, src, count_exp, false);
17674 count_exp = const0_rtx;
17680 rtx hot_label = gen_label_rtx ();
17681 jump_around_label = gen_label_rtx ();
17682 emit_cmp_and_jump_insns (count_exp, GEN_INT (dynamic_check - 1),
17683 LEU, 0, GET_MODE (count_exp), 1, hot_label);
17684 predict_jump (REG_BR_PROB_BASE * 90 / 100);
17685 emit_block_move_via_libcall (dst, src, count_exp, false);
17686 emit_jump (jump_around_label);
17687 emit_label (hot_label);
17691 /* Step 2: Alignment prologue. */
17693 if (desired_align > align)
17695 if (align_bytes == 0)
17697 /* Except for the first move in epilogue, we no longer know
17698 constant offset in aliasing info. It don't seems to worth
17699 the pain to maintain it for the first move, so throw away
17701 src = change_address (src, BLKmode, srcreg);
17702 dst = change_address (dst, BLKmode, destreg);
17703 expand_movmem_prologue (dst, src, destreg, srcreg, count_exp, align,
17708 /* If we know how many bytes need to be stored before dst is
17709 sufficiently aligned, maintain aliasing info accurately. */
17710 dst = expand_constant_movmem_prologue (dst, &src, destreg, srcreg,
17711 desired_align, align_bytes);
17712 count_exp = plus_constant (count_exp, -align_bytes);
17713 count -= align_bytes;
17715 if (need_zero_guard && !count)
17717 /* It is possible that we copied enough so the main loop will not
17719 emit_cmp_and_jump_insns (count_exp,
17720 GEN_INT (size_needed),
17721 LTU, 0, counter_mode (count_exp), 1, label);
17722 if (expected_size == -1
17723 || expected_size < (desired_align - align) / 2 + size_needed)
17724 predict_jump (REG_BR_PROB_BASE * 20 / 100);
17726 predict_jump (REG_BR_PROB_BASE * 60 / 100);
17729 if (label && size_needed == 1)
17731 emit_label (label);
17732 LABEL_NUSES (label) = 1;
17736 /* Step 3: Main loop. */
17742 gcc_unreachable ();
17744 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
17745 count_exp, QImode, 1, expected_size);
17748 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
17749 count_exp, Pmode, 1, expected_size);
17751 case unrolled_loop:
17752 /* Unroll only by factor of 2 in 32bit mode, since we don't have enough
17753 registers for 4 temporaries anyway. */
17754 expand_set_or_movmem_via_loop (dst, src, destreg, srcreg, NULL,
17755 count_exp, Pmode, TARGET_64BIT ? 4 : 2,
17758 case rep_prefix_8_byte:
17759 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
17762 case rep_prefix_4_byte:
17763 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
17766 case rep_prefix_1_byte:
17767 expand_movmem_via_rep_mov (dst, src, destreg, srcreg, count_exp,
17771 /* Adjust properly the offset of src and dest memory for aliasing. */
17772 if (CONST_INT_P (count_exp))
17774 src = adjust_automodify_address_nv (src, BLKmode, srcreg,
17775 (count / size_needed) * size_needed);
17776 dst = adjust_automodify_address_nv (dst, BLKmode, destreg,
17777 (count / size_needed) * size_needed);
17781 src = change_address (src, BLKmode, srcreg);
17782 dst = change_address (dst, BLKmode, destreg);
17785 /* Step 4: Epilogue to copy the remaining bytes. */
17789 /* When the main loop is done, COUNT_EXP might hold original count,
17790 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
17791 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
17792 bytes. Compensate if needed. */
17794 if (size_needed < epilogue_size_needed)
17797 expand_simple_binop (counter_mode (count_exp), AND, count_exp,
17798 GEN_INT (size_needed - 1), count_exp, 1,
17800 if (tmp != count_exp)
17801 emit_move_insn (count_exp, tmp);
17803 emit_label (label);
17804 LABEL_NUSES (label) = 1;
17807 if (count_exp != const0_rtx && epilogue_size_needed > 1)
17808 expand_movmem_epilogue (dst, src, destreg, srcreg, count_exp,
17809 epilogue_size_needed);
17810 if (jump_around_label)
17811 emit_label (jump_around_label);
17815 /* Helper function for memcpy. For QImode value 0xXY produce
17816 0xXYXYXYXY of wide specified by MODE. This is essentially
17817 a * 0x10101010, but we can do slightly better than
17818 synth_mult by unwinding the sequence by hand on CPUs with
17821 promote_duplicated_reg (enum machine_mode mode, rtx val)
17823 enum machine_mode valmode = GET_MODE (val);
17825 int nops = mode == DImode ? 3 : 2;
17827 gcc_assert (mode == SImode || mode == DImode);
17828 if (val == const0_rtx)
17829 return copy_to_mode_reg (mode, const0_rtx);
17830 if (CONST_INT_P (val))
17832 HOST_WIDE_INT v = INTVAL (val) & 255;
17836 if (mode == DImode)
17837 v |= (v << 16) << 16;
17838 return copy_to_mode_reg (mode, gen_int_mode (v, mode));
17841 if (valmode == VOIDmode)
17843 if (valmode != QImode)
17844 val = gen_lowpart (QImode, val);
17845 if (mode == QImode)
17847 if (!TARGET_PARTIAL_REG_STALL)
17849 if (ix86_cost->mult_init[mode == DImode ? 3 : 2]
17850 + ix86_cost->mult_bit * (mode == DImode ? 8 : 4)
17851 <= (ix86_cost->shift_const + ix86_cost->add) * nops
17852 + (COSTS_N_INSNS (TARGET_PARTIAL_REG_STALL == 0)))
17854 rtx reg = convert_modes (mode, QImode, val, true);
17855 tmp = promote_duplicated_reg (mode, const1_rtx);
17856 return expand_simple_binop (mode, MULT, reg, tmp, NULL, 1,
17861 rtx reg = convert_modes (mode, QImode, val, true);
17863 if (!TARGET_PARTIAL_REG_STALL)
17864 if (mode == SImode)
17865 emit_insn (gen_movsi_insv_1 (reg, reg));
17867 emit_insn (gen_movdi_insv_1_rex64 (reg, reg));
17870 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (8),
17871 NULL, 1, OPTAB_DIRECT);
17873 expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
17875 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (16),
17876 NULL, 1, OPTAB_DIRECT);
17877 reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
17878 if (mode == SImode)
17880 tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (32),
17881 NULL, 1, OPTAB_DIRECT);
17882 reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT);
17887 /* Duplicate value VAL using promote_duplicated_reg into maximal size that will
17888 be needed by main loop copying SIZE_NEEDED chunks and prologue getting
17889 alignment from ALIGN to DESIRED_ALIGN. */
17891 promote_duplicated_reg_to_size (rtx val, int size_needed, int desired_align, int align)
17896 && (size_needed > 4 || (desired_align > align && desired_align > 4)))
17897 promoted_val = promote_duplicated_reg (DImode, val);
17898 else if (size_needed > 2 || (desired_align > align && desired_align > 2))
17899 promoted_val = promote_duplicated_reg (SImode, val);
17900 else if (size_needed > 1 || (desired_align > align && desired_align > 1))
17901 promoted_val = promote_duplicated_reg (HImode, val);
17903 promoted_val = val;
17905 return promoted_val;
17908 /* Expand string clear operation (bzero). Use i386 string operations when
17909 profitable. See expand_movmem comment for explanation of individual
17910 steps performed. */
17912 ix86_expand_setmem (rtx dst, rtx count_exp, rtx val_exp, rtx align_exp,
17913 rtx expected_align_exp, rtx expected_size_exp)
17918 rtx jump_around_label = NULL;
17919 HOST_WIDE_INT align = 1;
17920 unsigned HOST_WIDE_INT count = 0;
17921 HOST_WIDE_INT expected_size = -1;
17922 int size_needed = 0, epilogue_size_needed;
17923 int desired_align = 0, align_bytes = 0;
17924 enum stringop_alg alg;
17925 rtx promoted_val = NULL;
17926 bool force_loopy_epilogue = false;
17928 bool need_zero_guard = false;
17930 if (CONST_INT_P (align_exp))
17931 align = INTVAL (align_exp);
17932 /* i386 can do misaligned access on reasonably increased cost. */
17933 if (CONST_INT_P (expected_align_exp)
17934 && INTVAL (expected_align_exp) > align)
17935 align = INTVAL (expected_align_exp);
17936 if (CONST_INT_P (count_exp))
17937 count = expected_size = INTVAL (count_exp);
17938 if (CONST_INT_P (expected_size_exp) && count == 0)
17939 expected_size = INTVAL (expected_size_exp);
17941 /* Make sure we don't need to care about overflow later on. */
17942 if (count > ((unsigned HOST_WIDE_INT) 1 << 30))
17945 /* Step 0: Decide on preferred algorithm, desired alignment and
17946 size of chunks to be copied by main loop. */
17948 alg = decide_alg (count, expected_size, true, &dynamic_check);
17949 desired_align = decide_alignment (align, alg, expected_size);
17951 if (!TARGET_ALIGN_STRINGOPS)
17952 align = desired_align;
17954 if (alg == libcall)
17956 gcc_assert (alg != no_stringop);
17958 count_exp = copy_to_mode_reg (counter_mode (count_exp), count_exp);
17959 destreg = copy_to_mode_reg (Pmode, XEXP (dst, 0));
17964 gcc_unreachable ();
17966 need_zero_guard = true;
17967 size_needed = GET_MODE_SIZE (Pmode);
17969 case unrolled_loop:
17970 need_zero_guard = true;
17971 size_needed = GET_MODE_SIZE (Pmode) * 4;
17973 case rep_prefix_8_byte:
17976 case rep_prefix_4_byte:
17979 case rep_prefix_1_byte:
17983 need_zero_guard = true;
17987 epilogue_size_needed = size_needed;
17989 /* Step 1: Prologue guard. */
17991 /* Alignment code needs count to be in register. */
17992 if (CONST_INT_P (count_exp) && desired_align > align)
17994 if (INTVAL (count_exp) > desired_align
17995 && INTVAL (count_exp) > size_needed)
17998 = get_mem_align_offset (dst, desired_align * BITS_PER_UNIT);
17999 if (align_bytes <= 0)
18002 align_bytes = desired_align - align_bytes;
18004 if (align_bytes == 0)
18006 enum machine_mode mode = SImode;
18007 if (TARGET_64BIT && (count & ~0xffffffff))
18009 count_exp = force_reg (mode, count_exp);
18012 /* Do the cheap promotion to allow better CSE across the
18013 main loop and epilogue (ie one load of the big constant in the
18014 front of all code. */
18015 if (CONST_INT_P (val_exp))
18016 promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed,
18017 desired_align, align);
18018 /* Ensure that alignment prologue won't copy past end of block. */
18019 if (size_needed > 1 || (desired_align > 1 && desired_align > align))
18021 epilogue_size_needed = MAX (size_needed - 1, desired_align - align);
18022 /* Epilogue always copies COUNT_EXP & (EPILOGUE_SIZE_NEEDED - 1) bytes.
18023 Make sure it is power of 2. */
18024 epilogue_size_needed = smallest_pow2_greater_than (epilogue_size_needed);
18026 /* To improve performance of small blocks, we jump around the VAL
18027 promoting mode. This mean that if the promoted VAL is not constant,
18028 we might not use it in the epilogue and have to use byte
18030 if (epilogue_size_needed > 2 && !promoted_val)
18031 force_loopy_epilogue = true;
18032 label = gen_label_rtx ();
18033 emit_cmp_and_jump_insns (count_exp,
18034 GEN_INT (epilogue_size_needed),
18035 LTU, 0, counter_mode (count_exp), 1, label);
18036 if (GET_CODE (count_exp) == CONST_INT)
18038 else if (expected_size == -1 || expected_size <= epilogue_size_needed)
18039 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18041 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18043 if (dynamic_check != -1)
18045 rtx hot_label = gen_label_rtx ();
18046 jump_around_label = gen_label_rtx ();
18047 emit_cmp_and_jump_insns (count_exp, GEN_INT (dynamic_check - 1),
18048 LEU, 0, counter_mode (count_exp), 1, hot_label);
18049 predict_jump (REG_BR_PROB_BASE * 90 / 100);
18050 set_storage_via_libcall (dst, count_exp, val_exp, false);
18051 emit_jump (jump_around_label);
18052 emit_label (hot_label);
18055 /* Step 2: Alignment prologue. */
18057 /* Do the expensive promotion once we branched off the small blocks. */
18059 promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed,
18060 desired_align, align);
18061 gcc_assert (desired_align >= 1 && align >= 1);
18063 if (desired_align > align)
18065 if (align_bytes == 0)
18067 /* Except for the first move in epilogue, we no longer know
18068 constant offset in aliasing info. It don't seems to worth
18069 the pain to maintain it for the first move, so throw away
18071 dst = change_address (dst, BLKmode, destreg);
18072 expand_setmem_prologue (dst, destreg, promoted_val, count_exp, align,
18077 /* If we know how many bytes need to be stored before dst is
18078 sufficiently aligned, maintain aliasing info accurately. */
18079 dst = expand_constant_setmem_prologue (dst, destreg, promoted_val,
18080 desired_align, align_bytes);
18081 count_exp = plus_constant (count_exp, -align_bytes);
18082 count -= align_bytes;
18084 if (need_zero_guard && !count)
18086 /* It is possible that we copied enough so the main loop will not
18088 emit_cmp_and_jump_insns (count_exp,
18089 GEN_INT (size_needed),
18090 LTU, 0, counter_mode (count_exp), 1, label);
18091 if (expected_size == -1
18092 || expected_size < (desired_align - align) / 2 + size_needed)
18093 predict_jump (REG_BR_PROB_BASE * 20 / 100);
18095 predict_jump (REG_BR_PROB_BASE * 60 / 100);
18098 if (label && size_needed == 1)
18100 emit_label (label);
18101 LABEL_NUSES (label) = 1;
18105 /* Step 3: Main loop. */
18111 gcc_unreachable ();
18113 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18114 count_exp, QImode, 1, expected_size);
18117 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18118 count_exp, Pmode, 1, expected_size);
18120 case unrolled_loop:
18121 expand_set_or_movmem_via_loop (dst, NULL, destreg, NULL, promoted_val,
18122 count_exp, Pmode, 4, expected_size);
18124 case rep_prefix_8_byte:
18125 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18128 case rep_prefix_4_byte:
18129 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18132 case rep_prefix_1_byte:
18133 expand_setmem_via_rep_stos (dst, destreg, promoted_val, count_exp,
18137 /* Adjust properly the offset of src and dest memory for aliasing. */
18138 if (CONST_INT_P (count_exp))
18139 dst = adjust_automodify_address_nv (dst, BLKmode, destreg,
18140 (count / size_needed) * size_needed);
18142 dst = change_address (dst, BLKmode, destreg);
18144 /* Step 4: Epilogue to copy the remaining bytes. */
18148 /* When the main loop is done, COUNT_EXP might hold original count,
18149 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
18150 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
18151 bytes. Compensate if needed. */
18153 if (size_needed < desired_align - align)
18156 expand_simple_binop (counter_mode (count_exp), AND, count_exp,
18157 GEN_INT (size_needed - 1), count_exp, 1,
18159 size_needed = desired_align - align + 1;
18160 if (tmp != count_exp)
18161 emit_move_insn (count_exp, tmp);
18163 emit_label (label);
18164 LABEL_NUSES (label) = 1;
18166 if (count_exp != const0_rtx && epilogue_size_needed > 1)
18168 if (force_loopy_epilogue)
18169 expand_setmem_epilogue_via_loop (dst, destreg, val_exp, count_exp,
18172 expand_setmem_epilogue (dst, destreg, promoted_val, count_exp,
18175 if (jump_around_label)
18176 emit_label (jump_around_label);
18180 /* Expand the appropriate insns for doing strlen if not just doing
18183 out = result, initialized with the start address
18184 align_rtx = alignment of the address.
18185 scratch = scratch register, initialized with the startaddress when
18186 not aligned, otherwise undefined
18188 This is just the body. It needs the initializations mentioned above and
18189 some address computing at the end. These things are done in i386.md. */
18192 ix86_expand_strlensi_unroll_1 (rtx out, rtx src, rtx align_rtx)
18196 rtx align_2_label = NULL_RTX;
18197 rtx align_3_label = NULL_RTX;
18198 rtx align_4_label = gen_label_rtx ();
18199 rtx end_0_label = gen_label_rtx ();
18201 rtx tmpreg = gen_reg_rtx (SImode);
18202 rtx scratch = gen_reg_rtx (SImode);
18206 if (CONST_INT_P (align_rtx))
18207 align = INTVAL (align_rtx);
18209 /* Loop to check 1..3 bytes for null to get an aligned pointer. */
18211 /* Is there a known alignment and is it less than 4? */
18214 rtx scratch1 = gen_reg_rtx (Pmode);
18215 emit_move_insn (scratch1, out);
18216 /* Is there a known alignment and is it not 2? */
18219 align_3_label = gen_label_rtx (); /* Label when aligned to 3-byte */
18220 align_2_label = gen_label_rtx (); /* Label when aligned to 2-byte */
18222 /* Leave just the 3 lower bits. */
18223 align_rtx = expand_binop (Pmode, and_optab, scratch1, GEN_INT (3),
18224 NULL_RTX, 0, OPTAB_WIDEN);
18226 emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL,
18227 Pmode, 1, align_4_label);
18228 emit_cmp_and_jump_insns (align_rtx, const2_rtx, EQ, NULL,
18229 Pmode, 1, align_2_label);
18230 emit_cmp_and_jump_insns (align_rtx, const2_rtx, GTU, NULL,
18231 Pmode, 1, align_3_label);
18235 /* Since the alignment is 2, we have to check 2 or 0 bytes;
18236 check if is aligned to 4 - byte. */
18238 align_rtx = expand_binop (Pmode, and_optab, scratch1, const2_rtx,
18239 NULL_RTX, 0, OPTAB_WIDEN);
18241 emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL,
18242 Pmode, 1, align_4_label);
18245 mem = change_address (src, QImode, out);
18247 /* Now compare the bytes. */
18249 /* Compare the first n unaligned byte on a byte per byte basis. */
18250 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL,
18251 QImode, 1, end_0_label);
18253 /* Increment the address. */
18254 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18256 /* Not needed with an alignment of 2 */
18259 emit_label (align_2_label);
18261 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1,
18264 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18266 emit_label (align_3_label);
18269 emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1,
18272 emit_insn ((*ix86_gen_add3) (out, out, const1_rtx));
18275 /* Generate loop to check 4 bytes at a time. It is not a good idea to
18276 align this loop. It gives only huge programs, but does not help to
18278 emit_label (align_4_label);
18280 mem = change_address (src, SImode, out);
18281 emit_move_insn (scratch, mem);
18282 emit_insn ((*ix86_gen_add3) (out, out, GEN_INT (4)));
18284 /* This formula yields a nonzero result iff one of the bytes is zero.
18285 This saves three branches inside loop and many cycles. */
18287 emit_insn (gen_addsi3 (tmpreg, scratch, GEN_INT (-0x01010101)));
18288 emit_insn (gen_one_cmplsi2 (scratch, scratch));
18289 emit_insn (gen_andsi3 (tmpreg, tmpreg, scratch));
18290 emit_insn (gen_andsi3 (tmpreg, tmpreg,
18291 gen_int_mode (0x80808080, SImode)));
18292 emit_cmp_and_jump_insns (tmpreg, const0_rtx, EQ, 0, SImode, 1,
18297 rtx reg = gen_reg_rtx (SImode);
18298 rtx reg2 = gen_reg_rtx (Pmode);
18299 emit_move_insn (reg, tmpreg);
18300 emit_insn (gen_lshrsi3 (reg, reg, GEN_INT (16)));
18302 /* If zero is not in the first two bytes, move two bytes forward. */
18303 emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080)));
18304 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
18305 tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
18306 emit_insn (gen_rtx_SET (VOIDmode, tmpreg,
18307 gen_rtx_IF_THEN_ELSE (SImode, tmp,
18310 /* Emit lea manually to avoid clobbering of flags. */
18311 emit_insn (gen_rtx_SET (SImode, reg2,
18312 gen_rtx_PLUS (Pmode, out, const2_rtx)));
18314 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
18315 tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
18316 emit_insn (gen_rtx_SET (VOIDmode, out,
18317 gen_rtx_IF_THEN_ELSE (Pmode, tmp,
18324 rtx end_2_label = gen_label_rtx ();
18325 /* Is zero in the first two bytes? */
18327 emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080)));
18328 tmp = gen_rtx_REG (CCNOmode, FLAGS_REG);
18329 tmp = gen_rtx_NE (VOIDmode, tmp, const0_rtx);
18330 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
18331 gen_rtx_LABEL_REF (VOIDmode, end_2_label),
18333 tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
18334 JUMP_LABEL (tmp) = end_2_label;
18336 /* Not in the first two. Move two bytes forward. */
18337 emit_insn (gen_lshrsi3 (tmpreg, tmpreg, GEN_INT (16)));
18338 emit_insn ((*ix86_gen_add3) (out, out, const2_rtx));
18340 emit_label (end_2_label);
18344 /* Avoid branch in fixing the byte. */
18345 tmpreg = gen_lowpart (QImode, tmpreg);
18346 emit_insn (gen_addqi3_cc (tmpreg, tmpreg, tmpreg));
18347 cmp = gen_rtx_LTU (Pmode, gen_rtx_REG (CCmode, FLAGS_REG), const0_rtx);
18348 emit_insn ((*ix86_gen_sub3_carry) (out, out, GEN_INT (3), cmp));
18350 emit_label (end_0_label);
18353 /* Expand strlen. */
18356 ix86_expand_strlen (rtx out, rtx src, rtx eoschar, rtx align)
18358 rtx addr, scratch1, scratch2, scratch3, scratch4;
18360 /* The generic case of strlen expander is long. Avoid it's
18361 expanding unless TARGET_INLINE_ALL_STRINGOPS. */
18363 if (TARGET_UNROLL_STRLEN && eoschar == const0_rtx && optimize > 1
18364 && !TARGET_INLINE_ALL_STRINGOPS
18365 && !optimize_insn_for_size_p ()
18366 && (!CONST_INT_P (align) || INTVAL (align) < 4))
18369 addr = force_reg (Pmode, XEXP (src, 0));
18370 scratch1 = gen_reg_rtx (Pmode);
18372 if (TARGET_UNROLL_STRLEN && eoschar == const0_rtx && optimize > 1
18373 && !optimize_insn_for_size_p ())
18375 /* Well it seems that some optimizer does not combine a call like
18376 foo(strlen(bar), strlen(bar));
18377 when the move and the subtraction is done here. It does calculate
18378 the length just once when these instructions are done inside of
18379 output_strlen_unroll(). But I think since &bar[strlen(bar)] is
18380 often used and I use one fewer register for the lifetime of
18381 output_strlen_unroll() this is better. */
18383 emit_move_insn (out, addr);
18385 ix86_expand_strlensi_unroll_1 (out, src, align);
18387 /* strlensi_unroll_1 returns the address of the zero at the end of
18388 the string, like memchr(), so compute the length by subtracting
18389 the start address. */
18390 emit_insn ((*ix86_gen_sub3) (out, out, addr));
18396 /* Can't use this if the user has appropriated eax, ecx, or edi. */
18397 if (fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])
18400 scratch2 = gen_reg_rtx (Pmode);
18401 scratch3 = gen_reg_rtx (Pmode);
18402 scratch4 = force_reg (Pmode, constm1_rtx);
18404 emit_move_insn (scratch3, addr);
18405 eoschar = force_reg (QImode, eoschar);
18407 src = replace_equiv_address_nv (src, scratch3);
18409 /* If .md starts supporting :P, this can be done in .md. */
18410 unspec = gen_rtx_UNSPEC (Pmode, gen_rtvec (4, src, eoschar, align,
18411 scratch4), UNSPEC_SCAS);
18412 emit_insn (gen_strlenqi_1 (scratch1, scratch3, unspec));
18413 emit_insn ((*ix86_gen_one_cmpl2) (scratch2, scratch1));
18414 emit_insn ((*ix86_gen_add3) (out, scratch2, constm1_rtx));
18419 /* For given symbol (function) construct code to compute address of it's PLT
18420 entry in large x86-64 PIC model. */
18422 construct_plt_address (rtx symbol)
18424 rtx tmp = gen_reg_rtx (Pmode);
18425 rtx unspec = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, symbol), UNSPEC_PLTOFF);
18427 gcc_assert (GET_CODE (symbol) == SYMBOL_REF);
18428 gcc_assert (ix86_cmodel == CM_LARGE_PIC);
18430 emit_move_insn (tmp, gen_rtx_CONST (Pmode, unspec));
18431 emit_insn (gen_adddi3 (tmp, tmp, pic_offset_table_rtx));
18436 ix86_expand_call (rtx retval, rtx fnaddr, rtx callarg1,
18438 rtx pop, int sibcall)
18440 rtx use = NULL, call;
18441 enum calling_abi function_call_abi;
18443 if (callarg2 && INTVAL (callarg2) == -2)
18444 function_call_abi = MS_ABI;
18446 function_call_abi = SYSV_ABI;
18447 if (pop == const0_rtx)
18449 gcc_assert (!TARGET_64BIT || !pop);
18451 if (TARGET_MACHO && !TARGET_64BIT)
18454 if (flag_pic && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF)
18455 fnaddr = machopic_indirect_call_target (fnaddr);
18460 /* Static functions and indirect calls don't need the pic register. */
18461 if (flag_pic && (!TARGET_64BIT || ix86_cmodel == CM_LARGE_PIC)
18462 && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF
18463 && ! SYMBOL_REF_LOCAL_P (XEXP (fnaddr, 0)))
18464 use_reg (&use, pic_offset_table_rtx);
18467 if (TARGET_64BIT && INTVAL (callarg2) >= 0)
18469 rtx al = gen_rtx_REG (QImode, AX_REG);
18470 emit_move_insn (al, callarg2);
18471 use_reg (&use, al);
18474 if (ix86_cmodel == CM_LARGE_PIC
18475 && GET_CODE (fnaddr) == MEM
18476 && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF
18477 && !local_symbolic_operand (XEXP (fnaddr, 0), VOIDmode))
18478 fnaddr = gen_rtx_MEM (QImode, construct_plt_address (XEXP (fnaddr, 0)));
18479 else if (! call_insn_operand (XEXP (fnaddr, 0), Pmode))
18481 fnaddr = copy_to_mode_reg (Pmode, XEXP (fnaddr, 0));
18482 fnaddr = gen_rtx_MEM (QImode, fnaddr);
18484 if (sibcall && TARGET_64BIT
18485 && !constant_call_address_operand (XEXP (fnaddr, 0), Pmode))
18488 addr = copy_to_mode_reg (Pmode, XEXP (fnaddr, 0));
18489 fnaddr = gen_rtx_REG (Pmode, R11_REG);
18490 emit_move_insn (fnaddr, addr);
18491 fnaddr = gen_rtx_MEM (QImode, fnaddr);
18494 call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1);
18496 call = gen_rtx_SET (VOIDmode, retval, call);
18499 pop = gen_rtx_PLUS (Pmode, stack_pointer_rtx, pop);
18500 pop = gen_rtx_SET (VOIDmode, stack_pointer_rtx, pop);
18501 call = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, call, pop));
18502 gcc_assert (ix86_cfun_abi () != MS_ABI || function_call_abi != SYSV_ABI);
18504 /* We need to represent that SI and DI registers are clobbered
18506 if (ix86_cfun_abi () == MS_ABI && function_call_abi == SYSV_ABI)
18508 static int clobbered_registers[] = {27, 28, 45, 46, 47, 48, 49, 50, 51,
18509 52, SI_REG, DI_REG};
18511 rtx vec[ARRAY_SIZE (clobbered_registers) + 2];
18512 rtx unspec = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, const0_rtx),
18513 UNSPEC_MS_TO_SYSV_CALL);
18517 for (i = 0; i < ARRAY_SIZE (clobbered_registers); i++)
18518 vec[i + 2] = gen_rtx_CLOBBER (SSE_REGNO_P (clobbered_registers[i])
18521 (SSE_REGNO_P (clobbered_registers[i])
18523 clobbered_registers[i]));
18525 call = gen_rtx_PARALLEL (VOIDmode,
18526 gen_rtvec_v (ARRAY_SIZE (clobbered_registers)
18530 call = emit_call_insn (call);
18532 CALL_INSN_FUNCTION_USAGE (call) = use;
18536 /* Clear stack slot assignments remembered from previous functions.
18537 This is called from INIT_EXPANDERS once before RTL is emitted for each
18540 static struct machine_function *
18541 ix86_init_machine_status (void)
18543 struct machine_function *f;
18545 f = GGC_CNEW (struct machine_function);
18546 f->use_fast_prologue_epilogue_nregs = -1;
18547 f->tls_descriptor_call_expanded_p = 0;
18548 f->call_abi = DEFAULT_ABI;
18553 /* Return a MEM corresponding to a stack slot with mode MODE.
18554 Allocate a new slot if necessary.
18556 The RTL for a function can have several slots available: N is
18557 which slot to use. */
18560 assign_386_stack_local (enum machine_mode mode, enum ix86_stack_slot n)
18562 struct stack_local_entry *s;
18564 gcc_assert (n < MAX_386_STACK_LOCALS);
18566 /* Virtual slot is valid only before vregs are instantiated. */
18567 gcc_assert ((n == SLOT_VIRTUAL) == !virtuals_instantiated);
18569 for (s = ix86_stack_locals; s; s = s->next)
18570 if (s->mode == mode && s->n == n)
18571 return copy_rtx (s->rtl);
18573 s = (struct stack_local_entry *)
18574 ggc_alloc (sizeof (struct stack_local_entry));
18577 s->rtl = assign_stack_local (mode, GET_MODE_SIZE (mode), 0);
18579 s->next = ix86_stack_locals;
18580 ix86_stack_locals = s;
18584 /* Construct the SYMBOL_REF for the tls_get_addr function. */
18586 static GTY(()) rtx ix86_tls_symbol;
18588 ix86_tls_get_addr (void)
18591 if (!ix86_tls_symbol)
18593 ix86_tls_symbol = gen_rtx_SYMBOL_REF (Pmode,
18594 (TARGET_ANY_GNU_TLS
18596 ? "___tls_get_addr"
18597 : "__tls_get_addr");
18600 return ix86_tls_symbol;
18603 /* Construct the SYMBOL_REF for the _TLS_MODULE_BASE_ symbol. */
18605 static GTY(()) rtx ix86_tls_module_base_symbol;
18607 ix86_tls_module_base (void)
18610 if (!ix86_tls_module_base_symbol)
18612 ix86_tls_module_base_symbol = gen_rtx_SYMBOL_REF (Pmode,
18613 "_TLS_MODULE_BASE_");
18614 SYMBOL_REF_FLAGS (ix86_tls_module_base_symbol)
18615 |= TLS_MODEL_GLOBAL_DYNAMIC << SYMBOL_FLAG_TLS_SHIFT;
18618 return ix86_tls_module_base_symbol;
18621 /* Calculate the length of the memory address in the instruction
18622 encoding. Does not include the one-byte modrm, opcode, or prefix. */
18625 memory_address_length (rtx addr)
18627 struct ix86_address parts;
18628 rtx base, index, disp;
18632 if (GET_CODE (addr) == PRE_DEC
18633 || GET_CODE (addr) == POST_INC
18634 || GET_CODE (addr) == PRE_MODIFY
18635 || GET_CODE (addr) == POST_MODIFY)
18638 ok = ix86_decompose_address (addr, &parts);
18641 if (parts.base && GET_CODE (parts.base) == SUBREG)
18642 parts.base = SUBREG_REG (parts.base);
18643 if (parts.index && GET_CODE (parts.index) == SUBREG)
18644 parts.index = SUBREG_REG (parts.index);
18647 index = parts.index;
18652 - esp as the base always wants an index,
18653 - ebp as the base always wants a displacement. */
18655 /* Register Indirect. */
18656 if (base && !index && !disp)
18658 /* esp (for its index) and ebp (for its displacement) need
18659 the two-byte modrm form. */
18660 if (addr == stack_pointer_rtx
18661 || addr == arg_pointer_rtx
18662 || addr == frame_pointer_rtx
18663 || addr == hard_frame_pointer_rtx)
18667 /* Direct Addressing. */
18668 else if (disp && !base && !index)
18673 /* Find the length of the displacement constant. */
18676 if (base && satisfies_constraint_K (disp))
18681 /* ebp always wants a displacement. */
18682 else if (base == hard_frame_pointer_rtx)
18685 /* An index requires the two-byte modrm form.... */
18687 /* ...like esp, which always wants an index. */
18688 || base == stack_pointer_rtx
18689 || base == arg_pointer_rtx
18690 || base == frame_pointer_rtx)
18697 /* Compute default value for "length_immediate" attribute. When SHORTFORM
18698 is set, expect that insn have 8bit immediate alternative. */
18700 ix86_attr_length_immediate_default (rtx insn, int shortform)
18704 extract_insn_cached (insn);
18705 for (i = recog_data.n_operands - 1; i >= 0; --i)
18706 if (CONSTANT_P (recog_data.operand[i]))
18709 if (shortform && satisfies_constraint_K (recog_data.operand[i]))
18713 switch (get_attr_mode (insn))
18724 /* Immediates for DImode instructions are encoded as 32bit sign extended values. */
18729 fatal_insn ("unknown insn mode", insn);
18735 /* Compute default value for "length_address" attribute. */
18737 ix86_attr_length_address_default (rtx insn)
18741 if (get_attr_type (insn) == TYPE_LEA)
18743 rtx set = PATTERN (insn);
18745 if (GET_CODE (set) == PARALLEL)
18746 set = XVECEXP (set, 0, 0);
18748 gcc_assert (GET_CODE (set) == SET);
18750 return memory_address_length (SET_SRC (set));
18753 extract_insn_cached (insn);
18754 for (i = recog_data.n_operands - 1; i >= 0; --i)
18755 if (MEM_P (recog_data.operand[i]))
18757 return memory_address_length (XEXP (recog_data.operand[i], 0));
18763 /* Compute default value for "length_vex" attribute. It includes
18764 2 or 3 byte VEX prefix and 1 opcode byte. */
18767 ix86_attr_length_vex_default (rtx insn, int has_0f_opcode,
18772 /* Only 0f opcode can use 2 byte VEX prefix and VEX W bit uses 3
18773 byte VEX prefix. */
18774 if (!has_0f_opcode || has_vex_w)
18777 /* We can always use 2 byte VEX prefix in 32bit. */
18781 extract_insn_cached (insn);
18783 for (i = recog_data.n_operands - 1; i >= 0; --i)
18784 if (REG_P (recog_data.operand[i]))
18786 /* REX.W bit uses 3 byte VEX prefix. */
18787 if (GET_MODE (recog_data.operand[i]) == DImode)
18792 /* REX.X or REX.B bits use 3 byte VEX prefix. */
18793 if (MEM_P (recog_data.operand[i])
18794 && x86_extended_reg_mentioned_p (recog_data.operand[i]))
18801 /* Return the maximum number of instructions a cpu can issue. */
18804 ix86_issue_rate (void)
18808 case PROCESSOR_PENTIUM:
18812 case PROCESSOR_PENTIUMPRO:
18813 case PROCESSOR_PENTIUM4:
18814 case PROCESSOR_ATHLON:
18816 case PROCESSOR_AMDFAM10:
18817 case PROCESSOR_NOCONA:
18818 case PROCESSOR_GENERIC32:
18819 case PROCESSOR_GENERIC64:
18822 case PROCESSOR_CORE2:
18830 /* A subroutine of ix86_adjust_cost -- return true iff INSN reads flags set
18831 by DEP_INSN and nothing set by DEP_INSN. */
18834 ix86_flags_dependent (rtx insn, rtx dep_insn, enum attr_type insn_type)
18838 /* Simplify the test for uninteresting insns. */
18839 if (insn_type != TYPE_SETCC
18840 && insn_type != TYPE_ICMOV
18841 && insn_type != TYPE_FCMOV
18842 && insn_type != TYPE_IBR)
18845 if ((set = single_set (dep_insn)) != 0)
18847 set = SET_DEST (set);
18850 else if (GET_CODE (PATTERN (dep_insn)) == PARALLEL
18851 && XVECLEN (PATTERN (dep_insn), 0) == 2
18852 && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 0)) == SET
18853 && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 1)) == SET)
18855 set = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0));
18856 set2 = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0));
18861 if (!REG_P (set) || REGNO (set) != FLAGS_REG)
18864 /* This test is true if the dependent insn reads the flags but
18865 not any other potentially set register. */
18866 if (!reg_overlap_mentioned_p (set, PATTERN (insn)))
18869 if (set2 && reg_overlap_mentioned_p (set2, PATTERN (insn)))
18875 /* A subroutine of ix86_adjust_cost -- return true iff INSN has a memory
18876 address with operands set by DEP_INSN. */
18879 ix86_agi_dependent (rtx insn, rtx dep_insn, enum attr_type insn_type)
18883 if (insn_type == TYPE_LEA
18886 addr = PATTERN (insn);
18888 if (GET_CODE (addr) == PARALLEL)
18889 addr = XVECEXP (addr, 0, 0);
18891 gcc_assert (GET_CODE (addr) == SET);
18893 addr = SET_SRC (addr);
18898 extract_insn_cached (insn);
18899 for (i = recog_data.n_operands - 1; i >= 0; --i)
18900 if (MEM_P (recog_data.operand[i]))
18902 addr = XEXP (recog_data.operand[i], 0);
18909 return modified_in_p (addr, dep_insn);
18913 ix86_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
18915 enum attr_type insn_type, dep_insn_type;
18916 enum attr_memory memory;
18918 int dep_insn_code_number;
18920 /* Anti and output dependencies have zero cost on all CPUs. */
18921 if (REG_NOTE_KIND (link) != 0)
18924 dep_insn_code_number = recog_memoized (dep_insn);
18926 /* If we can't recognize the insns, we can't really do anything. */
18927 if (dep_insn_code_number < 0 || recog_memoized (insn) < 0)
18930 insn_type = get_attr_type (insn);
18931 dep_insn_type = get_attr_type (dep_insn);
18935 case PROCESSOR_PENTIUM:
18936 /* Address Generation Interlock adds a cycle of latency. */
18937 if (ix86_agi_dependent (insn, dep_insn, insn_type))
18940 /* ??? Compares pair with jump/setcc. */
18941 if (ix86_flags_dependent (insn, dep_insn, insn_type))
18944 /* Floating point stores require value to be ready one cycle earlier. */
18945 if (insn_type == TYPE_FMOV
18946 && get_attr_memory (insn) == MEMORY_STORE
18947 && !ix86_agi_dependent (insn, dep_insn, insn_type))
18951 case PROCESSOR_PENTIUMPRO:
18952 memory = get_attr_memory (insn);
18954 /* INT->FP conversion is expensive. */
18955 if (get_attr_fp_int_src (dep_insn))
18958 /* There is one cycle extra latency between an FP op and a store. */
18959 if (insn_type == TYPE_FMOV
18960 && (set = single_set (dep_insn)) != NULL_RTX
18961 && (set2 = single_set (insn)) != NULL_RTX
18962 && rtx_equal_p (SET_DEST (set), SET_SRC (set2))
18963 && MEM_P (SET_DEST (set2)))
18966 /* Show ability of reorder buffer to hide latency of load by executing
18967 in parallel with previous instruction in case
18968 previous instruction is not needed to compute the address. */
18969 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
18970 && !ix86_agi_dependent (insn, dep_insn, insn_type))
18972 /* Claim moves to take one cycle, as core can issue one load
18973 at time and the next load can start cycle later. */
18974 if (dep_insn_type == TYPE_IMOV
18975 || dep_insn_type == TYPE_FMOV)
18983 memory = get_attr_memory (insn);
18985 /* The esp dependency is resolved before the instruction is really
18987 if ((insn_type == TYPE_PUSH || insn_type == TYPE_POP)
18988 && (dep_insn_type == TYPE_PUSH || dep_insn_type == TYPE_POP))
18991 /* INT->FP conversion is expensive. */
18992 if (get_attr_fp_int_src (dep_insn))
18995 /* Show ability of reorder buffer to hide latency of load by executing
18996 in parallel with previous instruction in case
18997 previous instruction is not needed to compute the address. */
18998 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
18999 && !ix86_agi_dependent (insn, dep_insn, insn_type))
19001 /* Claim moves to take one cycle, as core can issue one load
19002 at time and the next load can start cycle later. */
19003 if (dep_insn_type == TYPE_IMOV
19004 || dep_insn_type == TYPE_FMOV)
19013 case PROCESSOR_ATHLON:
19015 case PROCESSOR_AMDFAM10:
19016 case PROCESSOR_GENERIC32:
19017 case PROCESSOR_GENERIC64:
19018 memory = get_attr_memory (insn);
19020 /* Show ability of reorder buffer to hide latency of load by executing
19021 in parallel with previous instruction in case
19022 previous instruction is not needed to compute the address. */
19023 if ((memory == MEMORY_LOAD || memory == MEMORY_BOTH)
19024 && !ix86_agi_dependent (insn, dep_insn, insn_type))
19026 enum attr_unit unit = get_attr_unit (insn);
19029 /* Because of the difference between the length of integer and
19030 floating unit pipeline preparation stages, the memory operands
19031 for floating point are cheaper.
19033 ??? For Athlon it the difference is most probably 2. */
19034 if (unit == UNIT_INTEGER || unit == UNIT_UNKNOWN)
19037 loadcost = TARGET_ATHLON ? 2 : 0;
19039 if (cost >= loadcost)
19052 /* How many alternative schedules to try. This should be as wide as the
19053 scheduling freedom in the DFA, but no wider. Making this value too
19054 large results extra work for the scheduler. */
19057 ia32_multipass_dfa_lookahead (void)
19061 case PROCESSOR_PENTIUM:
19064 case PROCESSOR_PENTIUMPRO:
19074 /* Compute the alignment given to a constant that is being placed in memory.
19075 EXP is the constant and ALIGN is the alignment that the object would
19077 The value of this function is used instead of that alignment to align
19081 ix86_constant_alignment (tree exp, int align)
19083 if (TREE_CODE (exp) == REAL_CST || TREE_CODE (exp) == VECTOR_CST
19084 || TREE_CODE (exp) == INTEGER_CST)
19086 if (TYPE_MODE (TREE_TYPE (exp)) == DFmode && align < 64)
19088 else if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (exp))) && align < 128)
19091 else if (!optimize_size && TREE_CODE (exp) == STRING_CST
19092 && TREE_STRING_LENGTH (exp) >= 31 && align < BITS_PER_WORD)
19093 return BITS_PER_WORD;
19098 /* Compute the alignment for a static variable.
19099 TYPE is the data type, and ALIGN is the alignment that
19100 the object would ordinarily have. The value of this function is used
19101 instead of that alignment to align the object. */
19104 ix86_data_alignment (tree type, int align)
19106 int max_align = optimize_size ? BITS_PER_WORD : MIN (256, MAX_OFILE_ALIGNMENT);
19108 if (AGGREGATE_TYPE_P (type)
19109 && TYPE_SIZE (type)
19110 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19111 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= (unsigned) max_align
19112 || TREE_INT_CST_HIGH (TYPE_SIZE (type)))
19113 && align < max_align)
19116 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
19117 to 16byte boundary. */
19120 if (AGGREGATE_TYPE_P (type)
19121 && TYPE_SIZE (type)
19122 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19123 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= 128
19124 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 128)
19128 if (TREE_CODE (type) == ARRAY_TYPE)
19130 if (TYPE_MODE (TREE_TYPE (type)) == DFmode && align < 64)
19132 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type))) && align < 128)
19135 else if (TREE_CODE (type) == COMPLEX_TYPE)
19138 if (TYPE_MODE (type) == DCmode && align < 64)
19140 if ((TYPE_MODE (type) == XCmode
19141 || TYPE_MODE (type) == TCmode) && align < 128)
19144 else if ((TREE_CODE (type) == RECORD_TYPE
19145 || TREE_CODE (type) == UNION_TYPE
19146 || TREE_CODE (type) == QUAL_UNION_TYPE)
19147 && TYPE_FIELDS (type))
19149 if (DECL_MODE (TYPE_FIELDS (type)) == DFmode && align < 64)
19151 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type))) && align < 128)
19154 else if (TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == VECTOR_TYPE
19155 || TREE_CODE (type) == INTEGER_TYPE)
19157 if (TYPE_MODE (type) == DFmode && align < 64)
19159 if (ALIGN_MODE_128 (TYPE_MODE (type)) && align < 128)
19166 /* Compute the alignment for a local variable or a stack slot. TYPE is
19167 the data type, MODE is the widest mode available and ALIGN is the
19168 alignment that the object would ordinarily have. The value of this
19169 macro is used instead of that alignment to align the object. */
19172 ix86_local_alignment (tree type, enum machine_mode mode,
19173 unsigned int align)
19175 /* If TYPE is NULL, we are allocating a stack slot for caller-save
19176 register in MODE. We will return the largest alignment of XF
19180 if (mode == XFmode && align < GET_MODE_ALIGNMENT (DFmode))
19181 align = GET_MODE_ALIGNMENT (DFmode);
19185 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
19186 to 16byte boundary. */
19189 if (AGGREGATE_TYPE_P (type)
19190 && TYPE_SIZE (type)
19191 && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
19192 && (TREE_INT_CST_LOW (TYPE_SIZE (type)) >= 16
19193 || TREE_INT_CST_HIGH (TYPE_SIZE (type))) && align < 128)
19196 if (TREE_CODE (type) == ARRAY_TYPE)
19198 if (TYPE_MODE (TREE_TYPE (type)) == DFmode && align < 64)
19200 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type))) && align < 128)
19203 else if (TREE_CODE (type) == COMPLEX_TYPE)
19205 if (TYPE_MODE (type) == DCmode && align < 64)
19207 if ((TYPE_MODE (type) == XCmode
19208 || TYPE_MODE (type) == TCmode) && align < 128)
19211 else if ((TREE_CODE (type) == RECORD_TYPE
19212 || TREE_CODE (type) == UNION_TYPE
19213 || TREE_CODE (type) == QUAL_UNION_TYPE)
19214 && TYPE_FIELDS (type))
19216 if (DECL_MODE (TYPE_FIELDS (type)) == DFmode && align < 64)
19218 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type))) && align < 128)
19221 else if (TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == VECTOR_TYPE
19222 || TREE_CODE (type) == INTEGER_TYPE)
19225 if (TYPE_MODE (type) == DFmode && align < 64)
19227 if (ALIGN_MODE_128 (TYPE_MODE (type)) && align < 128)
19233 /* Emit RTL insns to initialize the variable parts of a trampoline.
19234 FNADDR is an RTX for the address of the function's pure code.
19235 CXT is an RTX for the static chain value for the function. */
19237 x86_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
19241 /* Compute offset from the end of the jmp to the target function. */
19242 rtx disp = expand_binop (SImode, sub_optab, fnaddr,
19243 plus_constant (tramp, 10),
19244 NULL_RTX, 1, OPTAB_DIRECT);
19245 emit_move_insn (gen_rtx_MEM (QImode, tramp),
19246 gen_int_mode (0xb9, QImode));
19247 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 1)), cxt);
19248 emit_move_insn (gen_rtx_MEM (QImode, plus_constant (tramp, 5)),
19249 gen_int_mode (0xe9, QImode));
19250 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 6)), disp);
19255 /* Try to load address using shorter movl instead of movabs.
19256 We may want to support movq for kernel mode, but kernel does not use
19257 trampolines at the moment. */
19258 if (x86_64_zext_immediate_operand (fnaddr, VOIDmode))
19260 fnaddr = copy_to_mode_reg (DImode, fnaddr);
19261 emit_move_insn (gen_rtx_MEM (HImode, plus_constant (tramp, offset)),
19262 gen_int_mode (0xbb41, HImode));
19263 emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, offset + 2)),
19264 gen_lowpart (SImode, fnaddr));
19269 emit_move_insn (gen_rtx_MEM (HImode, plus_constant (tramp, offset)),
19270 gen_int_mode (0xbb49, HImode));
19271 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, offset + 2)),
19275 /* Load static chain using movabs to r10. */
19276 emit_move_insn (gen_rtx_MEM (HImode, plus_constant (tramp, offset)),
19277 gen_int_mode (0xba49, HImode));
19278 emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, offset + 2)),
19281 /* Jump to the r11 */
19282 emit_move_insn (gen_rtx_MEM (HImode, plus_constant (tramp, offset)),
19283 gen_int_mode (0xff49, HImode));
19284 emit_move_insn (gen_rtx_MEM (QImode, plus_constant (tramp, offset+2)),
19285 gen_int_mode (0xe3, QImode));
19287 gcc_assert (offset <= TRAMPOLINE_SIZE);
19290 #ifdef ENABLE_EXECUTE_STACK
19291 emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__enable_execute_stack"),
19292 LCT_NORMAL, VOIDmode, 1, tramp, Pmode);
19296 /* Codes for all the SSE/MMX builtins. */
19299 IX86_BUILTIN_ADDPS,
19300 IX86_BUILTIN_ADDSS,
19301 IX86_BUILTIN_DIVPS,
19302 IX86_BUILTIN_DIVSS,
19303 IX86_BUILTIN_MULPS,
19304 IX86_BUILTIN_MULSS,
19305 IX86_BUILTIN_SUBPS,
19306 IX86_BUILTIN_SUBSS,
19308 IX86_BUILTIN_CMPEQPS,
19309 IX86_BUILTIN_CMPLTPS,
19310 IX86_BUILTIN_CMPLEPS,
19311 IX86_BUILTIN_CMPGTPS,
19312 IX86_BUILTIN_CMPGEPS,
19313 IX86_BUILTIN_CMPNEQPS,
19314 IX86_BUILTIN_CMPNLTPS,
19315 IX86_BUILTIN_CMPNLEPS,
19316 IX86_BUILTIN_CMPNGTPS,
19317 IX86_BUILTIN_CMPNGEPS,
19318 IX86_BUILTIN_CMPORDPS,
19319 IX86_BUILTIN_CMPUNORDPS,
19320 IX86_BUILTIN_CMPEQSS,
19321 IX86_BUILTIN_CMPLTSS,
19322 IX86_BUILTIN_CMPLESS,
19323 IX86_BUILTIN_CMPNEQSS,
19324 IX86_BUILTIN_CMPNLTSS,
19325 IX86_BUILTIN_CMPNLESS,
19326 IX86_BUILTIN_CMPNGTSS,
19327 IX86_BUILTIN_CMPNGESS,
19328 IX86_BUILTIN_CMPORDSS,
19329 IX86_BUILTIN_CMPUNORDSS,
19331 IX86_BUILTIN_COMIEQSS,
19332 IX86_BUILTIN_COMILTSS,
19333 IX86_BUILTIN_COMILESS,
19334 IX86_BUILTIN_COMIGTSS,
19335 IX86_BUILTIN_COMIGESS,
19336 IX86_BUILTIN_COMINEQSS,
19337 IX86_BUILTIN_UCOMIEQSS,
19338 IX86_BUILTIN_UCOMILTSS,
19339 IX86_BUILTIN_UCOMILESS,
19340 IX86_BUILTIN_UCOMIGTSS,
19341 IX86_BUILTIN_UCOMIGESS,
19342 IX86_BUILTIN_UCOMINEQSS,
19344 IX86_BUILTIN_CVTPI2PS,
19345 IX86_BUILTIN_CVTPS2PI,
19346 IX86_BUILTIN_CVTSI2SS,
19347 IX86_BUILTIN_CVTSI642SS,
19348 IX86_BUILTIN_CVTSS2SI,
19349 IX86_BUILTIN_CVTSS2SI64,
19350 IX86_BUILTIN_CVTTPS2PI,
19351 IX86_BUILTIN_CVTTSS2SI,
19352 IX86_BUILTIN_CVTTSS2SI64,
19354 IX86_BUILTIN_MAXPS,
19355 IX86_BUILTIN_MAXSS,
19356 IX86_BUILTIN_MINPS,
19357 IX86_BUILTIN_MINSS,
19359 IX86_BUILTIN_LOADUPS,
19360 IX86_BUILTIN_STOREUPS,
19361 IX86_BUILTIN_MOVSS,
19363 IX86_BUILTIN_MOVHLPS,
19364 IX86_BUILTIN_MOVLHPS,
19365 IX86_BUILTIN_LOADHPS,
19366 IX86_BUILTIN_LOADLPS,
19367 IX86_BUILTIN_STOREHPS,
19368 IX86_BUILTIN_STORELPS,
19370 IX86_BUILTIN_MASKMOVQ,
19371 IX86_BUILTIN_MOVMSKPS,
19372 IX86_BUILTIN_PMOVMSKB,
19374 IX86_BUILTIN_MOVNTPS,
19375 IX86_BUILTIN_MOVNTQ,
19377 IX86_BUILTIN_LOADDQU,
19378 IX86_BUILTIN_STOREDQU,
19380 IX86_BUILTIN_PACKSSWB,
19381 IX86_BUILTIN_PACKSSDW,
19382 IX86_BUILTIN_PACKUSWB,
19384 IX86_BUILTIN_PADDB,
19385 IX86_BUILTIN_PADDW,
19386 IX86_BUILTIN_PADDD,
19387 IX86_BUILTIN_PADDQ,
19388 IX86_BUILTIN_PADDSB,
19389 IX86_BUILTIN_PADDSW,
19390 IX86_BUILTIN_PADDUSB,
19391 IX86_BUILTIN_PADDUSW,
19392 IX86_BUILTIN_PSUBB,
19393 IX86_BUILTIN_PSUBW,
19394 IX86_BUILTIN_PSUBD,
19395 IX86_BUILTIN_PSUBQ,
19396 IX86_BUILTIN_PSUBSB,
19397 IX86_BUILTIN_PSUBSW,
19398 IX86_BUILTIN_PSUBUSB,
19399 IX86_BUILTIN_PSUBUSW,
19402 IX86_BUILTIN_PANDN,
19406 IX86_BUILTIN_PAVGB,
19407 IX86_BUILTIN_PAVGW,
19409 IX86_BUILTIN_PCMPEQB,
19410 IX86_BUILTIN_PCMPEQW,
19411 IX86_BUILTIN_PCMPEQD,
19412 IX86_BUILTIN_PCMPGTB,
19413 IX86_BUILTIN_PCMPGTW,
19414 IX86_BUILTIN_PCMPGTD,
19416 IX86_BUILTIN_PMADDWD,
19418 IX86_BUILTIN_PMAXSW,
19419 IX86_BUILTIN_PMAXUB,
19420 IX86_BUILTIN_PMINSW,
19421 IX86_BUILTIN_PMINUB,
19423 IX86_BUILTIN_PMULHUW,
19424 IX86_BUILTIN_PMULHW,
19425 IX86_BUILTIN_PMULLW,
19427 IX86_BUILTIN_PSADBW,
19428 IX86_BUILTIN_PSHUFW,
19430 IX86_BUILTIN_PSLLW,
19431 IX86_BUILTIN_PSLLD,
19432 IX86_BUILTIN_PSLLQ,
19433 IX86_BUILTIN_PSRAW,
19434 IX86_BUILTIN_PSRAD,
19435 IX86_BUILTIN_PSRLW,
19436 IX86_BUILTIN_PSRLD,
19437 IX86_BUILTIN_PSRLQ,
19438 IX86_BUILTIN_PSLLWI,
19439 IX86_BUILTIN_PSLLDI,
19440 IX86_BUILTIN_PSLLQI,
19441 IX86_BUILTIN_PSRAWI,
19442 IX86_BUILTIN_PSRADI,
19443 IX86_BUILTIN_PSRLWI,
19444 IX86_BUILTIN_PSRLDI,
19445 IX86_BUILTIN_PSRLQI,
19447 IX86_BUILTIN_PUNPCKHBW,
19448 IX86_BUILTIN_PUNPCKHWD,
19449 IX86_BUILTIN_PUNPCKHDQ,
19450 IX86_BUILTIN_PUNPCKLBW,
19451 IX86_BUILTIN_PUNPCKLWD,
19452 IX86_BUILTIN_PUNPCKLDQ,
19454 IX86_BUILTIN_SHUFPS,
19456 IX86_BUILTIN_RCPPS,
19457 IX86_BUILTIN_RCPSS,
19458 IX86_BUILTIN_RSQRTPS,
19459 IX86_BUILTIN_RSQRTPS_NR,
19460 IX86_BUILTIN_RSQRTSS,
19461 IX86_BUILTIN_RSQRTF,
19462 IX86_BUILTIN_SQRTPS,
19463 IX86_BUILTIN_SQRTPS_NR,
19464 IX86_BUILTIN_SQRTSS,
19466 IX86_BUILTIN_UNPCKHPS,
19467 IX86_BUILTIN_UNPCKLPS,
19469 IX86_BUILTIN_ANDPS,
19470 IX86_BUILTIN_ANDNPS,
19472 IX86_BUILTIN_XORPS,
19475 IX86_BUILTIN_LDMXCSR,
19476 IX86_BUILTIN_STMXCSR,
19477 IX86_BUILTIN_SFENCE,
19479 /* 3DNow! Original */
19480 IX86_BUILTIN_FEMMS,
19481 IX86_BUILTIN_PAVGUSB,
19482 IX86_BUILTIN_PF2ID,
19483 IX86_BUILTIN_PFACC,
19484 IX86_BUILTIN_PFADD,
19485 IX86_BUILTIN_PFCMPEQ,
19486 IX86_BUILTIN_PFCMPGE,
19487 IX86_BUILTIN_PFCMPGT,
19488 IX86_BUILTIN_PFMAX,
19489 IX86_BUILTIN_PFMIN,
19490 IX86_BUILTIN_PFMUL,
19491 IX86_BUILTIN_PFRCP,
19492 IX86_BUILTIN_PFRCPIT1,
19493 IX86_BUILTIN_PFRCPIT2,
19494 IX86_BUILTIN_PFRSQIT1,
19495 IX86_BUILTIN_PFRSQRT,
19496 IX86_BUILTIN_PFSUB,
19497 IX86_BUILTIN_PFSUBR,
19498 IX86_BUILTIN_PI2FD,
19499 IX86_BUILTIN_PMULHRW,
19501 /* 3DNow! Athlon Extensions */
19502 IX86_BUILTIN_PF2IW,
19503 IX86_BUILTIN_PFNACC,
19504 IX86_BUILTIN_PFPNACC,
19505 IX86_BUILTIN_PI2FW,
19506 IX86_BUILTIN_PSWAPDSI,
19507 IX86_BUILTIN_PSWAPDSF,
19510 IX86_BUILTIN_ADDPD,
19511 IX86_BUILTIN_ADDSD,
19512 IX86_BUILTIN_DIVPD,
19513 IX86_BUILTIN_DIVSD,
19514 IX86_BUILTIN_MULPD,
19515 IX86_BUILTIN_MULSD,
19516 IX86_BUILTIN_SUBPD,
19517 IX86_BUILTIN_SUBSD,
19519 IX86_BUILTIN_CMPEQPD,
19520 IX86_BUILTIN_CMPLTPD,
19521 IX86_BUILTIN_CMPLEPD,
19522 IX86_BUILTIN_CMPGTPD,
19523 IX86_BUILTIN_CMPGEPD,
19524 IX86_BUILTIN_CMPNEQPD,
19525 IX86_BUILTIN_CMPNLTPD,
19526 IX86_BUILTIN_CMPNLEPD,
19527 IX86_BUILTIN_CMPNGTPD,
19528 IX86_BUILTIN_CMPNGEPD,
19529 IX86_BUILTIN_CMPORDPD,
19530 IX86_BUILTIN_CMPUNORDPD,
19531 IX86_BUILTIN_CMPEQSD,
19532 IX86_BUILTIN_CMPLTSD,
19533 IX86_BUILTIN_CMPLESD,
19534 IX86_BUILTIN_CMPNEQSD,
19535 IX86_BUILTIN_CMPNLTSD,
19536 IX86_BUILTIN_CMPNLESD,
19537 IX86_BUILTIN_CMPORDSD,
19538 IX86_BUILTIN_CMPUNORDSD,
19540 IX86_BUILTIN_COMIEQSD,
19541 IX86_BUILTIN_COMILTSD,
19542 IX86_BUILTIN_COMILESD,
19543 IX86_BUILTIN_COMIGTSD,
19544 IX86_BUILTIN_COMIGESD,
19545 IX86_BUILTIN_COMINEQSD,
19546 IX86_BUILTIN_UCOMIEQSD,
19547 IX86_BUILTIN_UCOMILTSD,
19548 IX86_BUILTIN_UCOMILESD,
19549 IX86_BUILTIN_UCOMIGTSD,
19550 IX86_BUILTIN_UCOMIGESD,
19551 IX86_BUILTIN_UCOMINEQSD,
19553 IX86_BUILTIN_MAXPD,
19554 IX86_BUILTIN_MAXSD,
19555 IX86_BUILTIN_MINPD,
19556 IX86_BUILTIN_MINSD,
19558 IX86_BUILTIN_ANDPD,
19559 IX86_BUILTIN_ANDNPD,
19561 IX86_BUILTIN_XORPD,
19563 IX86_BUILTIN_SQRTPD,
19564 IX86_BUILTIN_SQRTSD,
19566 IX86_BUILTIN_UNPCKHPD,
19567 IX86_BUILTIN_UNPCKLPD,
19569 IX86_BUILTIN_SHUFPD,
19571 IX86_BUILTIN_LOADUPD,
19572 IX86_BUILTIN_STOREUPD,
19573 IX86_BUILTIN_MOVSD,
19575 IX86_BUILTIN_LOADHPD,
19576 IX86_BUILTIN_LOADLPD,
19578 IX86_BUILTIN_CVTDQ2PD,
19579 IX86_BUILTIN_CVTDQ2PS,
19581 IX86_BUILTIN_CVTPD2DQ,
19582 IX86_BUILTIN_CVTPD2PI,
19583 IX86_BUILTIN_CVTPD2PS,
19584 IX86_BUILTIN_CVTTPD2DQ,
19585 IX86_BUILTIN_CVTTPD2PI,
19587 IX86_BUILTIN_CVTPI2PD,
19588 IX86_BUILTIN_CVTSI2SD,
19589 IX86_BUILTIN_CVTSI642SD,
19591 IX86_BUILTIN_CVTSD2SI,
19592 IX86_BUILTIN_CVTSD2SI64,
19593 IX86_BUILTIN_CVTSD2SS,
19594 IX86_BUILTIN_CVTSS2SD,
19595 IX86_BUILTIN_CVTTSD2SI,
19596 IX86_BUILTIN_CVTTSD2SI64,
19598 IX86_BUILTIN_CVTPS2DQ,
19599 IX86_BUILTIN_CVTPS2PD,
19600 IX86_BUILTIN_CVTTPS2DQ,
19602 IX86_BUILTIN_MOVNTI,
19603 IX86_BUILTIN_MOVNTPD,
19604 IX86_BUILTIN_MOVNTDQ,
19606 IX86_BUILTIN_MOVQ128,
19609 IX86_BUILTIN_MASKMOVDQU,
19610 IX86_BUILTIN_MOVMSKPD,
19611 IX86_BUILTIN_PMOVMSKB128,
19613 IX86_BUILTIN_PACKSSWB128,
19614 IX86_BUILTIN_PACKSSDW128,
19615 IX86_BUILTIN_PACKUSWB128,
19617 IX86_BUILTIN_PADDB128,
19618 IX86_BUILTIN_PADDW128,
19619 IX86_BUILTIN_PADDD128,
19620 IX86_BUILTIN_PADDQ128,
19621 IX86_BUILTIN_PADDSB128,
19622 IX86_BUILTIN_PADDSW128,
19623 IX86_BUILTIN_PADDUSB128,
19624 IX86_BUILTIN_PADDUSW128,
19625 IX86_BUILTIN_PSUBB128,
19626 IX86_BUILTIN_PSUBW128,
19627 IX86_BUILTIN_PSUBD128,
19628 IX86_BUILTIN_PSUBQ128,
19629 IX86_BUILTIN_PSUBSB128,
19630 IX86_BUILTIN_PSUBSW128,
19631 IX86_BUILTIN_PSUBUSB128,
19632 IX86_BUILTIN_PSUBUSW128,
19634 IX86_BUILTIN_PAND128,
19635 IX86_BUILTIN_PANDN128,
19636 IX86_BUILTIN_POR128,
19637 IX86_BUILTIN_PXOR128,
19639 IX86_BUILTIN_PAVGB128,
19640 IX86_BUILTIN_PAVGW128,
19642 IX86_BUILTIN_PCMPEQB128,
19643 IX86_BUILTIN_PCMPEQW128,
19644 IX86_BUILTIN_PCMPEQD128,
19645 IX86_BUILTIN_PCMPGTB128,
19646 IX86_BUILTIN_PCMPGTW128,
19647 IX86_BUILTIN_PCMPGTD128,
19649 IX86_BUILTIN_PMADDWD128,
19651 IX86_BUILTIN_PMAXSW128,
19652 IX86_BUILTIN_PMAXUB128,
19653 IX86_BUILTIN_PMINSW128,
19654 IX86_BUILTIN_PMINUB128,
19656 IX86_BUILTIN_PMULUDQ,
19657 IX86_BUILTIN_PMULUDQ128,
19658 IX86_BUILTIN_PMULHUW128,
19659 IX86_BUILTIN_PMULHW128,
19660 IX86_BUILTIN_PMULLW128,
19662 IX86_BUILTIN_PSADBW128,
19663 IX86_BUILTIN_PSHUFHW,
19664 IX86_BUILTIN_PSHUFLW,
19665 IX86_BUILTIN_PSHUFD,
19667 IX86_BUILTIN_PSLLDQI128,
19668 IX86_BUILTIN_PSLLWI128,
19669 IX86_BUILTIN_PSLLDI128,
19670 IX86_BUILTIN_PSLLQI128,
19671 IX86_BUILTIN_PSRAWI128,
19672 IX86_BUILTIN_PSRADI128,
19673 IX86_BUILTIN_PSRLDQI128,
19674 IX86_BUILTIN_PSRLWI128,
19675 IX86_BUILTIN_PSRLDI128,
19676 IX86_BUILTIN_PSRLQI128,
19678 IX86_BUILTIN_PSLLDQ128,
19679 IX86_BUILTIN_PSLLW128,
19680 IX86_BUILTIN_PSLLD128,
19681 IX86_BUILTIN_PSLLQ128,
19682 IX86_BUILTIN_PSRAW128,
19683 IX86_BUILTIN_PSRAD128,
19684 IX86_BUILTIN_PSRLW128,
19685 IX86_BUILTIN_PSRLD128,
19686 IX86_BUILTIN_PSRLQ128,
19688 IX86_BUILTIN_PUNPCKHBW128,
19689 IX86_BUILTIN_PUNPCKHWD128,
19690 IX86_BUILTIN_PUNPCKHDQ128,
19691 IX86_BUILTIN_PUNPCKHQDQ128,
19692 IX86_BUILTIN_PUNPCKLBW128,
19693 IX86_BUILTIN_PUNPCKLWD128,
19694 IX86_BUILTIN_PUNPCKLDQ128,
19695 IX86_BUILTIN_PUNPCKLQDQ128,
19697 IX86_BUILTIN_CLFLUSH,
19698 IX86_BUILTIN_MFENCE,
19699 IX86_BUILTIN_LFENCE,
19702 IX86_BUILTIN_ADDSUBPS,
19703 IX86_BUILTIN_HADDPS,
19704 IX86_BUILTIN_HSUBPS,
19705 IX86_BUILTIN_MOVSHDUP,
19706 IX86_BUILTIN_MOVSLDUP,
19707 IX86_BUILTIN_ADDSUBPD,
19708 IX86_BUILTIN_HADDPD,
19709 IX86_BUILTIN_HSUBPD,
19710 IX86_BUILTIN_LDDQU,
19712 IX86_BUILTIN_MONITOR,
19713 IX86_BUILTIN_MWAIT,
19716 IX86_BUILTIN_PHADDW,
19717 IX86_BUILTIN_PHADDD,
19718 IX86_BUILTIN_PHADDSW,
19719 IX86_BUILTIN_PHSUBW,
19720 IX86_BUILTIN_PHSUBD,
19721 IX86_BUILTIN_PHSUBSW,
19722 IX86_BUILTIN_PMADDUBSW,
19723 IX86_BUILTIN_PMULHRSW,
19724 IX86_BUILTIN_PSHUFB,
19725 IX86_BUILTIN_PSIGNB,
19726 IX86_BUILTIN_PSIGNW,
19727 IX86_BUILTIN_PSIGND,
19728 IX86_BUILTIN_PALIGNR,
19729 IX86_BUILTIN_PABSB,
19730 IX86_BUILTIN_PABSW,
19731 IX86_BUILTIN_PABSD,
19733 IX86_BUILTIN_PHADDW128,
19734 IX86_BUILTIN_PHADDD128,
19735 IX86_BUILTIN_PHADDSW128,
19736 IX86_BUILTIN_PHSUBW128,
19737 IX86_BUILTIN_PHSUBD128,
19738 IX86_BUILTIN_PHSUBSW128,
19739 IX86_BUILTIN_PMADDUBSW128,
19740 IX86_BUILTIN_PMULHRSW128,
19741 IX86_BUILTIN_PSHUFB128,
19742 IX86_BUILTIN_PSIGNB128,
19743 IX86_BUILTIN_PSIGNW128,
19744 IX86_BUILTIN_PSIGND128,
19745 IX86_BUILTIN_PALIGNR128,
19746 IX86_BUILTIN_PABSB128,
19747 IX86_BUILTIN_PABSW128,
19748 IX86_BUILTIN_PABSD128,
19750 /* AMDFAM10 - SSE4A New Instructions. */
19751 IX86_BUILTIN_MOVNTSD,
19752 IX86_BUILTIN_MOVNTSS,
19753 IX86_BUILTIN_EXTRQI,
19754 IX86_BUILTIN_EXTRQ,
19755 IX86_BUILTIN_INSERTQI,
19756 IX86_BUILTIN_INSERTQ,
19759 IX86_BUILTIN_BLENDPD,
19760 IX86_BUILTIN_BLENDPS,
19761 IX86_BUILTIN_BLENDVPD,
19762 IX86_BUILTIN_BLENDVPS,
19763 IX86_BUILTIN_PBLENDVB128,
19764 IX86_BUILTIN_PBLENDW128,
19769 IX86_BUILTIN_INSERTPS128,
19771 IX86_BUILTIN_MOVNTDQA,
19772 IX86_BUILTIN_MPSADBW128,
19773 IX86_BUILTIN_PACKUSDW128,
19774 IX86_BUILTIN_PCMPEQQ,
19775 IX86_BUILTIN_PHMINPOSUW128,
19777 IX86_BUILTIN_PMAXSB128,
19778 IX86_BUILTIN_PMAXSD128,
19779 IX86_BUILTIN_PMAXUD128,
19780 IX86_BUILTIN_PMAXUW128,
19782 IX86_BUILTIN_PMINSB128,
19783 IX86_BUILTIN_PMINSD128,
19784 IX86_BUILTIN_PMINUD128,
19785 IX86_BUILTIN_PMINUW128,
19787 IX86_BUILTIN_PMOVSXBW128,
19788 IX86_BUILTIN_PMOVSXBD128,
19789 IX86_BUILTIN_PMOVSXBQ128,
19790 IX86_BUILTIN_PMOVSXWD128,
19791 IX86_BUILTIN_PMOVSXWQ128,
19792 IX86_BUILTIN_PMOVSXDQ128,
19794 IX86_BUILTIN_PMOVZXBW128,
19795 IX86_BUILTIN_PMOVZXBD128,
19796 IX86_BUILTIN_PMOVZXBQ128,
19797 IX86_BUILTIN_PMOVZXWD128,
19798 IX86_BUILTIN_PMOVZXWQ128,
19799 IX86_BUILTIN_PMOVZXDQ128,
19801 IX86_BUILTIN_PMULDQ128,
19802 IX86_BUILTIN_PMULLD128,
19804 IX86_BUILTIN_ROUNDPD,
19805 IX86_BUILTIN_ROUNDPS,
19806 IX86_BUILTIN_ROUNDSD,
19807 IX86_BUILTIN_ROUNDSS,
19809 IX86_BUILTIN_PTESTZ,
19810 IX86_BUILTIN_PTESTC,
19811 IX86_BUILTIN_PTESTNZC,
19813 IX86_BUILTIN_VEC_INIT_V2SI,
19814 IX86_BUILTIN_VEC_INIT_V4HI,
19815 IX86_BUILTIN_VEC_INIT_V8QI,
19816 IX86_BUILTIN_VEC_EXT_V2DF,
19817 IX86_BUILTIN_VEC_EXT_V2DI,
19818 IX86_BUILTIN_VEC_EXT_V4SF,
19819 IX86_BUILTIN_VEC_EXT_V4SI,
19820 IX86_BUILTIN_VEC_EXT_V8HI,
19821 IX86_BUILTIN_VEC_EXT_V2SI,
19822 IX86_BUILTIN_VEC_EXT_V4HI,
19823 IX86_BUILTIN_VEC_EXT_V16QI,
19824 IX86_BUILTIN_VEC_SET_V2DI,
19825 IX86_BUILTIN_VEC_SET_V4SF,
19826 IX86_BUILTIN_VEC_SET_V4SI,
19827 IX86_BUILTIN_VEC_SET_V8HI,
19828 IX86_BUILTIN_VEC_SET_V4HI,
19829 IX86_BUILTIN_VEC_SET_V16QI,
19831 IX86_BUILTIN_VEC_PACK_SFIX,
19834 IX86_BUILTIN_CRC32QI,
19835 IX86_BUILTIN_CRC32HI,
19836 IX86_BUILTIN_CRC32SI,
19837 IX86_BUILTIN_CRC32DI,
19839 IX86_BUILTIN_PCMPESTRI128,
19840 IX86_BUILTIN_PCMPESTRM128,
19841 IX86_BUILTIN_PCMPESTRA128,
19842 IX86_BUILTIN_PCMPESTRC128,
19843 IX86_BUILTIN_PCMPESTRO128,
19844 IX86_BUILTIN_PCMPESTRS128,
19845 IX86_BUILTIN_PCMPESTRZ128,
19846 IX86_BUILTIN_PCMPISTRI128,
19847 IX86_BUILTIN_PCMPISTRM128,
19848 IX86_BUILTIN_PCMPISTRA128,
19849 IX86_BUILTIN_PCMPISTRC128,
19850 IX86_BUILTIN_PCMPISTRO128,
19851 IX86_BUILTIN_PCMPISTRS128,
19852 IX86_BUILTIN_PCMPISTRZ128,
19854 IX86_BUILTIN_PCMPGTQ,
19856 /* AES instructions */
19857 IX86_BUILTIN_AESENC128,
19858 IX86_BUILTIN_AESENCLAST128,
19859 IX86_BUILTIN_AESDEC128,
19860 IX86_BUILTIN_AESDECLAST128,
19861 IX86_BUILTIN_AESIMC128,
19862 IX86_BUILTIN_AESKEYGENASSIST128,
19864 /* PCLMUL instruction */
19865 IX86_BUILTIN_PCLMULQDQ128,
19868 IX86_BUILTIN_ADDPD256,
19869 IX86_BUILTIN_ADDPS256,
19870 IX86_BUILTIN_ADDSUBPD256,
19871 IX86_BUILTIN_ADDSUBPS256,
19872 IX86_BUILTIN_ANDPD256,
19873 IX86_BUILTIN_ANDPS256,
19874 IX86_BUILTIN_ANDNPD256,
19875 IX86_BUILTIN_ANDNPS256,
19876 IX86_BUILTIN_BLENDPD256,
19877 IX86_BUILTIN_BLENDPS256,
19878 IX86_BUILTIN_BLENDVPD256,
19879 IX86_BUILTIN_BLENDVPS256,
19880 IX86_BUILTIN_DIVPD256,
19881 IX86_BUILTIN_DIVPS256,
19882 IX86_BUILTIN_DPPS256,
19883 IX86_BUILTIN_HADDPD256,
19884 IX86_BUILTIN_HADDPS256,
19885 IX86_BUILTIN_HSUBPD256,
19886 IX86_BUILTIN_HSUBPS256,
19887 IX86_BUILTIN_MAXPD256,
19888 IX86_BUILTIN_MAXPS256,
19889 IX86_BUILTIN_MINPD256,
19890 IX86_BUILTIN_MINPS256,
19891 IX86_BUILTIN_MULPD256,
19892 IX86_BUILTIN_MULPS256,
19893 IX86_BUILTIN_ORPD256,
19894 IX86_BUILTIN_ORPS256,
19895 IX86_BUILTIN_SHUFPD256,
19896 IX86_BUILTIN_SHUFPS256,
19897 IX86_BUILTIN_SUBPD256,
19898 IX86_BUILTIN_SUBPS256,
19899 IX86_BUILTIN_XORPD256,
19900 IX86_BUILTIN_XORPS256,
19901 IX86_BUILTIN_CMPSD,
19902 IX86_BUILTIN_CMPSS,
19903 IX86_BUILTIN_CMPPD,
19904 IX86_BUILTIN_CMPPS,
19905 IX86_BUILTIN_CMPPD256,
19906 IX86_BUILTIN_CMPPS256,
19907 IX86_BUILTIN_CVTDQ2PD256,
19908 IX86_BUILTIN_CVTDQ2PS256,
19909 IX86_BUILTIN_CVTPD2PS256,
19910 IX86_BUILTIN_CVTPS2DQ256,
19911 IX86_BUILTIN_CVTPS2PD256,
19912 IX86_BUILTIN_CVTTPD2DQ256,
19913 IX86_BUILTIN_CVTPD2DQ256,
19914 IX86_BUILTIN_CVTTPS2DQ256,
19915 IX86_BUILTIN_EXTRACTF128PD256,
19916 IX86_BUILTIN_EXTRACTF128PS256,
19917 IX86_BUILTIN_EXTRACTF128SI256,
19918 IX86_BUILTIN_VZEROALL,
19919 IX86_BUILTIN_VZEROUPPER,
19920 IX86_BUILTIN_VZEROUPPER_REX64,
19921 IX86_BUILTIN_VPERMILVARPD,
19922 IX86_BUILTIN_VPERMILVARPS,
19923 IX86_BUILTIN_VPERMILVARPD256,
19924 IX86_BUILTIN_VPERMILVARPS256,
19925 IX86_BUILTIN_VPERMILPD,
19926 IX86_BUILTIN_VPERMILPS,
19927 IX86_BUILTIN_VPERMILPD256,
19928 IX86_BUILTIN_VPERMILPS256,
19929 IX86_BUILTIN_VPERM2F128PD256,
19930 IX86_BUILTIN_VPERM2F128PS256,
19931 IX86_BUILTIN_VPERM2F128SI256,
19932 IX86_BUILTIN_VBROADCASTSS,
19933 IX86_BUILTIN_VBROADCASTSD256,
19934 IX86_BUILTIN_VBROADCASTSS256,
19935 IX86_BUILTIN_VBROADCASTPD256,
19936 IX86_BUILTIN_VBROADCASTPS256,
19937 IX86_BUILTIN_VINSERTF128PD256,
19938 IX86_BUILTIN_VINSERTF128PS256,
19939 IX86_BUILTIN_VINSERTF128SI256,
19940 IX86_BUILTIN_LOADUPD256,
19941 IX86_BUILTIN_LOADUPS256,
19942 IX86_BUILTIN_STOREUPD256,
19943 IX86_BUILTIN_STOREUPS256,
19944 IX86_BUILTIN_LDDQU256,
19945 IX86_BUILTIN_MOVNTDQ256,
19946 IX86_BUILTIN_MOVNTPD256,
19947 IX86_BUILTIN_MOVNTPS256,
19948 IX86_BUILTIN_LOADDQU256,
19949 IX86_BUILTIN_STOREDQU256,
19950 IX86_BUILTIN_MASKLOADPD,
19951 IX86_BUILTIN_MASKLOADPS,
19952 IX86_BUILTIN_MASKSTOREPD,
19953 IX86_BUILTIN_MASKSTOREPS,
19954 IX86_BUILTIN_MASKLOADPD256,
19955 IX86_BUILTIN_MASKLOADPS256,
19956 IX86_BUILTIN_MASKSTOREPD256,
19957 IX86_BUILTIN_MASKSTOREPS256,
19958 IX86_BUILTIN_MOVSHDUP256,
19959 IX86_BUILTIN_MOVSLDUP256,
19960 IX86_BUILTIN_MOVDDUP256,
19962 IX86_BUILTIN_SQRTPD256,
19963 IX86_BUILTIN_SQRTPS256,
19964 IX86_BUILTIN_SQRTPS_NR256,
19965 IX86_BUILTIN_RSQRTPS256,
19966 IX86_BUILTIN_RSQRTPS_NR256,
19968 IX86_BUILTIN_RCPPS256,
19970 IX86_BUILTIN_ROUNDPD256,
19971 IX86_BUILTIN_ROUNDPS256,
19973 IX86_BUILTIN_UNPCKHPD256,
19974 IX86_BUILTIN_UNPCKLPD256,
19975 IX86_BUILTIN_UNPCKHPS256,
19976 IX86_BUILTIN_UNPCKLPS256,
19978 IX86_BUILTIN_SI256_SI,
19979 IX86_BUILTIN_PS256_PS,
19980 IX86_BUILTIN_PD256_PD,
19981 IX86_BUILTIN_SI_SI256,
19982 IX86_BUILTIN_PS_PS256,
19983 IX86_BUILTIN_PD_PD256,
19985 IX86_BUILTIN_VTESTZPD,
19986 IX86_BUILTIN_VTESTCPD,
19987 IX86_BUILTIN_VTESTNZCPD,
19988 IX86_BUILTIN_VTESTZPS,
19989 IX86_BUILTIN_VTESTCPS,
19990 IX86_BUILTIN_VTESTNZCPS,
19991 IX86_BUILTIN_VTESTZPD256,
19992 IX86_BUILTIN_VTESTCPD256,
19993 IX86_BUILTIN_VTESTNZCPD256,
19994 IX86_BUILTIN_VTESTZPS256,
19995 IX86_BUILTIN_VTESTCPS256,
19996 IX86_BUILTIN_VTESTNZCPS256,
19997 IX86_BUILTIN_PTESTZ256,
19998 IX86_BUILTIN_PTESTC256,
19999 IX86_BUILTIN_PTESTNZC256,
20001 IX86_BUILTIN_MOVMSKPD256,
20002 IX86_BUILTIN_MOVMSKPS256,
20004 /* TFmode support builtins. */
20006 IX86_BUILTIN_FABSQ,
20007 IX86_BUILTIN_COPYSIGNQ,
20009 /* SSE5 instructions */
20010 IX86_BUILTIN_FMADDSS,
20011 IX86_BUILTIN_FMADDSD,
20012 IX86_BUILTIN_FMADDPS,
20013 IX86_BUILTIN_FMADDPD,
20014 IX86_BUILTIN_FMSUBSS,
20015 IX86_BUILTIN_FMSUBSD,
20016 IX86_BUILTIN_FMSUBPS,
20017 IX86_BUILTIN_FMSUBPD,
20018 IX86_BUILTIN_FNMADDSS,
20019 IX86_BUILTIN_FNMADDSD,
20020 IX86_BUILTIN_FNMADDPS,
20021 IX86_BUILTIN_FNMADDPD,
20022 IX86_BUILTIN_FNMSUBSS,
20023 IX86_BUILTIN_FNMSUBSD,
20024 IX86_BUILTIN_FNMSUBPS,
20025 IX86_BUILTIN_FNMSUBPD,
20026 IX86_BUILTIN_PCMOV,
20027 IX86_BUILTIN_PCMOV_V2DI,
20028 IX86_BUILTIN_PCMOV_V4SI,
20029 IX86_BUILTIN_PCMOV_V8HI,
20030 IX86_BUILTIN_PCMOV_V16QI,
20031 IX86_BUILTIN_PCMOV_V4SF,
20032 IX86_BUILTIN_PCMOV_V2DF,
20033 IX86_BUILTIN_PPERM,
20034 IX86_BUILTIN_PERMPS,
20035 IX86_BUILTIN_PERMPD,
20036 IX86_BUILTIN_PMACSSWW,
20037 IX86_BUILTIN_PMACSWW,
20038 IX86_BUILTIN_PMACSSWD,
20039 IX86_BUILTIN_PMACSWD,
20040 IX86_BUILTIN_PMACSSDD,
20041 IX86_BUILTIN_PMACSDD,
20042 IX86_BUILTIN_PMACSSDQL,
20043 IX86_BUILTIN_PMACSSDQH,
20044 IX86_BUILTIN_PMACSDQL,
20045 IX86_BUILTIN_PMACSDQH,
20046 IX86_BUILTIN_PMADCSSWD,
20047 IX86_BUILTIN_PMADCSWD,
20048 IX86_BUILTIN_PHADDBW,
20049 IX86_BUILTIN_PHADDBD,
20050 IX86_BUILTIN_PHADDBQ,
20051 IX86_BUILTIN_PHADDWD,
20052 IX86_BUILTIN_PHADDWQ,
20053 IX86_BUILTIN_PHADDDQ,
20054 IX86_BUILTIN_PHADDUBW,
20055 IX86_BUILTIN_PHADDUBD,
20056 IX86_BUILTIN_PHADDUBQ,
20057 IX86_BUILTIN_PHADDUWD,
20058 IX86_BUILTIN_PHADDUWQ,
20059 IX86_BUILTIN_PHADDUDQ,
20060 IX86_BUILTIN_PHSUBBW,
20061 IX86_BUILTIN_PHSUBWD,
20062 IX86_BUILTIN_PHSUBDQ,
20063 IX86_BUILTIN_PROTB,
20064 IX86_BUILTIN_PROTW,
20065 IX86_BUILTIN_PROTD,
20066 IX86_BUILTIN_PROTQ,
20067 IX86_BUILTIN_PROTB_IMM,
20068 IX86_BUILTIN_PROTW_IMM,
20069 IX86_BUILTIN_PROTD_IMM,
20070 IX86_BUILTIN_PROTQ_IMM,
20071 IX86_BUILTIN_PSHLB,
20072 IX86_BUILTIN_PSHLW,
20073 IX86_BUILTIN_PSHLD,
20074 IX86_BUILTIN_PSHLQ,
20075 IX86_BUILTIN_PSHAB,
20076 IX86_BUILTIN_PSHAW,
20077 IX86_BUILTIN_PSHAD,
20078 IX86_BUILTIN_PSHAQ,
20079 IX86_BUILTIN_FRCZSS,
20080 IX86_BUILTIN_FRCZSD,
20081 IX86_BUILTIN_FRCZPS,
20082 IX86_BUILTIN_FRCZPD,
20083 IX86_BUILTIN_CVTPH2PS,
20084 IX86_BUILTIN_CVTPS2PH,
20086 IX86_BUILTIN_COMEQSS,
20087 IX86_BUILTIN_COMNESS,
20088 IX86_BUILTIN_COMLTSS,
20089 IX86_BUILTIN_COMLESS,
20090 IX86_BUILTIN_COMGTSS,
20091 IX86_BUILTIN_COMGESS,
20092 IX86_BUILTIN_COMUEQSS,
20093 IX86_BUILTIN_COMUNESS,
20094 IX86_BUILTIN_COMULTSS,
20095 IX86_BUILTIN_COMULESS,
20096 IX86_BUILTIN_COMUGTSS,
20097 IX86_BUILTIN_COMUGESS,
20098 IX86_BUILTIN_COMORDSS,
20099 IX86_BUILTIN_COMUNORDSS,
20100 IX86_BUILTIN_COMFALSESS,
20101 IX86_BUILTIN_COMTRUESS,
20103 IX86_BUILTIN_COMEQSD,
20104 IX86_BUILTIN_COMNESD,
20105 IX86_BUILTIN_COMLTSD,
20106 IX86_BUILTIN_COMLESD,
20107 IX86_BUILTIN_COMGTSD,
20108 IX86_BUILTIN_COMGESD,
20109 IX86_BUILTIN_COMUEQSD,
20110 IX86_BUILTIN_COMUNESD,
20111 IX86_BUILTIN_COMULTSD,
20112 IX86_BUILTIN_COMULESD,
20113 IX86_BUILTIN_COMUGTSD,
20114 IX86_BUILTIN_COMUGESD,
20115 IX86_BUILTIN_COMORDSD,
20116 IX86_BUILTIN_COMUNORDSD,
20117 IX86_BUILTIN_COMFALSESD,
20118 IX86_BUILTIN_COMTRUESD,
20120 IX86_BUILTIN_COMEQPS,
20121 IX86_BUILTIN_COMNEPS,
20122 IX86_BUILTIN_COMLTPS,
20123 IX86_BUILTIN_COMLEPS,
20124 IX86_BUILTIN_COMGTPS,
20125 IX86_BUILTIN_COMGEPS,
20126 IX86_BUILTIN_COMUEQPS,
20127 IX86_BUILTIN_COMUNEPS,
20128 IX86_BUILTIN_COMULTPS,
20129 IX86_BUILTIN_COMULEPS,
20130 IX86_BUILTIN_COMUGTPS,
20131 IX86_BUILTIN_COMUGEPS,
20132 IX86_BUILTIN_COMORDPS,
20133 IX86_BUILTIN_COMUNORDPS,
20134 IX86_BUILTIN_COMFALSEPS,
20135 IX86_BUILTIN_COMTRUEPS,
20137 IX86_BUILTIN_COMEQPD,
20138 IX86_BUILTIN_COMNEPD,
20139 IX86_BUILTIN_COMLTPD,
20140 IX86_BUILTIN_COMLEPD,
20141 IX86_BUILTIN_COMGTPD,
20142 IX86_BUILTIN_COMGEPD,
20143 IX86_BUILTIN_COMUEQPD,
20144 IX86_BUILTIN_COMUNEPD,
20145 IX86_BUILTIN_COMULTPD,
20146 IX86_BUILTIN_COMULEPD,
20147 IX86_BUILTIN_COMUGTPD,
20148 IX86_BUILTIN_COMUGEPD,
20149 IX86_BUILTIN_COMORDPD,
20150 IX86_BUILTIN_COMUNORDPD,
20151 IX86_BUILTIN_COMFALSEPD,
20152 IX86_BUILTIN_COMTRUEPD,
20154 IX86_BUILTIN_PCOMEQUB,
20155 IX86_BUILTIN_PCOMNEUB,
20156 IX86_BUILTIN_PCOMLTUB,
20157 IX86_BUILTIN_PCOMLEUB,
20158 IX86_BUILTIN_PCOMGTUB,
20159 IX86_BUILTIN_PCOMGEUB,
20160 IX86_BUILTIN_PCOMFALSEUB,
20161 IX86_BUILTIN_PCOMTRUEUB,
20162 IX86_BUILTIN_PCOMEQUW,
20163 IX86_BUILTIN_PCOMNEUW,
20164 IX86_BUILTIN_PCOMLTUW,
20165 IX86_BUILTIN_PCOMLEUW,
20166 IX86_BUILTIN_PCOMGTUW,
20167 IX86_BUILTIN_PCOMGEUW,
20168 IX86_BUILTIN_PCOMFALSEUW,
20169 IX86_BUILTIN_PCOMTRUEUW,
20170 IX86_BUILTIN_PCOMEQUD,
20171 IX86_BUILTIN_PCOMNEUD,
20172 IX86_BUILTIN_PCOMLTUD,
20173 IX86_BUILTIN_PCOMLEUD,
20174 IX86_BUILTIN_PCOMGTUD,
20175 IX86_BUILTIN_PCOMGEUD,
20176 IX86_BUILTIN_PCOMFALSEUD,
20177 IX86_BUILTIN_PCOMTRUEUD,
20178 IX86_BUILTIN_PCOMEQUQ,
20179 IX86_BUILTIN_PCOMNEUQ,
20180 IX86_BUILTIN_PCOMLTUQ,
20181 IX86_BUILTIN_PCOMLEUQ,
20182 IX86_BUILTIN_PCOMGTUQ,
20183 IX86_BUILTIN_PCOMGEUQ,
20184 IX86_BUILTIN_PCOMFALSEUQ,
20185 IX86_BUILTIN_PCOMTRUEUQ,
20187 IX86_BUILTIN_PCOMEQB,
20188 IX86_BUILTIN_PCOMNEB,
20189 IX86_BUILTIN_PCOMLTB,
20190 IX86_BUILTIN_PCOMLEB,
20191 IX86_BUILTIN_PCOMGTB,
20192 IX86_BUILTIN_PCOMGEB,
20193 IX86_BUILTIN_PCOMFALSEB,
20194 IX86_BUILTIN_PCOMTRUEB,
20195 IX86_BUILTIN_PCOMEQW,
20196 IX86_BUILTIN_PCOMNEW,
20197 IX86_BUILTIN_PCOMLTW,
20198 IX86_BUILTIN_PCOMLEW,
20199 IX86_BUILTIN_PCOMGTW,
20200 IX86_BUILTIN_PCOMGEW,
20201 IX86_BUILTIN_PCOMFALSEW,
20202 IX86_BUILTIN_PCOMTRUEW,
20203 IX86_BUILTIN_PCOMEQD,
20204 IX86_BUILTIN_PCOMNED,
20205 IX86_BUILTIN_PCOMLTD,
20206 IX86_BUILTIN_PCOMLED,
20207 IX86_BUILTIN_PCOMGTD,
20208 IX86_BUILTIN_PCOMGED,
20209 IX86_BUILTIN_PCOMFALSED,
20210 IX86_BUILTIN_PCOMTRUED,
20211 IX86_BUILTIN_PCOMEQQ,
20212 IX86_BUILTIN_PCOMNEQ,
20213 IX86_BUILTIN_PCOMLTQ,
20214 IX86_BUILTIN_PCOMLEQ,
20215 IX86_BUILTIN_PCOMGTQ,
20216 IX86_BUILTIN_PCOMGEQ,
20217 IX86_BUILTIN_PCOMFALSEQ,
20218 IX86_BUILTIN_PCOMTRUEQ,
20223 /* Table for the ix86 builtin decls. */
20224 static GTY(()) tree ix86_builtins[(int) IX86_BUILTIN_MAX];
20226 /* Table of all of the builtin functions that are possible with different ISA's
20227 but are waiting to be built until a function is declared to use that
20229 struct builtin_isa GTY(())
20231 tree type; /* builtin type to use in the declaration */
20232 const char *name; /* function name */
20233 int isa; /* isa_flags this builtin is defined for */
20234 bool const_p; /* true if the declaration is constant */
20237 static GTY(()) struct builtin_isa ix86_builtins_isa[(int) IX86_BUILTIN_MAX];
20240 /* Add an ix86 target builtin function with CODE, NAME and TYPE. Save the MASK
20241 * of which isa_flags to use in the ix86_builtins_isa array. Stores the
20242 * function decl in the ix86_builtins array. Returns the function decl or
20243 * NULL_TREE, if the builtin was not added.
20245 * If the front end has a special hook for builtin functions, delay adding
20246 * builtin functions that aren't in the current ISA until the ISA is changed
20247 * with function specific optimization. Doing so, can save about 300K for the
20248 * default compiler. When the builtin is expanded, check at that time whether
20251 * If the front end doesn't have a special hook, record all builtins, even if
20252 * it isn't an instruction set in the current ISA in case the user uses
20253 * function specific options for a different ISA, so that we don't get scope
20254 * errors if a builtin is added in the middle of a function scope. */
20257 def_builtin (int mask, const char *name, tree type, enum ix86_builtins code)
20259 tree decl = NULL_TREE;
20261 if (!(mask & OPTION_MASK_ISA_64BIT) || TARGET_64BIT)
20263 ix86_builtins_isa[(int) code].isa = mask;
20265 if ((mask & ix86_isa_flags) != 0
20266 || (lang_hooks.builtin_function
20267 == lang_hooks.builtin_function_ext_scope))
20270 decl = add_builtin_function (name, type, code, BUILT_IN_MD, NULL,
20272 ix86_builtins[(int) code] = decl;
20273 ix86_builtins_isa[(int) code].type = NULL_TREE;
20277 ix86_builtins[(int) code] = NULL_TREE;
20278 ix86_builtins_isa[(int) code].const_p = false;
20279 ix86_builtins_isa[(int) code].type = type;
20280 ix86_builtins_isa[(int) code].name = name;
20287 /* Like def_builtin, but also marks the function decl "const". */
20290 def_builtin_const (int mask, const char *name, tree type,
20291 enum ix86_builtins code)
20293 tree decl = def_builtin (mask, name, type, code);
20295 TREE_READONLY (decl) = 1;
20297 ix86_builtins_isa[(int) code].const_p = true;
20302 /* Add any new builtin functions for a given ISA that may not have been
20303 declared. This saves a bit of space compared to adding all of the
20304 declarations to the tree, even if we didn't use them. */
20307 ix86_add_new_builtins (int isa)
20312 for (i = 0; i < (int)IX86_BUILTIN_MAX; i++)
20314 if ((ix86_builtins_isa[i].isa & isa) != 0
20315 && ix86_builtins_isa[i].type != NULL_TREE)
20317 decl = add_builtin_function_ext_scope (ix86_builtins_isa[i].name,
20318 ix86_builtins_isa[i].type,
20319 i, BUILT_IN_MD, NULL,
20322 ix86_builtins[i] = decl;
20323 ix86_builtins_isa[i].type = NULL_TREE;
20324 if (ix86_builtins_isa[i].const_p)
20325 TREE_READONLY (decl) = 1;
20330 /* Bits for builtin_description.flag. */
20332 /* Set when we don't support the comparison natively, and should
20333 swap_comparison in order to support it. */
20334 #define BUILTIN_DESC_SWAP_OPERANDS 1
20336 struct builtin_description
20338 const unsigned int mask;
20339 const enum insn_code icode;
20340 const char *const name;
20341 const enum ix86_builtins code;
20342 const enum rtx_code comparison;
20346 static const struct builtin_description bdesc_comi[] =
20348 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comieq", IX86_BUILTIN_COMIEQSS, UNEQ, 0 },
20349 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comilt", IX86_BUILTIN_COMILTSS, UNLT, 0 },
20350 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comile", IX86_BUILTIN_COMILESS, UNLE, 0 },
20351 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comigt", IX86_BUILTIN_COMIGTSS, GT, 0 },
20352 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comige", IX86_BUILTIN_COMIGESS, GE, 0 },
20353 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_comi, "__builtin_ia32_comineq", IX86_BUILTIN_COMINEQSS, LTGT, 0 },
20354 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomieq", IX86_BUILTIN_UCOMIEQSS, UNEQ, 0 },
20355 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomilt", IX86_BUILTIN_UCOMILTSS, UNLT, 0 },
20356 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomile", IX86_BUILTIN_UCOMILESS, UNLE, 0 },
20357 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomigt", IX86_BUILTIN_UCOMIGTSS, GT, 0 },
20358 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomige", IX86_BUILTIN_UCOMIGESS, GE, 0 },
20359 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_ucomi, "__builtin_ia32_ucomineq", IX86_BUILTIN_UCOMINEQSS, LTGT, 0 },
20360 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdeq", IX86_BUILTIN_COMIEQSD, UNEQ, 0 },
20361 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdlt", IX86_BUILTIN_COMILTSD, UNLT, 0 },
20362 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdle", IX86_BUILTIN_COMILESD, UNLE, 0 },
20363 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdgt", IX86_BUILTIN_COMIGTSD, GT, 0 },
20364 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdge", IX86_BUILTIN_COMIGESD, GE, 0 },
20365 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_comi, "__builtin_ia32_comisdneq", IX86_BUILTIN_COMINEQSD, LTGT, 0 },
20366 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdeq", IX86_BUILTIN_UCOMIEQSD, UNEQ, 0 },
20367 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdlt", IX86_BUILTIN_UCOMILTSD, UNLT, 0 },
20368 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdle", IX86_BUILTIN_UCOMILESD, UNLE, 0 },
20369 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdgt", IX86_BUILTIN_UCOMIGTSD, GT, 0 },
20370 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdge", IX86_BUILTIN_UCOMIGESD, GE, 0 },
20371 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ucomi, "__builtin_ia32_ucomisdneq", IX86_BUILTIN_UCOMINEQSD, LTGT, 0 },
20374 static const struct builtin_description bdesc_pcmpestr[] =
20377 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestri128", IX86_BUILTIN_PCMPESTRI128, UNKNOWN, 0 },
20378 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestrm128", IX86_BUILTIN_PCMPESTRM128, UNKNOWN, 0 },
20379 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestria128", IX86_BUILTIN_PCMPESTRA128, UNKNOWN, (int) CCAmode },
20380 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestric128", IX86_BUILTIN_PCMPESTRC128, UNKNOWN, (int) CCCmode },
20381 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestrio128", IX86_BUILTIN_PCMPESTRO128, UNKNOWN, (int) CCOmode },
20382 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestris128", IX86_BUILTIN_PCMPESTRS128, UNKNOWN, (int) CCSmode },
20383 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpestr, "__builtin_ia32_pcmpestriz128", IX86_BUILTIN_PCMPESTRZ128, UNKNOWN, (int) CCZmode },
20386 static const struct builtin_description bdesc_pcmpistr[] =
20389 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistri128", IX86_BUILTIN_PCMPISTRI128, UNKNOWN, 0 },
20390 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistrm128", IX86_BUILTIN_PCMPISTRM128, UNKNOWN, 0 },
20391 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistria128", IX86_BUILTIN_PCMPISTRA128, UNKNOWN, (int) CCAmode },
20392 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistric128", IX86_BUILTIN_PCMPISTRC128, UNKNOWN, (int) CCCmode },
20393 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistrio128", IX86_BUILTIN_PCMPISTRO128, UNKNOWN, (int) CCOmode },
20394 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistris128", IX86_BUILTIN_PCMPISTRS128, UNKNOWN, (int) CCSmode },
20395 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_pcmpistr, "__builtin_ia32_pcmpistriz128", IX86_BUILTIN_PCMPISTRZ128, UNKNOWN, (int) CCZmode },
20398 /* Special builtin types */
20399 enum ix86_special_builtin_type
20401 SPECIAL_FTYPE_UNKNOWN,
20403 V32QI_FTYPE_PCCHAR,
20404 V16QI_FTYPE_PCCHAR,
20406 V8SF_FTYPE_PCFLOAT,
20408 V4DF_FTYPE_PCDOUBLE,
20409 V4SF_FTYPE_PCFLOAT,
20410 V2DF_FTYPE_PCDOUBLE,
20411 V8SF_FTYPE_PCV8SF_V8SF,
20412 V4DF_FTYPE_PCV4DF_V4DF,
20413 V4SF_FTYPE_V4SF_PCV2SF,
20414 V4SF_FTYPE_PCV4SF_V4SF,
20415 V2DF_FTYPE_V2DF_PCDOUBLE,
20416 V2DF_FTYPE_PCV2DF_V2DF,
20418 VOID_FTYPE_PV2SF_V4SF,
20419 VOID_FTYPE_PV4DI_V4DI,
20420 VOID_FTYPE_PV2DI_V2DI,
20421 VOID_FTYPE_PCHAR_V32QI,
20422 VOID_FTYPE_PCHAR_V16QI,
20423 VOID_FTYPE_PFLOAT_V8SF,
20424 VOID_FTYPE_PFLOAT_V4SF,
20425 VOID_FTYPE_PDOUBLE_V4DF,
20426 VOID_FTYPE_PDOUBLE_V2DF,
20428 VOID_FTYPE_PINT_INT,
20429 VOID_FTYPE_PV8SF_V8SF_V8SF,
20430 VOID_FTYPE_PV4DF_V4DF_V4DF,
20431 VOID_FTYPE_PV4SF_V4SF_V4SF,
20432 VOID_FTYPE_PV2DF_V2DF_V2DF
20435 /* Builtin types */
20436 enum ix86_builtin_type
20439 FLOAT128_FTYPE_FLOAT128,
20441 FLOAT128_FTYPE_FLOAT128_FLOAT128,
20442 INT_FTYPE_V8SF_V8SF_PTEST,
20443 INT_FTYPE_V4DI_V4DI_PTEST,
20444 INT_FTYPE_V4DF_V4DF_PTEST,
20445 INT_FTYPE_V4SF_V4SF_PTEST,
20446 INT_FTYPE_V2DI_V2DI_PTEST,
20447 INT_FTYPE_V2DF_V2DF_PTEST,
20479 V4SF_FTYPE_V4SF_VEC_MERGE,
20488 V2DF_FTYPE_V2DF_VEC_MERGE,
20499 V16QI_FTYPE_V16QI_V16QI,
20500 V16QI_FTYPE_V8HI_V8HI,
20501 V8QI_FTYPE_V8QI_V8QI,
20502 V8QI_FTYPE_V4HI_V4HI,
20503 V8HI_FTYPE_V8HI_V8HI,
20504 V8HI_FTYPE_V8HI_V8HI_COUNT,
20505 V8HI_FTYPE_V16QI_V16QI,
20506 V8HI_FTYPE_V4SI_V4SI,
20507 V8HI_FTYPE_V8HI_SI_COUNT,
20508 V8SF_FTYPE_V8SF_V8SF,
20509 V8SF_FTYPE_V8SF_V8SI,
20510 V4SI_FTYPE_V4SI_V4SI,
20511 V4SI_FTYPE_V4SI_V4SI_COUNT,
20512 V4SI_FTYPE_V8HI_V8HI,
20513 V4SI_FTYPE_V4SF_V4SF,
20514 V4SI_FTYPE_V2DF_V2DF,
20515 V4SI_FTYPE_V4SI_SI_COUNT,
20516 V4HI_FTYPE_V4HI_V4HI,
20517 V4HI_FTYPE_V4HI_V4HI_COUNT,
20518 V4HI_FTYPE_V8QI_V8QI,
20519 V4HI_FTYPE_V2SI_V2SI,
20520 V4HI_FTYPE_V4HI_SI_COUNT,
20521 V4DF_FTYPE_V4DF_V4DF,
20522 V4DF_FTYPE_V4DF_V4DI,
20523 V4SF_FTYPE_V4SF_V4SF,
20524 V4SF_FTYPE_V4SF_V4SF_SWAP,
20525 V4SF_FTYPE_V4SF_V4SI,
20526 V4SF_FTYPE_V4SF_V2SI,
20527 V4SF_FTYPE_V4SF_V2DF,
20528 V4SF_FTYPE_V4SF_DI,
20529 V4SF_FTYPE_V4SF_SI,
20530 V2DI_FTYPE_V2DI_V2DI,
20531 V2DI_FTYPE_V2DI_V2DI_COUNT,
20532 V2DI_FTYPE_V16QI_V16QI,
20533 V2DI_FTYPE_V4SI_V4SI,
20534 V2DI_FTYPE_V2DI_V16QI,
20535 V2DI_FTYPE_V2DF_V2DF,
20536 V2DI_FTYPE_V2DI_SI_COUNT,
20537 V2SI_FTYPE_V2SI_V2SI,
20538 V2SI_FTYPE_V2SI_V2SI_COUNT,
20539 V2SI_FTYPE_V4HI_V4HI,
20540 V2SI_FTYPE_V2SF_V2SF,
20541 V2SI_FTYPE_V2SI_SI_COUNT,
20542 V2DF_FTYPE_V2DF_V2DF,
20543 V2DF_FTYPE_V2DF_V2DF_SWAP,
20544 V2DF_FTYPE_V2DF_V4SF,
20545 V2DF_FTYPE_V2DF_V2DI,
20546 V2DF_FTYPE_V2DF_DI,
20547 V2DF_FTYPE_V2DF_SI,
20548 V2SF_FTYPE_V2SF_V2SF,
20549 V1DI_FTYPE_V1DI_V1DI,
20550 V1DI_FTYPE_V1DI_V1DI_COUNT,
20551 V1DI_FTYPE_V8QI_V8QI,
20552 V1DI_FTYPE_V2SI_V2SI,
20553 V1DI_FTYPE_V1DI_SI_COUNT,
20554 UINT64_FTYPE_UINT64_UINT64,
20555 UINT_FTYPE_UINT_UINT,
20556 UINT_FTYPE_UINT_USHORT,
20557 UINT_FTYPE_UINT_UCHAR,
20558 V8HI_FTYPE_V8HI_INT,
20559 V4SI_FTYPE_V4SI_INT,
20560 V4HI_FTYPE_V4HI_INT,
20561 V8SF_FTYPE_V8SF_INT,
20562 V4SI_FTYPE_V8SI_INT,
20563 V4SF_FTYPE_V8SF_INT,
20564 V2DF_FTYPE_V4DF_INT,
20565 V4DF_FTYPE_V4DF_INT,
20566 V4SF_FTYPE_V4SF_INT,
20567 V2DI_FTYPE_V2DI_INT,
20568 V2DI2TI_FTYPE_V2DI_INT,
20569 V2DF_FTYPE_V2DF_INT,
20570 V16QI_FTYPE_V16QI_V16QI_V16QI,
20571 V8SF_FTYPE_V8SF_V8SF_V8SF,
20572 V4DF_FTYPE_V4DF_V4DF_V4DF,
20573 V4SF_FTYPE_V4SF_V4SF_V4SF,
20574 V2DF_FTYPE_V2DF_V2DF_V2DF,
20575 V16QI_FTYPE_V16QI_V16QI_INT,
20576 V8SI_FTYPE_V8SI_V8SI_INT,
20577 V8SI_FTYPE_V8SI_V4SI_INT,
20578 V8HI_FTYPE_V8HI_V8HI_INT,
20579 V8SF_FTYPE_V8SF_V8SF_INT,
20580 V8SF_FTYPE_V8SF_V4SF_INT,
20581 V4SI_FTYPE_V4SI_V4SI_INT,
20582 V4DF_FTYPE_V4DF_V4DF_INT,
20583 V4DF_FTYPE_V4DF_V2DF_INT,
20584 V4SF_FTYPE_V4SF_V4SF_INT,
20585 V2DI_FTYPE_V2DI_V2DI_INT,
20586 V2DI2TI_FTYPE_V2DI_V2DI_INT,
20587 V1DI2DI_FTYPE_V1DI_V1DI_INT,
20588 V2DF_FTYPE_V2DF_V2DF_INT,
20589 V2DI_FTYPE_V2DI_UINT_UINT,
20590 V2DI_FTYPE_V2DI_V2DI_UINT_UINT
20593 /* Special builtins with variable number of arguments. */
20594 static const struct builtin_description bdesc_special_args[] =
20597 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_emms, "__builtin_ia32_emms", IX86_BUILTIN_EMMS, UNKNOWN, (int) VOID_FTYPE_VOID },
20600 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_femms, "__builtin_ia32_femms", IX86_BUILTIN_FEMMS, UNKNOWN, (int) VOID_FTYPE_VOID },
20603 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movups, "__builtin_ia32_storeups", IX86_BUILTIN_STOREUPS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
20604 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movntv4sf, "__builtin_ia32_movntps", IX86_BUILTIN_MOVNTPS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
20605 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movups, "__builtin_ia32_loadups", IX86_BUILTIN_LOADUPS, UNKNOWN, (int) V4SF_FTYPE_PCFLOAT },
20607 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_loadhps_exp, "__builtin_ia32_loadhps", IX86_BUILTIN_LOADHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_PCV2SF },
20608 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_loadlps_exp, "__builtin_ia32_loadlps", IX86_BUILTIN_LOADLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_PCV2SF },
20609 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_storehps, "__builtin_ia32_storehps", IX86_BUILTIN_STOREHPS, UNKNOWN, (int) VOID_FTYPE_PV2SF_V4SF },
20610 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_storelps, "__builtin_ia32_storelps", IX86_BUILTIN_STORELPS, UNKNOWN, (int) VOID_FTYPE_PV2SF_V4SF },
20612 /* SSE or 3DNow!A */
20613 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_sse_sfence, "__builtin_ia32_sfence", IX86_BUILTIN_SFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
20614 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_sse_movntdi, "__builtin_ia32_movntq", IX86_BUILTIN_MOVNTQ, UNKNOWN, (int) VOID_FTYPE_PDI_DI },
20617 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_lfence, "__builtin_ia32_lfence", IX86_BUILTIN_LFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
20618 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_mfence, 0, IX86_BUILTIN_MFENCE, UNKNOWN, (int) VOID_FTYPE_VOID },
20619 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movupd, "__builtin_ia32_storeupd", IX86_BUILTIN_STOREUPD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
20620 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movdqu, "__builtin_ia32_storedqu", IX86_BUILTIN_STOREDQU, UNKNOWN, (int) VOID_FTYPE_PCHAR_V16QI },
20621 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntv2df, "__builtin_ia32_movntpd", IX86_BUILTIN_MOVNTPD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
20622 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntv2di, "__builtin_ia32_movntdq", IX86_BUILTIN_MOVNTDQ, UNKNOWN, (int) VOID_FTYPE_PV2DI_V2DI },
20623 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movntsi, "__builtin_ia32_movnti", IX86_BUILTIN_MOVNTI, UNKNOWN, (int) VOID_FTYPE_PINT_INT },
20624 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movupd, "__builtin_ia32_loadupd", IX86_BUILTIN_LOADUPD, UNKNOWN, (int) V2DF_FTYPE_PCDOUBLE },
20625 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movdqu, "__builtin_ia32_loaddqu", IX86_BUILTIN_LOADDQU, UNKNOWN, (int) V16QI_FTYPE_PCCHAR },
20627 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_loadhpd_exp, "__builtin_ia32_loadhpd", IX86_BUILTIN_LOADHPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_PCDOUBLE },
20628 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_loadlpd_exp, "__builtin_ia32_loadlpd", IX86_BUILTIN_LOADLPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_PCDOUBLE },
20631 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_lddqu, "__builtin_ia32_lddqu", IX86_BUILTIN_LDDQU, UNKNOWN, (int) V16QI_FTYPE_PCCHAR },
20634 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_movntdqa, "__builtin_ia32_movntdqa", IX86_BUILTIN_MOVNTDQA, UNKNOWN, (int) V2DI_FTYPE_PV2DI },
20637 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_vmmovntv2df, "__builtin_ia32_movntsd", IX86_BUILTIN_MOVNTSD, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V2DF },
20638 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_vmmovntv4sf, "__builtin_ia32_movntss", IX86_BUILTIN_MOVNTSS, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V4SF },
20641 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vzeroall, "__builtin_ia32_vzeroall", IX86_BUILTIN_VZEROALL, UNKNOWN, (int) VOID_FTYPE_VOID },
20642 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vzeroupper, 0, IX86_BUILTIN_VZEROUPPER, UNKNOWN, (int) VOID_FTYPE_VOID },
20643 { OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_64BIT, CODE_FOR_avx_vzeroupper_rex64, 0, IX86_BUILTIN_VZEROUPPER_REX64, UNKNOWN, (int) VOID_FTYPE_VOID },
20645 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastss, "__builtin_ia32_vbroadcastss", IX86_BUILTIN_VBROADCASTSS, UNKNOWN, (int) V4SF_FTYPE_PCFLOAT },
20646 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastsd256, "__builtin_ia32_vbroadcastsd256", IX86_BUILTIN_VBROADCASTSD256, UNKNOWN, (int) V4DF_FTYPE_PCDOUBLE },
20647 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastss256, "__builtin_ia32_vbroadcastss256", IX86_BUILTIN_VBROADCASTSS256, UNKNOWN, (int) V8SF_FTYPE_PCFLOAT },
20648 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastf128_pd256, "__builtin_ia32_vbroadcastf128_pd256", IX86_BUILTIN_VBROADCASTPD256, UNKNOWN, (int) V4DF_FTYPE_PCV2DF },
20649 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vbroadcastf128_ps256, "__builtin_ia32_vbroadcastf128_ps256", IX86_BUILTIN_VBROADCASTPS256, UNKNOWN, (int) V8SF_FTYPE_PCV4SF },
20651 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movupd256, "__builtin_ia32_loadupd256", IX86_BUILTIN_LOADUPD256, UNKNOWN, (int) V4DF_FTYPE_PCDOUBLE },
20652 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movups256, "__builtin_ia32_loadups256", IX86_BUILTIN_LOADUPS256, UNKNOWN, (int) V8SF_FTYPE_PCFLOAT },
20653 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movupd256, "__builtin_ia32_storeupd256", IX86_BUILTIN_STOREUPD256, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V4DF },
20654 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movups256, "__builtin_ia32_storeups256", IX86_BUILTIN_STOREUPS256, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V8SF },
20655 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movdqu256, "__builtin_ia32_loaddqu256", IX86_BUILTIN_LOADDQU256, UNKNOWN, (int) V32QI_FTYPE_PCCHAR },
20656 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movdqu256, "__builtin_ia32_storedqu256", IX86_BUILTIN_STOREDQU256, UNKNOWN, (int) VOID_FTYPE_PCHAR_V32QI },
20657 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_lddqu256, "__builtin_ia32_lddqu256", IX86_BUILTIN_LDDQU256, UNKNOWN, (int) V32QI_FTYPE_PCCHAR },
20659 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv4di, "__builtin_ia32_movntdq256", IX86_BUILTIN_MOVNTDQ256, UNKNOWN, (int) VOID_FTYPE_PV4DI_V4DI },
20660 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv4df, "__builtin_ia32_movntpd256", IX86_BUILTIN_MOVNTPD256, UNKNOWN, (int) VOID_FTYPE_PDOUBLE_V4DF },
20661 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movntv8sf, "__builtin_ia32_movntps256", IX86_BUILTIN_MOVNTPS256, UNKNOWN, (int) VOID_FTYPE_PFLOAT_V8SF },
20663 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadpd, "__builtin_ia32_maskloadpd", IX86_BUILTIN_MASKLOADPD, UNKNOWN, (int) V2DF_FTYPE_PCV2DF_V2DF },
20664 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadps, "__builtin_ia32_maskloadps", IX86_BUILTIN_MASKLOADPS, UNKNOWN, (int) V4SF_FTYPE_PCV4SF_V4SF },
20665 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadpd256, "__builtin_ia32_maskloadpd256", IX86_BUILTIN_MASKLOADPD256, UNKNOWN, (int) V4DF_FTYPE_PCV4DF_V4DF },
20666 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskloadps256, "__builtin_ia32_maskloadps256", IX86_BUILTIN_MASKLOADPS256, UNKNOWN, (int) V8SF_FTYPE_PCV8SF_V8SF },
20667 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstorepd, "__builtin_ia32_maskstorepd", IX86_BUILTIN_MASKSTOREPD, UNKNOWN, (int) VOID_FTYPE_PV2DF_V2DF_V2DF },
20668 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstoreps, "__builtin_ia32_maskstoreps", IX86_BUILTIN_MASKSTOREPS, UNKNOWN, (int) VOID_FTYPE_PV4SF_V4SF_V4SF },
20669 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstorepd256, "__builtin_ia32_maskstorepd256", IX86_BUILTIN_MASKSTOREPD256, UNKNOWN, (int) VOID_FTYPE_PV4DF_V4DF_V4DF },
20670 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_maskstoreps256, "__builtin_ia32_maskstoreps256", IX86_BUILTIN_MASKSTOREPS256, UNKNOWN, (int) VOID_FTYPE_PV8SF_V8SF_V8SF },
20673 /* Builtins with variable number of arguments. */
20674 static const struct builtin_description bdesc_args[] =
20677 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv8qi3, "__builtin_ia32_paddb", IX86_BUILTIN_PADDB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20678 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv4hi3, "__builtin_ia32_paddw", IX86_BUILTIN_PADDW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20679 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_addv2si3, "__builtin_ia32_paddd", IX86_BUILTIN_PADDD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20680 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv8qi3, "__builtin_ia32_psubb", IX86_BUILTIN_PSUBB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20681 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv4hi3, "__builtin_ia32_psubw", IX86_BUILTIN_PSUBW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20682 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_subv2si3, "__builtin_ia32_psubd", IX86_BUILTIN_PSUBD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20684 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ssaddv8qi3, "__builtin_ia32_paddsb", IX86_BUILTIN_PADDSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20685 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ssaddv4hi3, "__builtin_ia32_paddsw", IX86_BUILTIN_PADDSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20686 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_sssubv8qi3, "__builtin_ia32_psubsb", IX86_BUILTIN_PSUBSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20687 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_sssubv4hi3, "__builtin_ia32_psubsw", IX86_BUILTIN_PSUBSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20688 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_usaddv8qi3, "__builtin_ia32_paddusb", IX86_BUILTIN_PADDUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20689 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_usaddv4hi3, "__builtin_ia32_paddusw", IX86_BUILTIN_PADDUSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20690 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ussubv8qi3, "__builtin_ia32_psubusb", IX86_BUILTIN_PSUBUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20691 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ussubv4hi3, "__builtin_ia32_psubusw", IX86_BUILTIN_PSUBUSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20693 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_mulv4hi3, "__builtin_ia32_pmullw", IX86_BUILTIN_PMULLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20694 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_smulv4hi3_highpart, "__builtin_ia32_pmulhw", IX86_BUILTIN_PMULHW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20696 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_andv2si3, "__builtin_ia32_pand", IX86_BUILTIN_PAND, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20697 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_andnotv2si3, "__builtin_ia32_pandn", IX86_BUILTIN_PANDN, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20698 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_iorv2si3, "__builtin_ia32_por", IX86_BUILTIN_POR, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20699 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_xorv2si3, "__builtin_ia32_pxor", IX86_BUILTIN_PXOR, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20701 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv8qi3, "__builtin_ia32_pcmpeqb", IX86_BUILTIN_PCMPEQB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20702 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv4hi3, "__builtin_ia32_pcmpeqw", IX86_BUILTIN_PCMPEQW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20703 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_eqv2si3, "__builtin_ia32_pcmpeqd", IX86_BUILTIN_PCMPEQD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20704 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv8qi3, "__builtin_ia32_pcmpgtb", IX86_BUILTIN_PCMPGTB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20705 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv4hi3, "__builtin_ia32_pcmpgtw", IX86_BUILTIN_PCMPGTW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20706 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_gtv2si3, "__builtin_ia32_pcmpgtd", IX86_BUILTIN_PCMPGTD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20708 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhbw, "__builtin_ia32_punpckhbw", IX86_BUILTIN_PUNPCKHBW, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20709 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhwd, "__builtin_ia32_punpckhwd", IX86_BUILTIN_PUNPCKHWD, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20710 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckhdq, "__builtin_ia32_punpckhdq", IX86_BUILTIN_PUNPCKHDQ, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
20711 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpcklbw, "__builtin_ia32_punpcklbw", IX86_BUILTIN_PUNPCKLBW, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20712 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpcklwd, "__builtin_ia32_punpcklwd", IX86_BUILTIN_PUNPCKLWD, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI},
20713 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_punpckldq, "__builtin_ia32_punpckldq", IX86_BUILTIN_PUNPCKLDQ, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI},
20715 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packsswb, "__builtin_ia32_packsswb", IX86_BUILTIN_PACKSSWB, UNKNOWN, (int) V8QI_FTYPE_V4HI_V4HI },
20716 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packssdw, "__builtin_ia32_packssdw", IX86_BUILTIN_PACKSSDW, UNKNOWN, (int) V4HI_FTYPE_V2SI_V2SI },
20717 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_packuswb, "__builtin_ia32_packuswb", IX86_BUILTIN_PACKUSWB, UNKNOWN, (int) V8QI_FTYPE_V4HI_V4HI },
20719 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_pmaddwd, "__builtin_ia32_pmaddwd", IX86_BUILTIN_PMADDWD, UNKNOWN, (int) V2SI_FTYPE_V4HI_V4HI },
20721 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv4hi3, "__builtin_ia32_psllwi", IX86_BUILTIN_PSLLWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
20722 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv2si3, "__builtin_ia32_pslldi", IX86_BUILTIN_PSLLDI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
20723 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv1di3, "__builtin_ia32_psllqi", IX86_BUILTIN_PSLLQI, UNKNOWN, (int) V1DI_FTYPE_V1DI_SI_COUNT },
20724 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv4hi3, "__builtin_ia32_psllw", IX86_BUILTIN_PSLLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
20725 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv2si3, "__builtin_ia32_pslld", IX86_BUILTIN_PSLLD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
20726 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashlv1di3, "__builtin_ia32_psllq", IX86_BUILTIN_PSLLQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI_COUNT },
20728 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv4hi3, "__builtin_ia32_psrlwi", IX86_BUILTIN_PSRLWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
20729 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv2si3, "__builtin_ia32_psrldi", IX86_BUILTIN_PSRLDI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
20730 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv1di3, "__builtin_ia32_psrlqi", IX86_BUILTIN_PSRLQI, UNKNOWN, (int) V1DI_FTYPE_V1DI_SI_COUNT },
20731 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv4hi3, "__builtin_ia32_psrlw", IX86_BUILTIN_PSRLW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
20732 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv2si3, "__builtin_ia32_psrld", IX86_BUILTIN_PSRLD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
20733 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_lshrv1di3, "__builtin_ia32_psrlq", IX86_BUILTIN_PSRLQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI_COUNT },
20735 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv4hi3, "__builtin_ia32_psrawi", IX86_BUILTIN_PSRAWI, UNKNOWN, (int) V4HI_FTYPE_V4HI_SI_COUNT },
20736 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv2si3, "__builtin_ia32_psradi", IX86_BUILTIN_PSRADI, UNKNOWN, (int) V2SI_FTYPE_V2SI_SI_COUNT },
20737 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv4hi3, "__builtin_ia32_psraw", IX86_BUILTIN_PSRAW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI_COUNT },
20738 { OPTION_MASK_ISA_MMX, CODE_FOR_mmx_ashrv2si3, "__builtin_ia32_psrad", IX86_BUILTIN_PSRAD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI_COUNT },
20741 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_pf2id, "__builtin_ia32_pf2id", IX86_BUILTIN_PF2ID, UNKNOWN, (int) V2SI_FTYPE_V2SF },
20742 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_floatv2si2, "__builtin_ia32_pi2fd", IX86_BUILTIN_PI2FD, UNKNOWN, (int) V2SF_FTYPE_V2SI },
20743 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpv2sf2, "__builtin_ia32_pfrcp", IX86_BUILTIN_PFRCP, UNKNOWN, (int) V2SF_FTYPE_V2SF },
20744 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rsqrtv2sf2, "__builtin_ia32_pfrsqrt", IX86_BUILTIN_PFRSQRT, UNKNOWN, (int) V2SF_FTYPE_V2SF },
20746 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_uavgv8qi3, "__builtin_ia32_pavgusb", IX86_BUILTIN_PAVGUSB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
20747 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_haddv2sf3, "__builtin_ia32_pfacc", IX86_BUILTIN_PFACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20748 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_addv2sf3, "__builtin_ia32_pfadd", IX86_BUILTIN_PFADD, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20749 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_eqv2sf3, "__builtin_ia32_pfcmpeq", IX86_BUILTIN_PFCMPEQ, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
20750 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_gev2sf3, "__builtin_ia32_pfcmpge", IX86_BUILTIN_PFCMPGE, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
20751 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_gtv2sf3, "__builtin_ia32_pfcmpgt", IX86_BUILTIN_PFCMPGT, UNKNOWN, (int) V2SI_FTYPE_V2SF_V2SF },
20752 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_smaxv2sf3, "__builtin_ia32_pfmax", IX86_BUILTIN_PFMAX, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20753 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_sminv2sf3, "__builtin_ia32_pfmin", IX86_BUILTIN_PFMIN, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20754 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_mulv2sf3, "__builtin_ia32_pfmul", IX86_BUILTIN_PFMUL, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20755 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpit1v2sf3, "__builtin_ia32_pfrcpit1", IX86_BUILTIN_PFRCPIT1, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20756 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rcpit2v2sf3, "__builtin_ia32_pfrcpit2", IX86_BUILTIN_PFRCPIT2, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20757 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_rsqit1v2sf3, "__builtin_ia32_pfrsqit1", IX86_BUILTIN_PFRSQIT1, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20758 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_subv2sf3, "__builtin_ia32_pfsub", IX86_BUILTIN_PFSUB, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20759 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_subrv2sf3, "__builtin_ia32_pfsubr", IX86_BUILTIN_PFSUBR, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20760 { OPTION_MASK_ISA_3DNOW, CODE_FOR_mmx_pmulhrwv4hi3, "__builtin_ia32_pmulhrw", IX86_BUILTIN_PMULHRW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
20763 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pf2iw, "__builtin_ia32_pf2iw", IX86_BUILTIN_PF2IW, UNKNOWN, (int) V2SI_FTYPE_V2SF },
20764 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pi2fw, "__builtin_ia32_pi2fw", IX86_BUILTIN_PI2FW, UNKNOWN, (int) V2SF_FTYPE_V2SI },
20765 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pswapdv2si2, "__builtin_ia32_pswapdsi", IX86_BUILTIN_PSWAPDSI, UNKNOWN, (int) V2SI_FTYPE_V2SI },
20766 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pswapdv2sf2, "__builtin_ia32_pswapdsf", IX86_BUILTIN_PSWAPDSF, UNKNOWN, (int) V2SF_FTYPE_V2SF },
20767 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_hsubv2sf3, "__builtin_ia32_pfnacc", IX86_BUILTIN_PFNACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20768 { OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_addsubv2sf3, "__builtin_ia32_pfpnacc", IX86_BUILTIN_PFPNACC, UNKNOWN, (int) V2SF_FTYPE_V2SF_V2SF },
20771 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movmskps, "__builtin_ia32_movmskps", IX86_BUILTIN_MOVMSKPS, UNKNOWN, (int) INT_FTYPE_V4SF },
20772 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_sqrtv4sf2, "__builtin_ia32_sqrtps", IX86_BUILTIN_SQRTPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
20773 { OPTION_MASK_ISA_SSE, CODE_FOR_sqrtv4sf2, "__builtin_ia32_sqrtps_nr", IX86_BUILTIN_SQRTPS_NR, UNKNOWN, (int) V4SF_FTYPE_V4SF },
20774 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_rsqrtv4sf2, "__builtin_ia32_rsqrtps", IX86_BUILTIN_RSQRTPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
20775 { OPTION_MASK_ISA_SSE, CODE_FOR_rsqrtv4sf2, "__builtin_ia32_rsqrtps_nr", IX86_BUILTIN_RSQRTPS_NR, UNKNOWN, (int) V4SF_FTYPE_V4SF },
20776 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_rcpv4sf2, "__builtin_ia32_rcpps", IX86_BUILTIN_RCPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF },
20777 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtps2pi, "__builtin_ia32_cvtps2pi", IX86_BUILTIN_CVTPS2PI, UNKNOWN, (int) V2SI_FTYPE_V4SF },
20778 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtss2si, "__builtin_ia32_cvtss2si", IX86_BUILTIN_CVTSS2SI, UNKNOWN, (int) INT_FTYPE_V4SF },
20779 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvtss2siq, "__builtin_ia32_cvtss2si64", IX86_BUILTIN_CVTSS2SI64, UNKNOWN, (int) INT64_FTYPE_V4SF },
20780 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvttps2pi, "__builtin_ia32_cvttps2pi", IX86_BUILTIN_CVTTPS2PI, UNKNOWN, (int) V2SI_FTYPE_V4SF },
20781 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvttss2si, "__builtin_ia32_cvttss2si", IX86_BUILTIN_CVTTSS2SI, UNKNOWN, (int) INT_FTYPE_V4SF },
20782 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvttss2siq, "__builtin_ia32_cvttss2si64", IX86_BUILTIN_CVTTSS2SI64, UNKNOWN, (int) INT64_FTYPE_V4SF },
20784 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_shufps, "__builtin_ia32_shufps", IX86_BUILTIN_SHUFPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
20786 { OPTION_MASK_ISA_SSE, CODE_FOR_addv4sf3, "__builtin_ia32_addps", IX86_BUILTIN_ADDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20787 { OPTION_MASK_ISA_SSE, CODE_FOR_subv4sf3, "__builtin_ia32_subps", IX86_BUILTIN_SUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20788 { OPTION_MASK_ISA_SSE, CODE_FOR_mulv4sf3, "__builtin_ia32_mulps", IX86_BUILTIN_MULPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20789 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_divv4sf3, "__builtin_ia32_divps", IX86_BUILTIN_DIVPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20790 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmaddv4sf3, "__builtin_ia32_addss", IX86_BUILTIN_ADDSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20791 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsubv4sf3, "__builtin_ia32_subss", IX86_BUILTIN_SUBSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20792 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmulv4sf3, "__builtin_ia32_mulss", IX86_BUILTIN_MULSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20793 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmdivv4sf3, "__builtin_ia32_divss", IX86_BUILTIN_DIVSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20795 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpeqps", IX86_BUILTIN_CMPEQPS, EQ, (int) V4SF_FTYPE_V4SF_V4SF },
20796 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpltps", IX86_BUILTIN_CMPLTPS, LT, (int) V4SF_FTYPE_V4SF_V4SF },
20797 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpleps", IX86_BUILTIN_CMPLEPS, LE, (int) V4SF_FTYPE_V4SF_V4SF },
20798 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpgtps", IX86_BUILTIN_CMPGTPS, LT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
20799 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpgeps", IX86_BUILTIN_CMPGEPS, LE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
20800 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpunordps", IX86_BUILTIN_CMPUNORDPS, UNORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
20801 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpneqps", IX86_BUILTIN_CMPNEQPS, NE, (int) V4SF_FTYPE_V4SF_V4SF },
20802 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpnltps", IX86_BUILTIN_CMPNLTPS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF },
20803 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpnleps", IX86_BUILTIN_CMPNLEPS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF },
20804 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpngtps", IX86_BUILTIN_CMPNGTPS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
20805 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpngeps", IX86_BUILTIN_CMPNGEPS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP},
20806 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_maskcmpv4sf3, "__builtin_ia32_cmpordps", IX86_BUILTIN_CMPORDPS, ORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
20807 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpeqss", IX86_BUILTIN_CMPEQSS, EQ, (int) V4SF_FTYPE_V4SF_V4SF },
20808 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpltss", IX86_BUILTIN_CMPLTSS, LT, (int) V4SF_FTYPE_V4SF_V4SF },
20809 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpless", IX86_BUILTIN_CMPLESS, LE, (int) V4SF_FTYPE_V4SF_V4SF },
20810 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpunordss", IX86_BUILTIN_CMPUNORDSS, UNORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
20811 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpneqss", IX86_BUILTIN_CMPNEQSS, NE, (int) V4SF_FTYPE_V4SF_V4SF },
20812 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpnltss", IX86_BUILTIN_CMPNLTSS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF },
20813 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpnless", IX86_BUILTIN_CMPNLESS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF },
20814 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpngtss", IX86_BUILTIN_CMPNGTSS, UNGE, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
20815 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpngess", IX86_BUILTIN_CMPNGESS, UNGT, (int) V4SF_FTYPE_V4SF_V4SF_SWAP },
20816 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmmaskcmpv4sf3, "__builtin_ia32_cmpordss", IX86_BUILTIN_CMPORDSS, ORDERED, (int) V4SF_FTYPE_V4SF_V4SF },
20818 { OPTION_MASK_ISA_SSE, CODE_FOR_sminv4sf3, "__builtin_ia32_minps", IX86_BUILTIN_MINPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20819 { OPTION_MASK_ISA_SSE, CODE_FOR_smaxv4sf3, "__builtin_ia32_maxps", IX86_BUILTIN_MAXPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20820 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsminv4sf3, "__builtin_ia32_minss", IX86_BUILTIN_MINSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20821 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsmaxv4sf3, "__builtin_ia32_maxss", IX86_BUILTIN_MAXSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20823 { OPTION_MASK_ISA_SSE, CODE_FOR_andv4sf3, "__builtin_ia32_andps", IX86_BUILTIN_ANDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20824 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_andnotv4sf3, "__builtin_ia32_andnps", IX86_BUILTIN_ANDNPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20825 { OPTION_MASK_ISA_SSE, CODE_FOR_iorv4sf3, "__builtin_ia32_orps", IX86_BUILTIN_ORPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20826 { OPTION_MASK_ISA_SSE, CODE_FOR_xorv4sf3, "__builtin_ia32_xorps", IX86_BUILTIN_XORPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20828 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movss, "__builtin_ia32_movss", IX86_BUILTIN_MOVSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20829 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movhlps_exp, "__builtin_ia32_movhlps", IX86_BUILTIN_MOVHLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20830 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_movlhps_exp, "__builtin_ia32_movlhps", IX86_BUILTIN_MOVLHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20831 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_unpckhps, "__builtin_ia32_unpckhps", IX86_BUILTIN_UNPCKHPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20832 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_unpcklps, "__builtin_ia32_unpcklps", IX86_BUILTIN_UNPCKLPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
20834 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtpi2ps, "__builtin_ia32_cvtpi2ps", IX86_BUILTIN_CVTPI2PS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V2SI },
20835 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_cvtsi2ss, "__builtin_ia32_cvtsi2ss", IX86_BUILTIN_CVTSI2SS, UNKNOWN, (int) V4SF_FTYPE_V4SF_SI },
20836 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_64BIT, CODE_FOR_sse_cvtsi2ssq, "__builtin_ia32_cvtsi642ss", IX86_BUILTIN_CVTSI642SS, UNKNOWN, V4SF_FTYPE_V4SF_DI },
20838 { OPTION_MASK_ISA_SSE, CODE_FOR_rsqrtsf2, "__builtin_ia32_rsqrtf", IX86_BUILTIN_RSQRTF, UNKNOWN, (int) FLOAT_FTYPE_FLOAT },
20840 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmsqrtv4sf2, "__builtin_ia32_sqrtss", IX86_BUILTIN_SQRTSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
20841 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmrsqrtv4sf2, "__builtin_ia32_rsqrtss", IX86_BUILTIN_RSQRTSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
20842 { OPTION_MASK_ISA_SSE, CODE_FOR_sse_vmrcpv4sf2, "__builtin_ia32_rcpss", IX86_BUILTIN_RCPSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_VEC_MERGE },
20844 /* SSE MMX or 3Dnow!A */
20845 { 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 },
20846 { 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 },
20847 { 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 },
20849 { 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 },
20850 { 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 },
20851 { 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 },
20852 { 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 },
20854 { 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 },
20855 { OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, CODE_FOR_mmx_pmovmskb, "__builtin_ia32_pmovmskb", IX86_BUILTIN_PMOVMSKB, UNKNOWN, (int) INT_FTYPE_V8QI },
20857 { 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 },
20860 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_shufpd, "__builtin_ia32_shufpd", IX86_BUILTIN_SHUFPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
20862 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movmskpd, "__builtin_ia32_movmskpd", IX86_BUILTIN_MOVMSKPD, UNKNOWN, (int) INT_FTYPE_V2DF },
20863 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pmovmskb, "__builtin_ia32_pmovmskb128", IX86_BUILTIN_PMOVMSKB128, UNKNOWN, (int) INT_FTYPE_V16QI },
20864 { OPTION_MASK_ISA_SSE2, CODE_FOR_sqrtv2df2, "__builtin_ia32_sqrtpd", IX86_BUILTIN_SQRTPD, UNKNOWN, (int) V2DF_FTYPE_V2DF },
20865 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtdq2pd, "__builtin_ia32_cvtdq2pd", IX86_BUILTIN_CVTDQ2PD, UNKNOWN, (int) V2DF_FTYPE_V4SI },
20866 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtdq2ps, "__builtin_ia32_cvtdq2ps", IX86_BUILTIN_CVTDQ2PS, UNKNOWN, (int) V4SF_FTYPE_V4SI },
20868 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2dq, "__builtin_ia32_cvtpd2dq", IX86_BUILTIN_CVTPD2DQ, UNKNOWN, (int) V4SI_FTYPE_V2DF },
20869 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2pi, "__builtin_ia32_cvtpd2pi", IX86_BUILTIN_CVTPD2PI, UNKNOWN, (int) V2SI_FTYPE_V2DF },
20870 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpd2ps, "__builtin_ia32_cvtpd2ps", IX86_BUILTIN_CVTPD2PS, UNKNOWN, (int) V4SF_FTYPE_V2DF },
20871 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttpd2dq, "__builtin_ia32_cvttpd2dq", IX86_BUILTIN_CVTTPD2DQ, UNKNOWN, (int) V4SI_FTYPE_V2DF },
20872 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttpd2pi, "__builtin_ia32_cvttpd2pi", IX86_BUILTIN_CVTTPD2PI, UNKNOWN, (int) V2SI_FTYPE_V2DF },
20874 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtpi2pd, "__builtin_ia32_cvtpi2pd", IX86_BUILTIN_CVTPI2PD, UNKNOWN, (int) V2DF_FTYPE_V2SI },
20876 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsd2si, "__builtin_ia32_cvtsd2si", IX86_BUILTIN_CVTSD2SI, UNKNOWN, (int) INT_FTYPE_V2DF },
20877 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttsd2si, "__builtin_ia32_cvttsd2si", IX86_BUILTIN_CVTTSD2SI, UNKNOWN, (int) INT_FTYPE_V2DF },
20878 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvtsd2siq, "__builtin_ia32_cvtsd2si64", IX86_BUILTIN_CVTSD2SI64, UNKNOWN, (int) INT64_FTYPE_V2DF },
20879 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvttsd2siq, "__builtin_ia32_cvttsd2si64", IX86_BUILTIN_CVTTSD2SI64, UNKNOWN, (int) INT64_FTYPE_V2DF },
20881 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtps2dq, "__builtin_ia32_cvtps2dq", IX86_BUILTIN_CVTPS2DQ, UNKNOWN, (int) V4SI_FTYPE_V4SF },
20882 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtps2pd, "__builtin_ia32_cvtps2pd", IX86_BUILTIN_CVTPS2PD, UNKNOWN, (int) V2DF_FTYPE_V4SF },
20883 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvttps2dq, "__builtin_ia32_cvttps2dq", IX86_BUILTIN_CVTTPS2DQ, UNKNOWN, (int) V4SI_FTYPE_V4SF },
20885 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv2df3, "__builtin_ia32_addpd", IX86_BUILTIN_ADDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20886 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv2df3, "__builtin_ia32_subpd", IX86_BUILTIN_SUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20887 { OPTION_MASK_ISA_SSE2, CODE_FOR_mulv2df3, "__builtin_ia32_mulpd", IX86_BUILTIN_MULPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20888 { OPTION_MASK_ISA_SSE2, CODE_FOR_divv2df3, "__builtin_ia32_divpd", IX86_BUILTIN_DIVPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20889 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmaddv2df3, "__builtin_ia32_addsd", IX86_BUILTIN_ADDSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20890 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsubv2df3, "__builtin_ia32_subsd", IX86_BUILTIN_SUBSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20891 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmulv2df3, "__builtin_ia32_mulsd", IX86_BUILTIN_MULSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20892 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmdivv2df3, "__builtin_ia32_divsd", IX86_BUILTIN_DIVSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20894 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpeqpd", IX86_BUILTIN_CMPEQPD, EQ, (int) V2DF_FTYPE_V2DF_V2DF },
20895 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpltpd", IX86_BUILTIN_CMPLTPD, LT, (int) V2DF_FTYPE_V2DF_V2DF },
20896 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmplepd", IX86_BUILTIN_CMPLEPD, LE, (int) V2DF_FTYPE_V2DF_V2DF },
20897 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpgtpd", IX86_BUILTIN_CMPGTPD, LT, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
20898 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpgepd", IX86_BUILTIN_CMPGEPD, LE, (int) V2DF_FTYPE_V2DF_V2DF_SWAP},
20899 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpunordpd", IX86_BUILTIN_CMPUNORDPD, UNORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
20900 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpneqpd", IX86_BUILTIN_CMPNEQPD, NE, (int) V2DF_FTYPE_V2DF_V2DF },
20901 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpnltpd", IX86_BUILTIN_CMPNLTPD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF },
20902 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpnlepd", IX86_BUILTIN_CMPNLEPD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF },
20903 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpngtpd", IX86_BUILTIN_CMPNGTPD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
20904 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpngepd", IX86_BUILTIN_CMPNGEPD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF_SWAP },
20905 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_maskcmpv2df3, "__builtin_ia32_cmpordpd", IX86_BUILTIN_CMPORDPD, ORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
20906 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpeqsd", IX86_BUILTIN_CMPEQSD, EQ, (int) V2DF_FTYPE_V2DF_V2DF },
20907 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpltsd", IX86_BUILTIN_CMPLTSD, LT, (int) V2DF_FTYPE_V2DF_V2DF },
20908 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmplesd", IX86_BUILTIN_CMPLESD, LE, (int) V2DF_FTYPE_V2DF_V2DF },
20909 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpunordsd", IX86_BUILTIN_CMPUNORDSD, UNORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
20910 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpneqsd", IX86_BUILTIN_CMPNEQSD, NE, (int) V2DF_FTYPE_V2DF_V2DF },
20911 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpnltsd", IX86_BUILTIN_CMPNLTSD, UNGE, (int) V2DF_FTYPE_V2DF_V2DF },
20912 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpnlesd", IX86_BUILTIN_CMPNLESD, UNGT, (int) V2DF_FTYPE_V2DF_V2DF },
20913 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmmaskcmpv2df3, "__builtin_ia32_cmpordsd", IX86_BUILTIN_CMPORDSD, ORDERED, (int) V2DF_FTYPE_V2DF_V2DF },
20915 { OPTION_MASK_ISA_SSE2, CODE_FOR_sminv2df3, "__builtin_ia32_minpd", IX86_BUILTIN_MINPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20916 { OPTION_MASK_ISA_SSE2, CODE_FOR_smaxv2df3, "__builtin_ia32_maxpd", IX86_BUILTIN_MAXPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20917 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsminv2df3, "__builtin_ia32_minsd", IX86_BUILTIN_MINSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20918 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsmaxv2df3, "__builtin_ia32_maxsd", IX86_BUILTIN_MAXSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20920 { OPTION_MASK_ISA_SSE2, CODE_FOR_andv2df3, "__builtin_ia32_andpd", IX86_BUILTIN_ANDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20921 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_andnotv2df3, "__builtin_ia32_andnpd", IX86_BUILTIN_ANDNPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20922 { OPTION_MASK_ISA_SSE2, CODE_FOR_iorv2df3, "__builtin_ia32_orpd", IX86_BUILTIN_ORPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20923 { OPTION_MASK_ISA_SSE2, CODE_FOR_xorv2df3, "__builtin_ia32_xorpd", IX86_BUILTIN_XORPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20925 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_movsd, "__builtin_ia32_movsd", IX86_BUILTIN_MOVSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20926 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_unpckhpd_exp, "__builtin_ia32_unpckhpd", IX86_BUILTIN_UNPCKHPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20927 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_unpcklpd_exp, "__builtin_ia32_unpcklpd", IX86_BUILTIN_UNPCKLPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
20929 { 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 },
20931 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv16qi3, "__builtin_ia32_paddb128", IX86_BUILTIN_PADDB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20932 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv8hi3, "__builtin_ia32_paddw128", IX86_BUILTIN_PADDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20933 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv4si3, "__builtin_ia32_paddd128", IX86_BUILTIN_PADDD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20934 { OPTION_MASK_ISA_SSE2, CODE_FOR_addv2di3, "__builtin_ia32_paddq128", IX86_BUILTIN_PADDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20935 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv16qi3, "__builtin_ia32_psubb128", IX86_BUILTIN_PSUBB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20936 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv8hi3, "__builtin_ia32_psubw128", IX86_BUILTIN_PSUBW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20937 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv4si3, "__builtin_ia32_psubd128", IX86_BUILTIN_PSUBD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20938 { OPTION_MASK_ISA_SSE2, CODE_FOR_subv2di3, "__builtin_ia32_psubq128", IX86_BUILTIN_PSUBQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20940 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ssaddv16qi3, "__builtin_ia32_paddsb128", IX86_BUILTIN_PADDSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20941 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ssaddv8hi3, "__builtin_ia32_paddsw128", IX86_BUILTIN_PADDSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20942 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_sssubv16qi3, "__builtin_ia32_psubsb128", IX86_BUILTIN_PSUBSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20943 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_sssubv8hi3, "__builtin_ia32_psubsw128", IX86_BUILTIN_PSUBSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20944 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_usaddv16qi3, "__builtin_ia32_paddusb128", IX86_BUILTIN_PADDUSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20945 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_usaddv8hi3, "__builtin_ia32_paddusw128", IX86_BUILTIN_PADDUSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20946 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ussubv16qi3, "__builtin_ia32_psubusb128", IX86_BUILTIN_PSUBUSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20947 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ussubv8hi3, "__builtin_ia32_psubusw128", IX86_BUILTIN_PSUBUSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20949 { OPTION_MASK_ISA_SSE2, CODE_FOR_mulv8hi3, "__builtin_ia32_pmullw128", IX86_BUILTIN_PMULLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20950 { OPTION_MASK_ISA_SSE2, CODE_FOR_smulv8hi3_highpart, "__builtin_ia32_pmulhw128", IX86_BUILTIN_PMULHW128, UNKNOWN,(int) V8HI_FTYPE_V8HI_V8HI },
20952 { OPTION_MASK_ISA_SSE2, CODE_FOR_andv2di3, "__builtin_ia32_pand128", IX86_BUILTIN_PAND128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20953 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_andnotv2di3, "__builtin_ia32_pandn128", IX86_BUILTIN_PANDN128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20954 { OPTION_MASK_ISA_SSE2, CODE_FOR_iorv2di3, "__builtin_ia32_por128", IX86_BUILTIN_POR128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20955 { OPTION_MASK_ISA_SSE2, CODE_FOR_xorv2di3, "__builtin_ia32_pxor128", IX86_BUILTIN_PXOR128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20957 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_uavgv16qi3, "__builtin_ia32_pavgb128", IX86_BUILTIN_PAVGB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20958 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_uavgv8hi3, "__builtin_ia32_pavgw128", IX86_BUILTIN_PAVGW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20960 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv16qi3, "__builtin_ia32_pcmpeqb128", IX86_BUILTIN_PCMPEQB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20961 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv8hi3, "__builtin_ia32_pcmpeqw128", IX86_BUILTIN_PCMPEQW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20962 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_eqv4si3, "__builtin_ia32_pcmpeqd128", IX86_BUILTIN_PCMPEQD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20963 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv16qi3, "__builtin_ia32_pcmpgtb128", IX86_BUILTIN_PCMPGTB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20964 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv8hi3, "__builtin_ia32_pcmpgtw128", IX86_BUILTIN_PCMPGTW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20965 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_gtv4si3, "__builtin_ia32_pcmpgtd128", IX86_BUILTIN_PCMPGTD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20967 { OPTION_MASK_ISA_SSE2, CODE_FOR_umaxv16qi3, "__builtin_ia32_pmaxub128", IX86_BUILTIN_PMAXUB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20968 { OPTION_MASK_ISA_SSE2, CODE_FOR_smaxv8hi3, "__builtin_ia32_pmaxsw128", IX86_BUILTIN_PMAXSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20969 { OPTION_MASK_ISA_SSE2, CODE_FOR_uminv16qi3, "__builtin_ia32_pminub128", IX86_BUILTIN_PMINUB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20970 { OPTION_MASK_ISA_SSE2, CODE_FOR_sminv8hi3, "__builtin_ia32_pminsw128", IX86_BUILTIN_PMINSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20972 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpckhbw, "__builtin_ia32_punpckhbw128", IX86_BUILTIN_PUNPCKHBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20973 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpckhwd, "__builtin_ia32_punpckhwd128", IX86_BUILTIN_PUNPCKHWD128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20974 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpckhdq, "__builtin_ia32_punpckhdq128", IX86_BUILTIN_PUNPCKHDQ128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20975 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpckhqdq, "__builtin_ia32_punpckhqdq128", IX86_BUILTIN_PUNPCKHQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20976 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpcklbw, "__builtin_ia32_punpcklbw128", IX86_BUILTIN_PUNPCKLBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
20977 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpcklwd, "__builtin_ia32_punpcklwd128", IX86_BUILTIN_PUNPCKLWD128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20978 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpckldq, "__builtin_ia32_punpckldq128", IX86_BUILTIN_PUNPCKLDQ128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
20979 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_punpcklqdq, "__builtin_ia32_punpcklqdq128", IX86_BUILTIN_PUNPCKLQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
20981 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packsswb, "__builtin_ia32_packsswb128", IX86_BUILTIN_PACKSSWB128, UNKNOWN, (int) V16QI_FTYPE_V8HI_V8HI },
20982 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packssdw, "__builtin_ia32_packssdw128", IX86_BUILTIN_PACKSSDW128, UNKNOWN, (int) V8HI_FTYPE_V4SI_V4SI },
20983 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_packuswb, "__builtin_ia32_packuswb128", IX86_BUILTIN_PACKUSWB128, UNKNOWN, (int) V16QI_FTYPE_V8HI_V8HI },
20985 { OPTION_MASK_ISA_SSE2, CODE_FOR_umulv8hi3_highpart, "__builtin_ia32_pmulhuw128", IX86_BUILTIN_PMULHUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
20986 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_psadbw, "__builtin_ia32_psadbw128", IX86_BUILTIN_PSADBW128, UNKNOWN, (int) V2DI_FTYPE_V16QI_V16QI },
20988 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_umulv1siv1di3, "__builtin_ia32_pmuludq", IX86_BUILTIN_PMULUDQ, UNKNOWN, (int) V1DI_FTYPE_V2SI_V2SI },
20989 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_umulv2siv2di3, "__builtin_ia32_pmuludq128", IX86_BUILTIN_PMULUDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI_V4SI },
20991 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pmaddwd, "__builtin_ia32_pmaddwd128", IX86_BUILTIN_PMADDWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI_V8HI },
20993 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsi2sd, "__builtin_ia32_cvtsi2sd", IX86_BUILTIN_CVTSI2SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_SI },
20994 { OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse2_cvtsi2sdq, "__builtin_ia32_cvtsi642sd", IX86_BUILTIN_CVTSI642SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_DI },
20995 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtsd2ss, "__builtin_ia32_cvtsd2ss", IX86_BUILTIN_CVTSD2SS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V2DF },
20996 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_cvtss2sd, "__builtin_ia32_cvtss2sd", IX86_BUILTIN_CVTSS2SD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V4SF },
20998 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_ashlti3, "__builtin_ia32_pslldqi128", IX86_BUILTIN_PSLLDQI128, UNKNOWN, (int) V2DI2TI_FTYPE_V2DI_INT },
20999 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv8hi3, "__builtin_ia32_psllwi128", IX86_BUILTIN_PSLLWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21000 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv4si3, "__builtin_ia32_pslldi128", IX86_BUILTIN_PSLLDI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21001 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv2di3, "__builtin_ia32_psllqi128", IX86_BUILTIN_PSLLQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_SI_COUNT },
21002 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv8hi3, "__builtin_ia32_psllw128", IX86_BUILTIN_PSLLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21003 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv4si3, "__builtin_ia32_pslld128", IX86_BUILTIN_PSLLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21004 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashlv2di3, "__builtin_ia32_psllq128", IX86_BUILTIN_PSLLQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_COUNT },
21006 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_lshrti3, "__builtin_ia32_psrldqi128", IX86_BUILTIN_PSRLDQI128, UNKNOWN, (int) V2DI2TI_FTYPE_V2DI_INT },
21007 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv8hi3, "__builtin_ia32_psrlwi128", IX86_BUILTIN_PSRLWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21008 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv4si3, "__builtin_ia32_psrldi128", IX86_BUILTIN_PSRLDI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21009 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv2di3, "__builtin_ia32_psrlqi128", IX86_BUILTIN_PSRLQI128, UNKNOWN, (int) V2DI_FTYPE_V2DI_SI_COUNT },
21010 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv8hi3, "__builtin_ia32_psrlw128", IX86_BUILTIN_PSRLW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21011 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv4si3, "__builtin_ia32_psrld128", IX86_BUILTIN_PSRLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21012 { OPTION_MASK_ISA_SSE2, CODE_FOR_lshrv2di3, "__builtin_ia32_psrlq128", IX86_BUILTIN_PSRLQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_COUNT },
21014 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv8hi3, "__builtin_ia32_psrawi128", IX86_BUILTIN_PSRAWI128, UNKNOWN, (int) V8HI_FTYPE_V8HI_SI_COUNT },
21015 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv4si3, "__builtin_ia32_psradi128", IX86_BUILTIN_PSRADI128, UNKNOWN, (int) V4SI_FTYPE_V4SI_SI_COUNT },
21016 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv8hi3, "__builtin_ia32_psraw128", IX86_BUILTIN_PSRAW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_COUNT },
21017 { OPTION_MASK_ISA_SSE2, CODE_FOR_ashrv4si3, "__builtin_ia32_psrad128", IX86_BUILTIN_PSRAD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI_COUNT },
21019 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshufd, "__builtin_ia32_pshufd", IX86_BUILTIN_PSHUFD, UNKNOWN, (int) V4SI_FTYPE_V4SI_INT },
21020 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshuflw, "__builtin_ia32_pshuflw", IX86_BUILTIN_PSHUFLW, UNKNOWN, (int) V8HI_FTYPE_V8HI_INT },
21021 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_pshufhw, "__builtin_ia32_pshufhw", IX86_BUILTIN_PSHUFHW, UNKNOWN, (int) V8HI_FTYPE_V8HI_INT },
21023 { OPTION_MASK_ISA_SSE2, CODE_FOR_sse2_vmsqrtv2df2, "__builtin_ia32_sqrtsd", IX86_BUILTIN_SQRTSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_VEC_MERGE },
21025 { OPTION_MASK_ISA_SSE2, CODE_FOR_abstf2, 0, IX86_BUILTIN_FABSQ, UNKNOWN, (int) FLOAT128_FTYPE_FLOAT128 },
21026 { OPTION_MASK_ISA_SSE2, CODE_FOR_copysigntf3, 0, IX86_BUILTIN_COPYSIGNQ, UNKNOWN, (int) FLOAT128_FTYPE_FLOAT128_FLOAT128 },
21028 { OPTION_MASK_ISA_SSE, CODE_FOR_sse2_movq128, "__builtin_ia32_movq128", IX86_BUILTIN_MOVQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI },
21031 { OPTION_MASK_ISA_SSE2, CODE_FOR_mmx_addv1di3, "__builtin_ia32_paddq", IX86_BUILTIN_PADDQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI },
21032 { OPTION_MASK_ISA_SSE2, CODE_FOR_mmx_subv1di3, "__builtin_ia32_psubq", IX86_BUILTIN_PSUBQ, UNKNOWN, (int) V1DI_FTYPE_V1DI_V1DI },
21035 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_movshdup, "__builtin_ia32_movshdup", IX86_BUILTIN_MOVSHDUP, UNKNOWN, (int) V4SF_FTYPE_V4SF},
21036 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_movsldup, "__builtin_ia32_movsldup", IX86_BUILTIN_MOVSLDUP, UNKNOWN, (int) V4SF_FTYPE_V4SF },
21038 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_addsubv4sf3, "__builtin_ia32_addsubps", IX86_BUILTIN_ADDSUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21039 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_addsubv2df3, "__builtin_ia32_addsubpd", IX86_BUILTIN_ADDSUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21040 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_haddv4sf3, "__builtin_ia32_haddps", IX86_BUILTIN_HADDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21041 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_haddv2df3, "__builtin_ia32_haddpd", IX86_BUILTIN_HADDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21042 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_hsubv4sf3, "__builtin_ia32_hsubps", IX86_BUILTIN_HSUBPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF },
21043 { OPTION_MASK_ISA_SSE3, CODE_FOR_sse3_hsubv2df3, "__builtin_ia32_hsubpd", IX86_BUILTIN_HSUBPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF },
21046 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv16qi2, "__builtin_ia32_pabsb128", IX86_BUILTIN_PABSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI },
21047 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv8qi2, "__builtin_ia32_pabsb", IX86_BUILTIN_PABSB, UNKNOWN, (int) V8QI_FTYPE_V8QI },
21048 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv8hi2, "__builtin_ia32_pabsw128", IX86_BUILTIN_PABSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI },
21049 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv4hi2, "__builtin_ia32_pabsw", IX86_BUILTIN_PABSW, UNKNOWN, (int) V4HI_FTYPE_V4HI },
21050 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv4si2, "__builtin_ia32_pabsd128", IX86_BUILTIN_PABSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI },
21051 { OPTION_MASK_ISA_SSSE3, CODE_FOR_absv2si2, "__builtin_ia32_pabsd", IX86_BUILTIN_PABSD, UNKNOWN, (int) V2SI_FTYPE_V2SI },
21053 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddwv8hi3, "__builtin_ia32_phaddw128", IX86_BUILTIN_PHADDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21054 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddwv4hi3, "__builtin_ia32_phaddw", IX86_BUILTIN_PHADDW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21055 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phadddv4si3, "__builtin_ia32_phaddd128", IX86_BUILTIN_PHADDD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21056 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phadddv2si3, "__builtin_ia32_phaddd", IX86_BUILTIN_PHADDD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21057 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddswv8hi3, "__builtin_ia32_phaddsw128", IX86_BUILTIN_PHADDSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21058 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phaddswv4hi3, "__builtin_ia32_phaddsw", IX86_BUILTIN_PHADDSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21059 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubwv8hi3, "__builtin_ia32_phsubw128", IX86_BUILTIN_PHSUBW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21060 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubwv4hi3, "__builtin_ia32_phsubw", IX86_BUILTIN_PHSUBW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21061 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubdv4si3, "__builtin_ia32_phsubd128", IX86_BUILTIN_PHSUBD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21062 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubdv2si3, "__builtin_ia32_phsubd", IX86_BUILTIN_PHSUBD, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21063 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubswv8hi3, "__builtin_ia32_phsubsw128", IX86_BUILTIN_PHSUBSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21064 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_phsubswv4hi3, "__builtin_ia32_phsubsw", IX86_BUILTIN_PHSUBSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21065 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmaddubsw128, "__builtin_ia32_pmaddubsw128", IX86_BUILTIN_PMADDUBSW128, UNKNOWN, (int) V8HI_FTYPE_V16QI_V16QI },
21066 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmaddubsw, "__builtin_ia32_pmaddubsw", IX86_BUILTIN_PMADDUBSW, UNKNOWN, (int) V4HI_FTYPE_V8QI_V8QI },
21067 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmulhrswv8hi3, "__builtin_ia32_pmulhrsw128", IX86_BUILTIN_PMULHRSW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21068 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pmulhrswv4hi3, "__builtin_ia32_pmulhrsw", IX86_BUILTIN_PMULHRSW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21069 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pshufbv16qi3, "__builtin_ia32_pshufb128", IX86_BUILTIN_PSHUFB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21070 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_pshufbv8qi3, "__builtin_ia32_pshufb", IX86_BUILTIN_PSHUFB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21071 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv16qi3, "__builtin_ia32_psignb128", IX86_BUILTIN_PSIGNB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21072 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv8qi3, "__builtin_ia32_psignb", IX86_BUILTIN_PSIGNB, UNKNOWN, (int) V8QI_FTYPE_V8QI_V8QI },
21073 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv8hi3, "__builtin_ia32_psignw128", IX86_BUILTIN_PSIGNW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21074 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv4hi3, "__builtin_ia32_psignw", IX86_BUILTIN_PSIGNW, UNKNOWN, (int) V4HI_FTYPE_V4HI_V4HI },
21075 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv4si3, "__builtin_ia32_psignd128", IX86_BUILTIN_PSIGND128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21076 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_psignv2si3, "__builtin_ia32_psignd", IX86_BUILTIN_PSIGND, UNKNOWN, (int) V2SI_FTYPE_V2SI_V2SI },
21079 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_palignrti, "__builtin_ia32_palignr128", IX86_BUILTIN_PALIGNR128, UNKNOWN, (int) V2DI2TI_FTYPE_V2DI_V2DI_INT },
21080 { OPTION_MASK_ISA_SSSE3, CODE_FOR_ssse3_palignrdi, "__builtin_ia32_palignr", IX86_BUILTIN_PALIGNR, UNKNOWN, (int) V1DI2DI_FTYPE_V1DI_V1DI_INT },
21083 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendpd, "__builtin_ia32_blendpd", IX86_BUILTIN_BLENDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21084 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendps, "__builtin_ia32_blendps", IX86_BUILTIN_BLENDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21085 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendvpd, "__builtin_ia32_blendvpd", IX86_BUILTIN_BLENDVPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_V2DF },
21086 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_blendvps, "__builtin_ia32_blendvps", IX86_BUILTIN_BLENDVPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_V4SF },
21087 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_dppd, "__builtin_ia32_dppd", IX86_BUILTIN_DPPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21088 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_dpps, "__builtin_ia32_dpps", IX86_BUILTIN_DPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21089 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_insertps, "__builtin_ia32_insertps128", IX86_BUILTIN_INSERTPS128, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21090 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_mpsadbw, "__builtin_ia32_mpsadbw128", IX86_BUILTIN_MPSADBW128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI_INT },
21091 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_pblendvb, "__builtin_ia32_pblendvb128", IX86_BUILTIN_PBLENDVB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI_V16QI },
21092 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_pblendw, "__builtin_ia32_pblendw128", IX86_BUILTIN_PBLENDW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI_INT },
21094 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv8qiv8hi2, "__builtin_ia32_pmovsxbw128", IX86_BUILTIN_PMOVSXBW128, UNKNOWN, (int) V8HI_FTYPE_V16QI },
21095 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv4qiv4si2, "__builtin_ia32_pmovsxbd128", IX86_BUILTIN_PMOVSXBD128, UNKNOWN, (int) V4SI_FTYPE_V16QI },
21096 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2qiv2di2, "__builtin_ia32_pmovsxbq128", IX86_BUILTIN_PMOVSXBQ128, UNKNOWN, (int) V2DI_FTYPE_V16QI },
21097 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv4hiv4si2, "__builtin_ia32_pmovsxwd128", IX86_BUILTIN_PMOVSXWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI },
21098 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2hiv2di2, "__builtin_ia32_pmovsxwq128", IX86_BUILTIN_PMOVSXWQ128, UNKNOWN, (int) V2DI_FTYPE_V8HI },
21099 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_extendv2siv2di2, "__builtin_ia32_pmovsxdq128", IX86_BUILTIN_PMOVSXDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI },
21100 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv8qiv8hi2, "__builtin_ia32_pmovzxbw128", IX86_BUILTIN_PMOVZXBW128, UNKNOWN, (int) V8HI_FTYPE_V16QI },
21101 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv4qiv4si2, "__builtin_ia32_pmovzxbd128", IX86_BUILTIN_PMOVZXBD128, UNKNOWN, (int) V4SI_FTYPE_V16QI },
21102 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2qiv2di2, "__builtin_ia32_pmovzxbq128", IX86_BUILTIN_PMOVZXBQ128, UNKNOWN, (int) V2DI_FTYPE_V16QI },
21103 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv4hiv4si2, "__builtin_ia32_pmovzxwd128", IX86_BUILTIN_PMOVZXWD128, UNKNOWN, (int) V4SI_FTYPE_V8HI },
21104 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2hiv2di2, "__builtin_ia32_pmovzxwq128", IX86_BUILTIN_PMOVZXWQ128, UNKNOWN, (int) V2DI_FTYPE_V8HI },
21105 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_zero_extendv2siv2di2, "__builtin_ia32_pmovzxdq128", IX86_BUILTIN_PMOVZXDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI },
21106 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_phminposuw, "__builtin_ia32_phminposuw128", IX86_BUILTIN_PHMINPOSUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI },
21108 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_packusdw, "__builtin_ia32_packusdw128", IX86_BUILTIN_PACKUSDW128, UNKNOWN, (int) V8HI_FTYPE_V4SI_V4SI },
21109 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_eqv2di3, "__builtin_ia32_pcmpeqq", IX86_BUILTIN_PCMPEQQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21110 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_smaxv16qi3, "__builtin_ia32_pmaxsb128", IX86_BUILTIN_PMAXSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21111 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_smaxv4si3, "__builtin_ia32_pmaxsd128", IX86_BUILTIN_PMAXSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21112 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_umaxv4si3, "__builtin_ia32_pmaxud128", IX86_BUILTIN_PMAXUD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21113 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_umaxv8hi3, "__builtin_ia32_pmaxuw128", IX86_BUILTIN_PMAXUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21114 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sminv16qi3, "__builtin_ia32_pminsb128", IX86_BUILTIN_PMINSB128, UNKNOWN, (int) V16QI_FTYPE_V16QI_V16QI },
21115 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sminv4si3, "__builtin_ia32_pminsd128", IX86_BUILTIN_PMINSD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21116 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_uminv4si3, "__builtin_ia32_pminud128", IX86_BUILTIN_PMINUD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21117 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_uminv8hi3, "__builtin_ia32_pminuw128", IX86_BUILTIN_PMINUW128, UNKNOWN, (int) V8HI_FTYPE_V8HI_V8HI },
21118 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_sse4_1_mulv2siv2di3, "__builtin_ia32_pmuldq128", IX86_BUILTIN_PMULDQ128, UNKNOWN, (int) V2DI_FTYPE_V4SI_V4SI },
21119 { OPTION_MASK_ISA_SSE4_1, CODE_FOR_mulv4si3, "__builtin_ia32_pmulld128", IX86_BUILTIN_PMULLD128, UNKNOWN, (int) V4SI_FTYPE_V4SI_V4SI },
21121 /* SSE4.1 and SSE5 */
21122 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd, "__builtin_ia32_roundpd", IX86_BUILTIN_ROUNDPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_INT },
21123 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_roundps", IX86_BUILTIN_ROUNDPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_INT },
21124 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundsd, "__builtin_ia32_roundsd", IX86_BUILTIN_ROUNDSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21125 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundss, "__builtin_ia32_roundss", IX86_BUILTIN_ROUNDSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21127 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestz128", IX86_BUILTIN_PTESTZ, EQ, (int) INT_FTYPE_V2DI_V2DI_PTEST },
21128 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestc128", IX86_BUILTIN_PTESTC, LTU, (int) INT_FTYPE_V2DI_V2DI_PTEST },
21129 { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestnzc128", IX86_BUILTIN_PTESTNZC, GTU, (int) INT_FTYPE_V2DI_V2DI_PTEST },
21132 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_gtv2di3, "__builtin_ia32_pcmpgtq", IX86_BUILTIN_PCMPGTQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21133 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_crc32qi, "__builtin_ia32_crc32qi", IX86_BUILTIN_CRC32QI, UNKNOWN, (int) UINT_FTYPE_UINT_UCHAR },
21134 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_crc32hi, "__builtin_ia32_crc32hi", IX86_BUILTIN_CRC32HI, UNKNOWN, (int) UINT_FTYPE_UINT_USHORT },
21135 { OPTION_MASK_ISA_SSE4_2, CODE_FOR_sse4_2_crc32si, "__builtin_ia32_crc32si", IX86_BUILTIN_CRC32SI, UNKNOWN, (int) UINT_FTYPE_UINT_UINT },
21136 { OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_64BIT, CODE_FOR_sse4_2_crc32di, "__builtin_ia32_crc32di", IX86_BUILTIN_CRC32DI, UNKNOWN, (int) UINT64_FTYPE_UINT64_UINT64 },
21139 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_extrqi, "__builtin_ia32_extrqi", IX86_BUILTIN_EXTRQI, UNKNOWN, (int) V2DI_FTYPE_V2DI_UINT_UINT },
21140 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_extrq, "__builtin_ia32_extrq", IX86_BUILTIN_EXTRQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V16QI },
21141 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_insertqi, "__builtin_ia32_insertqi", IX86_BUILTIN_INSERTQI, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_UINT_UINT },
21142 { OPTION_MASK_ISA_SSE4A, CODE_FOR_sse4a_insertq, "__builtin_ia32_insertq", IX86_BUILTIN_INSERTQ, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21145 { OPTION_MASK_ISA_SSE2, CODE_FOR_aeskeygenassist, 0, IX86_BUILTIN_AESKEYGENASSIST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_INT },
21146 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesimc, 0, IX86_BUILTIN_AESIMC128, UNKNOWN, (int) V2DI_FTYPE_V2DI },
21148 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesenc, 0, IX86_BUILTIN_AESENC128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21149 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesenclast, 0, IX86_BUILTIN_AESENCLAST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21150 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesdec, 0, IX86_BUILTIN_AESDEC128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21151 { OPTION_MASK_ISA_SSE2, CODE_FOR_aesdeclast, 0, IX86_BUILTIN_AESDECLAST128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI },
21154 { OPTION_MASK_ISA_SSE2, CODE_FOR_pclmulqdq, 0, IX86_BUILTIN_PCLMULQDQ128, UNKNOWN, (int) V2DI_FTYPE_V2DI_V2DI_INT },
21157 { OPTION_MASK_ISA_AVX, CODE_FOR_addv4df3, "__builtin_ia32_addpd256", IX86_BUILTIN_ADDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21158 { OPTION_MASK_ISA_AVX, CODE_FOR_addv8sf3, "__builtin_ia32_addps256", IX86_BUILTIN_ADDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21159 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_addsubv4df3, "__builtin_ia32_addsubpd256", IX86_BUILTIN_ADDSUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21160 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_addsubv8sf3, "__builtin_ia32_addsubps256", IX86_BUILTIN_ADDSUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21161 { OPTION_MASK_ISA_AVX, CODE_FOR_andv4df3, "__builtin_ia32_andpd256", IX86_BUILTIN_ANDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21162 { OPTION_MASK_ISA_AVX, CODE_FOR_andv8sf3, "__builtin_ia32_andps256", IX86_BUILTIN_ANDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21163 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_andnotv4df3, "__builtin_ia32_andnpd256", IX86_BUILTIN_ANDNPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21164 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_andnotv8sf3, "__builtin_ia32_andnps256", IX86_BUILTIN_ANDNPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21165 { OPTION_MASK_ISA_AVX, CODE_FOR_divv4df3, "__builtin_ia32_divpd256", IX86_BUILTIN_DIVPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21166 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_divv8sf3, "__builtin_ia32_divps256", IX86_BUILTIN_DIVPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21167 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_haddv4df3, "__builtin_ia32_haddpd256", IX86_BUILTIN_HADDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21168 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_hsubv8sf3, "__builtin_ia32_hsubps256", IX86_BUILTIN_HSUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21169 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_hsubv4df3, "__builtin_ia32_hsubpd256", IX86_BUILTIN_HSUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21170 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_haddv8sf3, "__builtin_ia32_haddps256", IX86_BUILTIN_HADDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21171 { OPTION_MASK_ISA_AVX, CODE_FOR_smaxv4df3, "__builtin_ia32_maxpd256", IX86_BUILTIN_MAXPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21172 { OPTION_MASK_ISA_AVX, CODE_FOR_smaxv8sf3, "__builtin_ia32_maxps256", IX86_BUILTIN_MAXPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21173 { OPTION_MASK_ISA_AVX, CODE_FOR_sminv4df3, "__builtin_ia32_minpd256", IX86_BUILTIN_MINPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21174 { OPTION_MASK_ISA_AVX, CODE_FOR_sminv8sf3, "__builtin_ia32_minps256", IX86_BUILTIN_MINPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21175 { OPTION_MASK_ISA_AVX, CODE_FOR_mulv4df3, "__builtin_ia32_mulpd256", IX86_BUILTIN_MULPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21176 { OPTION_MASK_ISA_AVX, CODE_FOR_mulv8sf3, "__builtin_ia32_mulps256", IX86_BUILTIN_MULPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21177 { OPTION_MASK_ISA_AVX, CODE_FOR_iorv4df3, "__builtin_ia32_orpd256", IX86_BUILTIN_ORPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21178 { OPTION_MASK_ISA_AVX, CODE_FOR_iorv8sf3, "__builtin_ia32_orps256", IX86_BUILTIN_ORPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21179 { OPTION_MASK_ISA_AVX, CODE_FOR_subv4df3, "__builtin_ia32_subpd256", IX86_BUILTIN_SUBPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21180 { OPTION_MASK_ISA_AVX, CODE_FOR_subv8sf3, "__builtin_ia32_subps256", IX86_BUILTIN_SUBPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21181 { OPTION_MASK_ISA_AVX, CODE_FOR_xorv4df3, "__builtin_ia32_xorpd256", IX86_BUILTIN_XORPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21182 { OPTION_MASK_ISA_AVX, CODE_FOR_xorv8sf3, "__builtin_ia32_xorps256", IX86_BUILTIN_XORPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21184 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv2df3, "__builtin_ia32_vpermilvarpd", IX86_BUILTIN_VPERMILVARPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DI },
21185 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv4sf3, "__builtin_ia32_vpermilvarps", IX86_BUILTIN_VPERMILVARPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SI },
21186 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv4df3, "__builtin_ia32_vpermilvarpd256", IX86_BUILTIN_VPERMILVARPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DI },
21187 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilvarv8sf3, "__builtin_ia32_vpermilvarps256", IX86_BUILTIN_VPERMILVARPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SI },
21189 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendpd256, "__builtin_ia32_blendpd256", IX86_BUILTIN_BLENDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
21190 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendps256, "__builtin_ia32_blendps256", IX86_BUILTIN_BLENDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
21191 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendvpd256, "__builtin_ia32_blendvpd256", IX86_BUILTIN_BLENDVPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_V4DF },
21192 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_blendvps256, "__builtin_ia32_blendvps256", IX86_BUILTIN_BLENDVPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_V8SF },
21193 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_dpps256, "__builtin_ia32_dpps256", IX86_BUILTIN_DPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
21194 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_shufpd256, "__builtin_ia32_shufpd256", IX86_BUILTIN_SHUFPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
21195 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_shufps256, "__builtin_ia32_shufps256", IX86_BUILTIN_SHUFPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
21196 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmpsdv2df3, "__builtin_ia32_cmpsd", IX86_BUILTIN_CMPSD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21197 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmpssv4sf3, "__builtin_ia32_cmpss", IX86_BUILTIN_CMPSS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21198 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppdv2df3, "__builtin_ia32_cmppd", IX86_BUILTIN_CMPPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_V2DF_INT },
21199 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppsv4sf3, "__builtin_ia32_cmpps", IX86_BUILTIN_CMPPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_V4SF_INT },
21200 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppdv4df3, "__builtin_ia32_cmppd256", IX86_BUILTIN_CMPPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
21201 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cmppsv8sf3, "__builtin_ia32_cmpps256", IX86_BUILTIN_CMPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
21202 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v4df, "__builtin_ia32_vextractf128_pd256", IX86_BUILTIN_EXTRACTF128PD256, UNKNOWN, (int) V2DF_FTYPE_V4DF_INT },
21203 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v8sf, "__builtin_ia32_vextractf128_ps256", IX86_BUILTIN_EXTRACTF128PS256, UNKNOWN, (int) V4SF_FTYPE_V8SF_INT },
21204 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vextractf128v8si, "__builtin_ia32_vextractf128_si256", IX86_BUILTIN_EXTRACTF128SI256, UNKNOWN, (int) V4SI_FTYPE_V8SI_INT },
21205 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtdq2pd256, "__builtin_ia32_cvtdq2pd256", IX86_BUILTIN_CVTDQ2PD256, UNKNOWN, (int) V4DF_FTYPE_V4SI },
21206 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtdq2ps256, "__builtin_ia32_cvtdq2ps256", IX86_BUILTIN_CVTDQ2PS256, UNKNOWN, (int) V8SF_FTYPE_V8SI },
21207 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtpd2ps256, "__builtin_ia32_cvtpd2ps256", IX86_BUILTIN_CVTPD2PS256, UNKNOWN, (int) V4SF_FTYPE_V4DF },
21208 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtps2dq256, "__builtin_ia32_cvtps2dq256", IX86_BUILTIN_CVTPS2DQ256, UNKNOWN, (int) V8SI_FTYPE_V8SF },
21209 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtps2pd256, "__builtin_ia32_cvtps2pd256", IX86_BUILTIN_CVTPS2PD256, UNKNOWN, (int) V4DF_FTYPE_V4SF },
21210 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvttpd2dq256, "__builtin_ia32_cvttpd2dq256", IX86_BUILTIN_CVTTPD2DQ256, UNKNOWN, (int) V4SI_FTYPE_V4DF },
21211 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvtpd2dq256, "__builtin_ia32_cvtpd2dq256", IX86_BUILTIN_CVTPD2DQ256, UNKNOWN, (int) V4SI_FTYPE_V4DF },
21212 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_cvttps2dq256, "__builtin_ia32_cvttps2dq256", IX86_BUILTIN_CVTTPS2DQ256, UNKNOWN, (int) V8SI_FTYPE_V8SF },
21213 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v4df3, "__builtin_ia32_vperm2f128_pd256", IX86_BUILTIN_VPERM2F128PD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF_INT },
21214 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v8sf3, "__builtin_ia32_vperm2f128_ps256", IX86_BUILTIN_VPERM2F128PS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF_INT },
21215 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vperm2f128v8si3, "__builtin_ia32_vperm2f128_si256", IX86_BUILTIN_VPERM2F128SI256, UNKNOWN, (int) V8SI_FTYPE_V8SI_V8SI_INT },
21216 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv2df, "__builtin_ia32_vpermilpd", IX86_BUILTIN_VPERMILPD, UNKNOWN, (int) V2DF_FTYPE_V2DF_INT },
21217 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv4sf, "__builtin_ia32_vpermilps", IX86_BUILTIN_VPERMILPS, UNKNOWN, (int) V4SF_FTYPE_V4SF_INT },
21218 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv4df, "__builtin_ia32_vpermilpd256", IX86_BUILTIN_VPERMILPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_INT },
21219 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vpermilv8sf, "__builtin_ia32_vpermilps256", IX86_BUILTIN_VPERMILPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_INT },
21220 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v4df, "__builtin_ia32_vinsertf128_pd256", IX86_BUILTIN_VINSERTF128PD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V2DF_INT },
21221 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v8sf, "__builtin_ia32_vinsertf128_ps256", IX86_BUILTIN_VINSERTF128PS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V4SF_INT },
21222 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vinsertf128v8si, "__builtin_ia32_vinsertf128_si256", IX86_BUILTIN_VINSERTF128SI256, UNKNOWN, (int) V8SI_FTYPE_V8SI_V4SI_INT },
21224 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movshdup256, "__builtin_ia32_movshdup256", IX86_BUILTIN_MOVSHDUP256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21225 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movsldup256, "__builtin_ia32_movsldup256", IX86_BUILTIN_MOVSLDUP256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21226 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movddup256, "__builtin_ia32_movddup256", IX86_BUILTIN_MOVDDUP256, UNKNOWN, (int) V4DF_FTYPE_V4DF },
21228 { OPTION_MASK_ISA_AVX, CODE_FOR_sqrtv4df2, "__builtin_ia32_sqrtpd256", IX86_BUILTIN_SQRTPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF },
21229 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_sqrtv8sf2, "__builtin_ia32_sqrtps256", IX86_BUILTIN_SQRTPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21230 { OPTION_MASK_ISA_AVX, CODE_FOR_sqrtv8sf2, "__builtin_ia32_sqrtps_nr256", IX86_BUILTIN_SQRTPS_NR256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21231 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_rsqrtv8sf2, "__builtin_ia32_rsqrtps256", IX86_BUILTIN_RSQRTPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21232 { OPTION_MASK_ISA_AVX, CODE_FOR_rsqrtv8sf2, "__builtin_ia32_rsqrtps_nr256", IX86_BUILTIN_RSQRTPS_NR256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21234 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_rcpv8sf2, "__builtin_ia32_rcpps256", IX86_BUILTIN_RCPPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
21236 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundpd256, "__builtin_ia32_roundpd256", IX86_BUILTIN_ROUNDPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_INT },
21237 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_roundps256", IX86_BUILTIN_ROUNDPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_INT },
21239 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpckhpd256, "__builtin_ia32_unpckhpd256", IX86_BUILTIN_UNPCKHPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21240 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpcklpd256, "__builtin_ia32_unpcklpd256", IX86_BUILTIN_UNPCKLPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
21241 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpckhps256, "__builtin_ia32_unpckhps256", IX86_BUILTIN_UNPCKHPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21242 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpcklps256, "__builtin_ia32_unpcklps256", IX86_BUILTIN_UNPCKLPS256, UNKNOWN, (int) V8SF_FTYPE_V8SF_V8SF },
21244 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_si256_si, "__builtin_ia32_si256_si", IX86_BUILTIN_SI256_SI, UNKNOWN, (int) V8SI_FTYPE_V4SI },
21245 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ps256_ps, "__builtin_ia32_ps256_ps", IX86_BUILTIN_PS256_PS, UNKNOWN, (int) V8SF_FTYPE_V4SF },
21246 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_pd256_pd, "__builtin_ia32_pd256_pd", IX86_BUILTIN_PD256_PD, UNKNOWN, (int) V4DF_FTYPE_V2DF },
21247 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_si_si256, "__builtin_ia32_si_si256", IX86_BUILTIN_SI_SI256, UNKNOWN, (int) V4SI_FTYPE_V8SI },
21248 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ps_ps256, "__builtin_ia32_ps_ps256", IX86_BUILTIN_PS_PS256, UNKNOWN, (int) V4SF_FTYPE_V8SF },
21249 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_pd_pd256, "__builtin_ia32_pd_pd256", IX86_BUILTIN_PD_PD256, UNKNOWN, (int) V2DF_FTYPE_V4DF },
21251 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestzpd", IX86_BUILTIN_VTESTZPD, EQ, (int) INT_FTYPE_V2DF_V2DF_PTEST },
21252 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestcpd", IX86_BUILTIN_VTESTCPD, LTU, (int) INT_FTYPE_V2DF_V2DF_PTEST },
21253 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd, "__builtin_ia32_vtestnzcpd", IX86_BUILTIN_VTESTNZCPD, GTU, (int) INT_FTYPE_V2DF_V2DF_PTEST },
21254 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestzps", IX86_BUILTIN_VTESTZPS, EQ, (int) INT_FTYPE_V4SF_V4SF_PTEST },
21255 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestcps", IX86_BUILTIN_VTESTCPS, LTU, (int) INT_FTYPE_V4SF_V4SF_PTEST },
21256 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps, "__builtin_ia32_vtestnzcps", IX86_BUILTIN_VTESTNZCPS, GTU, (int) INT_FTYPE_V4SF_V4SF_PTEST },
21257 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestzpd256", IX86_BUILTIN_VTESTZPD256, EQ, (int) INT_FTYPE_V4DF_V4DF_PTEST },
21258 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestcpd256", IX86_BUILTIN_VTESTCPD256, LTU, (int) INT_FTYPE_V4DF_V4DF_PTEST },
21259 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestpd256, "__builtin_ia32_vtestnzcpd256", IX86_BUILTIN_VTESTNZCPD256, GTU, (int) INT_FTYPE_V4DF_V4DF_PTEST },
21260 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestzps256", IX86_BUILTIN_VTESTZPS256, EQ, (int) INT_FTYPE_V8SF_V8SF_PTEST },
21261 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestcps256", IX86_BUILTIN_VTESTCPS256, LTU, (int) INT_FTYPE_V8SF_V8SF_PTEST },
21262 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_vtestps256, "__builtin_ia32_vtestnzcps256", IX86_BUILTIN_VTESTNZCPS256, GTU, (int) INT_FTYPE_V8SF_V8SF_PTEST },
21263 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestz256", IX86_BUILTIN_PTESTZ256, EQ, (int) INT_FTYPE_V4DI_V4DI_PTEST },
21264 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestc256", IX86_BUILTIN_PTESTC256, LTU, (int) INT_FTYPE_V4DI_V4DI_PTEST },
21265 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_ptest256, "__builtin_ia32_ptestnzc256", IX86_BUILTIN_PTESTNZC256, GTU, (int) INT_FTYPE_V4DI_V4DI_PTEST },
21267 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movmskpd256, "__builtin_ia32_movmskpd256", IX86_BUILTIN_MOVMSKPD256, UNKNOWN, (int) INT_FTYPE_V4DF },
21268 { OPTION_MASK_ISA_AVX, CODE_FOR_avx_movmskps256, "__builtin_ia32_movmskps256", IX86_BUILTIN_MOVMSKPS256, UNKNOWN, (int) INT_FTYPE_V8SF },
21272 enum multi_arg_type {
21282 MULTI_ARG_3_PERMPS,
21283 MULTI_ARG_3_PERMPD,
21290 MULTI_ARG_2_DI_IMM,
21291 MULTI_ARG_2_SI_IMM,
21292 MULTI_ARG_2_HI_IMM,
21293 MULTI_ARG_2_QI_IMM,
21294 MULTI_ARG_2_SF_CMP,
21295 MULTI_ARG_2_DF_CMP,
21296 MULTI_ARG_2_DI_CMP,
21297 MULTI_ARG_2_SI_CMP,
21298 MULTI_ARG_2_HI_CMP,
21299 MULTI_ARG_2_QI_CMP,
21322 static const struct builtin_description bdesc_multi_arg[] =
21324 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfmaddv4sf4, "__builtin_ia32_fmaddss", IX86_BUILTIN_FMADDSS, 0, (int)MULTI_ARG_3_SF },
21325 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfmaddv2df4, "__builtin_ia32_fmaddsd", IX86_BUILTIN_FMADDSD, 0, (int)MULTI_ARG_3_DF },
21326 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fmaddv4sf4, "__builtin_ia32_fmaddps", IX86_BUILTIN_FMADDPS, 0, (int)MULTI_ARG_3_SF },
21327 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fmaddv2df4, "__builtin_ia32_fmaddpd", IX86_BUILTIN_FMADDPD, 0, (int)MULTI_ARG_3_DF },
21328 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfmsubv4sf4, "__builtin_ia32_fmsubss", IX86_BUILTIN_FMSUBSS, 0, (int)MULTI_ARG_3_SF },
21329 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfmsubv2df4, "__builtin_ia32_fmsubsd", IX86_BUILTIN_FMSUBSD, 0, (int)MULTI_ARG_3_DF },
21330 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fmsubv4sf4, "__builtin_ia32_fmsubps", IX86_BUILTIN_FMSUBPS, 0, (int)MULTI_ARG_3_SF },
21331 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fmsubv2df4, "__builtin_ia32_fmsubpd", IX86_BUILTIN_FMSUBPD, 0, (int)MULTI_ARG_3_DF },
21332 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfnmaddv4sf4, "__builtin_ia32_fnmaddss", IX86_BUILTIN_FNMADDSS, 0, (int)MULTI_ARG_3_SF },
21333 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfnmaddv2df4, "__builtin_ia32_fnmaddsd", IX86_BUILTIN_FNMADDSD, 0, (int)MULTI_ARG_3_DF },
21334 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fnmaddv4sf4, "__builtin_ia32_fnmaddps", IX86_BUILTIN_FNMADDPS, 0, (int)MULTI_ARG_3_SF },
21335 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fnmaddv2df4, "__builtin_ia32_fnmaddpd", IX86_BUILTIN_FNMADDPD, 0, (int)MULTI_ARG_3_DF },
21336 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfnmsubv4sf4, "__builtin_ia32_fnmsubss", IX86_BUILTIN_FNMSUBSS, 0, (int)MULTI_ARG_3_SF },
21337 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_vmfnmsubv2df4, "__builtin_ia32_fnmsubsd", IX86_BUILTIN_FNMSUBSD, 0, (int)MULTI_ARG_3_DF },
21338 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fnmsubv4sf4, "__builtin_ia32_fnmsubps", IX86_BUILTIN_FNMSUBPS, 0, (int)MULTI_ARG_3_SF },
21339 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5i_fnmsubv2df4, "__builtin_ia32_fnmsubpd", IX86_BUILTIN_FNMSUBPD, 0, (int)MULTI_ARG_3_DF },
21340 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v2di, "__builtin_ia32_pcmov", IX86_BUILTIN_PCMOV, 0, (int)MULTI_ARG_3_DI },
21341 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v2di, "__builtin_ia32_pcmov_v2di", IX86_BUILTIN_PCMOV_V2DI, 0, (int)MULTI_ARG_3_DI },
21342 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v4si, "__builtin_ia32_pcmov_v4si", IX86_BUILTIN_PCMOV_V4SI, 0, (int)MULTI_ARG_3_SI },
21343 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v8hi, "__builtin_ia32_pcmov_v8hi", IX86_BUILTIN_PCMOV_V8HI, 0, (int)MULTI_ARG_3_HI },
21344 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v16qi, "__builtin_ia32_pcmov_v16qi",IX86_BUILTIN_PCMOV_V16QI,0, (int)MULTI_ARG_3_QI },
21345 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v2df, "__builtin_ia32_pcmov_v2df", IX86_BUILTIN_PCMOV_V2DF, 0, (int)MULTI_ARG_3_DF },
21346 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcmov_v4sf, "__builtin_ia32_pcmov_v4sf", IX86_BUILTIN_PCMOV_V4SF, 0, (int)MULTI_ARG_3_SF },
21347 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pperm, "__builtin_ia32_pperm", IX86_BUILTIN_PPERM, 0, (int)MULTI_ARG_3_QI },
21348 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_permv4sf, "__builtin_ia32_permps", IX86_BUILTIN_PERMPS, 0, (int)MULTI_ARG_3_PERMPS },
21349 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_permv2df, "__builtin_ia32_permpd", IX86_BUILTIN_PERMPD, 0, (int)MULTI_ARG_3_PERMPD },
21350 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacssww, "__builtin_ia32_pmacssww", IX86_BUILTIN_PMACSSWW, 0, (int)MULTI_ARG_3_HI },
21351 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacsww, "__builtin_ia32_pmacsww", IX86_BUILTIN_PMACSWW, 0, (int)MULTI_ARG_3_HI },
21352 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacsswd, "__builtin_ia32_pmacsswd", IX86_BUILTIN_PMACSSWD, 0, (int)MULTI_ARG_3_HI_SI },
21353 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacswd, "__builtin_ia32_pmacswd", IX86_BUILTIN_PMACSWD, 0, (int)MULTI_ARG_3_HI_SI },
21354 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacssdd, "__builtin_ia32_pmacssdd", IX86_BUILTIN_PMACSSDD, 0, (int)MULTI_ARG_3_SI },
21355 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacsdd, "__builtin_ia32_pmacsdd", IX86_BUILTIN_PMACSDD, 0, (int)MULTI_ARG_3_SI },
21356 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacssdql, "__builtin_ia32_pmacssdql", IX86_BUILTIN_PMACSSDQL, 0, (int)MULTI_ARG_3_SI_DI },
21357 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacssdqh, "__builtin_ia32_pmacssdqh", IX86_BUILTIN_PMACSSDQH, 0, (int)MULTI_ARG_3_SI_DI },
21358 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacsdql, "__builtin_ia32_pmacsdql", IX86_BUILTIN_PMACSDQL, 0, (int)MULTI_ARG_3_SI_DI },
21359 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmacsdqh, "__builtin_ia32_pmacsdqh", IX86_BUILTIN_PMACSDQH, 0, (int)MULTI_ARG_3_SI_DI },
21360 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmadcsswd, "__builtin_ia32_pmadcsswd", IX86_BUILTIN_PMADCSSWD, 0, (int)MULTI_ARG_3_HI_SI },
21361 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pmadcswd, "__builtin_ia32_pmadcswd", IX86_BUILTIN_PMADCSWD, 0, (int)MULTI_ARG_3_HI_SI },
21362 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vrotlv2di3, "__builtin_ia32_protq", IX86_BUILTIN_PROTQ, 0, (int)MULTI_ARG_2_DI },
21363 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vrotlv4si3, "__builtin_ia32_protd", IX86_BUILTIN_PROTD, 0, (int)MULTI_ARG_2_SI },
21364 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vrotlv8hi3, "__builtin_ia32_protw", IX86_BUILTIN_PROTW, 0, (int)MULTI_ARG_2_HI },
21365 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vrotlv16qi3, "__builtin_ia32_protb", IX86_BUILTIN_PROTB, 0, (int)MULTI_ARG_2_QI },
21366 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_rotlv2di3, "__builtin_ia32_protqi", IX86_BUILTIN_PROTQ_IMM, 0, (int)MULTI_ARG_2_DI_IMM },
21367 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_rotlv4si3, "__builtin_ia32_protdi", IX86_BUILTIN_PROTD_IMM, 0, (int)MULTI_ARG_2_SI_IMM },
21368 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_rotlv8hi3, "__builtin_ia32_protwi", IX86_BUILTIN_PROTW_IMM, 0, (int)MULTI_ARG_2_HI_IMM },
21369 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_rotlv16qi3, "__builtin_ia32_protbi", IX86_BUILTIN_PROTB_IMM, 0, (int)MULTI_ARG_2_QI_IMM },
21370 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_ashlv2di3, "__builtin_ia32_pshaq", IX86_BUILTIN_PSHAQ, 0, (int)MULTI_ARG_2_DI },
21371 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_ashlv4si3, "__builtin_ia32_pshad", IX86_BUILTIN_PSHAD, 0, (int)MULTI_ARG_2_SI },
21372 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_ashlv8hi3, "__builtin_ia32_pshaw", IX86_BUILTIN_PSHAW, 0, (int)MULTI_ARG_2_HI },
21373 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_ashlv16qi3, "__builtin_ia32_pshab", IX86_BUILTIN_PSHAB, 0, (int)MULTI_ARG_2_QI },
21374 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_lshlv2di3, "__builtin_ia32_pshlq", IX86_BUILTIN_PSHLQ, 0, (int)MULTI_ARG_2_DI },
21375 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_lshlv4si3, "__builtin_ia32_pshld", IX86_BUILTIN_PSHLD, 0, (int)MULTI_ARG_2_SI },
21376 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_lshlv8hi3, "__builtin_ia32_pshlw", IX86_BUILTIN_PSHLW, 0, (int)MULTI_ARG_2_HI },
21377 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_lshlv16qi3, "__builtin_ia32_pshlb", IX86_BUILTIN_PSHLB, 0, (int)MULTI_ARG_2_QI },
21378 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmfrczv4sf2, "__builtin_ia32_frczss", IX86_BUILTIN_FRCZSS, 0, (int)MULTI_ARG_2_SF },
21379 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmfrczv2df2, "__builtin_ia32_frczsd", IX86_BUILTIN_FRCZSD, 0, (int)MULTI_ARG_2_DF },
21380 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_frczv4sf2, "__builtin_ia32_frczps", IX86_BUILTIN_FRCZPS, 0, (int)MULTI_ARG_1_SF },
21381 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_frczv2df2, "__builtin_ia32_frczpd", IX86_BUILTIN_FRCZPD, 0, (int)MULTI_ARG_1_DF },
21382 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_cvtph2ps, "__builtin_ia32_cvtph2ps", IX86_BUILTIN_CVTPH2PS, 0, (int)MULTI_ARG_1_PH2PS },
21383 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_cvtps2ph, "__builtin_ia32_cvtps2ph", IX86_BUILTIN_CVTPS2PH, 0, (int)MULTI_ARG_1_PS2PH },
21384 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddbw, "__builtin_ia32_phaddbw", IX86_BUILTIN_PHADDBW, 0, (int)MULTI_ARG_1_QI_HI },
21385 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddbd, "__builtin_ia32_phaddbd", IX86_BUILTIN_PHADDBD, 0, (int)MULTI_ARG_1_QI_SI },
21386 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddbq, "__builtin_ia32_phaddbq", IX86_BUILTIN_PHADDBQ, 0, (int)MULTI_ARG_1_QI_DI },
21387 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddwd, "__builtin_ia32_phaddwd", IX86_BUILTIN_PHADDWD, 0, (int)MULTI_ARG_1_HI_SI },
21388 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddwq, "__builtin_ia32_phaddwq", IX86_BUILTIN_PHADDWQ, 0, (int)MULTI_ARG_1_HI_DI },
21389 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phadddq, "__builtin_ia32_phadddq", IX86_BUILTIN_PHADDDQ, 0, (int)MULTI_ARG_1_SI_DI },
21390 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddubw, "__builtin_ia32_phaddubw", IX86_BUILTIN_PHADDUBW, 0, (int)MULTI_ARG_1_QI_HI },
21391 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddubd, "__builtin_ia32_phaddubd", IX86_BUILTIN_PHADDUBD, 0, (int)MULTI_ARG_1_QI_SI },
21392 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddubq, "__builtin_ia32_phaddubq", IX86_BUILTIN_PHADDUBQ, 0, (int)MULTI_ARG_1_QI_DI },
21393 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phadduwd, "__builtin_ia32_phadduwd", IX86_BUILTIN_PHADDUWD, 0, (int)MULTI_ARG_1_HI_SI },
21394 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phadduwq, "__builtin_ia32_phadduwq", IX86_BUILTIN_PHADDUWQ, 0, (int)MULTI_ARG_1_HI_DI },
21395 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phaddudq, "__builtin_ia32_phaddudq", IX86_BUILTIN_PHADDUDQ, 0, (int)MULTI_ARG_1_SI_DI },
21396 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phsubbw, "__builtin_ia32_phsubbw", IX86_BUILTIN_PHSUBBW, 0, (int)MULTI_ARG_1_QI_HI },
21397 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phsubwd, "__builtin_ia32_phsubwd", IX86_BUILTIN_PHSUBWD, 0, (int)MULTI_ARG_1_HI_SI },
21398 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_phsubdq, "__builtin_ia32_phsubdq", IX86_BUILTIN_PHSUBDQ, 0, (int)MULTI_ARG_1_SI_DI },
21400 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comeqss", IX86_BUILTIN_COMEQSS, EQ, (int)MULTI_ARG_2_SF_CMP },
21401 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comness", IX86_BUILTIN_COMNESS, NE, (int)MULTI_ARG_2_SF_CMP },
21402 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comneqss", IX86_BUILTIN_COMNESS, NE, (int)MULTI_ARG_2_SF_CMP },
21403 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comltss", IX86_BUILTIN_COMLTSS, LT, (int)MULTI_ARG_2_SF_CMP },
21404 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comless", IX86_BUILTIN_COMLESS, LE, (int)MULTI_ARG_2_SF_CMP },
21405 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comgtss", IX86_BUILTIN_COMGTSS, GT, (int)MULTI_ARG_2_SF_CMP },
21406 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comgess", IX86_BUILTIN_COMGESS, GE, (int)MULTI_ARG_2_SF_CMP },
21407 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comueqss", IX86_BUILTIN_COMUEQSS, UNEQ, (int)MULTI_ARG_2_SF_CMP },
21408 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comuness", IX86_BUILTIN_COMUNESS, LTGT, (int)MULTI_ARG_2_SF_CMP },
21409 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comuneqss", IX86_BUILTIN_COMUNESS, LTGT, (int)MULTI_ARG_2_SF_CMP },
21410 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comunltss", IX86_BUILTIN_COMULTSS, UNLT, (int)MULTI_ARG_2_SF_CMP },
21411 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comunless", IX86_BUILTIN_COMULESS, UNLE, (int)MULTI_ARG_2_SF_CMP },
21412 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comungtss", IX86_BUILTIN_COMUGTSS, UNGT, (int)MULTI_ARG_2_SF_CMP },
21413 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comungess", IX86_BUILTIN_COMUGESS, UNGE, (int)MULTI_ARG_2_SF_CMP },
21414 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comordss", IX86_BUILTIN_COMORDSS, ORDERED, (int)MULTI_ARG_2_SF_CMP },
21415 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv4sf3, "__builtin_ia32_comunordss", IX86_BUILTIN_COMUNORDSS, UNORDERED, (int)MULTI_ARG_2_SF_CMP },
21417 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comeqsd", IX86_BUILTIN_COMEQSD, EQ, (int)MULTI_ARG_2_DF_CMP },
21418 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comnesd", IX86_BUILTIN_COMNESD, NE, (int)MULTI_ARG_2_DF_CMP },
21419 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comneqsd", IX86_BUILTIN_COMNESD, NE, (int)MULTI_ARG_2_DF_CMP },
21420 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comltsd", IX86_BUILTIN_COMLTSD, LT, (int)MULTI_ARG_2_DF_CMP },
21421 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comlesd", IX86_BUILTIN_COMLESD, LE, (int)MULTI_ARG_2_DF_CMP },
21422 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comgtsd", IX86_BUILTIN_COMGTSD, GT, (int)MULTI_ARG_2_DF_CMP },
21423 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comgesd", IX86_BUILTIN_COMGESD, GE, (int)MULTI_ARG_2_DF_CMP },
21424 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comueqsd", IX86_BUILTIN_COMUEQSD, UNEQ, (int)MULTI_ARG_2_DF_CMP },
21425 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comunesd", IX86_BUILTIN_COMUNESD, LTGT, (int)MULTI_ARG_2_DF_CMP },
21426 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comuneqsd", IX86_BUILTIN_COMUNESD, LTGT, (int)MULTI_ARG_2_DF_CMP },
21427 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comunltsd", IX86_BUILTIN_COMULTSD, UNLT, (int)MULTI_ARG_2_DF_CMP },
21428 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comunlesd", IX86_BUILTIN_COMULESD, UNLE, (int)MULTI_ARG_2_DF_CMP },
21429 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comungtsd", IX86_BUILTIN_COMUGTSD, UNGT, (int)MULTI_ARG_2_DF_CMP },
21430 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comungesd", IX86_BUILTIN_COMUGESD, UNGE, (int)MULTI_ARG_2_DF_CMP },
21431 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comordsd", IX86_BUILTIN_COMORDSD, ORDERED, (int)MULTI_ARG_2_DF_CMP },
21432 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_vmmaskcmpv2df3, "__builtin_ia32_comunordsd", IX86_BUILTIN_COMUNORDSD, UNORDERED, (int)MULTI_ARG_2_DF_CMP },
21434 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comeqps", IX86_BUILTIN_COMEQPS, EQ, (int)MULTI_ARG_2_SF_CMP },
21435 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comneps", IX86_BUILTIN_COMNEPS, NE, (int)MULTI_ARG_2_SF_CMP },
21436 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comneqps", IX86_BUILTIN_COMNEPS, NE, (int)MULTI_ARG_2_SF_CMP },
21437 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comltps", IX86_BUILTIN_COMLTPS, LT, (int)MULTI_ARG_2_SF_CMP },
21438 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comleps", IX86_BUILTIN_COMLEPS, LE, (int)MULTI_ARG_2_SF_CMP },
21439 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comgtps", IX86_BUILTIN_COMGTPS, GT, (int)MULTI_ARG_2_SF_CMP },
21440 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comgeps", IX86_BUILTIN_COMGEPS, GE, (int)MULTI_ARG_2_SF_CMP },
21441 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comueqps", IX86_BUILTIN_COMUEQPS, UNEQ, (int)MULTI_ARG_2_SF_CMP },
21442 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comuneps", IX86_BUILTIN_COMUNEPS, LTGT, (int)MULTI_ARG_2_SF_CMP },
21443 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comuneqps", IX86_BUILTIN_COMUNEPS, LTGT, (int)MULTI_ARG_2_SF_CMP },
21444 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comunltps", IX86_BUILTIN_COMULTPS, UNLT, (int)MULTI_ARG_2_SF_CMP },
21445 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comunleps", IX86_BUILTIN_COMULEPS, UNLE, (int)MULTI_ARG_2_SF_CMP },
21446 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comungtps", IX86_BUILTIN_COMUGTPS, UNGT, (int)MULTI_ARG_2_SF_CMP },
21447 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comungeps", IX86_BUILTIN_COMUGEPS, UNGE, (int)MULTI_ARG_2_SF_CMP },
21448 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comordps", IX86_BUILTIN_COMORDPS, ORDERED, (int)MULTI_ARG_2_SF_CMP },
21449 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4sf3, "__builtin_ia32_comunordps", IX86_BUILTIN_COMUNORDPS, UNORDERED, (int)MULTI_ARG_2_SF_CMP },
21451 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comeqpd", IX86_BUILTIN_COMEQPD, EQ, (int)MULTI_ARG_2_DF_CMP },
21452 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comnepd", IX86_BUILTIN_COMNEPD, NE, (int)MULTI_ARG_2_DF_CMP },
21453 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comneqpd", IX86_BUILTIN_COMNEPD, NE, (int)MULTI_ARG_2_DF_CMP },
21454 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comltpd", IX86_BUILTIN_COMLTPD, LT, (int)MULTI_ARG_2_DF_CMP },
21455 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comlepd", IX86_BUILTIN_COMLEPD, LE, (int)MULTI_ARG_2_DF_CMP },
21456 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comgtpd", IX86_BUILTIN_COMGTPD, GT, (int)MULTI_ARG_2_DF_CMP },
21457 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comgepd", IX86_BUILTIN_COMGEPD, GE, (int)MULTI_ARG_2_DF_CMP },
21458 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comueqpd", IX86_BUILTIN_COMUEQPD, UNEQ, (int)MULTI_ARG_2_DF_CMP },
21459 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comunepd", IX86_BUILTIN_COMUNEPD, LTGT, (int)MULTI_ARG_2_DF_CMP },
21460 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comuneqpd", IX86_BUILTIN_COMUNEPD, LTGT, (int)MULTI_ARG_2_DF_CMP },
21461 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comunltpd", IX86_BUILTIN_COMULTPD, UNLT, (int)MULTI_ARG_2_DF_CMP },
21462 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comunlepd", IX86_BUILTIN_COMULEPD, UNLE, (int)MULTI_ARG_2_DF_CMP },
21463 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comungtpd", IX86_BUILTIN_COMUGTPD, UNGT, (int)MULTI_ARG_2_DF_CMP },
21464 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comungepd", IX86_BUILTIN_COMUGEPD, UNGE, (int)MULTI_ARG_2_DF_CMP },
21465 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comordpd", IX86_BUILTIN_COMORDPD, ORDERED, (int)MULTI_ARG_2_DF_CMP },
21466 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2df3, "__builtin_ia32_comunordpd", IX86_BUILTIN_COMUNORDPD, UNORDERED, (int)MULTI_ARG_2_DF_CMP },
21468 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomeqb", IX86_BUILTIN_PCOMEQB, EQ, (int)MULTI_ARG_2_QI_CMP },
21469 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomneb", IX86_BUILTIN_PCOMNEB, NE, (int)MULTI_ARG_2_QI_CMP },
21470 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomneqb", IX86_BUILTIN_PCOMNEB, NE, (int)MULTI_ARG_2_QI_CMP },
21471 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomltb", IX86_BUILTIN_PCOMLTB, LT, (int)MULTI_ARG_2_QI_CMP },
21472 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomleb", IX86_BUILTIN_PCOMLEB, LE, (int)MULTI_ARG_2_QI_CMP },
21473 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomgtb", IX86_BUILTIN_PCOMGTB, GT, (int)MULTI_ARG_2_QI_CMP },
21474 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv16qi3, "__builtin_ia32_pcomgeb", IX86_BUILTIN_PCOMGEB, GE, (int)MULTI_ARG_2_QI_CMP },
21476 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomeqw", IX86_BUILTIN_PCOMEQW, EQ, (int)MULTI_ARG_2_HI_CMP },
21477 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomnew", IX86_BUILTIN_PCOMNEW, NE, (int)MULTI_ARG_2_HI_CMP },
21478 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomneqw", IX86_BUILTIN_PCOMNEW, NE, (int)MULTI_ARG_2_HI_CMP },
21479 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomltw", IX86_BUILTIN_PCOMLTW, LT, (int)MULTI_ARG_2_HI_CMP },
21480 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomlew", IX86_BUILTIN_PCOMLEW, LE, (int)MULTI_ARG_2_HI_CMP },
21481 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomgtw", IX86_BUILTIN_PCOMGTW, GT, (int)MULTI_ARG_2_HI_CMP },
21482 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv8hi3, "__builtin_ia32_pcomgew", IX86_BUILTIN_PCOMGEW, GE, (int)MULTI_ARG_2_HI_CMP },
21484 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomeqd", IX86_BUILTIN_PCOMEQD, EQ, (int)MULTI_ARG_2_SI_CMP },
21485 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomned", IX86_BUILTIN_PCOMNED, NE, (int)MULTI_ARG_2_SI_CMP },
21486 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomneqd", IX86_BUILTIN_PCOMNED, NE, (int)MULTI_ARG_2_SI_CMP },
21487 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomltd", IX86_BUILTIN_PCOMLTD, LT, (int)MULTI_ARG_2_SI_CMP },
21488 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomled", IX86_BUILTIN_PCOMLED, LE, (int)MULTI_ARG_2_SI_CMP },
21489 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomgtd", IX86_BUILTIN_PCOMGTD, GT, (int)MULTI_ARG_2_SI_CMP },
21490 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv4si3, "__builtin_ia32_pcomged", IX86_BUILTIN_PCOMGED, GE, (int)MULTI_ARG_2_SI_CMP },
21492 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomeqq", IX86_BUILTIN_PCOMEQQ, EQ, (int)MULTI_ARG_2_DI_CMP },
21493 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomneq", IX86_BUILTIN_PCOMNEQ, NE, (int)MULTI_ARG_2_DI_CMP },
21494 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomneqq", IX86_BUILTIN_PCOMNEQ, NE, (int)MULTI_ARG_2_DI_CMP },
21495 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomltq", IX86_BUILTIN_PCOMLTQ, LT, (int)MULTI_ARG_2_DI_CMP },
21496 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomleq", IX86_BUILTIN_PCOMLEQ, LE, (int)MULTI_ARG_2_DI_CMP },
21497 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomgtq", IX86_BUILTIN_PCOMGTQ, GT, (int)MULTI_ARG_2_DI_CMP },
21498 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmpv2di3, "__builtin_ia32_pcomgeq", IX86_BUILTIN_PCOMGEQ, GE, (int)MULTI_ARG_2_DI_CMP },
21500 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v16qi3,"__builtin_ia32_pcomequb", IX86_BUILTIN_PCOMEQUB, EQ, (int)MULTI_ARG_2_QI_CMP },
21501 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v16qi3,"__builtin_ia32_pcomneub", IX86_BUILTIN_PCOMNEUB, NE, (int)MULTI_ARG_2_QI_CMP },
21502 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v16qi3,"__builtin_ia32_pcomnequb", IX86_BUILTIN_PCOMNEUB, NE, (int)MULTI_ARG_2_QI_CMP },
21503 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv16qi3, "__builtin_ia32_pcomltub", IX86_BUILTIN_PCOMLTUB, LTU, (int)MULTI_ARG_2_QI_CMP },
21504 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv16qi3, "__builtin_ia32_pcomleub", IX86_BUILTIN_PCOMLEUB, LEU, (int)MULTI_ARG_2_QI_CMP },
21505 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv16qi3, "__builtin_ia32_pcomgtub", IX86_BUILTIN_PCOMGTUB, GTU, (int)MULTI_ARG_2_QI_CMP },
21506 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv16qi3, "__builtin_ia32_pcomgeub", IX86_BUILTIN_PCOMGEUB, GEU, (int)MULTI_ARG_2_QI_CMP },
21508 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v8hi3, "__builtin_ia32_pcomequw", IX86_BUILTIN_PCOMEQUW, EQ, (int)MULTI_ARG_2_HI_CMP },
21509 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v8hi3, "__builtin_ia32_pcomneuw", IX86_BUILTIN_PCOMNEUW, NE, (int)MULTI_ARG_2_HI_CMP },
21510 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v8hi3, "__builtin_ia32_pcomnequw", IX86_BUILTIN_PCOMNEUW, NE, (int)MULTI_ARG_2_HI_CMP },
21511 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv8hi3, "__builtin_ia32_pcomltuw", IX86_BUILTIN_PCOMLTUW, LTU, (int)MULTI_ARG_2_HI_CMP },
21512 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv8hi3, "__builtin_ia32_pcomleuw", IX86_BUILTIN_PCOMLEUW, LEU, (int)MULTI_ARG_2_HI_CMP },
21513 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv8hi3, "__builtin_ia32_pcomgtuw", IX86_BUILTIN_PCOMGTUW, GTU, (int)MULTI_ARG_2_HI_CMP },
21514 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv8hi3, "__builtin_ia32_pcomgeuw", IX86_BUILTIN_PCOMGEUW, GEU, (int)MULTI_ARG_2_HI_CMP },
21516 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v4si3, "__builtin_ia32_pcomequd", IX86_BUILTIN_PCOMEQUD, EQ, (int)MULTI_ARG_2_SI_CMP },
21517 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v4si3, "__builtin_ia32_pcomneud", IX86_BUILTIN_PCOMNEUD, NE, (int)MULTI_ARG_2_SI_CMP },
21518 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v4si3, "__builtin_ia32_pcomnequd", IX86_BUILTIN_PCOMNEUD, NE, (int)MULTI_ARG_2_SI_CMP },
21519 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv4si3, "__builtin_ia32_pcomltud", IX86_BUILTIN_PCOMLTUD, LTU, (int)MULTI_ARG_2_SI_CMP },
21520 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv4si3, "__builtin_ia32_pcomleud", IX86_BUILTIN_PCOMLEUD, LEU, (int)MULTI_ARG_2_SI_CMP },
21521 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv4si3, "__builtin_ia32_pcomgtud", IX86_BUILTIN_PCOMGTUD, GTU, (int)MULTI_ARG_2_SI_CMP },
21522 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv4si3, "__builtin_ia32_pcomgeud", IX86_BUILTIN_PCOMGEUD, GEU, (int)MULTI_ARG_2_SI_CMP },
21524 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v2di3, "__builtin_ia32_pcomequq", IX86_BUILTIN_PCOMEQUQ, EQ, (int)MULTI_ARG_2_DI_CMP },
21525 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v2di3, "__builtin_ia32_pcomneuq", IX86_BUILTIN_PCOMNEUQ, NE, (int)MULTI_ARG_2_DI_CMP },
21526 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_uns2v2di3, "__builtin_ia32_pcomnequq", IX86_BUILTIN_PCOMNEUQ, NE, (int)MULTI_ARG_2_DI_CMP },
21527 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv2di3, "__builtin_ia32_pcomltuq", IX86_BUILTIN_PCOMLTUQ, LTU, (int)MULTI_ARG_2_DI_CMP },
21528 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv2di3, "__builtin_ia32_pcomleuq", IX86_BUILTIN_PCOMLEUQ, LEU, (int)MULTI_ARG_2_DI_CMP },
21529 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv2di3, "__builtin_ia32_pcomgtuq", IX86_BUILTIN_PCOMGTUQ, GTU, (int)MULTI_ARG_2_DI_CMP },
21530 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_maskcmp_unsv2di3, "__builtin_ia32_pcomgeuq", IX86_BUILTIN_PCOMGEUQ, GEU, (int)MULTI_ARG_2_DI_CMP },
21532 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv4sf3, "__builtin_ia32_comfalsess", IX86_BUILTIN_COMFALSESS, COM_FALSE_S, (int)MULTI_ARG_2_SF_TF },
21533 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv4sf3, "__builtin_ia32_comtruess", IX86_BUILTIN_COMTRUESS, COM_TRUE_S, (int)MULTI_ARG_2_SF_TF },
21534 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv4sf3, "__builtin_ia32_comfalseps", IX86_BUILTIN_COMFALSEPS, COM_FALSE_P, (int)MULTI_ARG_2_SF_TF },
21535 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv4sf3, "__builtin_ia32_comtrueps", IX86_BUILTIN_COMTRUEPS, COM_TRUE_P, (int)MULTI_ARG_2_SF_TF },
21536 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv2df3, "__builtin_ia32_comfalsesd", IX86_BUILTIN_COMFALSESD, COM_FALSE_S, (int)MULTI_ARG_2_DF_TF },
21537 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv2df3, "__builtin_ia32_comtruesd", IX86_BUILTIN_COMTRUESD, COM_TRUE_S, (int)MULTI_ARG_2_DF_TF },
21538 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv2df3, "__builtin_ia32_comfalsepd", IX86_BUILTIN_COMFALSEPD, COM_FALSE_P, (int)MULTI_ARG_2_DF_TF },
21539 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_com_tfv2df3, "__builtin_ia32_comtruepd", IX86_BUILTIN_COMTRUEPD, COM_TRUE_P, (int)MULTI_ARG_2_DF_TF },
21541 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv16qi3, "__builtin_ia32_pcomfalseb", IX86_BUILTIN_PCOMFALSEB, PCOM_FALSE, (int)MULTI_ARG_2_QI_TF },
21542 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv8hi3, "__builtin_ia32_pcomfalsew", IX86_BUILTIN_PCOMFALSEW, PCOM_FALSE, (int)MULTI_ARG_2_HI_TF },
21543 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv4si3, "__builtin_ia32_pcomfalsed", IX86_BUILTIN_PCOMFALSED, PCOM_FALSE, (int)MULTI_ARG_2_SI_TF },
21544 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv2di3, "__builtin_ia32_pcomfalseq", IX86_BUILTIN_PCOMFALSEQ, PCOM_FALSE, (int)MULTI_ARG_2_DI_TF },
21545 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv16qi3, "__builtin_ia32_pcomfalseub",IX86_BUILTIN_PCOMFALSEUB,PCOM_FALSE, (int)MULTI_ARG_2_QI_TF },
21546 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv8hi3, "__builtin_ia32_pcomfalseuw",IX86_BUILTIN_PCOMFALSEUW,PCOM_FALSE, (int)MULTI_ARG_2_HI_TF },
21547 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv4si3, "__builtin_ia32_pcomfalseud",IX86_BUILTIN_PCOMFALSEUD,PCOM_FALSE, (int)MULTI_ARG_2_SI_TF },
21548 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv2di3, "__builtin_ia32_pcomfalseuq",IX86_BUILTIN_PCOMFALSEUQ,PCOM_FALSE, (int)MULTI_ARG_2_DI_TF },
21550 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv16qi3, "__builtin_ia32_pcomtrueb", IX86_BUILTIN_PCOMTRUEB, PCOM_TRUE, (int)MULTI_ARG_2_QI_TF },
21551 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv8hi3, "__builtin_ia32_pcomtruew", IX86_BUILTIN_PCOMTRUEW, PCOM_TRUE, (int)MULTI_ARG_2_HI_TF },
21552 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv4si3, "__builtin_ia32_pcomtrued", IX86_BUILTIN_PCOMTRUED, PCOM_TRUE, (int)MULTI_ARG_2_SI_TF },
21553 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv2di3, "__builtin_ia32_pcomtrueq", IX86_BUILTIN_PCOMTRUEQ, PCOM_TRUE, (int)MULTI_ARG_2_DI_TF },
21554 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv16qi3, "__builtin_ia32_pcomtrueub", IX86_BUILTIN_PCOMTRUEUB, PCOM_TRUE, (int)MULTI_ARG_2_QI_TF },
21555 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv8hi3, "__builtin_ia32_pcomtrueuw", IX86_BUILTIN_PCOMTRUEUW, PCOM_TRUE, (int)MULTI_ARG_2_HI_TF },
21556 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv4si3, "__builtin_ia32_pcomtrueud", IX86_BUILTIN_PCOMTRUEUD, PCOM_TRUE, (int)MULTI_ARG_2_SI_TF },
21557 { OPTION_MASK_ISA_SSE5, CODE_FOR_sse5_pcom_tfv2di3, "__builtin_ia32_pcomtrueuq", IX86_BUILTIN_PCOMTRUEUQ, PCOM_TRUE, (int)MULTI_ARG_2_DI_TF },
21560 /* Set up all the MMX/SSE builtins, even builtins for instructions that are not
21561 in the current target ISA to allow the user to compile particular modules
21562 with different target specific options that differ from the command line
21565 ix86_init_mmx_sse_builtins (void)
21567 const struct builtin_description * d;
21570 tree V16QI_type_node = build_vector_type_for_mode (char_type_node, V16QImode);
21571 tree V2SI_type_node = build_vector_type_for_mode (intSI_type_node, V2SImode);
21572 tree V1DI_type_node
21573 = build_vector_type_for_mode (long_long_integer_type_node, V1DImode);
21574 tree V2SF_type_node = build_vector_type_for_mode (float_type_node, V2SFmode);
21575 tree V2DI_type_node
21576 = build_vector_type_for_mode (long_long_integer_type_node, V2DImode);
21577 tree V2DF_type_node = build_vector_type_for_mode (double_type_node, V2DFmode);
21578 tree V4SF_type_node = build_vector_type_for_mode (float_type_node, V4SFmode);
21579 tree V4SI_type_node = build_vector_type_for_mode (intSI_type_node, V4SImode);
21580 tree V4HI_type_node = build_vector_type_for_mode (intHI_type_node, V4HImode);
21581 tree V8QI_type_node = build_vector_type_for_mode (char_type_node, V8QImode);
21582 tree V8HI_type_node = build_vector_type_for_mode (intHI_type_node, V8HImode);
21584 tree pchar_type_node = build_pointer_type (char_type_node);
21585 tree pcchar_type_node
21586 = build_pointer_type (build_type_variant (char_type_node, 1, 0));
21587 tree pfloat_type_node = build_pointer_type (float_type_node);
21588 tree pcfloat_type_node
21589 = build_pointer_type (build_type_variant (float_type_node, 1, 0));
21590 tree pv2sf_type_node = build_pointer_type (V2SF_type_node);
21591 tree pcv2sf_type_node
21592 = build_pointer_type (build_type_variant (V2SF_type_node, 1, 0));
21593 tree pv2di_type_node = build_pointer_type (V2DI_type_node);
21594 tree pdi_type_node = build_pointer_type (long_long_unsigned_type_node);
21597 tree int_ftype_v4sf_v4sf
21598 = build_function_type_list (integer_type_node,
21599 V4SF_type_node, V4SF_type_node, NULL_TREE);
21600 tree v4si_ftype_v4sf_v4sf
21601 = build_function_type_list (V4SI_type_node,
21602 V4SF_type_node, V4SF_type_node, NULL_TREE);
21603 /* MMX/SSE/integer conversions. */
21604 tree int_ftype_v4sf
21605 = build_function_type_list (integer_type_node,
21606 V4SF_type_node, NULL_TREE);
21607 tree int64_ftype_v4sf
21608 = build_function_type_list (long_long_integer_type_node,
21609 V4SF_type_node, NULL_TREE);
21610 tree int_ftype_v8qi
21611 = build_function_type_list (integer_type_node, V8QI_type_node, NULL_TREE);
21612 tree v4sf_ftype_v4sf_int
21613 = build_function_type_list (V4SF_type_node,
21614 V4SF_type_node, integer_type_node, NULL_TREE);
21615 tree v4sf_ftype_v4sf_int64
21616 = build_function_type_list (V4SF_type_node,
21617 V4SF_type_node, long_long_integer_type_node,
21619 tree v4sf_ftype_v4sf_v2si
21620 = build_function_type_list (V4SF_type_node,
21621 V4SF_type_node, V2SI_type_node, NULL_TREE);
21623 /* Miscellaneous. */
21624 tree v8qi_ftype_v4hi_v4hi
21625 = build_function_type_list (V8QI_type_node,
21626 V4HI_type_node, V4HI_type_node, NULL_TREE);
21627 tree v4hi_ftype_v2si_v2si
21628 = build_function_type_list (V4HI_type_node,
21629 V2SI_type_node, V2SI_type_node, NULL_TREE);
21630 tree v4sf_ftype_v4sf_v4sf_int
21631 = build_function_type_list (V4SF_type_node,
21632 V4SF_type_node, V4SF_type_node,
21633 integer_type_node, NULL_TREE);
21634 tree v2si_ftype_v4hi_v4hi
21635 = build_function_type_list (V2SI_type_node,
21636 V4HI_type_node, V4HI_type_node, NULL_TREE);
21637 tree v4hi_ftype_v4hi_int
21638 = build_function_type_list (V4HI_type_node,
21639 V4HI_type_node, integer_type_node, NULL_TREE);
21640 tree v2si_ftype_v2si_int
21641 = build_function_type_list (V2SI_type_node,
21642 V2SI_type_node, integer_type_node, NULL_TREE);
21643 tree v1di_ftype_v1di_int
21644 = build_function_type_list (V1DI_type_node,
21645 V1DI_type_node, integer_type_node, NULL_TREE);
21647 tree void_ftype_void
21648 = build_function_type (void_type_node, void_list_node);
21649 tree void_ftype_unsigned
21650 = build_function_type_list (void_type_node, unsigned_type_node, NULL_TREE);
21651 tree void_ftype_unsigned_unsigned
21652 = build_function_type_list (void_type_node, unsigned_type_node,
21653 unsigned_type_node, NULL_TREE);
21654 tree void_ftype_pcvoid_unsigned_unsigned
21655 = build_function_type_list (void_type_node, const_ptr_type_node,
21656 unsigned_type_node, unsigned_type_node,
21658 tree unsigned_ftype_void
21659 = build_function_type (unsigned_type_node, void_list_node);
21660 tree v2si_ftype_v4sf
21661 = build_function_type_list (V2SI_type_node, V4SF_type_node, NULL_TREE);
21662 /* Loads/stores. */
21663 tree void_ftype_v8qi_v8qi_pchar
21664 = build_function_type_list (void_type_node,
21665 V8QI_type_node, V8QI_type_node,
21666 pchar_type_node, NULL_TREE);
21667 tree v4sf_ftype_pcfloat
21668 = build_function_type_list (V4SF_type_node, pcfloat_type_node, NULL_TREE);
21669 tree v4sf_ftype_v4sf_pcv2sf
21670 = build_function_type_list (V4SF_type_node,
21671 V4SF_type_node, pcv2sf_type_node, NULL_TREE);
21672 tree void_ftype_pv2sf_v4sf
21673 = build_function_type_list (void_type_node,
21674 pv2sf_type_node, V4SF_type_node, NULL_TREE);
21675 tree void_ftype_pfloat_v4sf
21676 = build_function_type_list (void_type_node,
21677 pfloat_type_node, V4SF_type_node, NULL_TREE);
21678 tree void_ftype_pdi_di
21679 = build_function_type_list (void_type_node,
21680 pdi_type_node, long_long_unsigned_type_node,
21682 tree void_ftype_pv2di_v2di
21683 = build_function_type_list (void_type_node,
21684 pv2di_type_node, V2DI_type_node, NULL_TREE);
21685 /* Normal vector unops. */
21686 tree v4sf_ftype_v4sf
21687 = build_function_type_list (V4SF_type_node, V4SF_type_node, NULL_TREE);
21688 tree v16qi_ftype_v16qi
21689 = build_function_type_list (V16QI_type_node, V16QI_type_node, NULL_TREE);
21690 tree v8hi_ftype_v8hi
21691 = build_function_type_list (V8HI_type_node, V8HI_type_node, NULL_TREE);
21692 tree v4si_ftype_v4si
21693 = build_function_type_list (V4SI_type_node, V4SI_type_node, NULL_TREE);
21694 tree v8qi_ftype_v8qi
21695 = build_function_type_list (V8QI_type_node, V8QI_type_node, NULL_TREE);
21696 tree v4hi_ftype_v4hi
21697 = build_function_type_list (V4HI_type_node, V4HI_type_node, NULL_TREE);
21699 /* Normal vector binops. */
21700 tree v4sf_ftype_v4sf_v4sf
21701 = build_function_type_list (V4SF_type_node,
21702 V4SF_type_node, V4SF_type_node, NULL_TREE);
21703 tree v8qi_ftype_v8qi_v8qi
21704 = build_function_type_list (V8QI_type_node,
21705 V8QI_type_node, V8QI_type_node, NULL_TREE);
21706 tree v4hi_ftype_v4hi_v4hi
21707 = build_function_type_list (V4HI_type_node,
21708 V4HI_type_node, V4HI_type_node, NULL_TREE);
21709 tree v2si_ftype_v2si_v2si
21710 = build_function_type_list (V2SI_type_node,
21711 V2SI_type_node, V2SI_type_node, NULL_TREE);
21712 tree v1di_ftype_v1di_v1di
21713 = build_function_type_list (V1DI_type_node,
21714 V1DI_type_node, V1DI_type_node, NULL_TREE);
21715 tree v1di_ftype_v1di_v1di_int
21716 = build_function_type_list (V1DI_type_node,
21717 V1DI_type_node, V1DI_type_node,
21718 integer_type_node, NULL_TREE);
21719 tree v2si_ftype_v2sf
21720 = build_function_type_list (V2SI_type_node, V2SF_type_node, NULL_TREE);
21721 tree v2sf_ftype_v2si
21722 = build_function_type_list (V2SF_type_node, V2SI_type_node, NULL_TREE);
21723 tree v2si_ftype_v2si
21724 = build_function_type_list (V2SI_type_node, V2SI_type_node, NULL_TREE);
21725 tree v2sf_ftype_v2sf
21726 = build_function_type_list (V2SF_type_node, V2SF_type_node, NULL_TREE);
21727 tree v2sf_ftype_v2sf_v2sf
21728 = build_function_type_list (V2SF_type_node,
21729 V2SF_type_node, V2SF_type_node, NULL_TREE);
21730 tree v2si_ftype_v2sf_v2sf
21731 = build_function_type_list (V2SI_type_node,
21732 V2SF_type_node, V2SF_type_node, NULL_TREE);
21733 tree pint_type_node = build_pointer_type (integer_type_node);
21734 tree pdouble_type_node = build_pointer_type (double_type_node);
21735 tree pcdouble_type_node = build_pointer_type (
21736 build_type_variant (double_type_node, 1, 0));
21737 tree int_ftype_v2df_v2df
21738 = build_function_type_list (integer_type_node,
21739 V2DF_type_node, V2DF_type_node, NULL_TREE);
21741 tree void_ftype_pcvoid
21742 = build_function_type_list (void_type_node, const_ptr_type_node, NULL_TREE);
21743 tree v4sf_ftype_v4si
21744 = build_function_type_list (V4SF_type_node, V4SI_type_node, NULL_TREE);
21745 tree v4si_ftype_v4sf
21746 = build_function_type_list (V4SI_type_node, V4SF_type_node, NULL_TREE);
21747 tree v2df_ftype_v4si
21748 = build_function_type_list (V2DF_type_node, V4SI_type_node, NULL_TREE);
21749 tree v4si_ftype_v2df
21750 = build_function_type_list (V4SI_type_node, V2DF_type_node, NULL_TREE);
21751 tree v4si_ftype_v2df_v2df
21752 = build_function_type_list (V4SI_type_node,
21753 V2DF_type_node, V2DF_type_node, NULL_TREE);
21754 tree v2si_ftype_v2df
21755 = build_function_type_list (V2SI_type_node, V2DF_type_node, NULL_TREE);
21756 tree v4sf_ftype_v2df
21757 = build_function_type_list (V4SF_type_node, V2DF_type_node, NULL_TREE);
21758 tree v2df_ftype_v2si
21759 = build_function_type_list (V2DF_type_node, V2SI_type_node, NULL_TREE);
21760 tree v2df_ftype_v4sf
21761 = build_function_type_list (V2DF_type_node, V4SF_type_node, NULL_TREE);
21762 tree int_ftype_v2df
21763 = build_function_type_list (integer_type_node, V2DF_type_node, NULL_TREE);
21764 tree int64_ftype_v2df
21765 = build_function_type_list (long_long_integer_type_node,
21766 V2DF_type_node, NULL_TREE);
21767 tree v2df_ftype_v2df_int
21768 = build_function_type_list (V2DF_type_node,
21769 V2DF_type_node, integer_type_node, NULL_TREE);
21770 tree v2df_ftype_v2df_int64
21771 = build_function_type_list (V2DF_type_node,
21772 V2DF_type_node, long_long_integer_type_node,
21774 tree v4sf_ftype_v4sf_v2df
21775 = build_function_type_list (V4SF_type_node,
21776 V4SF_type_node, V2DF_type_node, NULL_TREE);
21777 tree v2df_ftype_v2df_v4sf
21778 = build_function_type_list (V2DF_type_node,
21779 V2DF_type_node, V4SF_type_node, NULL_TREE);
21780 tree v2df_ftype_v2df_v2df_int
21781 = build_function_type_list (V2DF_type_node,
21782 V2DF_type_node, V2DF_type_node,
21785 tree v2df_ftype_v2df_pcdouble
21786 = build_function_type_list (V2DF_type_node,
21787 V2DF_type_node, pcdouble_type_node, NULL_TREE);
21788 tree void_ftype_pdouble_v2df
21789 = build_function_type_list (void_type_node,
21790 pdouble_type_node, V2DF_type_node, NULL_TREE);
21791 tree void_ftype_pint_int
21792 = build_function_type_list (void_type_node,
21793 pint_type_node, integer_type_node, NULL_TREE);
21794 tree void_ftype_v16qi_v16qi_pchar
21795 = build_function_type_list (void_type_node,
21796 V16QI_type_node, V16QI_type_node,
21797 pchar_type_node, NULL_TREE);
21798 tree v2df_ftype_pcdouble
21799 = build_function_type_list (V2DF_type_node, pcdouble_type_node, NULL_TREE);
21800 tree v2df_ftype_v2df_v2df
21801 = build_function_type_list (V2DF_type_node,
21802 V2DF_type_node, V2DF_type_node, NULL_TREE);
21803 tree v16qi_ftype_v16qi_v16qi
21804 = build_function_type_list (V16QI_type_node,
21805 V16QI_type_node, V16QI_type_node, NULL_TREE);
21806 tree v8hi_ftype_v8hi_v8hi
21807 = build_function_type_list (V8HI_type_node,
21808 V8HI_type_node, V8HI_type_node, NULL_TREE);
21809 tree v4si_ftype_v4si_v4si
21810 = build_function_type_list (V4SI_type_node,
21811 V4SI_type_node, V4SI_type_node, NULL_TREE);
21812 tree v2di_ftype_v2di_v2di
21813 = build_function_type_list (V2DI_type_node,
21814 V2DI_type_node, V2DI_type_node, NULL_TREE);
21815 tree v2di_ftype_v2df_v2df
21816 = build_function_type_list (V2DI_type_node,
21817 V2DF_type_node, V2DF_type_node, NULL_TREE);
21818 tree v2df_ftype_v2df
21819 = build_function_type_list (V2DF_type_node, V2DF_type_node, NULL_TREE);
21820 tree v2di_ftype_v2di_int
21821 = build_function_type_list (V2DI_type_node,
21822 V2DI_type_node, integer_type_node, NULL_TREE);
21823 tree v2di_ftype_v2di_v2di_int
21824 = build_function_type_list (V2DI_type_node, V2DI_type_node,
21825 V2DI_type_node, integer_type_node, NULL_TREE);
21826 tree v4si_ftype_v4si_int
21827 = build_function_type_list (V4SI_type_node,
21828 V4SI_type_node, integer_type_node, NULL_TREE);
21829 tree v8hi_ftype_v8hi_int
21830 = build_function_type_list (V8HI_type_node,
21831 V8HI_type_node, integer_type_node, NULL_TREE);
21832 tree v4si_ftype_v8hi_v8hi
21833 = build_function_type_list (V4SI_type_node,
21834 V8HI_type_node, V8HI_type_node, NULL_TREE);
21835 tree v1di_ftype_v8qi_v8qi
21836 = build_function_type_list (V1DI_type_node,
21837 V8QI_type_node, V8QI_type_node, NULL_TREE);
21838 tree v1di_ftype_v2si_v2si
21839 = build_function_type_list (V1DI_type_node,
21840 V2SI_type_node, V2SI_type_node, NULL_TREE);
21841 tree v2di_ftype_v16qi_v16qi
21842 = build_function_type_list (V2DI_type_node,
21843 V16QI_type_node, V16QI_type_node, NULL_TREE);
21844 tree v2di_ftype_v4si_v4si
21845 = build_function_type_list (V2DI_type_node,
21846 V4SI_type_node, V4SI_type_node, NULL_TREE);
21847 tree int_ftype_v16qi
21848 = build_function_type_list (integer_type_node, V16QI_type_node, NULL_TREE);
21849 tree v16qi_ftype_pcchar
21850 = build_function_type_list (V16QI_type_node, pcchar_type_node, NULL_TREE);
21851 tree void_ftype_pchar_v16qi
21852 = build_function_type_list (void_type_node,
21853 pchar_type_node, V16QI_type_node, NULL_TREE);
21855 tree v2di_ftype_v2di_unsigned_unsigned
21856 = build_function_type_list (V2DI_type_node, V2DI_type_node,
21857 unsigned_type_node, unsigned_type_node,
21859 tree v2di_ftype_v2di_v2di_unsigned_unsigned
21860 = build_function_type_list (V2DI_type_node, V2DI_type_node, V2DI_type_node,
21861 unsigned_type_node, unsigned_type_node,
21863 tree v2di_ftype_v2di_v16qi
21864 = build_function_type_list (V2DI_type_node, V2DI_type_node, V16QI_type_node,
21866 tree v2df_ftype_v2df_v2df_v2df
21867 = build_function_type_list (V2DF_type_node,
21868 V2DF_type_node, V2DF_type_node,
21869 V2DF_type_node, NULL_TREE);
21870 tree v4sf_ftype_v4sf_v4sf_v4sf
21871 = build_function_type_list (V4SF_type_node,
21872 V4SF_type_node, V4SF_type_node,
21873 V4SF_type_node, NULL_TREE);
21874 tree v8hi_ftype_v16qi
21875 = build_function_type_list (V8HI_type_node, V16QI_type_node,
21877 tree v4si_ftype_v16qi
21878 = build_function_type_list (V4SI_type_node, V16QI_type_node,
21880 tree v2di_ftype_v16qi
21881 = build_function_type_list (V2DI_type_node, V16QI_type_node,
21883 tree v4si_ftype_v8hi
21884 = build_function_type_list (V4SI_type_node, V8HI_type_node,
21886 tree v2di_ftype_v8hi
21887 = build_function_type_list (V2DI_type_node, V8HI_type_node,
21889 tree v2di_ftype_v4si
21890 = build_function_type_list (V2DI_type_node, V4SI_type_node,
21892 tree v2di_ftype_pv2di
21893 = build_function_type_list (V2DI_type_node, pv2di_type_node,
21895 tree v16qi_ftype_v16qi_v16qi_int
21896 = build_function_type_list (V16QI_type_node, V16QI_type_node,
21897 V16QI_type_node, integer_type_node,
21899 tree v16qi_ftype_v16qi_v16qi_v16qi
21900 = build_function_type_list (V16QI_type_node, V16QI_type_node,
21901 V16QI_type_node, V16QI_type_node,
21903 tree v8hi_ftype_v8hi_v8hi_int
21904 = build_function_type_list (V8HI_type_node, V8HI_type_node,
21905 V8HI_type_node, integer_type_node,
21907 tree v4si_ftype_v4si_v4si_int
21908 = build_function_type_list (V4SI_type_node, V4SI_type_node,
21909 V4SI_type_node, integer_type_node,
21911 tree int_ftype_v2di_v2di
21912 = build_function_type_list (integer_type_node,
21913 V2DI_type_node, V2DI_type_node,
21915 tree int_ftype_v16qi_int_v16qi_int_int
21916 = build_function_type_list (integer_type_node,
21923 tree v16qi_ftype_v16qi_int_v16qi_int_int
21924 = build_function_type_list (V16QI_type_node,
21931 tree int_ftype_v16qi_v16qi_int
21932 = build_function_type_list (integer_type_node,
21938 /* SSE5 instructions */
21939 tree v2di_ftype_v2di_v2di_v2di
21940 = build_function_type_list (V2DI_type_node,
21946 tree v4si_ftype_v4si_v4si_v4si
21947 = build_function_type_list (V4SI_type_node,
21953 tree v4si_ftype_v4si_v4si_v2di
21954 = build_function_type_list (V4SI_type_node,
21960 tree v8hi_ftype_v8hi_v8hi_v8hi
21961 = build_function_type_list (V8HI_type_node,
21967 tree v8hi_ftype_v8hi_v8hi_v4si
21968 = build_function_type_list (V8HI_type_node,
21974 tree v2df_ftype_v2df_v2df_v16qi
21975 = build_function_type_list (V2DF_type_node,
21981 tree v4sf_ftype_v4sf_v4sf_v16qi
21982 = build_function_type_list (V4SF_type_node,
21988 tree v2di_ftype_v2di_si
21989 = build_function_type_list (V2DI_type_node,
21994 tree v4si_ftype_v4si_si
21995 = build_function_type_list (V4SI_type_node,
22000 tree v8hi_ftype_v8hi_si
22001 = build_function_type_list (V8HI_type_node,
22006 tree v16qi_ftype_v16qi_si
22007 = build_function_type_list (V16QI_type_node,
22011 tree v4sf_ftype_v4hi
22012 = build_function_type_list (V4SF_type_node,
22016 tree v4hi_ftype_v4sf
22017 = build_function_type_list (V4HI_type_node,
22021 tree v2di_ftype_v2di
22022 = build_function_type_list (V2DI_type_node, V2DI_type_node, NULL_TREE);
22024 tree v16qi_ftype_v8hi_v8hi
22025 = build_function_type_list (V16QI_type_node,
22026 V8HI_type_node, V8HI_type_node,
22028 tree v8hi_ftype_v4si_v4si
22029 = build_function_type_list (V8HI_type_node,
22030 V4SI_type_node, V4SI_type_node,
22032 tree v8hi_ftype_v16qi_v16qi
22033 = build_function_type_list (V8HI_type_node,
22034 V16QI_type_node, V16QI_type_node,
22036 tree v4hi_ftype_v8qi_v8qi
22037 = build_function_type_list (V4HI_type_node,
22038 V8QI_type_node, V8QI_type_node,
22040 tree unsigned_ftype_unsigned_uchar
22041 = build_function_type_list (unsigned_type_node,
22042 unsigned_type_node,
22043 unsigned_char_type_node,
22045 tree unsigned_ftype_unsigned_ushort
22046 = build_function_type_list (unsigned_type_node,
22047 unsigned_type_node,
22048 short_unsigned_type_node,
22050 tree unsigned_ftype_unsigned_unsigned
22051 = build_function_type_list (unsigned_type_node,
22052 unsigned_type_node,
22053 unsigned_type_node,
22055 tree uint64_ftype_uint64_uint64
22056 = build_function_type_list (long_long_unsigned_type_node,
22057 long_long_unsigned_type_node,
22058 long_long_unsigned_type_node,
22060 tree float_ftype_float
22061 = build_function_type_list (float_type_node,
22066 tree V32QI_type_node = build_vector_type_for_mode (char_type_node,
22068 tree V8SI_type_node = build_vector_type_for_mode (intSI_type_node,
22070 tree V8SF_type_node = build_vector_type_for_mode (float_type_node,
22072 tree V4DI_type_node = build_vector_type_for_mode (long_long_integer_type_node,
22074 tree V4DF_type_node = build_vector_type_for_mode (double_type_node,
22076 tree v8sf_ftype_v8sf
22077 = build_function_type_list (V8SF_type_node,
22080 tree v8si_ftype_v8sf
22081 = build_function_type_list (V8SI_type_node,
22084 tree v8sf_ftype_v8si
22085 = build_function_type_list (V8SF_type_node,
22088 tree v4si_ftype_v4df
22089 = build_function_type_list (V4SI_type_node,
22092 tree v4df_ftype_v4df
22093 = build_function_type_list (V4DF_type_node,
22096 tree v4df_ftype_v4si
22097 = build_function_type_list (V4DF_type_node,
22100 tree v4df_ftype_v4sf
22101 = build_function_type_list (V4DF_type_node,
22104 tree v4sf_ftype_v4df
22105 = build_function_type_list (V4SF_type_node,
22108 tree v8sf_ftype_v8sf_v8sf
22109 = build_function_type_list (V8SF_type_node,
22110 V8SF_type_node, V8SF_type_node,
22112 tree v4df_ftype_v4df_v4df
22113 = build_function_type_list (V4DF_type_node,
22114 V4DF_type_node, V4DF_type_node,
22116 tree v8sf_ftype_v8sf_int
22117 = build_function_type_list (V8SF_type_node,
22118 V8SF_type_node, integer_type_node,
22120 tree v4si_ftype_v8si_int
22121 = build_function_type_list (V4SI_type_node,
22122 V8SI_type_node, integer_type_node,
22124 tree v4df_ftype_v4df_int
22125 = build_function_type_list (V4DF_type_node,
22126 V4DF_type_node, integer_type_node,
22128 tree v4sf_ftype_v8sf_int
22129 = build_function_type_list (V4SF_type_node,
22130 V8SF_type_node, integer_type_node,
22132 tree v2df_ftype_v4df_int
22133 = build_function_type_list (V2DF_type_node,
22134 V4DF_type_node, integer_type_node,
22136 tree v8sf_ftype_v8sf_v8sf_int
22137 = build_function_type_list (V8SF_type_node,
22138 V8SF_type_node, V8SF_type_node,
22141 tree v8sf_ftype_v8sf_v8sf_v8sf
22142 = build_function_type_list (V8SF_type_node,
22143 V8SF_type_node, V8SF_type_node,
22146 tree v4df_ftype_v4df_v4df_v4df
22147 = build_function_type_list (V4DF_type_node,
22148 V4DF_type_node, V4DF_type_node,
22151 tree v8si_ftype_v8si_v8si_int
22152 = build_function_type_list (V8SI_type_node,
22153 V8SI_type_node, V8SI_type_node,
22156 tree v4df_ftype_v4df_v4df_int
22157 = build_function_type_list (V4DF_type_node,
22158 V4DF_type_node, V4DF_type_node,
22161 tree v8sf_ftype_pcfloat
22162 = build_function_type_list (V8SF_type_node,
22165 tree v4df_ftype_pcdouble
22166 = build_function_type_list (V4DF_type_node,
22167 pcdouble_type_node,
22169 tree pcv4sf_type_node
22170 = build_pointer_type (build_type_variant (V4SF_type_node, 1, 0));
22171 tree pcv2df_type_node
22172 = build_pointer_type (build_type_variant (V2DF_type_node, 1, 0));
22173 tree v8sf_ftype_pcv4sf
22174 = build_function_type_list (V8SF_type_node,
22177 tree v4df_ftype_pcv2df
22178 = build_function_type_list (V4DF_type_node,
22181 tree v32qi_ftype_pcchar
22182 = build_function_type_list (V32QI_type_node,
22185 tree void_ftype_pchar_v32qi
22186 = build_function_type_list (void_type_node,
22187 pchar_type_node, V32QI_type_node,
22189 tree v8si_ftype_v8si_v4si_int
22190 = build_function_type_list (V8SI_type_node,
22191 V8SI_type_node, V4SI_type_node,
22194 tree pv4di_type_node = build_pointer_type (V4DI_type_node);
22195 tree void_ftype_pv4di_v4di
22196 = build_function_type_list (void_type_node,
22197 pv4di_type_node, V4DI_type_node,
22199 tree v8sf_ftype_v8sf_v4sf_int
22200 = build_function_type_list (V8SF_type_node,
22201 V8SF_type_node, V4SF_type_node,
22204 tree v4df_ftype_v4df_v2df_int
22205 = build_function_type_list (V4DF_type_node,
22206 V4DF_type_node, V2DF_type_node,
22209 tree void_ftype_pfloat_v8sf
22210 = build_function_type_list (void_type_node,
22211 pfloat_type_node, V8SF_type_node,
22213 tree void_ftype_pdouble_v4df
22214 = build_function_type_list (void_type_node,
22215 pdouble_type_node, V4DF_type_node,
22217 tree pv8sf_type_node = build_pointer_type (V8SF_type_node);
22218 tree pv4sf_type_node = build_pointer_type (V4SF_type_node);
22219 tree pv4df_type_node = build_pointer_type (V4DF_type_node);
22220 tree pv2df_type_node = build_pointer_type (V2DF_type_node);
22221 tree pcv8sf_type_node
22222 = build_pointer_type (build_type_variant (V8SF_type_node, 1, 0));
22223 tree pcv4df_type_node
22224 = build_pointer_type (build_type_variant (V4DF_type_node, 1, 0));
22225 tree v8sf_ftype_pcv8sf_v8sf
22226 = build_function_type_list (V8SF_type_node,
22227 pcv8sf_type_node, V8SF_type_node,
22229 tree v4df_ftype_pcv4df_v4df
22230 = build_function_type_list (V4DF_type_node,
22231 pcv4df_type_node, V4DF_type_node,
22233 tree v4sf_ftype_pcv4sf_v4sf
22234 = build_function_type_list (V4SF_type_node,
22235 pcv4sf_type_node, V4SF_type_node,
22237 tree v2df_ftype_pcv2df_v2df
22238 = build_function_type_list (V2DF_type_node,
22239 pcv2df_type_node, V2DF_type_node,
22241 tree void_ftype_pv8sf_v8sf_v8sf
22242 = build_function_type_list (void_type_node,
22243 pv8sf_type_node, V8SF_type_node,
22246 tree void_ftype_pv4df_v4df_v4df
22247 = build_function_type_list (void_type_node,
22248 pv4df_type_node, V4DF_type_node,
22251 tree void_ftype_pv4sf_v4sf_v4sf
22252 = build_function_type_list (void_type_node,
22253 pv4sf_type_node, V4SF_type_node,
22256 tree void_ftype_pv2df_v2df_v2df
22257 = build_function_type_list (void_type_node,
22258 pv2df_type_node, V2DF_type_node,
22261 tree v4df_ftype_v2df
22262 = build_function_type_list (V4DF_type_node,
22265 tree v8sf_ftype_v4sf
22266 = build_function_type_list (V8SF_type_node,
22269 tree v8si_ftype_v4si
22270 = build_function_type_list (V8SI_type_node,
22273 tree v2df_ftype_v4df
22274 = build_function_type_list (V2DF_type_node,
22277 tree v4sf_ftype_v8sf
22278 = build_function_type_list (V4SF_type_node,
22281 tree v4si_ftype_v8si
22282 = build_function_type_list (V4SI_type_node,
22285 tree int_ftype_v4df
22286 = build_function_type_list (integer_type_node,
22289 tree int_ftype_v8sf
22290 = build_function_type_list (integer_type_node,
22293 tree int_ftype_v8sf_v8sf
22294 = build_function_type_list (integer_type_node,
22295 V8SF_type_node, V8SF_type_node,
22297 tree int_ftype_v4di_v4di
22298 = build_function_type_list (integer_type_node,
22299 V4DI_type_node, V4DI_type_node,
22301 tree int_ftype_v4df_v4df
22302 = build_function_type_list (integer_type_node,
22303 V4DF_type_node, V4DF_type_node,
22305 tree v8sf_ftype_v8sf_v8si
22306 = build_function_type_list (V8SF_type_node,
22307 V8SF_type_node, V8SI_type_node,
22309 tree v4df_ftype_v4df_v4di
22310 = build_function_type_list (V4DF_type_node,
22311 V4DF_type_node, V4DI_type_node,
22313 tree v4sf_ftype_v4sf_v4si
22314 = build_function_type_list (V4SF_type_node,
22315 V4SF_type_node, V4SI_type_node, NULL_TREE);
22316 tree v2df_ftype_v2df_v2di
22317 = build_function_type_list (V2DF_type_node,
22318 V2DF_type_node, V2DI_type_node, NULL_TREE);
22322 /* Add all special builtins with variable number of operands. */
22323 for (i = 0, d = bdesc_special_args;
22324 i < ARRAY_SIZE (bdesc_special_args);
22332 switch ((enum ix86_special_builtin_type) d->flag)
22334 case VOID_FTYPE_VOID:
22335 type = void_ftype_void;
22337 case V32QI_FTYPE_PCCHAR:
22338 type = v32qi_ftype_pcchar;
22340 case V16QI_FTYPE_PCCHAR:
22341 type = v16qi_ftype_pcchar;
22343 case V8SF_FTYPE_PCV4SF:
22344 type = v8sf_ftype_pcv4sf;
22346 case V8SF_FTYPE_PCFLOAT:
22347 type = v8sf_ftype_pcfloat;
22349 case V4DF_FTYPE_PCV2DF:
22350 type = v4df_ftype_pcv2df;
22352 case V4DF_FTYPE_PCDOUBLE:
22353 type = v4df_ftype_pcdouble;
22355 case V4SF_FTYPE_PCFLOAT:
22356 type = v4sf_ftype_pcfloat;
22358 case V2DI_FTYPE_PV2DI:
22359 type = v2di_ftype_pv2di;
22361 case V2DF_FTYPE_PCDOUBLE:
22362 type = v2df_ftype_pcdouble;
22364 case V8SF_FTYPE_PCV8SF_V8SF:
22365 type = v8sf_ftype_pcv8sf_v8sf;
22367 case V4DF_FTYPE_PCV4DF_V4DF:
22368 type = v4df_ftype_pcv4df_v4df;
22370 case V4SF_FTYPE_V4SF_PCV2SF:
22371 type = v4sf_ftype_v4sf_pcv2sf;
22373 case V4SF_FTYPE_PCV4SF_V4SF:
22374 type = v4sf_ftype_pcv4sf_v4sf;
22376 case V2DF_FTYPE_V2DF_PCDOUBLE:
22377 type = v2df_ftype_v2df_pcdouble;
22379 case V2DF_FTYPE_PCV2DF_V2DF:
22380 type = v2df_ftype_pcv2df_v2df;
22382 case VOID_FTYPE_PV2SF_V4SF:
22383 type = void_ftype_pv2sf_v4sf;
22385 case VOID_FTYPE_PV4DI_V4DI:
22386 type = void_ftype_pv4di_v4di;
22388 case VOID_FTYPE_PV2DI_V2DI:
22389 type = void_ftype_pv2di_v2di;
22391 case VOID_FTYPE_PCHAR_V32QI:
22392 type = void_ftype_pchar_v32qi;
22394 case VOID_FTYPE_PCHAR_V16QI:
22395 type = void_ftype_pchar_v16qi;
22397 case VOID_FTYPE_PFLOAT_V8SF:
22398 type = void_ftype_pfloat_v8sf;
22400 case VOID_FTYPE_PFLOAT_V4SF:
22401 type = void_ftype_pfloat_v4sf;
22403 case VOID_FTYPE_PDOUBLE_V4DF:
22404 type = void_ftype_pdouble_v4df;
22406 case VOID_FTYPE_PDOUBLE_V2DF:
22407 type = void_ftype_pdouble_v2df;
22409 case VOID_FTYPE_PDI_DI:
22410 type = void_ftype_pdi_di;
22412 case VOID_FTYPE_PINT_INT:
22413 type = void_ftype_pint_int;
22415 case VOID_FTYPE_PV8SF_V8SF_V8SF:
22416 type = void_ftype_pv8sf_v8sf_v8sf;
22418 case VOID_FTYPE_PV4DF_V4DF_V4DF:
22419 type = void_ftype_pv4df_v4df_v4df;
22421 case VOID_FTYPE_PV4SF_V4SF_V4SF:
22422 type = void_ftype_pv4sf_v4sf_v4sf;
22424 case VOID_FTYPE_PV2DF_V2DF_V2DF:
22425 type = void_ftype_pv2df_v2df_v2df;
22428 gcc_unreachable ();
22431 def_builtin (d->mask, d->name, type, d->code);
22434 /* Add all builtins with variable number of operands. */
22435 for (i = 0, d = bdesc_args;
22436 i < ARRAY_SIZE (bdesc_args);
22444 switch ((enum ix86_builtin_type) d->flag)
22446 case FLOAT_FTYPE_FLOAT:
22447 type = float_ftype_float;
22449 case INT_FTYPE_V8SF_V8SF_PTEST:
22450 type = int_ftype_v8sf_v8sf;
22452 case INT_FTYPE_V4DI_V4DI_PTEST:
22453 type = int_ftype_v4di_v4di;
22455 case INT_FTYPE_V4DF_V4DF_PTEST:
22456 type = int_ftype_v4df_v4df;
22458 case INT_FTYPE_V4SF_V4SF_PTEST:
22459 type = int_ftype_v4sf_v4sf;
22461 case INT_FTYPE_V2DI_V2DI_PTEST:
22462 type = int_ftype_v2di_v2di;
22464 case INT_FTYPE_V2DF_V2DF_PTEST:
22465 type = int_ftype_v2df_v2df;
22467 case INT64_FTYPE_V4SF:
22468 type = int64_ftype_v4sf;
22470 case INT64_FTYPE_V2DF:
22471 type = int64_ftype_v2df;
22473 case INT_FTYPE_V16QI:
22474 type = int_ftype_v16qi;
22476 case INT_FTYPE_V8QI:
22477 type = int_ftype_v8qi;
22479 case INT_FTYPE_V8SF:
22480 type = int_ftype_v8sf;
22482 case INT_FTYPE_V4DF:
22483 type = int_ftype_v4df;
22485 case INT_FTYPE_V4SF:
22486 type = int_ftype_v4sf;
22488 case INT_FTYPE_V2DF:
22489 type = int_ftype_v2df;
22491 case V16QI_FTYPE_V16QI:
22492 type = v16qi_ftype_v16qi;
22494 case V8SI_FTYPE_V8SF:
22495 type = v8si_ftype_v8sf;
22497 case V8SI_FTYPE_V4SI:
22498 type = v8si_ftype_v4si;
22500 case V8HI_FTYPE_V8HI:
22501 type = v8hi_ftype_v8hi;
22503 case V8HI_FTYPE_V16QI:
22504 type = v8hi_ftype_v16qi;
22506 case V8QI_FTYPE_V8QI:
22507 type = v8qi_ftype_v8qi;
22509 case V8SF_FTYPE_V8SF:
22510 type = v8sf_ftype_v8sf;
22512 case V8SF_FTYPE_V8SI:
22513 type = v8sf_ftype_v8si;
22515 case V8SF_FTYPE_V4SF:
22516 type = v8sf_ftype_v4sf;
22518 case V4SI_FTYPE_V4DF:
22519 type = v4si_ftype_v4df;
22521 case V4SI_FTYPE_V4SI:
22522 type = v4si_ftype_v4si;
22524 case V4SI_FTYPE_V16QI:
22525 type = v4si_ftype_v16qi;
22527 case V4SI_FTYPE_V8SI:
22528 type = v4si_ftype_v8si;
22530 case V4SI_FTYPE_V8HI:
22531 type = v4si_ftype_v8hi;
22533 case V4SI_FTYPE_V4SF:
22534 type = v4si_ftype_v4sf;
22536 case V4SI_FTYPE_V2DF:
22537 type = v4si_ftype_v2df;
22539 case V4HI_FTYPE_V4HI:
22540 type = v4hi_ftype_v4hi;
22542 case V4DF_FTYPE_V4DF:
22543 type = v4df_ftype_v4df;
22545 case V4DF_FTYPE_V4SI:
22546 type = v4df_ftype_v4si;
22548 case V4DF_FTYPE_V4SF:
22549 type = v4df_ftype_v4sf;
22551 case V4DF_FTYPE_V2DF:
22552 type = v4df_ftype_v2df;
22554 case V4SF_FTYPE_V4SF:
22555 case V4SF_FTYPE_V4SF_VEC_MERGE:
22556 type = v4sf_ftype_v4sf;
22558 case V4SF_FTYPE_V8SF:
22559 type = v4sf_ftype_v8sf;
22561 case V4SF_FTYPE_V4SI:
22562 type = v4sf_ftype_v4si;
22564 case V4SF_FTYPE_V4DF:
22565 type = v4sf_ftype_v4df;
22567 case V4SF_FTYPE_V2DF:
22568 type = v4sf_ftype_v2df;
22570 case V2DI_FTYPE_V2DI:
22571 type = v2di_ftype_v2di;
22573 case V2DI_FTYPE_V16QI:
22574 type = v2di_ftype_v16qi;
22576 case V2DI_FTYPE_V8HI:
22577 type = v2di_ftype_v8hi;
22579 case V2DI_FTYPE_V4SI:
22580 type = v2di_ftype_v4si;
22582 case V2SI_FTYPE_V2SI:
22583 type = v2si_ftype_v2si;
22585 case V2SI_FTYPE_V4SF:
22586 type = v2si_ftype_v4sf;
22588 case V2SI_FTYPE_V2DF:
22589 type = v2si_ftype_v2df;
22591 case V2SI_FTYPE_V2SF:
22592 type = v2si_ftype_v2sf;
22594 case V2DF_FTYPE_V4DF:
22595 type = v2df_ftype_v4df;
22597 case V2DF_FTYPE_V4SF:
22598 type = v2df_ftype_v4sf;
22600 case V2DF_FTYPE_V2DF:
22601 case V2DF_FTYPE_V2DF_VEC_MERGE:
22602 type = v2df_ftype_v2df;
22604 case V2DF_FTYPE_V2SI:
22605 type = v2df_ftype_v2si;
22607 case V2DF_FTYPE_V4SI:
22608 type = v2df_ftype_v4si;
22610 case V2SF_FTYPE_V2SF:
22611 type = v2sf_ftype_v2sf;
22613 case V2SF_FTYPE_V2SI:
22614 type = v2sf_ftype_v2si;
22616 case V16QI_FTYPE_V16QI_V16QI:
22617 type = v16qi_ftype_v16qi_v16qi;
22619 case V16QI_FTYPE_V8HI_V8HI:
22620 type = v16qi_ftype_v8hi_v8hi;
22622 case V8QI_FTYPE_V8QI_V8QI:
22623 type = v8qi_ftype_v8qi_v8qi;
22625 case V8QI_FTYPE_V4HI_V4HI:
22626 type = v8qi_ftype_v4hi_v4hi;
22628 case V8HI_FTYPE_V8HI_V8HI:
22629 case V8HI_FTYPE_V8HI_V8HI_COUNT:
22630 type = v8hi_ftype_v8hi_v8hi;
22632 case V8HI_FTYPE_V16QI_V16QI:
22633 type = v8hi_ftype_v16qi_v16qi;
22635 case V8HI_FTYPE_V4SI_V4SI:
22636 type = v8hi_ftype_v4si_v4si;
22638 case V8HI_FTYPE_V8HI_SI_COUNT:
22639 type = v8hi_ftype_v8hi_int;
22641 case V8SF_FTYPE_V8SF_V8SF:
22642 type = v8sf_ftype_v8sf_v8sf;
22644 case V8SF_FTYPE_V8SF_V8SI:
22645 type = v8sf_ftype_v8sf_v8si;
22647 case V4SI_FTYPE_V4SI_V4SI:
22648 case V4SI_FTYPE_V4SI_V4SI_COUNT:
22649 type = v4si_ftype_v4si_v4si;
22651 case V4SI_FTYPE_V8HI_V8HI:
22652 type = v4si_ftype_v8hi_v8hi;
22654 case V4SI_FTYPE_V4SF_V4SF:
22655 type = v4si_ftype_v4sf_v4sf;
22657 case V4SI_FTYPE_V2DF_V2DF:
22658 type = v4si_ftype_v2df_v2df;
22660 case V4SI_FTYPE_V4SI_SI_COUNT:
22661 type = v4si_ftype_v4si_int;
22663 case V4HI_FTYPE_V4HI_V4HI:
22664 case V4HI_FTYPE_V4HI_V4HI_COUNT:
22665 type = v4hi_ftype_v4hi_v4hi;
22667 case V4HI_FTYPE_V8QI_V8QI:
22668 type = v4hi_ftype_v8qi_v8qi;
22670 case V4HI_FTYPE_V2SI_V2SI:
22671 type = v4hi_ftype_v2si_v2si;
22673 case V4HI_FTYPE_V4HI_SI_COUNT:
22674 type = v4hi_ftype_v4hi_int;
22676 case V4DF_FTYPE_V4DF_V4DF:
22677 type = v4df_ftype_v4df_v4df;
22679 case V4DF_FTYPE_V4DF_V4DI:
22680 type = v4df_ftype_v4df_v4di;
22682 case V4SF_FTYPE_V4SF_V4SF:
22683 case V4SF_FTYPE_V4SF_V4SF_SWAP:
22684 type = v4sf_ftype_v4sf_v4sf;
22686 case V4SF_FTYPE_V4SF_V4SI:
22687 type = v4sf_ftype_v4sf_v4si;
22689 case V4SF_FTYPE_V4SF_V2SI:
22690 type = v4sf_ftype_v4sf_v2si;
22692 case V4SF_FTYPE_V4SF_V2DF:
22693 type = v4sf_ftype_v4sf_v2df;
22695 case V4SF_FTYPE_V4SF_DI:
22696 type = v4sf_ftype_v4sf_int64;
22698 case V4SF_FTYPE_V4SF_SI:
22699 type = v4sf_ftype_v4sf_int;
22701 case V2DI_FTYPE_V2DI_V2DI:
22702 case V2DI_FTYPE_V2DI_V2DI_COUNT:
22703 type = v2di_ftype_v2di_v2di;
22705 case V2DI_FTYPE_V16QI_V16QI:
22706 type = v2di_ftype_v16qi_v16qi;
22708 case V2DI_FTYPE_V4SI_V4SI:
22709 type = v2di_ftype_v4si_v4si;
22711 case V2DI_FTYPE_V2DI_V16QI:
22712 type = v2di_ftype_v2di_v16qi;
22714 case V2DI_FTYPE_V2DF_V2DF:
22715 type = v2di_ftype_v2df_v2df;
22717 case V2DI_FTYPE_V2DI_SI_COUNT:
22718 type = v2di_ftype_v2di_int;
22720 case V2SI_FTYPE_V2SI_V2SI:
22721 case V2SI_FTYPE_V2SI_V2SI_COUNT:
22722 type = v2si_ftype_v2si_v2si;
22724 case V2SI_FTYPE_V4HI_V4HI:
22725 type = v2si_ftype_v4hi_v4hi;
22727 case V2SI_FTYPE_V2SF_V2SF:
22728 type = v2si_ftype_v2sf_v2sf;
22730 case V2SI_FTYPE_V2SI_SI_COUNT:
22731 type = v2si_ftype_v2si_int;
22733 case V2DF_FTYPE_V2DF_V2DF:
22734 case V2DF_FTYPE_V2DF_V2DF_SWAP:
22735 type = v2df_ftype_v2df_v2df;
22737 case V2DF_FTYPE_V2DF_V4SF:
22738 type = v2df_ftype_v2df_v4sf;
22740 case V2DF_FTYPE_V2DF_V2DI:
22741 type = v2df_ftype_v2df_v2di;
22743 case V2DF_FTYPE_V2DF_DI:
22744 type = v2df_ftype_v2df_int64;
22746 case V2DF_FTYPE_V2DF_SI:
22747 type = v2df_ftype_v2df_int;
22749 case V2SF_FTYPE_V2SF_V2SF:
22750 type = v2sf_ftype_v2sf_v2sf;
22752 case V1DI_FTYPE_V1DI_V1DI:
22753 case V1DI_FTYPE_V1DI_V1DI_COUNT:
22754 type = v1di_ftype_v1di_v1di;
22756 case V1DI_FTYPE_V8QI_V8QI:
22757 type = v1di_ftype_v8qi_v8qi;
22759 case V1DI_FTYPE_V2SI_V2SI:
22760 type = v1di_ftype_v2si_v2si;
22762 case V1DI_FTYPE_V1DI_SI_COUNT:
22763 type = v1di_ftype_v1di_int;
22765 case UINT64_FTYPE_UINT64_UINT64:
22766 type = uint64_ftype_uint64_uint64;
22768 case UINT_FTYPE_UINT_UINT:
22769 type = unsigned_ftype_unsigned_unsigned;
22771 case UINT_FTYPE_UINT_USHORT:
22772 type = unsigned_ftype_unsigned_ushort;
22774 case UINT_FTYPE_UINT_UCHAR:
22775 type = unsigned_ftype_unsigned_uchar;
22777 case V8HI_FTYPE_V8HI_INT:
22778 type = v8hi_ftype_v8hi_int;
22780 case V8SF_FTYPE_V8SF_INT:
22781 type = v8sf_ftype_v8sf_int;
22783 case V4SI_FTYPE_V4SI_INT:
22784 type = v4si_ftype_v4si_int;
22786 case V4SI_FTYPE_V8SI_INT:
22787 type = v4si_ftype_v8si_int;
22789 case V4HI_FTYPE_V4HI_INT:
22790 type = v4hi_ftype_v4hi_int;
22792 case V4DF_FTYPE_V4DF_INT:
22793 type = v4df_ftype_v4df_int;
22795 case V4SF_FTYPE_V4SF_INT:
22796 type = v4sf_ftype_v4sf_int;
22798 case V4SF_FTYPE_V8SF_INT:
22799 type = v4sf_ftype_v8sf_int;
22801 case V2DI_FTYPE_V2DI_INT:
22802 case V2DI2TI_FTYPE_V2DI_INT:
22803 type = v2di_ftype_v2di_int;
22805 case V2DF_FTYPE_V2DF_INT:
22806 type = v2df_ftype_v2df_int;
22808 case V2DF_FTYPE_V4DF_INT:
22809 type = v2df_ftype_v4df_int;
22811 case V16QI_FTYPE_V16QI_V16QI_V16QI:
22812 type = v16qi_ftype_v16qi_v16qi_v16qi;
22814 case V8SF_FTYPE_V8SF_V8SF_V8SF:
22815 type = v8sf_ftype_v8sf_v8sf_v8sf;
22817 case V4DF_FTYPE_V4DF_V4DF_V4DF:
22818 type = v4df_ftype_v4df_v4df_v4df;
22820 case V4SF_FTYPE_V4SF_V4SF_V4SF:
22821 type = v4sf_ftype_v4sf_v4sf_v4sf;
22823 case V2DF_FTYPE_V2DF_V2DF_V2DF:
22824 type = v2df_ftype_v2df_v2df_v2df;
22826 case V16QI_FTYPE_V16QI_V16QI_INT:
22827 type = v16qi_ftype_v16qi_v16qi_int;
22829 case V8SI_FTYPE_V8SI_V8SI_INT:
22830 type = v8si_ftype_v8si_v8si_int;
22832 case V8SI_FTYPE_V8SI_V4SI_INT:
22833 type = v8si_ftype_v8si_v4si_int;
22835 case V8HI_FTYPE_V8HI_V8HI_INT:
22836 type = v8hi_ftype_v8hi_v8hi_int;
22838 case V8SF_FTYPE_V8SF_V8SF_INT:
22839 type = v8sf_ftype_v8sf_v8sf_int;
22841 case V8SF_FTYPE_V8SF_V4SF_INT:
22842 type = v8sf_ftype_v8sf_v4sf_int;
22844 case V4SI_FTYPE_V4SI_V4SI_INT:
22845 type = v4si_ftype_v4si_v4si_int;
22847 case V4DF_FTYPE_V4DF_V4DF_INT:
22848 type = v4df_ftype_v4df_v4df_int;
22850 case V4DF_FTYPE_V4DF_V2DF_INT:
22851 type = v4df_ftype_v4df_v2df_int;
22853 case V4SF_FTYPE_V4SF_V4SF_INT:
22854 type = v4sf_ftype_v4sf_v4sf_int;
22856 case V2DI_FTYPE_V2DI_V2DI_INT:
22857 case V2DI2TI_FTYPE_V2DI_V2DI_INT:
22858 type = v2di_ftype_v2di_v2di_int;
22860 case V2DF_FTYPE_V2DF_V2DF_INT:
22861 type = v2df_ftype_v2df_v2df_int;
22863 case V2DI_FTYPE_V2DI_UINT_UINT:
22864 type = v2di_ftype_v2di_unsigned_unsigned;
22866 case V2DI_FTYPE_V2DI_V2DI_UINT_UINT:
22867 type = v2di_ftype_v2di_v2di_unsigned_unsigned;
22869 case V1DI2DI_FTYPE_V1DI_V1DI_INT:
22870 type = v1di_ftype_v1di_v1di_int;
22873 gcc_unreachable ();
22876 def_builtin_const (d->mask, d->name, type, d->code);
22879 /* pcmpestr[im] insns. */
22880 for (i = 0, d = bdesc_pcmpestr;
22881 i < ARRAY_SIZE (bdesc_pcmpestr);
22884 if (d->code == IX86_BUILTIN_PCMPESTRM128)
22885 ftype = v16qi_ftype_v16qi_int_v16qi_int_int;
22887 ftype = int_ftype_v16qi_int_v16qi_int_int;
22888 def_builtin_const (d->mask, d->name, ftype, d->code);
22891 /* pcmpistr[im] insns. */
22892 for (i = 0, d = bdesc_pcmpistr;
22893 i < ARRAY_SIZE (bdesc_pcmpistr);
22896 if (d->code == IX86_BUILTIN_PCMPISTRM128)
22897 ftype = v16qi_ftype_v16qi_v16qi_int;
22899 ftype = int_ftype_v16qi_v16qi_int;
22900 def_builtin_const (d->mask, d->name, ftype, d->code);
22903 /* comi/ucomi insns. */
22904 for (i = 0, d = bdesc_comi; i < ARRAY_SIZE (bdesc_comi); i++, d++)
22905 if (d->mask == OPTION_MASK_ISA_SSE2)
22906 def_builtin_const (d->mask, d->name, int_ftype_v2df_v2df, d->code);
22908 def_builtin_const (d->mask, d->name, int_ftype_v4sf_v4sf, d->code);
22911 def_builtin (OPTION_MASK_ISA_SSE, "__builtin_ia32_ldmxcsr", void_ftype_unsigned, IX86_BUILTIN_LDMXCSR);
22912 def_builtin (OPTION_MASK_ISA_SSE, "__builtin_ia32_stmxcsr", unsigned_ftype_void, IX86_BUILTIN_STMXCSR);
22914 /* SSE or 3DNow!A */
22915 def_builtin (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, "__builtin_ia32_maskmovq", void_ftype_v8qi_v8qi_pchar, IX86_BUILTIN_MASKMOVQ);
22918 def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_maskmovdqu", void_ftype_v16qi_v16qi_pchar, IX86_BUILTIN_MASKMOVDQU);
22920 def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_clflush", void_ftype_pcvoid, IX86_BUILTIN_CLFLUSH);
22921 x86_mfence = def_builtin (OPTION_MASK_ISA_SSE2, "__builtin_ia32_mfence", void_ftype_void, IX86_BUILTIN_MFENCE);
22924 def_builtin (OPTION_MASK_ISA_SSE3, "__builtin_ia32_monitor", void_ftype_pcvoid_unsigned_unsigned, IX86_BUILTIN_MONITOR);
22925 def_builtin (OPTION_MASK_ISA_SSE3, "__builtin_ia32_mwait", void_ftype_unsigned_unsigned, IX86_BUILTIN_MWAIT);
22928 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesenc128", v2di_ftype_v2di_v2di, IX86_BUILTIN_AESENC128);
22929 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesenclast128", v2di_ftype_v2di_v2di, IX86_BUILTIN_AESENCLAST128);
22930 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesdec128", v2di_ftype_v2di_v2di, IX86_BUILTIN_AESDEC128);
22931 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesdeclast128", v2di_ftype_v2di_v2di, IX86_BUILTIN_AESDECLAST128);
22932 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aesimc128", v2di_ftype_v2di, IX86_BUILTIN_AESIMC128);
22933 def_builtin_const (OPTION_MASK_ISA_AES, "__builtin_ia32_aeskeygenassist128", v2di_ftype_v2di_int, IX86_BUILTIN_AESKEYGENASSIST128);
22936 def_builtin_const (OPTION_MASK_ISA_PCLMUL, "__builtin_ia32_pclmulqdq128", v2di_ftype_v2di_v2di_int, IX86_BUILTIN_PCLMULQDQ128);
22939 def_builtin (OPTION_MASK_ISA_AVX, "__builtin_ia32_vzeroupper", void_ftype_void,
22940 TARGET_64BIT ? IX86_BUILTIN_VZEROUPPER_REX64 : IX86_BUILTIN_VZEROUPPER);
22942 /* Access to the vec_init patterns. */
22943 ftype = build_function_type_list (V2SI_type_node, integer_type_node,
22944 integer_type_node, NULL_TREE);
22945 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v2si", ftype, IX86_BUILTIN_VEC_INIT_V2SI);
22947 ftype = build_function_type_list (V4HI_type_node, short_integer_type_node,
22948 short_integer_type_node,
22949 short_integer_type_node,
22950 short_integer_type_node, NULL_TREE);
22951 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v4hi", ftype, IX86_BUILTIN_VEC_INIT_V4HI);
22953 ftype = build_function_type_list (V8QI_type_node, char_type_node,
22954 char_type_node, char_type_node,
22955 char_type_node, char_type_node,
22956 char_type_node, char_type_node,
22957 char_type_node, NULL_TREE);
22958 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_init_v8qi", ftype, IX86_BUILTIN_VEC_INIT_V8QI);
22960 /* Access to the vec_extract patterns. */
22961 ftype = build_function_type_list (double_type_node, V2DF_type_node,
22962 integer_type_node, NULL_TREE);
22963 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v2df", ftype, IX86_BUILTIN_VEC_EXT_V2DF);
22965 ftype = build_function_type_list (long_long_integer_type_node,
22966 V2DI_type_node, integer_type_node,
22968 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v2di", ftype, IX86_BUILTIN_VEC_EXT_V2DI);
22970 ftype = build_function_type_list (float_type_node, V4SF_type_node,
22971 integer_type_node, NULL_TREE);
22972 def_builtin_const (OPTION_MASK_ISA_SSE, "__builtin_ia32_vec_ext_v4sf", ftype, IX86_BUILTIN_VEC_EXT_V4SF);
22974 ftype = build_function_type_list (intSI_type_node, V4SI_type_node,
22975 integer_type_node, NULL_TREE);
22976 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v4si", ftype, IX86_BUILTIN_VEC_EXT_V4SI);
22978 ftype = build_function_type_list (intHI_type_node, V8HI_type_node,
22979 integer_type_node, NULL_TREE);
22980 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v8hi", ftype, IX86_BUILTIN_VEC_EXT_V8HI);
22982 ftype = build_function_type_list (intHI_type_node, V4HI_type_node,
22983 integer_type_node, NULL_TREE);
22984 def_builtin_const (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, "__builtin_ia32_vec_ext_v4hi", ftype, IX86_BUILTIN_VEC_EXT_V4HI);
22986 ftype = build_function_type_list (intSI_type_node, V2SI_type_node,
22987 integer_type_node, NULL_TREE);
22988 def_builtin_const (OPTION_MASK_ISA_MMX, "__builtin_ia32_vec_ext_v2si", ftype, IX86_BUILTIN_VEC_EXT_V2SI);
22990 ftype = build_function_type_list (intQI_type_node, V16QI_type_node,
22991 integer_type_node, NULL_TREE);
22992 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_ext_v16qi", ftype, IX86_BUILTIN_VEC_EXT_V16QI);
22994 /* Access to the vec_set patterns. */
22995 ftype = build_function_type_list (V2DI_type_node, V2DI_type_node,
22997 integer_type_node, NULL_TREE);
22998 def_builtin_const (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_64BIT, "__builtin_ia32_vec_set_v2di", ftype, IX86_BUILTIN_VEC_SET_V2DI);
23000 ftype = build_function_type_list (V4SF_type_node, V4SF_type_node,
23002 integer_type_node, NULL_TREE);
23003 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v4sf", ftype, IX86_BUILTIN_VEC_SET_V4SF);
23005 ftype = build_function_type_list (V4SI_type_node, V4SI_type_node,
23007 integer_type_node, NULL_TREE);
23008 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v4si", ftype, IX86_BUILTIN_VEC_SET_V4SI);
23010 ftype = build_function_type_list (V8HI_type_node, V8HI_type_node,
23012 integer_type_node, NULL_TREE);
23013 def_builtin_const (OPTION_MASK_ISA_SSE2, "__builtin_ia32_vec_set_v8hi", ftype, IX86_BUILTIN_VEC_SET_V8HI);
23015 ftype = build_function_type_list (V4HI_type_node, V4HI_type_node,
23017 integer_type_node, NULL_TREE);
23018 def_builtin_const (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A, "__builtin_ia32_vec_set_v4hi", ftype, IX86_BUILTIN_VEC_SET_V4HI);
23020 ftype = build_function_type_list (V16QI_type_node, V16QI_type_node,
23022 integer_type_node, NULL_TREE);
23023 def_builtin_const (OPTION_MASK_ISA_SSE4_1, "__builtin_ia32_vec_set_v16qi", ftype, IX86_BUILTIN_VEC_SET_V16QI);
23025 /* Add SSE5 multi-arg argument instructions */
23026 for (i = 0, d = bdesc_multi_arg; i < ARRAY_SIZE (bdesc_multi_arg); i++, d++)
23028 tree mtype = NULL_TREE;
23033 switch ((enum multi_arg_type)d->flag)
23035 case MULTI_ARG_3_SF: mtype = v4sf_ftype_v4sf_v4sf_v4sf; break;
23036 case MULTI_ARG_3_DF: mtype = v2df_ftype_v2df_v2df_v2df; break;
23037 case MULTI_ARG_3_DI: mtype = v2di_ftype_v2di_v2di_v2di; break;
23038 case MULTI_ARG_3_SI: mtype = v4si_ftype_v4si_v4si_v4si; break;
23039 case MULTI_ARG_3_SI_DI: mtype = v4si_ftype_v4si_v4si_v2di; break;
23040 case MULTI_ARG_3_HI: mtype = v8hi_ftype_v8hi_v8hi_v8hi; break;
23041 case MULTI_ARG_3_HI_SI: mtype = v8hi_ftype_v8hi_v8hi_v4si; break;
23042 case MULTI_ARG_3_QI: mtype = v16qi_ftype_v16qi_v16qi_v16qi; break;
23043 case MULTI_ARG_3_PERMPS: mtype = v4sf_ftype_v4sf_v4sf_v16qi; break;
23044 case MULTI_ARG_3_PERMPD: mtype = v2df_ftype_v2df_v2df_v16qi; break;
23045 case MULTI_ARG_2_SF: mtype = v4sf_ftype_v4sf_v4sf; break;
23046 case MULTI_ARG_2_DF: mtype = v2df_ftype_v2df_v2df; break;
23047 case MULTI_ARG_2_DI: mtype = v2di_ftype_v2di_v2di; break;
23048 case MULTI_ARG_2_SI: mtype = v4si_ftype_v4si_v4si; break;
23049 case MULTI_ARG_2_HI: mtype = v8hi_ftype_v8hi_v8hi; break;
23050 case MULTI_ARG_2_QI: mtype = v16qi_ftype_v16qi_v16qi; break;
23051 case MULTI_ARG_2_DI_IMM: mtype = v2di_ftype_v2di_si; break;
23052 case MULTI_ARG_2_SI_IMM: mtype = v4si_ftype_v4si_si; break;
23053 case MULTI_ARG_2_HI_IMM: mtype = v8hi_ftype_v8hi_si; break;
23054 case MULTI_ARG_2_QI_IMM: mtype = v16qi_ftype_v16qi_si; break;
23055 case MULTI_ARG_2_SF_CMP: mtype = v4sf_ftype_v4sf_v4sf; break;
23056 case MULTI_ARG_2_DF_CMP: mtype = v2df_ftype_v2df_v2df; break;
23057 case MULTI_ARG_2_DI_CMP: mtype = v2di_ftype_v2di_v2di; break;
23058 case MULTI_ARG_2_SI_CMP: mtype = v4si_ftype_v4si_v4si; break;
23059 case MULTI_ARG_2_HI_CMP: mtype = v8hi_ftype_v8hi_v8hi; break;
23060 case MULTI_ARG_2_QI_CMP: mtype = v16qi_ftype_v16qi_v16qi; break;
23061 case MULTI_ARG_2_SF_TF: mtype = v4sf_ftype_v4sf_v4sf; break;
23062 case MULTI_ARG_2_DF_TF: mtype = v2df_ftype_v2df_v2df; break;
23063 case MULTI_ARG_2_DI_TF: mtype = v2di_ftype_v2di_v2di; break;
23064 case MULTI_ARG_2_SI_TF: mtype = v4si_ftype_v4si_v4si; break;
23065 case MULTI_ARG_2_HI_TF: mtype = v8hi_ftype_v8hi_v8hi; break;
23066 case MULTI_ARG_2_QI_TF: mtype = v16qi_ftype_v16qi_v16qi; break;
23067 case MULTI_ARG_1_SF: mtype = v4sf_ftype_v4sf; break;
23068 case MULTI_ARG_1_DF: mtype = v2df_ftype_v2df; break;
23069 case MULTI_ARG_1_DI: mtype = v2di_ftype_v2di; break;
23070 case MULTI_ARG_1_SI: mtype = v4si_ftype_v4si; break;
23071 case MULTI_ARG_1_HI: mtype = v8hi_ftype_v8hi; break;
23072 case MULTI_ARG_1_QI: mtype = v16qi_ftype_v16qi; break;
23073 case MULTI_ARG_1_SI_DI: mtype = v2di_ftype_v4si; break;
23074 case MULTI_ARG_1_HI_DI: mtype = v2di_ftype_v8hi; break;
23075 case MULTI_ARG_1_HI_SI: mtype = v4si_ftype_v8hi; break;
23076 case MULTI_ARG_1_QI_DI: mtype = v2di_ftype_v16qi; break;
23077 case MULTI_ARG_1_QI_SI: mtype = v4si_ftype_v16qi; break;
23078 case MULTI_ARG_1_QI_HI: mtype = v8hi_ftype_v16qi; break;
23079 case MULTI_ARG_1_PH2PS: mtype = v4sf_ftype_v4hi; break;
23080 case MULTI_ARG_1_PS2PH: mtype = v4hi_ftype_v4sf; break;
23081 case MULTI_ARG_UNKNOWN:
23083 gcc_unreachable ();
23087 def_builtin_const (d->mask, d->name, mtype, d->code);
23091 /* Internal method for ix86_init_builtins. */
23094 ix86_init_builtins_va_builtins_abi (void)
23096 tree ms_va_ref, sysv_va_ref;
23097 tree fnvoid_va_end_ms, fnvoid_va_end_sysv;
23098 tree fnvoid_va_start_ms, fnvoid_va_start_sysv;
23099 tree fnvoid_va_copy_ms, fnvoid_va_copy_sysv;
23100 tree fnattr_ms = NULL_TREE, fnattr_sysv = NULL_TREE;
23104 fnattr_ms = build_tree_list (get_identifier ("ms_abi"), NULL_TREE);
23105 fnattr_sysv = build_tree_list (get_identifier ("sysv_abi"), NULL_TREE);
23106 ms_va_ref = build_reference_type (ms_va_list_type_node);
23108 build_pointer_type (TREE_TYPE (sysv_va_list_type_node));
23111 build_function_type_list (void_type_node, ms_va_ref, NULL_TREE);
23112 fnvoid_va_start_ms =
23113 build_varargs_function_type_list (void_type_node, ms_va_ref, NULL_TREE);
23114 fnvoid_va_end_sysv =
23115 build_function_type_list (void_type_node, sysv_va_ref, NULL_TREE);
23116 fnvoid_va_start_sysv =
23117 build_varargs_function_type_list (void_type_node, sysv_va_ref,
23119 fnvoid_va_copy_ms =
23120 build_function_type_list (void_type_node, ms_va_ref, ms_va_list_type_node,
23122 fnvoid_va_copy_sysv =
23123 build_function_type_list (void_type_node, sysv_va_ref,
23124 sysv_va_ref, NULL_TREE);
23126 add_builtin_function ("__builtin_ms_va_start", fnvoid_va_start_ms,
23127 BUILT_IN_VA_START, BUILT_IN_NORMAL, NULL, fnattr_ms);
23128 add_builtin_function ("__builtin_ms_va_end", fnvoid_va_end_ms,
23129 BUILT_IN_VA_END, BUILT_IN_NORMAL, NULL, fnattr_ms);
23130 add_builtin_function ("__builtin_ms_va_copy", fnvoid_va_copy_ms,
23131 BUILT_IN_VA_COPY, BUILT_IN_NORMAL, NULL, fnattr_ms);
23132 add_builtin_function ("__builtin_sysv_va_start", fnvoid_va_start_sysv,
23133 BUILT_IN_VA_START, BUILT_IN_NORMAL, NULL, fnattr_sysv);
23134 add_builtin_function ("__builtin_sysv_va_end", fnvoid_va_end_sysv,
23135 BUILT_IN_VA_END, BUILT_IN_NORMAL, NULL, fnattr_sysv);
23136 add_builtin_function ("__builtin_sysv_va_copy", fnvoid_va_copy_sysv,
23137 BUILT_IN_VA_COPY, BUILT_IN_NORMAL, NULL, fnattr_sysv);
23141 ix86_init_builtins (void)
23143 tree float128_type_node = make_node (REAL_TYPE);
23146 /* The __float80 type. */
23147 if (TYPE_MODE (long_double_type_node) == XFmode)
23148 (*lang_hooks.types.register_builtin_type) (long_double_type_node,
23152 /* The __float80 type. */
23153 tree float80_type_node = make_node (REAL_TYPE);
23155 TYPE_PRECISION (float80_type_node) = 80;
23156 layout_type (float80_type_node);
23157 (*lang_hooks.types.register_builtin_type) (float80_type_node,
23161 /* The __float128 type. */
23162 TYPE_PRECISION (float128_type_node) = 128;
23163 layout_type (float128_type_node);
23164 (*lang_hooks.types.register_builtin_type) (float128_type_node,
23167 /* TFmode support builtins. */
23168 ftype = build_function_type (float128_type_node, void_list_node);
23169 decl = add_builtin_function ("__builtin_infq", ftype,
23170 IX86_BUILTIN_INFQ, BUILT_IN_MD,
23172 ix86_builtins[(int) IX86_BUILTIN_INFQ] = decl;
23174 /* We will expand them to normal call if SSE2 isn't available since
23175 they are used by libgcc. */
23176 ftype = build_function_type_list (float128_type_node,
23177 float128_type_node,
23179 decl = add_builtin_function ("__builtin_fabsq", ftype,
23180 IX86_BUILTIN_FABSQ, BUILT_IN_MD,
23181 "__fabstf2", NULL_TREE);
23182 ix86_builtins[(int) IX86_BUILTIN_FABSQ] = decl;
23183 TREE_READONLY (decl) = 1;
23185 ftype = build_function_type_list (float128_type_node,
23186 float128_type_node,
23187 float128_type_node,
23189 decl = add_builtin_function ("__builtin_copysignq", ftype,
23190 IX86_BUILTIN_COPYSIGNQ, BUILT_IN_MD,
23191 "__copysigntf3", NULL_TREE);
23192 ix86_builtins[(int) IX86_BUILTIN_COPYSIGNQ] = decl;
23193 TREE_READONLY (decl) = 1;
23195 ix86_init_mmx_sse_builtins ();
23197 ix86_init_builtins_va_builtins_abi ();
23200 /* Errors in the source file can cause expand_expr to return const0_rtx
23201 where we expect a vector. To avoid crashing, use one of the vector
23202 clear instructions. */
23204 safe_vector_operand (rtx x, enum machine_mode mode)
23206 if (x == const0_rtx)
23207 x = CONST0_RTX (mode);
23211 /* Subroutine of ix86_expand_builtin to take care of binop insns. */
23214 ix86_expand_binop_builtin (enum insn_code icode, tree exp, rtx target)
23217 tree arg0 = CALL_EXPR_ARG (exp, 0);
23218 tree arg1 = CALL_EXPR_ARG (exp, 1);
23219 rtx op0 = expand_normal (arg0);
23220 rtx op1 = expand_normal (arg1);
23221 enum machine_mode tmode = insn_data[icode].operand[0].mode;
23222 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
23223 enum machine_mode mode1 = insn_data[icode].operand[2].mode;
23225 if (VECTOR_MODE_P (mode0))
23226 op0 = safe_vector_operand (op0, mode0);
23227 if (VECTOR_MODE_P (mode1))
23228 op1 = safe_vector_operand (op1, mode1);
23230 if (optimize || !target
23231 || GET_MODE (target) != tmode
23232 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
23233 target = gen_reg_rtx (tmode);
23235 if (GET_MODE (op1) == SImode && mode1 == TImode)
23237 rtx x = gen_reg_rtx (V4SImode);
23238 emit_insn (gen_sse2_loadd (x, op1));
23239 op1 = gen_lowpart (TImode, x);
23242 if (!(*insn_data[icode].operand[1].predicate) (op0, mode0))
23243 op0 = copy_to_mode_reg (mode0, op0);
23244 if (!(*insn_data[icode].operand[2].predicate) (op1, mode1))
23245 op1 = copy_to_mode_reg (mode1, op1);
23247 pat = GEN_FCN (icode) (target, op0, op1);
23256 /* Subroutine of ix86_expand_builtin to take care of 2-4 argument insns. */
23259 ix86_expand_multi_arg_builtin (enum insn_code icode, tree exp, rtx target,
23260 enum multi_arg_type m_type,
23261 enum insn_code sub_code)
23266 bool comparison_p = false;
23268 bool last_arg_constant = false;
23269 int num_memory = 0;
23272 enum machine_mode mode;
23275 enum machine_mode tmode = insn_data[icode].operand[0].mode;
23279 case MULTI_ARG_3_SF:
23280 case MULTI_ARG_3_DF:
23281 case MULTI_ARG_3_DI:
23282 case MULTI_ARG_3_SI:
23283 case MULTI_ARG_3_SI_DI:
23284 case MULTI_ARG_3_HI:
23285 case MULTI_ARG_3_HI_SI:
23286 case MULTI_ARG_3_QI:
23287 case MULTI_ARG_3_PERMPS:
23288 case MULTI_ARG_3_PERMPD:
23292 case MULTI_ARG_2_SF:
23293 case MULTI_ARG_2_DF:
23294 case MULTI_ARG_2_DI:
23295 case MULTI_ARG_2_SI:
23296 case MULTI_ARG_2_HI:
23297 case MULTI_ARG_2_QI:
23301 case MULTI_ARG_2_DI_IMM:
23302 case MULTI_ARG_2_SI_IMM:
23303 case MULTI_ARG_2_HI_IMM:
23304 case MULTI_ARG_2_QI_IMM:
23306 last_arg_constant = true;
23309 case MULTI_ARG_1_SF:
23310 case MULTI_ARG_1_DF:
23311 case MULTI_ARG_1_DI:
23312 case MULTI_ARG_1_SI:
23313 case MULTI_ARG_1_HI:
23314 case MULTI_ARG_1_QI:
23315 case MULTI_ARG_1_SI_DI:
23316 case MULTI_ARG_1_HI_DI:
23317 case MULTI_ARG_1_HI_SI:
23318 case MULTI_ARG_1_QI_DI:
23319 case MULTI_ARG_1_QI_SI:
23320 case MULTI_ARG_1_QI_HI:
23321 case MULTI_ARG_1_PH2PS:
23322 case MULTI_ARG_1_PS2PH:
23326 case MULTI_ARG_2_SF_CMP:
23327 case MULTI_ARG_2_DF_CMP:
23328 case MULTI_ARG_2_DI_CMP:
23329 case MULTI_ARG_2_SI_CMP:
23330 case MULTI_ARG_2_HI_CMP:
23331 case MULTI_ARG_2_QI_CMP:
23333 comparison_p = true;
23336 case MULTI_ARG_2_SF_TF:
23337 case MULTI_ARG_2_DF_TF:
23338 case MULTI_ARG_2_DI_TF:
23339 case MULTI_ARG_2_SI_TF:
23340 case MULTI_ARG_2_HI_TF:
23341 case MULTI_ARG_2_QI_TF:
23346 case MULTI_ARG_UNKNOWN:
23348 gcc_unreachable ();
23351 if (optimize || !target
23352 || GET_MODE (target) != tmode
23353 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
23354 target = gen_reg_rtx (tmode);
23356 gcc_assert (nargs <= 4);
23358 for (i = 0; i < nargs; i++)
23360 tree arg = CALL_EXPR_ARG (exp, i);
23361 rtx op = expand_normal (arg);
23362 int adjust = (comparison_p) ? 1 : 0;
23363 enum machine_mode mode = insn_data[icode].operand[i+adjust+1].mode;
23365 if (last_arg_constant && i == nargs-1)
23367 if (GET_CODE (op) != CONST_INT)
23369 error ("last argument must be an immediate");
23370 return gen_reg_rtx (tmode);
23375 if (VECTOR_MODE_P (mode))
23376 op = safe_vector_operand (op, mode);
23378 /* If we aren't optimizing, only allow one memory operand to be
23380 if (memory_operand (op, mode))
23383 gcc_assert (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode);
23386 || ! (*insn_data[icode].operand[i+adjust+1].predicate) (op, mode)
23388 op = force_reg (mode, op);
23392 args[i].mode = mode;
23398 pat = GEN_FCN (icode) (target, args[0].op);
23403 pat = GEN_FCN (icode) (target, args[0].op, args[1].op,
23404 GEN_INT ((int)sub_code));
23405 else if (! comparison_p)
23406 pat = GEN_FCN (icode) (target, args[0].op, args[1].op);
23409 rtx cmp_op = gen_rtx_fmt_ee (sub_code, GET_MODE (target),
23413 pat = GEN_FCN (icode) (target, cmp_op, args[0].op, args[1].op);
23418 pat = GEN_FCN (icode) (target, args[0].op, args[1].op, args[2].op);
23422 gcc_unreachable ();
23432 /* Subroutine of ix86_expand_args_builtin to take care of scalar unop
23433 insns with vec_merge. */
23436 ix86_expand_unop_vec_merge_builtin (enum insn_code icode, tree exp,
23440 tree arg0 = CALL_EXPR_ARG (exp, 0);
23441 rtx op1, op0 = expand_normal (arg0);
23442 enum machine_mode tmode = insn_data[icode].operand[0].mode;
23443 enum machine_mode mode0 = insn_data[icode].operand[1].mode;
23445 if (optimize || !target
23446 || GET_MODE (target) != tmode
23447 || ! (*insn_data[icode].operand[0].predicate) (target, tmode))
23448 target = gen_reg_rtx (tmode);
23450 if (VECTOR_MODE_P (mode0))
23451 op0 = safe_vector_operand (op0, mode0);
23453 if ((optimize && !register_operand (op0, mode0))
23454 || ! (*insn_data[icode].operand[1].predicate) (op0, mode0))
23455 op0 = copy_to_mode_reg (mode0, op0);
23458 if (! (*insn_data[icode].operand[2].predicate) (op1, mode0))
23459 op1 = copy_to_mode_reg (mode0, op1);
23461 pat = GEN_FCN (icode) (target, op0, op1);
23468 /* Subroutine of ix86_expand_builtin to take care of comparison insns. */
23471 ix86_expand_sse_compare (const struct builtin_description *d,
23472 tree exp, rtx target, bool swap)
23475 tree arg0 = CALL_EXPR_ARG (exp, 0);
23476 tree arg1 = CALL_EXPR_ARG (exp, 1);
23477 rtx op0 = expand_normal (arg0);
23478 rtx op1 = expand_normal (arg1);
23480 enum machine_mode tmode = insn_data[d->icode].operand[0].mode;
23481 enum machine_mode mode0 = insn_data[d->icode].operand[1].mode;
23482 enum machine_mode mode1 = insn_data[d->icode].operand[2].mode;
23483 enum rtx_code comparison = d->comparison;
23485 if (VECTOR_MODE_P (mode0))
23486 op0 = safe_vector_operand (op0, mode0);
23487 if (VECTOR_MODE_P (mode1))
23488 op1 = safe_vector_operand (op1, mode1);
23490 /* Swap operands if we have a comparison that isn't available in
23494 rtx tmp = gen_reg_rtx (mode1);
23495 emit_move_insn (tmp, op1);
23500 if (optimize || !target
23501 || GET_MODE (target) != tmode
23502 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode))
23503 target = gen_reg_rtx (tmode);
23505 if ((optimize && !register_operand (op0, mode0))
23506 || ! (*insn_data[d->icode].operand[1].predicate) (op0, mode0))
23507 op0 = copy_to_mode_reg (mode0, op0);
23508 if ((optimize && !register_operand (op1, mode1))
23509 || ! (*insn_data[d->icode].operand[2].predicate) (op1, mode1))
23510 op1 = copy_to_mode_reg (mode1, op1);
23512 op2 = gen_rtx_fmt_ee (comparison, mode0, op0, op1);
23513 pat = GEN_FCN (d->icode) (target, op0, op1, op2);
23520 /* Subroutine of ix86_expand_builtin to take care of comi insns. */
23523 ix86_expand_sse_comi (const struct builtin_description *d, tree exp,
23527 tree arg0 = CALL_EXPR_ARG (exp, 0);
23528 tree arg1 = CALL_EXPR_ARG (exp, 1);
23529 rtx op0 = expand_normal (arg0);
23530 rtx op1 = expand_normal (arg1);
23531 enum machine_mode mode0 = insn_data[d->icode].operand[0].mode;
23532 enum machine_mode mode1 = insn_data[d->icode].operand[1].mode;
23533 enum rtx_code comparison = d->comparison;
23535 if (VECTOR_MODE_P (mode0))
23536 op0 = safe_vector_operand (op0, mode0);
23537 if (VECTOR_MODE_P (mode1))
23538 op1 = safe_vector_operand (op1, mode1);
23540 /* Swap operands if we have a comparison that isn't available in
23542 if (d->flag & BUILTIN_DESC_SWAP_OPERANDS)
23549 target = gen_reg_rtx (SImode);
23550 emit_move_insn (target, const0_rtx);
23551 target = gen_rtx_SUBREG (QImode, target, 0);
23553 if ((optimize && !register_operand (op0, mode0))
23554 || !(*insn_data[d->icode].operand[0].predicate) (op0, mode0))
23555 op0 = copy_to_mode_reg (mode0, op0);
23556 if ((optimize && !register_operand (op1, mode1))
23557 || !(*insn_data[d->icode].operand[1].predicate) (op1, mode1))
23558 op1 = copy_to_mode_reg (mode1, op1);
23560 pat = GEN_FCN (d->icode) (op0, op1);
23564 emit_insn (gen_rtx_SET (VOIDmode,
23565 gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23566 gen_rtx_fmt_ee (comparison, QImode,
23570 return SUBREG_REG (target);
23573 /* Subroutine of ix86_expand_builtin to take care of ptest insns. */
23576 ix86_expand_sse_ptest (const struct builtin_description *d, tree exp,
23580 tree arg0 = CALL_EXPR_ARG (exp, 0);
23581 tree arg1 = CALL_EXPR_ARG (exp, 1);
23582 rtx op0 = expand_normal (arg0);
23583 rtx op1 = expand_normal (arg1);
23584 enum machine_mode mode0 = insn_data[d->icode].operand[0].mode;
23585 enum machine_mode mode1 = insn_data[d->icode].operand[1].mode;
23586 enum rtx_code comparison = d->comparison;
23588 if (VECTOR_MODE_P (mode0))
23589 op0 = safe_vector_operand (op0, mode0);
23590 if (VECTOR_MODE_P (mode1))
23591 op1 = safe_vector_operand (op1, mode1);
23593 target = gen_reg_rtx (SImode);
23594 emit_move_insn (target, const0_rtx);
23595 target = gen_rtx_SUBREG (QImode, target, 0);
23597 if ((optimize && !register_operand (op0, mode0))
23598 || !(*insn_data[d->icode].operand[0].predicate) (op0, mode0))
23599 op0 = copy_to_mode_reg (mode0, op0);
23600 if ((optimize && !register_operand (op1, mode1))
23601 || !(*insn_data[d->icode].operand[1].predicate) (op1, mode1))
23602 op1 = copy_to_mode_reg (mode1, op1);
23604 pat = GEN_FCN (d->icode) (op0, op1);
23608 emit_insn (gen_rtx_SET (VOIDmode,
23609 gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23610 gen_rtx_fmt_ee (comparison, QImode,
23614 return SUBREG_REG (target);
23617 /* Subroutine of ix86_expand_builtin to take care of pcmpestr[im] insns. */
23620 ix86_expand_sse_pcmpestr (const struct builtin_description *d,
23621 tree exp, rtx target)
23624 tree arg0 = CALL_EXPR_ARG (exp, 0);
23625 tree arg1 = CALL_EXPR_ARG (exp, 1);
23626 tree arg2 = CALL_EXPR_ARG (exp, 2);
23627 tree arg3 = CALL_EXPR_ARG (exp, 3);
23628 tree arg4 = CALL_EXPR_ARG (exp, 4);
23629 rtx scratch0, scratch1;
23630 rtx op0 = expand_normal (arg0);
23631 rtx op1 = expand_normal (arg1);
23632 rtx op2 = expand_normal (arg2);
23633 rtx op3 = expand_normal (arg3);
23634 rtx op4 = expand_normal (arg4);
23635 enum machine_mode tmode0, tmode1, modev2, modei3, modev4, modei5, modeimm;
23637 tmode0 = insn_data[d->icode].operand[0].mode;
23638 tmode1 = insn_data[d->icode].operand[1].mode;
23639 modev2 = insn_data[d->icode].operand[2].mode;
23640 modei3 = insn_data[d->icode].operand[3].mode;
23641 modev4 = insn_data[d->icode].operand[4].mode;
23642 modei5 = insn_data[d->icode].operand[5].mode;
23643 modeimm = insn_data[d->icode].operand[6].mode;
23645 if (VECTOR_MODE_P (modev2))
23646 op0 = safe_vector_operand (op0, modev2);
23647 if (VECTOR_MODE_P (modev4))
23648 op2 = safe_vector_operand (op2, modev4);
23650 if (! (*insn_data[d->icode].operand[2].predicate) (op0, modev2))
23651 op0 = copy_to_mode_reg (modev2, op0);
23652 if (! (*insn_data[d->icode].operand[3].predicate) (op1, modei3))
23653 op1 = copy_to_mode_reg (modei3, op1);
23654 if ((optimize && !register_operand (op2, modev4))
23655 || !(*insn_data[d->icode].operand[4].predicate) (op2, modev4))
23656 op2 = copy_to_mode_reg (modev4, op2);
23657 if (! (*insn_data[d->icode].operand[5].predicate) (op3, modei5))
23658 op3 = copy_to_mode_reg (modei5, op3);
23660 if (! (*insn_data[d->icode].operand[6].predicate) (op4, modeimm))
23662 error ("the fifth argument must be a 8-bit immediate");
23666 if (d->code == IX86_BUILTIN_PCMPESTRI128)
23668 if (optimize || !target
23669 || GET_MODE (target) != tmode0
23670 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode0))
23671 target = gen_reg_rtx (tmode0);
23673 scratch1 = gen_reg_rtx (tmode1);
23675 pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2, op3, op4);
23677 else if (d->code == IX86_BUILTIN_PCMPESTRM128)
23679 if (optimize || !target
23680 || GET_MODE (target) != tmode1
23681 || ! (*insn_data[d->icode].operand[1].predicate) (target, tmode1))
23682 target = gen_reg_rtx (tmode1);
23684 scratch0 = gen_reg_rtx (tmode0);
23686 pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2, op3, op4);
23690 gcc_assert (d->flag);
23692 scratch0 = gen_reg_rtx (tmode0);
23693 scratch1 = gen_reg_rtx (tmode1);
23695 pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2, op3, op4);
23705 target = gen_reg_rtx (SImode);
23706 emit_move_insn (target, const0_rtx);
23707 target = gen_rtx_SUBREG (QImode, target, 0);
23710 (gen_rtx_SET (VOIDmode, gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23711 gen_rtx_fmt_ee (EQ, QImode,
23712 gen_rtx_REG ((enum machine_mode) d->flag,
23715 return SUBREG_REG (target);
23722 /* Subroutine of ix86_expand_builtin to take care of pcmpistr[im] insns. */
23725 ix86_expand_sse_pcmpistr (const struct builtin_description *d,
23726 tree exp, rtx target)
23729 tree arg0 = CALL_EXPR_ARG (exp, 0);
23730 tree arg1 = CALL_EXPR_ARG (exp, 1);
23731 tree arg2 = CALL_EXPR_ARG (exp, 2);
23732 rtx scratch0, scratch1;
23733 rtx op0 = expand_normal (arg0);
23734 rtx op1 = expand_normal (arg1);
23735 rtx op2 = expand_normal (arg2);
23736 enum machine_mode tmode0, tmode1, modev2, modev3, modeimm;
23738 tmode0 = insn_data[d->icode].operand[0].mode;
23739 tmode1 = insn_data[d->icode].operand[1].mode;
23740 modev2 = insn_data[d->icode].operand[2].mode;
23741 modev3 = insn_data[d->icode].operand[3].mode;
23742 modeimm = insn_data[d->icode].operand[4].mode;
23744 if (VECTOR_MODE_P (modev2))
23745 op0 = safe_vector_operand (op0, modev2);
23746 if (VECTOR_MODE_P (modev3))
23747 op1 = safe_vector_operand (op1, modev3);
23749 if (! (*insn_data[d->icode].operand[2].predicate) (op0, modev2))
23750 op0 = copy_to_mode_reg (modev2, op0);
23751 if ((optimize && !register_operand (op1, modev3))
23752 || !(*insn_data[d->icode].operand[3].predicate) (op1, modev3))
23753 op1 = copy_to_mode_reg (modev3, op1);
23755 if (! (*insn_data[d->icode].operand[4].predicate) (op2, modeimm))
23757 error ("the third argument must be a 8-bit immediate");
23761 if (d->code == IX86_BUILTIN_PCMPISTRI128)
23763 if (optimize || !target
23764 || GET_MODE (target) != tmode0
23765 || ! (*insn_data[d->icode].operand[0].predicate) (target, tmode0))
23766 target = gen_reg_rtx (tmode0);
23768 scratch1 = gen_reg_rtx (tmode1);
23770 pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2);
23772 else if (d->code == IX86_BUILTIN_PCMPISTRM128)
23774 if (optimize || !target
23775 || GET_MODE (target) != tmode1
23776 || ! (*insn_data[d->icode].operand[1].predicate) (target, tmode1))
23777 target = gen_reg_rtx (tmode1);
23779 scratch0 = gen_reg_rtx (tmode0);
23781 pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2);
23785 gcc_assert (d->flag);
23787 scratch0 = gen_reg_rtx (tmode0);
23788 scratch1 = gen_reg_rtx (tmode1);
23790 pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2);
23800 target = gen_reg_rtx (SImode);
23801 emit_move_insn (target, const0_rtx);
23802 target = gen_rtx_SUBREG (QImode, target, 0);
23805 (gen_rtx_SET (VOIDmode, gen_rtx_STRICT_LOW_PART (VOIDmode, target),
23806 gen_rtx_fmt_ee (EQ, QImode,
23807 gen_rtx_REG ((enum machine_mode) d->flag,
23810 return SUBREG_REG (target);
23816 /* Subroutine of ix86_expand_builtin to take care of insns with
23817 variable number of operands. */
23820 ix86_expand_args_builtin (const struct builtin_description *d,
23821 tree exp, rtx target)
23823 rtx pat, real_target;
23824 unsigned int i, nargs;
23825 unsigned int nargs_constant = 0;
23826 int num_memory = 0;
23830 enum machine_mode mode;
23832 bool last_arg_count = false;
23833 enum insn_code icode = d->icode;
23834 const struct insn_data *insn_p = &insn_data[icode];
23835 enum machine_mode tmode = insn_p->operand[0].mode;
23836 enum machine_mode rmode = VOIDmode;
23838 enum rtx_code comparison = d->comparison;
23840 switch ((enum ix86_builtin_type) d->flag)
23842 case INT_FTYPE_V8SF_V8SF_PTEST:
23843 case INT_FTYPE_V4DI_V4DI_PTEST:
23844 case INT_FTYPE_V4DF_V4DF_PTEST:
23845 case INT_FTYPE_V4SF_V4SF_PTEST:
23846 case INT_FTYPE_V2DI_V2DI_PTEST:
23847 case INT_FTYPE_V2DF_V2DF_PTEST:
23848 return ix86_expand_sse_ptest (d, exp, target);
23849 case FLOAT128_FTYPE_FLOAT128:
23850 case FLOAT_FTYPE_FLOAT:
23851 case INT64_FTYPE_V4SF:
23852 case INT64_FTYPE_V2DF:
23853 case INT_FTYPE_V16QI:
23854 case INT_FTYPE_V8QI:
23855 case INT_FTYPE_V8SF:
23856 case INT_FTYPE_V4DF:
23857 case INT_FTYPE_V4SF:
23858 case INT_FTYPE_V2DF:
23859 case V16QI_FTYPE_V16QI:
23860 case V8SI_FTYPE_V8SF:
23861 case V8SI_FTYPE_V4SI:
23862 case V8HI_FTYPE_V8HI:
23863 case V8HI_FTYPE_V16QI:
23864 case V8QI_FTYPE_V8QI:
23865 case V8SF_FTYPE_V8SF:
23866 case V8SF_FTYPE_V8SI:
23867 case V8SF_FTYPE_V4SF:
23868 case V4SI_FTYPE_V4SI:
23869 case V4SI_FTYPE_V16QI:
23870 case V4SI_FTYPE_V4SF:
23871 case V4SI_FTYPE_V8SI:
23872 case V4SI_FTYPE_V8HI:
23873 case V4SI_FTYPE_V4DF:
23874 case V4SI_FTYPE_V2DF:
23875 case V4HI_FTYPE_V4HI:
23876 case V4DF_FTYPE_V4DF:
23877 case V4DF_FTYPE_V4SI:
23878 case V4DF_FTYPE_V4SF:
23879 case V4DF_FTYPE_V2DF:
23880 case V4SF_FTYPE_V4SF:
23881 case V4SF_FTYPE_V4SI:
23882 case V4SF_FTYPE_V8SF:
23883 case V4SF_FTYPE_V4DF:
23884 case V4SF_FTYPE_V2DF:
23885 case V2DI_FTYPE_V2DI:
23886 case V2DI_FTYPE_V16QI:
23887 case V2DI_FTYPE_V8HI:
23888 case V2DI_FTYPE_V4SI:
23889 case V2DF_FTYPE_V2DF:
23890 case V2DF_FTYPE_V4SI:
23891 case V2DF_FTYPE_V4DF:
23892 case V2DF_FTYPE_V4SF:
23893 case V2DF_FTYPE_V2SI:
23894 case V2SI_FTYPE_V2SI:
23895 case V2SI_FTYPE_V4SF:
23896 case V2SI_FTYPE_V2SF:
23897 case V2SI_FTYPE_V2DF:
23898 case V2SF_FTYPE_V2SF:
23899 case V2SF_FTYPE_V2SI:
23902 case V4SF_FTYPE_V4SF_VEC_MERGE:
23903 case V2DF_FTYPE_V2DF_VEC_MERGE:
23904 return ix86_expand_unop_vec_merge_builtin (icode, exp, target);
23905 case FLOAT128_FTYPE_FLOAT128_FLOAT128:
23906 case V16QI_FTYPE_V16QI_V16QI:
23907 case V16QI_FTYPE_V8HI_V8HI:
23908 case V8QI_FTYPE_V8QI_V8QI:
23909 case V8QI_FTYPE_V4HI_V4HI:
23910 case V8HI_FTYPE_V8HI_V8HI:
23911 case V8HI_FTYPE_V16QI_V16QI:
23912 case V8HI_FTYPE_V4SI_V4SI:
23913 case V8SF_FTYPE_V8SF_V8SF:
23914 case V8SF_FTYPE_V8SF_V8SI:
23915 case V4SI_FTYPE_V4SI_V4SI:
23916 case V4SI_FTYPE_V8HI_V8HI:
23917 case V4SI_FTYPE_V4SF_V4SF:
23918 case V4SI_FTYPE_V2DF_V2DF:
23919 case V4HI_FTYPE_V4HI_V4HI:
23920 case V4HI_FTYPE_V8QI_V8QI:
23921 case V4HI_FTYPE_V2SI_V2SI:
23922 case V4DF_FTYPE_V4DF_V4DF:
23923 case V4DF_FTYPE_V4DF_V4DI:
23924 case V4SF_FTYPE_V4SF_V4SF:
23925 case V4SF_FTYPE_V4SF_V4SI:
23926 case V4SF_FTYPE_V4SF_V2SI:
23927 case V4SF_FTYPE_V4SF_V2DF:
23928 case V4SF_FTYPE_V4SF_DI:
23929 case V4SF_FTYPE_V4SF_SI:
23930 case V2DI_FTYPE_V2DI_V2DI:
23931 case V2DI_FTYPE_V16QI_V16QI:
23932 case V2DI_FTYPE_V4SI_V4SI:
23933 case V2DI_FTYPE_V2DI_V16QI:
23934 case V2DI_FTYPE_V2DF_V2DF:
23935 case V2SI_FTYPE_V2SI_V2SI:
23936 case V2SI_FTYPE_V4HI_V4HI:
23937 case V2SI_FTYPE_V2SF_V2SF:
23938 case V2DF_FTYPE_V2DF_V2DF:
23939 case V2DF_FTYPE_V2DF_V4SF:
23940 case V2DF_FTYPE_V2DF_V2DI:
23941 case V2DF_FTYPE_V2DF_DI:
23942 case V2DF_FTYPE_V2DF_SI:
23943 case V2SF_FTYPE_V2SF_V2SF:
23944 case V1DI_FTYPE_V1DI_V1DI:
23945 case V1DI_FTYPE_V8QI_V8QI:
23946 case V1DI_FTYPE_V2SI_V2SI:
23947 if (comparison == UNKNOWN)
23948 return ix86_expand_binop_builtin (icode, exp, target);
23951 case V4SF_FTYPE_V4SF_V4SF_SWAP:
23952 case V2DF_FTYPE_V2DF_V2DF_SWAP:
23953 gcc_assert (comparison != UNKNOWN);
23957 case V8HI_FTYPE_V8HI_V8HI_COUNT:
23958 case V8HI_FTYPE_V8HI_SI_COUNT:
23959 case V4SI_FTYPE_V4SI_V4SI_COUNT:
23960 case V4SI_FTYPE_V4SI_SI_COUNT:
23961 case V4HI_FTYPE_V4HI_V4HI_COUNT:
23962 case V4HI_FTYPE_V4HI_SI_COUNT:
23963 case V2DI_FTYPE_V2DI_V2DI_COUNT:
23964 case V2DI_FTYPE_V2DI_SI_COUNT:
23965 case V2SI_FTYPE_V2SI_V2SI_COUNT:
23966 case V2SI_FTYPE_V2SI_SI_COUNT:
23967 case V1DI_FTYPE_V1DI_V1DI_COUNT:
23968 case V1DI_FTYPE_V1DI_SI_COUNT:
23970 last_arg_count = true;
23972 case UINT64_FTYPE_UINT64_UINT64:
23973 case UINT_FTYPE_UINT_UINT:
23974 case UINT_FTYPE_UINT_USHORT:
23975 case UINT_FTYPE_UINT_UCHAR:
23978 case V2DI2TI_FTYPE_V2DI_INT:
23981 nargs_constant = 1;
23983 case V8HI_FTYPE_V8HI_INT:
23984 case V8SF_FTYPE_V8SF_INT:
23985 case V4SI_FTYPE_V4SI_INT:
23986 case V4SI_FTYPE_V8SI_INT:
23987 case V4HI_FTYPE_V4HI_INT:
23988 case V4DF_FTYPE_V4DF_INT:
23989 case V4SF_FTYPE_V4SF_INT:
23990 case V4SF_FTYPE_V8SF_INT:
23991 case V2DI_FTYPE_V2DI_INT:
23992 case V2DF_FTYPE_V2DF_INT:
23993 case V2DF_FTYPE_V4DF_INT:
23995 nargs_constant = 1;
23997 case V16QI_FTYPE_V16QI_V16QI_V16QI:
23998 case V8SF_FTYPE_V8SF_V8SF_V8SF:
23999 case V4DF_FTYPE_V4DF_V4DF_V4DF:
24000 case V4SF_FTYPE_V4SF_V4SF_V4SF:
24001 case V2DF_FTYPE_V2DF_V2DF_V2DF:
24004 case V16QI_FTYPE_V16QI_V16QI_INT:
24005 case V8HI_FTYPE_V8HI_V8HI_INT:
24006 case V8SI_FTYPE_V8SI_V8SI_INT:
24007 case V8SI_FTYPE_V8SI_V4SI_INT:
24008 case V8SF_FTYPE_V8SF_V8SF_INT:
24009 case V8SF_FTYPE_V8SF_V4SF_INT:
24010 case V4SI_FTYPE_V4SI_V4SI_INT:
24011 case V4DF_FTYPE_V4DF_V4DF_INT:
24012 case V4DF_FTYPE_V4DF_V2DF_INT:
24013 case V4SF_FTYPE_V4SF_V4SF_INT:
24014 case V2DI_FTYPE_V2DI_V2DI_INT:
24015 case V2DF_FTYPE_V2DF_V2DF_INT:
24017 nargs_constant = 1;
24019 case V2DI2TI_FTYPE_V2DI_V2DI_INT:
24022 nargs_constant = 1;
24024 case V1DI2DI_FTYPE_V1DI_V1DI_INT:
24027 nargs_constant = 1;
24029 case V2DI_FTYPE_V2DI_UINT_UINT:
24031 nargs_constant = 2;
24033 case V2DI_FTYPE_V2DI_V2DI_UINT_UINT:
24035 nargs_constant = 2;
24038 gcc_unreachable ();
24041 gcc_assert (nargs <= ARRAY_SIZE (args));
24043 if (comparison != UNKNOWN)
24045 gcc_assert (nargs == 2);
24046 return ix86_expand_sse_compare (d, exp, target, swap);
24049 if (rmode == VOIDmode || rmode == tmode)
24053 || GET_MODE (target) != tmode
24054 || ! (*insn_p->operand[0].predicate) (target, tmode))
24055 target = gen_reg_rtx (tmode);
24056 real_target = target;
24060 target = gen_reg_rtx (rmode);
24061 real_target = simplify_gen_subreg (tmode, target, rmode, 0);
24064 for (i = 0; i < nargs; i++)
24066 tree arg = CALL_EXPR_ARG (exp, i);
24067 rtx op = expand_normal (arg);
24068 enum machine_mode mode = insn_p->operand[i + 1].mode;
24069 bool match = (*insn_p->operand[i + 1].predicate) (op, mode);
24071 if (last_arg_count && (i + 1) == nargs)
24073 /* SIMD shift insns take either an 8-bit immediate or
24074 register as count. But builtin functions take int as
24075 count. If count doesn't match, we put it in register. */
24078 op = simplify_gen_subreg (SImode, op, GET_MODE (op), 0);
24079 if (!(*insn_p->operand[i + 1].predicate) (op, mode))
24080 op = copy_to_reg (op);
24083 else if ((nargs - i) <= nargs_constant)
24088 case CODE_FOR_sse4_1_roundpd:
24089 case CODE_FOR_sse4_1_roundps:
24090 case CODE_FOR_sse4_1_roundsd:
24091 case CODE_FOR_sse4_1_roundss:
24092 case CODE_FOR_sse4_1_blendps:
24093 case CODE_FOR_avx_blendpd256:
24094 case CODE_FOR_avx_vpermilv4df:
24095 case CODE_FOR_avx_roundpd256:
24096 case CODE_FOR_avx_roundps256:
24097 error ("the last argument must be a 4-bit immediate");
24100 case CODE_FOR_sse4_1_blendpd:
24101 case CODE_FOR_avx_vpermilv2df:
24102 error ("the last argument must be a 2-bit immediate");
24105 case CODE_FOR_avx_vextractf128v4df:
24106 case CODE_FOR_avx_vextractf128v8sf:
24107 case CODE_FOR_avx_vextractf128v8si:
24108 case CODE_FOR_avx_vinsertf128v4df:
24109 case CODE_FOR_avx_vinsertf128v8sf:
24110 case CODE_FOR_avx_vinsertf128v8si:
24111 error ("the last argument must be a 1-bit immediate");
24114 case CODE_FOR_avx_cmpsdv2df3:
24115 case CODE_FOR_avx_cmpssv4sf3:
24116 case CODE_FOR_avx_cmppdv2df3:
24117 case CODE_FOR_avx_cmppsv4sf3:
24118 case CODE_FOR_avx_cmppdv4df3:
24119 case CODE_FOR_avx_cmppsv8sf3:
24120 error ("the last argument must be a 5-bit immediate");
24124 switch (nargs_constant)
24127 if ((nargs - i) == nargs_constant)
24129 error ("the next to last argument must be an 8-bit immediate");
24133 error ("the last argument must be an 8-bit immediate");
24136 gcc_unreachable ();
24143 if (VECTOR_MODE_P (mode))
24144 op = safe_vector_operand (op, mode);
24146 /* If we aren't optimizing, only allow one memory operand to
24148 if (memory_operand (op, mode))
24151 if (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode)
24153 if (optimize || !match || num_memory > 1)
24154 op = copy_to_mode_reg (mode, op);
24158 op = copy_to_reg (op);
24159 op = simplify_gen_subreg (mode, op, GET_MODE (op), 0);
24164 args[i].mode = mode;
24170 pat = GEN_FCN (icode) (real_target, args[0].op);
24173 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op);
24176 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op,
24180 pat = GEN_FCN (icode) (real_target, args[0].op, args[1].op,
24181 args[2].op, args[3].op);
24184 gcc_unreachable ();
24194 /* Subroutine of ix86_expand_builtin to take care of special insns
24195 with variable number of operands. */
24198 ix86_expand_special_args_builtin (const struct builtin_description *d,
24199 tree exp, rtx target)
24203 unsigned int i, nargs, arg_adjust, memory;
24207 enum machine_mode mode;
24209 enum insn_code icode = d->icode;
24210 bool last_arg_constant = false;
24211 const struct insn_data *insn_p = &insn_data[icode];
24212 enum machine_mode tmode = insn_p->operand[0].mode;
24213 enum { load, store } klass;
24215 switch ((enum ix86_special_builtin_type) d->flag)
24217 case VOID_FTYPE_VOID:
24218 emit_insn (GEN_FCN (icode) (target));
24220 case V2DI_FTYPE_PV2DI:
24221 case V32QI_FTYPE_PCCHAR:
24222 case V16QI_FTYPE_PCCHAR:
24223 case V8SF_FTYPE_PCV4SF:
24224 case V8SF_FTYPE_PCFLOAT:
24225 case V4SF_FTYPE_PCFLOAT:
24226 case V4DF_FTYPE_PCV2DF:
24227 case V4DF_FTYPE_PCDOUBLE:
24228 case V2DF_FTYPE_PCDOUBLE:
24233 case VOID_FTYPE_PV2SF_V4SF:
24234 case VOID_FTYPE_PV4DI_V4DI:
24235 case VOID_FTYPE_PV2DI_V2DI:
24236 case VOID_FTYPE_PCHAR_V32QI:
24237 case VOID_FTYPE_PCHAR_V16QI:
24238 case VOID_FTYPE_PFLOAT_V8SF:
24239 case VOID_FTYPE_PFLOAT_V4SF:
24240 case VOID_FTYPE_PDOUBLE_V4DF:
24241 case VOID_FTYPE_PDOUBLE_V2DF:
24242 case VOID_FTYPE_PDI_DI:
24243 case VOID_FTYPE_PINT_INT:
24246 /* Reserve memory operand for target. */
24247 memory = ARRAY_SIZE (args);
24249 case V4SF_FTYPE_V4SF_PCV2SF:
24250 case V2DF_FTYPE_V2DF_PCDOUBLE:
24255 case V8SF_FTYPE_PCV8SF_V8SF:
24256 case V4DF_FTYPE_PCV4DF_V4DF:
24257 case V4SF_FTYPE_PCV4SF_V4SF:
24258 case V2DF_FTYPE_PCV2DF_V2DF:
24263 case VOID_FTYPE_PV8SF_V8SF_V8SF:
24264 case VOID_FTYPE_PV4DF_V4DF_V4DF:
24265 case VOID_FTYPE_PV4SF_V4SF_V4SF:
24266 case VOID_FTYPE_PV2DF_V2DF_V2DF:
24269 /* Reserve memory operand for target. */
24270 memory = ARRAY_SIZE (args);
24273 gcc_unreachable ();
24276 gcc_assert (nargs <= ARRAY_SIZE (args));
24278 if (klass == store)
24280 arg = CALL_EXPR_ARG (exp, 0);
24281 op = expand_normal (arg);
24282 gcc_assert (target == 0);
24283 target = gen_rtx_MEM (tmode, copy_to_mode_reg (Pmode, op));
24291 || GET_MODE (target) != tmode
24292 || ! (*insn_p->operand[0].predicate) (target, tmode))
24293 target = gen_reg_rtx (tmode);
24296 for (i = 0; i < nargs; i++)
24298 enum machine_mode mode = insn_p->operand[i + 1].mode;
24301 arg = CALL_EXPR_ARG (exp, i + arg_adjust);
24302 op = expand_normal (arg);
24303 match = (*insn_p->operand[i + 1].predicate) (op, mode);
24305 if (last_arg_constant && (i + 1) == nargs)
24311 error ("the last argument must be an 8-bit immediate");
24319 /* This must be the memory operand. */
24320 op = gen_rtx_MEM (mode, copy_to_mode_reg (Pmode, op));
24321 gcc_assert (GET_MODE (op) == mode
24322 || GET_MODE (op) == VOIDmode);
24326 /* This must be register. */
24327 if (VECTOR_MODE_P (mode))
24328 op = safe_vector_operand (op, mode);
24330 gcc_assert (GET_MODE (op) == mode
24331 || GET_MODE (op) == VOIDmode);
24332 op = copy_to_mode_reg (mode, op);
24337 args[i].mode = mode;
24343 pat = GEN_FCN (icode) (target, args[0].op);
24346 pat = GEN_FCN (icode) (target, args[0].op, args[1].op);
24349 gcc_unreachable ();
24355 return klass == store ? 0 : target;
24358 /* Return the integer constant in ARG. Constrain it to be in the range
24359 of the subparts of VEC_TYPE; issue an error if not. */
24362 get_element_number (tree vec_type, tree arg)
24364 unsigned HOST_WIDE_INT elt, max = TYPE_VECTOR_SUBPARTS (vec_type) - 1;
24366 if (!host_integerp (arg, 1)
24367 || (elt = tree_low_cst (arg, 1), elt > max))
24369 error ("selector must be an integer constant in the range 0..%wi", max);
24376 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
24377 ix86_expand_vector_init. We DO have language-level syntax for this, in
24378 the form of (type){ init-list }. Except that since we can't place emms
24379 instructions from inside the compiler, we can't allow the use of MMX
24380 registers unless the user explicitly asks for it. So we do *not* define
24381 vec_set/vec_extract/vec_init patterns for MMX modes in mmx.md. Instead
24382 we have builtins invoked by mmintrin.h that gives us license to emit
24383 these sorts of instructions. */
24386 ix86_expand_vec_init_builtin (tree type, tree exp, rtx target)
24388 enum machine_mode tmode = TYPE_MODE (type);
24389 enum machine_mode inner_mode = GET_MODE_INNER (tmode);
24390 int i, n_elt = GET_MODE_NUNITS (tmode);
24391 rtvec v = rtvec_alloc (n_elt);
24393 gcc_assert (VECTOR_MODE_P (tmode));
24394 gcc_assert (call_expr_nargs (exp) == n_elt);
24396 for (i = 0; i < n_elt; ++i)
24398 rtx x = expand_normal (CALL_EXPR_ARG (exp, i));
24399 RTVEC_ELT (v, i) = gen_lowpart (inner_mode, x);
24402 if (!target || !register_operand (target, tmode))
24403 target = gen_reg_rtx (tmode);
24405 ix86_expand_vector_init (true, target, gen_rtx_PARALLEL (tmode, v));
24409 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
24410 ix86_expand_vector_extract. They would be redundant (for non-MMX) if we
24411 had a language-level syntax for referencing vector elements. */
24414 ix86_expand_vec_ext_builtin (tree exp, rtx target)
24416 enum machine_mode tmode, mode0;
24421 arg0 = CALL_EXPR_ARG (exp, 0);
24422 arg1 = CALL_EXPR_ARG (exp, 1);
24424 op0 = expand_normal (arg0);
24425 elt = get_element_number (TREE_TYPE (arg0), arg1);
24427 tmode = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
24428 mode0 = TYPE_MODE (TREE_TYPE (arg0));
24429 gcc_assert (VECTOR_MODE_P (mode0));
24431 op0 = force_reg (mode0, op0);
24433 if (optimize || !target || !register_operand (target, tmode))
24434 target = gen_reg_rtx (tmode);
24436 ix86_expand_vector_extract (true, target, op0, elt);
24441 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
24442 ix86_expand_vector_set. They would be redundant (for non-MMX) if we had
24443 a language-level syntax for referencing vector elements. */
24446 ix86_expand_vec_set_builtin (tree exp)
24448 enum machine_mode tmode, mode1;
24449 tree arg0, arg1, arg2;
24451 rtx op0, op1, target;
24453 arg0 = CALL_EXPR_ARG (exp, 0);
24454 arg1 = CALL_EXPR_ARG (exp, 1);
24455 arg2 = CALL_EXPR_ARG (exp, 2);
24457 tmode = TYPE_MODE (TREE_TYPE (arg0));
24458 mode1 = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0)));
24459 gcc_assert (VECTOR_MODE_P (tmode));
24461 op0 = expand_expr (arg0, NULL_RTX, tmode, EXPAND_NORMAL);
24462 op1 = expand_expr (arg1, NULL_RTX, mode1, EXPAND_NORMAL);
24463 elt = get_element_number (TREE_TYPE (arg0), arg2);
24465 if (GET_MODE (op1) != mode1 && GET_MODE (op1) != VOIDmode)
24466 op1 = convert_modes (mode1, GET_MODE (op1), op1, true);
24468 op0 = force_reg (tmode, op0);
24469 op1 = force_reg (mode1, op1);
24471 /* OP0 is the source of these builtin functions and shouldn't be
24472 modified. Create a copy, use it and return it as target. */
24473 target = gen_reg_rtx (tmode);
24474 emit_move_insn (target, op0);
24475 ix86_expand_vector_set (true, target, op1, elt);
24480 /* Expand an expression EXP that calls a built-in function,
24481 with result going to TARGET if that's convenient
24482 (and in mode MODE if that's convenient).
24483 SUBTARGET may be used as the target for computing one of EXP's operands.
24484 IGNORE is nonzero if the value is to be ignored. */
24487 ix86_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
24488 enum machine_mode mode ATTRIBUTE_UNUSED,
24489 int ignore ATTRIBUTE_UNUSED)
24491 const struct builtin_description *d;
24493 enum insn_code icode;
24494 tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
24495 tree arg0, arg1, arg2;
24496 rtx op0, op1, op2, pat;
24497 enum machine_mode mode0, mode1, mode2;
24498 unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
24500 /* Determine whether the builtin function is available under the current ISA.
24501 Originally the builtin was not created if it wasn't applicable to the
24502 current ISA based on the command line switches. With function specific
24503 options, we need to check in the context of the function making the call
24504 whether it is supported. */
24505 if (ix86_builtins_isa[fcode].isa
24506 && !(ix86_builtins_isa[fcode].isa & ix86_isa_flags))
24508 char *opts = ix86_target_string (ix86_builtins_isa[fcode].isa, 0, NULL,
24509 NULL, NULL, false);
24512 error ("%qE needs unknown isa option", fndecl);
24515 gcc_assert (opts != NULL);
24516 error ("%qE needs isa option %s", fndecl, opts);
24524 case IX86_BUILTIN_MASKMOVQ:
24525 case IX86_BUILTIN_MASKMOVDQU:
24526 icode = (fcode == IX86_BUILTIN_MASKMOVQ
24527 ? CODE_FOR_mmx_maskmovq
24528 : CODE_FOR_sse2_maskmovdqu);
24529 /* Note the arg order is different from the operand order. */
24530 arg1 = CALL_EXPR_ARG (exp, 0);
24531 arg2 = CALL_EXPR_ARG (exp, 1);
24532 arg0 = CALL_EXPR_ARG (exp, 2);
24533 op0 = expand_normal (arg0);
24534 op1 = expand_normal (arg1);
24535 op2 = expand_normal (arg2);
24536 mode0 = insn_data[icode].operand[0].mode;
24537 mode1 = insn_data[icode].operand[1].mode;
24538 mode2 = insn_data[icode].operand[2].mode;
24540 op0 = force_reg (Pmode, op0);
24541 op0 = gen_rtx_MEM (mode1, op0);
24543 if (! (*insn_data[icode].operand[0].predicate) (op0, mode0))
24544 op0 = copy_to_mode_reg (mode0, op0);
24545 if (! (*insn_data[icode].operand[1].predicate) (op1, mode1))
24546 op1 = copy_to_mode_reg (mode1, op1);
24547 if (! (*insn_data[icode].operand[2].predicate) (op2, mode2))
24548 op2 = copy_to_mode_reg (mode2, op2);
24549 pat = GEN_FCN (icode) (op0, op1, op2);
24555 case IX86_BUILTIN_LDMXCSR:
24556 op0 = expand_normal (CALL_EXPR_ARG (exp, 0));
24557 target = assign_386_stack_local (SImode, SLOT_VIRTUAL);
24558 emit_move_insn (target, op0);
24559 emit_insn (gen_sse_ldmxcsr (target));
24562 case IX86_BUILTIN_STMXCSR:
24563 target = assign_386_stack_local (SImode, SLOT_VIRTUAL);
24564 emit_insn (gen_sse_stmxcsr (target));
24565 return copy_to_mode_reg (SImode, target);
24567 case IX86_BUILTIN_CLFLUSH:
24568 arg0 = CALL_EXPR_ARG (exp, 0);
24569 op0 = expand_normal (arg0);
24570 icode = CODE_FOR_sse2_clflush;
24571 if (! (*insn_data[icode].operand[0].predicate) (op0, Pmode))
24572 op0 = copy_to_mode_reg (Pmode, op0);
24574 emit_insn (gen_sse2_clflush (op0));
24577 case IX86_BUILTIN_MONITOR:
24578 arg0 = CALL_EXPR_ARG (exp, 0);
24579 arg1 = CALL_EXPR_ARG (exp, 1);
24580 arg2 = CALL_EXPR_ARG (exp, 2);
24581 op0 = expand_normal (arg0);
24582 op1 = expand_normal (arg1);
24583 op2 = expand_normal (arg2);
24585 op0 = copy_to_mode_reg (Pmode, op0);
24587 op1 = copy_to_mode_reg (SImode, op1);
24589 op2 = copy_to_mode_reg (SImode, op2);
24590 emit_insn ((*ix86_gen_monitor) (op0, op1, op2));
24593 case IX86_BUILTIN_MWAIT:
24594 arg0 = CALL_EXPR_ARG (exp, 0);
24595 arg1 = CALL_EXPR_ARG (exp, 1);
24596 op0 = expand_normal (arg0);
24597 op1 = expand_normal (arg1);
24599 op0 = copy_to_mode_reg (SImode, op0);
24601 op1 = copy_to_mode_reg (SImode, op1);
24602 emit_insn (gen_sse3_mwait (op0, op1));
24605 case IX86_BUILTIN_VEC_INIT_V2SI:
24606 case IX86_BUILTIN_VEC_INIT_V4HI:
24607 case IX86_BUILTIN_VEC_INIT_V8QI:
24608 return ix86_expand_vec_init_builtin (TREE_TYPE (exp), exp, target);
24610 case IX86_BUILTIN_VEC_EXT_V2DF:
24611 case IX86_BUILTIN_VEC_EXT_V2DI:
24612 case IX86_BUILTIN_VEC_EXT_V4SF:
24613 case IX86_BUILTIN_VEC_EXT_V4SI:
24614 case IX86_BUILTIN_VEC_EXT_V8HI:
24615 case IX86_BUILTIN_VEC_EXT_V2SI:
24616 case IX86_BUILTIN_VEC_EXT_V4HI:
24617 case IX86_BUILTIN_VEC_EXT_V16QI:
24618 return ix86_expand_vec_ext_builtin (exp, target);
24620 case IX86_BUILTIN_VEC_SET_V2DI:
24621 case IX86_BUILTIN_VEC_SET_V4SF:
24622 case IX86_BUILTIN_VEC_SET_V4SI:
24623 case IX86_BUILTIN_VEC_SET_V8HI:
24624 case IX86_BUILTIN_VEC_SET_V4HI:
24625 case IX86_BUILTIN_VEC_SET_V16QI:
24626 return ix86_expand_vec_set_builtin (exp);
24628 case IX86_BUILTIN_INFQ:
24630 REAL_VALUE_TYPE inf;
24634 tmp = CONST_DOUBLE_FROM_REAL_VALUE (inf, mode);
24636 tmp = validize_mem (force_const_mem (mode, tmp));
24639 target = gen_reg_rtx (mode);
24641 emit_move_insn (target, tmp);
24649 for (i = 0, d = bdesc_special_args;
24650 i < ARRAY_SIZE (bdesc_special_args);
24652 if (d->code == fcode)
24653 return ix86_expand_special_args_builtin (d, exp, target);
24655 for (i = 0, d = bdesc_args;
24656 i < ARRAY_SIZE (bdesc_args);
24658 if (d->code == fcode)
24661 case IX86_BUILTIN_FABSQ:
24662 case IX86_BUILTIN_COPYSIGNQ:
24664 /* Emit a normal call if SSE2 isn't available. */
24665 return expand_call (exp, target, ignore);
24667 return ix86_expand_args_builtin (d, exp, target);
24670 for (i = 0, d = bdesc_comi; i < ARRAY_SIZE (bdesc_comi); i++, d++)
24671 if (d->code == fcode)
24672 return ix86_expand_sse_comi (d, exp, target);
24674 for (i = 0, d = bdesc_pcmpestr;
24675 i < ARRAY_SIZE (bdesc_pcmpestr);
24677 if (d->code == fcode)
24678 return ix86_expand_sse_pcmpestr (d, exp, target);
24680 for (i = 0, d = bdesc_pcmpistr;
24681 i < ARRAY_SIZE (bdesc_pcmpistr);
24683 if (d->code == fcode)
24684 return ix86_expand_sse_pcmpistr (d, exp, target);
24686 for (i = 0, d = bdesc_multi_arg; i < ARRAY_SIZE (bdesc_multi_arg); i++, d++)
24687 if (d->code == fcode)
24688 return ix86_expand_multi_arg_builtin (d->icode, exp, target,
24689 (enum multi_arg_type)d->flag,
24692 gcc_unreachable ();
24695 /* Returns a function decl for a vectorized version of the builtin function
24696 with builtin function code FN and the result vector type TYPE, or NULL_TREE
24697 if it is not available. */
24700 ix86_builtin_vectorized_function (unsigned int fn, tree type_out,
24703 enum machine_mode in_mode, out_mode;
24706 if (TREE_CODE (type_out) != VECTOR_TYPE
24707 || TREE_CODE (type_in) != VECTOR_TYPE)
24710 out_mode = TYPE_MODE (TREE_TYPE (type_out));
24711 out_n = TYPE_VECTOR_SUBPARTS (type_out);
24712 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24713 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24717 case BUILT_IN_SQRT:
24718 if (out_mode == DFmode && out_n == 2
24719 && in_mode == DFmode && in_n == 2)
24720 return ix86_builtins[IX86_BUILTIN_SQRTPD];
24723 case BUILT_IN_SQRTF:
24724 if (out_mode == SFmode && out_n == 4
24725 && in_mode == SFmode && in_n == 4)
24726 return ix86_builtins[IX86_BUILTIN_SQRTPS_NR];
24729 case BUILT_IN_LRINT:
24730 if (out_mode == SImode && out_n == 4
24731 && in_mode == DFmode && in_n == 2)
24732 return ix86_builtins[IX86_BUILTIN_VEC_PACK_SFIX];
24735 case BUILT_IN_LRINTF:
24736 if (out_mode == SImode && out_n == 4
24737 && in_mode == SFmode && in_n == 4)
24738 return ix86_builtins[IX86_BUILTIN_CVTPS2DQ];
24745 /* Dispatch to a handler for a vectorization library. */
24746 if (ix86_veclib_handler)
24747 return (*ix86_veclib_handler)(fn, type_out, type_in);
24752 /* Handler for an SVML-style interface to
24753 a library with vectorized intrinsics. */
24756 ix86_veclibabi_svml (enum built_in_function fn, tree type_out, tree type_in)
24759 tree fntype, new_fndecl, args;
24762 enum machine_mode el_mode, in_mode;
24765 /* The SVML is suitable for unsafe math only. */
24766 if (!flag_unsafe_math_optimizations)
24769 el_mode = TYPE_MODE (TREE_TYPE (type_out));
24770 n = TYPE_VECTOR_SUBPARTS (type_out);
24771 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24772 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24773 if (el_mode != in_mode
24781 case BUILT_IN_LOG10:
24783 case BUILT_IN_TANH:
24785 case BUILT_IN_ATAN:
24786 case BUILT_IN_ATAN2:
24787 case BUILT_IN_ATANH:
24788 case BUILT_IN_CBRT:
24789 case BUILT_IN_SINH:
24791 case BUILT_IN_ASINH:
24792 case BUILT_IN_ASIN:
24793 case BUILT_IN_COSH:
24795 case BUILT_IN_ACOSH:
24796 case BUILT_IN_ACOS:
24797 if (el_mode != DFmode || n != 2)
24801 case BUILT_IN_EXPF:
24802 case BUILT_IN_LOGF:
24803 case BUILT_IN_LOG10F:
24804 case BUILT_IN_POWF:
24805 case BUILT_IN_TANHF:
24806 case BUILT_IN_TANF:
24807 case BUILT_IN_ATANF:
24808 case BUILT_IN_ATAN2F:
24809 case BUILT_IN_ATANHF:
24810 case BUILT_IN_CBRTF:
24811 case BUILT_IN_SINHF:
24812 case BUILT_IN_SINF:
24813 case BUILT_IN_ASINHF:
24814 case BUILT_IN_ASINF:
24815 case BUILT_IN_COSHF:
24816 case BUILT_IN_COSF:
24817 case BUILT_IN_ACOSHF:
24818 case BUILT_IN_ACOSF:
24819 if (el_mode != SFmode || n != 4)
24827 bname = IDENTIFIER_POINTER (DECL_NAME (implicit_built_in_decls[fn]));
24829 if (fn == BUILT_IN_LOGF)
24830 strcpy (name, "vmlsLn4");
24831 else if (fn == BUILT_IN_LOG)
24832 strcpy (name, "vmldLn2");
24835 sprintf (name, "vmls%s", bname+10);
24836 name[strlen (name)-1] = '4';
24839 sprintf (name, "vmld%s2", bname+10);
24841 /* Convert to uppercase. */
24845 for (args = DECL_ARGUMENTS (implicit_built_in_decls[fn]); args;
24846 args = TREE_CHAIN (args))
24850 fntype = build_function_type_list (type_out, type_in, NULL);
24852 fntype = build_function_type_list (type_out, type_in, type_in, NULL);
24854 /* Build a function declaration for the vectorized function. */
24855 new_fndecl = build_decl (FUNCTION_DECL, get_identifier (name), fntype);
24856 TREE_PUBLIC (new_fndecl) = 1;
24857 DECL_EXTERNAL (new_fndecl) = 1;
24858 DECL_IS_NOVOPS (new_fndecl) = 1;
24859 TREE_READONLY (new_fndecl) = 1;
24864 /* Handler for an ACML-style interface to
24865 a library with vectorized intrinsics. */
24868 ix86_veclibabi_acml (enum built_in_function fn, tree type_out, tree type_in)
24870 char name[20] = "__vr.._";
24871 tree fntype, new_fndecl, args;
24874 enum machine_mode el_mode, in_mode;
24877 /* The ACML is 64bits only and suitable for unsafe math only as
24878 it does not correctly support parts of IEEE with the required
24879 precision such as denormals. */
24881 || !flag_unsafe_math_optimizations)
24884 el_mode = TYPE_MODE (TREE_TYPE (type_out));
24885 n = TYPE_VECTOR_SUBPARTS (type_out);
24886 in_mode = TYPE_MODE (TREE_TYPE (type_in));
24887 in_n = TYPE_VECTOR_SUBPARTS (type_in);
24888 if (el_mode != in_mode
24898 case BUILT_IN_LOG2:
24899 case BUILT_IN_LOG10:
24902 if (el_mode != DFmode
24907 case BUILT_IN_SINF:
24908 case BUILT_IN_COSF:
24909 case BUILT_IN_EXPF:
24910 case BUILT_IN_POWF:
24911 case BUILT_IN_LOGF:
24912 case BUILT_IN_LOG2F:
24913 case BUILT_IN_LOG10F:
24916 if (el_mode != SFmode
24925 bname = IDENTIFIER_POINTER (DECL_NAME (implicit_built_in_decls[fn]));
24926 sprintf (name + 7, "%s", bname+10);
24929 for (args = DECL_ARGUMENTS (implicit_built_in_decls[fn]); args;
24930 args = TREE_CHAIN (args))
24934 fntype = build_function_type_list (type_out, type_in, NULL);
24936 fntype = build_function_type_list (type_out, type_in, type_in, NULL);
24938 /* Build a function declaration for the vectorized function. */
24939 new_fndecl = build_decl (FUNCTION_DECL, get_identifier (name), fntype);
24940 TREE_PUBLIC (new_fndecl) = 1;
24941 DECL_EXTERNAL (new_fndecl) = 1;
24942 DECL_IS_NOVOPS (new_fndecl) = 1;
24943 TREE_READONLY (new_fndecl) = 1;
24949 /* Returns a decl of a function that implements conversion of an integer vector
24950 into a floating-point vector, or vice-versa. TYPE is the type of the integer
24951 side of the conversion.
24952 Return NULL_TREE if it is not available. */
24955 ix86_vectorize_builtin_conversion (unsigned int code, tree type)
24957 if (TREE_CODE (type) != VECTOR_TYPE)
24963 switch (TYPE_MODE (type))
24966 return ix86_builtins[IX86_BUILTIN_CVTDQ2PS];
24971 case FIX_TRUNC_EXPR:
24972 switch (TYPE_MODE (type))
24975 return ix86_builtins[IX86_BUILTIN_CVTTPS2DQ];
24985 /* Returns a code for a target-specific builtin that implements
24986 reciprocal of the function, or NULL_TREE if not available. */
24989 ix86_builtin_reciprocal (unsigned int fn, bool md_fn,
24990 bool sqrt ATTRIBUTE_UNUSED)
24992 if (! (TARGET_SSE_MATH && TARGET_RECIP && !optimize_insn_for_size_p ()
24993 && flag_finite_math_only && !flag_trapping_math
24994 && flag_unsafe_math_optimizations))
24998 /* Machine dependent builtins. */
25001 /* Vectorized version of sqrt to rsqrt conversion. */
25002 case IX86_BUILTIN_SQRTPS_NR:
25003 return ix86_builtins[IX86_BUILTIN_RSQRTPS_NR];
25009 /* Normal builtins. */
25012 /* Sqrt to rsqrt conversion. */
25013 case BUILT_IN_SQRTF:
25014 return ix86_builtins[IX86_BUILTIN_RSQRTF];
25021 /* Store OPERAND to the memory after reload is completed. This means
25022 that we can't easily use assign_stack_local. */
25024 ix86_force_to_memory (enum machine_mode mode, rtx operand)
25028 gcc_assert (reload_completed);
25029 if (!TARGET_64BIT_MS_ABI && TARGET_RED_ZONE)
25031 result = gen_rtx_MEM (mode,
25032 gen_rtx_PLUS (Pmode,
25034 GEN_INT (-RED_ZONE_SIZE)));
25035 emit_move_insn (result, operand);
25037 else if ((TARGET_64BIT_MS_ABI || !TARGET_RED_ZONE) && TARGET_64BIT)
25043 operand = gen_lowpart (DImode, operand);
25047 gen_rtx_SET (VOIDmode,
25048 gen_rtx_MEM (DImode,
25049 gen_rtx_PRE_DEC (DImode,
25050 stack_pointer_rtx)),
25054 gcc_unreachable ();
25056 result = gen_rtx_MEM (mode, stack_pointer_rtx);
25065 split_di (&operand, 1, operands, operands + 1);
25067 gen_rtx_SET (VOIDmode,
25068 gen_rtx_MEM (SImode,
25069 gen_rtx_PRE_DEC (Pmode,
25070 stack_pointer_rtx)),
25073 gen_rtx_SET (VOIDmode,
25074 gen_rtx_MEM (SImode,
25075 gen_rtx_PRE_DEC (Pmode,
25076 stack_pointer_rtx)),
25081 /* Store HImodes as SImodes. */
25082 operand = gen_lowpart (SImode, operand);
25086 gen_rtx_SET (VOIDmode,
25087 gen_rtx_MEM (GET_MODE (operand),
25088 gen_rtx_PRE_DEC (SImode,
25089 stack_pointer_rtx)),
25093 gcc_unreachable ();
25095 result = gen_rtx_MEM (mode, stack_pointer_rtx);
25100 /* Free operand from the memory. */
25102 ix86_free_from_memory (enum machine_mode mode)
25104 if (!TARGET_RED_ZONE || TARGET_64BIT_MS_ABI)
25108 if (mode == DImode || TARGET_64BIT)
25112 /* Use LEA to deallocate stack space. In peephole2 it will be converted
25113 to pop or add instruction if registers are available. */
25114 emit_insn (gen_rtx_SET (VOIDmode, stack_pointer_rtx,
25115 gen_rtx_PLUS (Pmode, stack_pointer_rtx,
25120 /* Put float CONST_DOUBLE in the constant pool instead of fp regs.
25121 QImode must go into class Q_REGS.
25122 Narrow ALL_REGS to GENERAL_REGS. This supports allowing movsf and
25123 movdf to do mem-to-mem moves through integer regs. */
25125 ix86_preferred_reload_class (rtx x, enum reg_class regclass)
25127 enum machine_mode mode = GET_MODE (x);
25129 /* We're only allowed to return a subclass of CLASS. Many of the
25130 following checks fail for NO_REGS, so eliminate that early. */
25131 if (regclass == NO_REGS)
25134 /* All classes can load zeros. */
25135 if (x == CONST0_RTX (mode))
25138 /* Force constants into memory if we are loading a (nonzero) constant into
25139 an MMX or SSE register. This is because there are no MMX/SSE instructions
25140 to load from a constant. */
25142 && (MAYBE_MMX_CLASS_P (regclass) || MAYBE_SSE_CLASS_P (regclass)))
25145 /* Prefer SSE regs only, if we can use them for math. */
25146 if (TARGET_SSE_MATH && !TARGET_MIX_SSE_I387 && SSE_FLOAT_MODE_P (mode))
25147 return SSE_CLASS_P (regclass) ? regclass : NO_REGS;
25149 /* Floating-point constants need more complex checks. */
25150 if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) != VOIDmode)
25152 /* General regs can load everything. */
25153 if (reg_class_subset_p (regclass, GENERAL_REGS))
25156 /* Floats can load 0 and 1 plus some others. Note that we eliminated
25157 zero above. We only want to wind up preferring 80387 registers if
25158 we plan on doing computation with them. */
25160 && standard_80387_constant_p (x))
25162 /* Limit class to non-sse. */
25163 if (regclass == FLOAT_SSE_REGS)
25165 if (regclass == FP_TOP_SSE_REGS)
25167 if (regclass == FP_SECOND_SSE_REGS)
25168 return FP_SECOND_REG;
25169 if (regclass == FLOAT_INT_REGS || regclass == FLOAT_REGS)
25176 /* Generally when we see PLUS here, it's the function invariant
25177 (plus soft-fp const_int). Which can only be computed into general
25179 if (GET_CODE (x) == PLUS)
25180 return reg_class_subset_p (regclass, GENERAL_REGS) ? regclass : NO_REGS;
25182 /* QImode constants are easy to load, but non-constant QImode data
25183 must go into Q_REGS. */
25184 if (GET_MODE (x) == QImode && !CONSTANT_P (x))
25186 if (reg_class_subset_p (regclass, Q_REGS))
25188 if (reg_class_subset_p (Q_REGS, regclass))
25196 /* Discourage putting floating-point values in SSE registers unless
25197 SSE math is being used, and likewise for the 387 registers. */
25199 ix86_preferred_output_reload_class (rtx x, enum reg_class regclass)
25201 enum machine_mode mode = GET_MODE (x);
25203 /* Restrict the output reload class to the register bank that we are doing
25204 math on. If we would like not to return a subset of CLASS, reject this
25205 alternative: if reload cannot do this, it will still use its choice. */
25206 mode = GET_MODE (x);
25207 if (TARGET_SSE_MATH && SSE_FLOAT_MODE_P (mode))
25208 return MAYBE_SSE_CLASS_P (regclass) ? SSE_REGS : NO_REGS;
25210 if (X87_FLOAT_MODE_P (mode))
25212 if (regclass == FP_TOP_SSE_REGS)
25214 else if (regclass == FP_SECOND_SSE_REGS)
25215 return FP_SECOND_REG;
25217 return FLOAT_CLASS_P (regclass) ? regclass : NO_REGS;
25223 static enum reg_class
25224 ix86_secondary_reload (bool in_p, rtx x, enum reg_class rclass,
25225 enum machine_mode mode,
25226 secondary_reload_info *sri ATTRIBUTE_UNUSED)
25228 /* QImode spills from non-QI registers require
25229 intermediate register on 32bit targets. */
25230 if (!in_p && mode == QImode && !TARGET_64BIT
25231 && (rclass == GENERAL_REGS
25232 || rclass == LEGACY_REGS
25233 || rclass == INDEX_REGS))
25242 if (regno >= FIRST_PSEUDO_REGISTER || GET_CODE (x) == SUBREG)
25243 regno = true_regnum (x);
25245 /* Return Q_REGS if the operand is in memory. */
25253 /* If we are copying between general and FP registers, we need a memory
25254 location. The same is true for SSE and MMX registers.
25256 To optimize register_move_cost performance, allow inline variant.
25258 The macro can't work reliably when one of the CLASSES is class containing
25259 registers from multiple units (SSE, MMX, integer). We avoid this by never
25260 combining those units in single alternative in the machine description.
25261 Ensure that this constraint holds to avoid unexpected surprises.
25263 When STRICT is false, we are being called from REGISTER_MOVE_COST, so do not
25264 enforce these sanity checks. */
25267 inline_secondary_memory_needed (enum reg_class class1, enum reg_class class2,
25268 enum machine_mode mode, int strict)
25270 if (MAYBE_FLOAT_CLASS_P (class1) != FLOAT_CLASS_P (class1)
25271 || MAYBE_FLOAT_CLASS_P (class2) != FLOAT_CLASS_P (class2)
25272 || MAYBE_SSE_CLASS_P (class1) != SSE_CLASS_P (class1)
25273 || MAYBE_SSE_CLASS_P (class2) != SSE_CLASS_P (class2)
25274 || MAYBE_MMX_CLASS_P (class1) != MMX_CLASS_P (class1)
25275 || MAYBE_MMX_CLASS_P (class2) != MMX_CLASS_P (class2))
25277 gcc_assert (!strict);
25281 if (FLOAT_CLASS_P (class1) != FLOAT_CLASS_P (class2))
25284 /* ??? This is a lie. We do have moves between mmx/general, and for
25285 mmx/sse2. But by saying we need secondary memory we discourage the
25286 register allocator from using the mmx registers unless needed. */
25287 if (MMX_CLASS_P (class1) != MMX_CLASS_P (class2))
25290 if (SSE_CLASS_P (class1) != SSE_CLASS_P (class2))
25292 /* SSE1 doesn't have any direct moves from other classes. */
25296 /* If the target says that inter-unit moves are more expensive
25297 than moving through memory, then don't generate them. */
25298 if (!TARGET_INTER_UNIT_MOVES)
25301 /* Between SSE and general, we have moves no larger than word size. */
25302 if (GET_MODE_SIZE (mode) > UNITS_PER_WORD)
25310 ix86_secondary_memory_needed (enum reg_class class1, enum reg_class class2,
25311 enum machine_mode mode, int strict)
25313 return inline_secondary_memory_needed (class1, class2, mode, strict);
25316 /* Return true if the registers in CLASS cannot represent the change from
25317 modes FROM to TO. */
25320 ix86_cannot_change_mode_class (enum machine_mode from, enum machine_mode to,
25321 enum reg_class regclass)
25326 /* x87 registers can't do subreg at all, as all values are reformatted
25327 to extended precision. */
25328 if (MAYBE_FLOAT_CLASS_P (regclass))
25331 if (MAYBE_SSE_CLASS_P (regclass) || MAYBE_MMX_CLASS_P (regclass))
25333 /* Vector registers do not support QI or HImode loads. If we don't
25334 disallow a change to these modes, reload will assume it's ok to
25335 drop the subreg from (subreg:SI (reg:HI 100) 0). This affects
25336 the vec_dupv4hi pattern. */
25337 if (GET_MODE_SIZE (from) < 4)
25340 /* Vector registers do not support subreg with nonzero offsets, which
25341 are otherwise valid for integer registers. Since we can't see
25342 whether we have a nonzero offset from here, prohibit all
25343 nonparadoxical subregs changing size. */
25344 if (GET_MODE_SIZE (to) < GET_MODE_SIZE (from))
25351 /* Return the cost of moving data of mode M between a
25352 register and memory. A value of 2 is the default; this cost is
25353 relative to those in `REGISTER_MOVE_COST'.
25355 This function is used extensively by register_move_cost that is used to
25356 build tables at startup. Make it inline in this case.
25357 When IN is 2, return maximum of in and out move cost.
25359 If moving between registers and memory is more expensive than
25360 between two registers, you should define this macro to express the
25363 Model also increased moving costs of QImode registers in non
25367 inline_memory_move_cost (enum machine_mode mode, enum reg_class regclass,
25371 if (FLOAT_CLASS_P (regclass))
25389 return MAX (ix86_cost->fp_load [index], ix86_cost->fp_store [index]);
25390 return in ? ix86_cost->fp_load [index] : ix86_cost->fp_store [index];
25392 if (SSE_CLASS_P (regclass))
25395 switch (GET_MODE_SIZE (mode))
25410 return MAX (ix86_cost->sse_load [index], ix86_cost->sse_store [index]);
25411 return in ? ix86_cost->sse_load [index] : ix86_cost->sse_store [index];
25413 if (MMX_CLASS_P (regclass))
25416 switch (GET_MODE_SIZE (mode))
25428 return MAX (ix86_cost->mmx_load [index], ix86_cost->mmx_store [index]);
25429 return in ? ix86_cost->mmx_load [index] : ix86_cost->mmx_store [index];
25431 switch (GET_MODE_SIZE (mode))
25434 if (Q_CLASS_P (regclass) || TARGET_64BIT)
25437 return ix86_cost->int_store[0];
25438 if (TARGET_PARTIAL_REG_DEPENDENCY
25439 && optimize_function_for_speed_p (cfun))
25440 cost = ix86_cost->movzbl_load;
25442 cost = ix86_cost->int_load[0];
25444 return MAX (cost, ix86_cost->int_store[0]);
25450 return MAX (ix86_cost->movzbl_load, ix86_cost->int_store[0] + 4);
25452 return ix86_cost->movzbl_load;
25454 return ix86_cost->int_store[0] + 4;
25459 return MAX (ix86_cost->int_load[1], ix86_cost->int_store[1]);
25460 return in ? ix86_cost->int_load[1] : ix86_cost->int_store[1];
25462 /* Compute number of 32bit moves needed. TFmode is moved as XFmode. */
25463 if (mode == TFmode)
25466 cost = MAX (ix86_cost->int_load[2] , ix86_cost->int_store[2]);
25468 cost = ix86_cost->int_load[2];
25470 cost = ix86_cost->int_store[2];
25471 return (cost * (((int) GET_MODE_SIZE (mode)
25472 + UNITS_PER_WORD - 1) / UNITS_PER_WORD));
25477 ix86_memory_move_cost (enum machine_mode mode, enum reg_class regclass, int in)
25479 return inline_memory_move_cost (mode, regclass, in);
25483 /* Return the cost of moving data from a register in class CLASS1 to
25484 one in class CLASS2.
25486 It is not required that the cost always equal 2 when FROM is the same as TO;
25487 on some machines it is expensive to move between registers if they are not
25488 general registers. */
25491 ix86_register_move_cost (enum machine_mode mode, enum reg_class class1,
25492 enum reg_class class2)
25494 /* In case we require secondary memory, compute cost of the store followed
25495 by load. In order to avoid bad register allocation choices, we need
25496 for this to be *at least* as high as the symmetric MEMORY_MOVE_COST. */
25498 if (inline_secondary_memory_needed (class1, class2, mode, 0))
25502 cost += inline_memory_move_cost (mode, class1, 2);
25503 cost += inline_memory_move_cost (mode, class2, 2);
25505 /* In case of copying from general_purpose_register we may emit multiple
25506 stores followed by single load causing memory size mismatch stall.
25507 Count this as arbitrarily high cost of 20. */
25508 if (CLASS_MAX_NREGS (class1, mode) > CLASS_MAX_NREGS (class2, mode))
25511 /* In the case of FP/MMX moves, the registers actually overlap, and we
25512 have to switch modes in order to treat them differently. */
25513 if ((MMX_CLASS_P (class1) && MAYBE_FLOAT_CLASS_P (class2))
25514 || (MMX_CLASS_P (class2) && MAYBE_FLOAT_CLASS_P (class1)))
25520 /* Moves between SSE/MMX and integer unit are expensive. */
25521 if (MMX_CLASS_P (class1) != MMX_CLASS_P (class2)
25522 || SSE_CLASS_P (class1) != SSE_CLASS_P (class2))
25524 /* ??? By keeping returned value relatively high, we limit the number
25525 of moves between integer and MMX/SSE registers for all targets.
25526 Additionally, high value prevents problem with x86_modes_tieable_p(),
25527 where integer modes in MMX/SSE registers are not tieable
25528 because of missing QImode and HImode moves to, from or between
25529 MMX/SSE registers. */
25530 return MAX (8, ix86_cost->mmxsse_to_integer);
25532 if (MAYBE_FLOAT_CLASS_P (class1))
25533 return ix86_cost->fp_move;
25534 if (MAYBE_SSE_CLASS_P (class1))
25535 return ix86_cost->sse_move;
25536 if (MAYBE_MMX_CLASS_P (class1))
25537 return ix86_cost->mmx_move;
25541 /* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
25544 ix86_hard_regno_mode_ok (int regno, enum machine_mode mode)
25546 /* Flags and only flags can only hold CCmode values. */
25547 if (CC_REGNO_P (regno))
25548 return GET_MODE_CLASS (mode) == MODE_CC;
25549 if (GET_MODE_CLASS (mode) == MODE_CC
25550 || GET_MODE_CLASS (mode) == MODE_RANDOM
25551 || GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
25553 if (FP_REGNO_P (regno))
25554 return VALID_FP_MODE_P (mode);
25555 if (SSE_REGNO_P (regno))
25557 /* We implement the move patterns for all vector modes into and
25558 out of SSE registers, even when no operation instructions
25559 are available. OImode move is available only when AVX is
25561 return ((TARGET_AVX && mode == OImode)
25562 || VALID_AVX256_REG_MODE (mode)
25563 || VALID_SSE_REG_MODE (mode)
25564 || VALID_SSE2_REG_MODE (mode)
25565 || VALID_MMX_REG_MODE (mode)
25566 || VALID_MMX_REG_MODE_3DNOW (mode));
25568 if (MMX_REGNO_P (regno))
25570 /* We implement the move patterns for 3DNOW modes even in MMX mode,
25571 so if the register is available at all, then we can move data of
25572 the given mode into or out of it. */
25573 return (VALID_MMX_REG_MODE (mode)
25574 || VALID_MMX_REG_MODE_3DNOW (mode));
25577 if (mode == QImode)
25579 /* Take care for QImode values - they can be in non-QI regs,
25580 but then they do cause partial register stalls. */
25581 if (regno < 4 || TARGET_64BIT)
25583 if (!TARGET_PARTIAL_REG_STALL)
25585 return reload_in_progress || reload_completed;
25587 /* We handle both integer and floats in the general purpose registers. */
25588 else if (VALID_INT_MODE_P (mode))
25590 else if (VALID_FP_MODE_P (mode))
25592 else if (VALID_DFP_MODE_P (mode))
25594 /* Lots of MMX code casts 8 byte vector modes to DImode. If we then go
25595 on to use that value in smaller contexts, this can easily force a
25596 pseudo to be allocated to GENERAL_REGS. Since this is no worse than
25597 supporting DImode, allow it. */
25598 else if (VALID_MMX_REG_MODE_3DNOW (mode) || VALID_MMX_REG_MODE (mode))
25604 /* A subroutine of ix86_modes_tieable_p. Return true if MODE is a
25605 tieable integer mode. */
25608 ix86_tieable_integer_mode_p (enum machine_mode mode)
25617 return TARGET_64BIT || !TARGET_PARTIAL_REG_STALL;
25620 return TARGET_64BIT;
25627 /* Return true if MODE1 is accessible in a register that can hold MODE2
25628 without copying. That is, all register classes that can hold MODE2
25629 can also hold MODE1. */
25632 ix86_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2)
25634 if (mode1 == mode2)
25637 if (ix86_tieable_integer_mode_p (mode1)
25638 && ix86_tieable_integer_mode_p (mode2))
25641 /* MODE2 being XFmode implies fp stack or general regs, which means we
25642 can tie any smaller floating point modes to it. Note that we do not
25643 tie this with TFmode. */
25644 if (mode2 == XFmode)
25645 return mode1 == SFmode || mode1 == DFmode;
25647 /* MODE2 being DFmode implies fp stack, general or sse regs, which means
25648 that we can tie it with SFmode. */
25649 if (mode2 == DFmode)
25650 return mode1 == SFmode;
25652 /* If MODE2 is only appropriate for an SSE register, then tie with
25653 any other mode acceptable to SSE registers. */
25654 if (GET_MODE_SIZE (mode2) == 16
25655 && ix86_hard_regno_mode_ok (FIRST_SSE_REG, mode2))
25656 return (GET_MODE_SIZE (mode1) == 16
25657 && ix86_hard_regno_mode_ok (FIRST_SSE_REG, mode1));
25659 /* If MODE2 is appropriate for an MMX register, then tie
25660 with any other mode acceptable to MMX registers. */
25661 if (GET_MODE_SIZE (mode2) == 8
25662 && ix86_hard_regno_mode_ok (FIRST_MMX_REG, mode2))
25663 return (GET_MODE_SIZE (mode1) == 8
25664 && ix86_hard_regno_mode_ok (FIRST_MMX_REG, mode1));
25669 /* Compute a (partial) cost for rtx X. Return true if the complete
25670 cost has been computed, and false if subexpressions should be
25671 scanned. In either case, *TOTAL contains the cost result. */
25674 ix86_rtx_costs (rtx x, int code, int outer_code_i, int *total, bool speed)
25676 enum rtx_code outer_code = (enum rtx_code) outer_code_i;
25677 enum machine_mode mode = GET_MODE (x);
25678 const struct processor_costs *cost = speed ? ix86_cost : &ix86_size_cost;
25686 if (TARGET_64BIT && !x86_64_immediate_operand (x, VOIDmode))
25688 else if (TARGET_64BIT && !x86_64_zext_immediate_operand (x, VOIDmode))
25690 else if (flag_pic && SYMBOLIC_CONST (x)
25692 || (!GET_CODE (x) != LABEL_REF
25693 && (GET_CODE (x) != SYMBOL_REF
25694 || !SYMBOL_REF_LOCAL_P (x)))))
25701 if (mode == VOIDmode)
25704 switch (standard_80387_constant_p (x))
25709 default: /* Other constants */
25714 /* Start with (MEM (SYMBOL_REF)), since that's where
25715 it'll probably end up. Add a penalty for size. */
25716 *total = (COSTS_N_INSNS (1)
25717 + (flag_pic != 0 && !TARGET_64BIT)
25718 + (mode == SFmode ? 0 : mode == DFmode ? 1 : 2));
25724 /* The zero extensions is often completely free on x86_64, so make
25725 it as cheap as possible. */
25726 if (TARGET_64BIT && mode == DImode
25727 && GET_MODE (XEXP (x, 0)) == SImode)
25729 else if (TARGET_ZERO_EXTEND_WITH_AND)
25730 *total = cost->add;
25732 *total = cost->movzx;
25736 *total = cost->movsx;
25740 if (CONST_INT_P (XEXP (x, 1))
25741 && (GET_MODE (XEXP (x, 0)) != DImode || TARGET_64BIT))
25743 HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
25746 *total = cost->add;
25749 if ((value == 2 || value == 3)
25750 && cost->lea <= cost->shift_const)
25752 *total = cost->lea;
25762 if (!TARGET_64BIT && GET_MODE (XEXP (x, 0)) == DImode)
25764 if (CONST_INT_P (XEXP (x, 1)))
25766 if (INTVAL (XEXP (x, 1)) > 32)
25767 *total = cost->shift_const + COSTS_N_INSNS (2);
25769 *total = cost->shift_const * 2;
25773 if (GET_CODE (XEXP (x, 1)) == AND)
25774 *total = cost->shift_var * 2;
25776 *total = cost->shift_var * 6 + COSTS_N_INSNS (2);
25781 if (CONST_INT_P (XEXP (x, 1)))
25782 *total = cost->shift_const;
25784 *total = cost->shift_var;
25789 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25791 /* ??? SSE scalar cost should be used here. */
25792 *total = cost->fmul;
25795 else if (X87_FLOAT_MODE_P (mode))
25797 *total = cost->fmul;
25800 else if (FLOAT_MODE_P (mode))
25802 /* ??? SSE vector cost should be used here. */
25803 *total = cost->fmul;
25808 rtx op0 = XEXP (x, 0);
25809 rtx op1 = XEXP (x, 1);
25811 if (CONST_INT_P (XEXP (x, 1)))
25813 unsigned HOST_WIDE_INT value = INTVAL (XEXP (x, 1));
25814 for (nbits = 0; value != 0; value &= value - 1)
25818 /* This is arbitrary. */
25821 /* Compute costs correctly for widening multiplication. */
25822 if ((GET_CODE (op0) == SIGN_EXTEND || GET_CODE (op0) == ZERO_EXTEND)
25823 && GET_MODE_SIZE (GET_MODE (XEXP (op0, 0))) * 2
25824 == GET_MODE_SIZE (mode))
25826 int is_mulwiden = 0;
25827 enum machine_mode inner_mode = GET_MODE (op0);
25829 if (GET_CODE (op0) == GET_CODE (op1))
25830 is_mulwiden = 1, op1 = XEXP (op1, 0);
25831 else if (CONST_INT_P (op1))
25833 if (GET_CODE (op0) == SIGN_EXTEND)
25834 is_mulwiden = trunc_int_for_mode (INTVAL (op1), inner_mode)
25837 is_mulwiden = !(INTVAL (op1) & ~GET_MODE_MASK (inner_mode));
25841 op0 = XEXP (op0, 0), mode = GET_MODE (op0);
25844 *total = (cost->mult_init[MODE_INDEX (mode)]
25845 + nbits * cost->mult_bit
25846 + rtx_cost (op0, outer_code, speed) + rtx_cost (op1, outer_code, speed));
25855 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25856 /* ??? SSE cost should be used here. */
25857 *total = cost->fdiv;
25858 else if (X87_FLOAT_MODE_P (mode))
25859 *total = cost->fdiv;
25860 else if (FLOAT_MODE_P (mode))
25861 /* ??? SSE vector cost should be used here. */
25862 *total = cost->fdiv;
25864 *total = cost->divide[MODE_INDEX (mode)];
25868 if (GET_MODE_CLASS (mode) == MODE_INT
25869 && GET_MODE_BITSIZE (mode) <= GET_MODE_BITSIZE (Pmode))
25871 if (GET_CODE (XEXP (x, 0)) == PLUS
25872 && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
25873 && CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 0), 1))
25874 && CONSTANT_P (XEXP (x, 1)))
25876 HOST_WIDE_INT val = INTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1));
25877 if (val == 2 || val == 4 || val == 8)
25879 *total = cost->lea;
25880 *total += rtx_cost (XEXP (XEXP (x, 0), 1), outer_code, speed);
25881 *total += rtx_cost (XEXP (XEXP (XEXP (x, 0), 0), 0),
25882 outer_code, speed);
25883 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25887 else if (GET_CODE (XEXP (x, 0)) == MULT
25888 && CONST_INT_P (XEXP (XEXP (x, 0), 1)))
25890 HOST_WIDE_INT val = INTVAL (XEXP (XEXP (x, 0), 1));
25891 if (val == 2 || val == 4 || val == 8)
25893 *total = cost->lea;
25894 *total += rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed);
25895 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25899 else if (GET_CODE (XEXP (x, 0)) == PLUS)
25901 *total = cost->lea;
25902 *total += rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed);
25903 *total += rtx_cost (XEXP (XEXP (x, 0), 1), outer_code, speed);
25904 *total += rtx_cost (XEXP (x, 1), outer_code, speed);
25911 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25913 /* ??? SSE cost should be used here. */
25914 *total = cost->fadd;
25917 else if (X87_FLOAT_MODE_P (mode))
25919 *total = cost->fadd;
25922 else if (FLOAT_MODE_P (mode))
25924 /* ??? SSE vector cost should be used here. */
25925 *total = cost->fadd;
25933 if (!TARGET_64BIT && mode == DImode)
25935 *total = (cost->add * 2
25936 + (rtx_cost (XEXP (x, 0), outer_code, speed)
25937 << (GET_MODE (XEXP (x, 0)) != DImode))
25938 + (rtx_cost (XEXP (x, 1), outer_code, speed)
25939 << (GET_MODE (XEXP (x, 1)) != DImode)));
25945 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25947 /* ??? SSE cost should be used here. */
25948 *total = cost->fchs;
25951 else if (X87_FLOAT_MODE_P (mode))
25953 *total = cost->fchs;
25956 else if (FLOAT_MODE_P (mode))
25958 /* ??? SSE vector cost should be used here. */
25959 *total = cost->fchs;
25965 if (!TARGET_64BIT && mode == DImode)
25966 *total = cost->add * 2;
25968 *total = cost->add;
25972 if (GET_CODE (XEXP (x, 0)) == ZERO_EXTRACT
25973 && XEXP (XEXP (x, 0), 1) == const1_rtx
25974 && CONST_INT_P (XEXP (XEXP (x, 0), 2))
25975 && XEXP (x, 1) == const0_rtx)
25977 /* This kind of construct is implemented using test[bwl].
25978 Treat it as if we had an AND. */
25979 *total = (cost->add
25980 + rtx_cost (XEXP (XEXP (x, 0), 0), outer_code, speed)
25981 + rtx_cost (const1_rtx, outer_code, speed));
25987 if (!(SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH))
25992 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
25993 /* ??? SSE cost should be used here. */
25994 *total = cost->fabs;
25995 else if (X87_FLOAT_MODE_P (mode))
25996 *total = cost->fabs;
25997 else if (FLOAT_MODE_P (mode))
25998 /* ??? SSE vector cost should be used here. */
25999 *total = cost->fabs;
26003 if (SSE_FLOAT_MODE_P (mode) && TARGET_SSE_MATH)
26004 /* ??? SSE cost should be used here. */
26005 *total = cost->fsqrt;
26006 else if (X87_FLOAT_MODE_P (mode))
26007 *total = cost->fsqrt;
26008 else if (FLOAT_MODE_P (mode))
26009 /* ??? SSE vector cost should be used here. */
26010 *total = cost->fsqrt;
26014 if (XINT (x, 1) == UNSPEC_TP)
26025 static int current_machopic_label_num;
26027 /* Given a symbol name and its associated stub, write out the
26028 definition of the stub. */
26031 machopic_output_stub (FILE *file, const char *symb, const char *stub)
26033 unsigned int length;
26034 char *binder_name, *symbol_name, lazy_ptr_name[32];
26035 int label = ++current_machopic_label_num;
26037 /* For 64-bit we shouldn't get here. */
26038 gcc_assert (!TARGET_64BIT);
26040 /* Lose our funky encoding stuff so it doesn't contaminate the stub. */
26041 symb = (*targetm.strip_name_encoding) (symb);
26043 length = strlen (stub);
26044 binder_name = XALLOCAVEC (char, length + 32);
26045 GEN_BINDER_NAME_FOR_STUB (binder_name, stub, length);
26047 length = strlen (symb);
26048 symbol_name = XALLOCAVEC (char, length + 32);
26049 GEN_SYMBOL_NAME_FOR_SYMBOL (symbol_name, symb, length);
26051 sprintf (lazy_ptr_name, "L%d$lz", label);
26054 switch_to_section (darwin_sections[machopic_picsymbol_stub_section]);
26056 switch_to_section (darwin_sections[machopic_symbol_stub_section]);
26058 fprintf (file, "%s:\n", stub);
26059 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
26063 fprintf (file, "\tcall\tLPC$%d\nLPC$%d:\tpopl\t%%eax\n", label, label);
26064 fprintf (file, "\tmovl\t%s-LPC$%d(%%eax),%%edx\n", lazy_ptr_name, label);
26065 fprintf (file, "\tjmp\t*%%edx\n");
26068 fprintf (file, "\tjmp\t*%s\n", lazy_ptr_name);
26070 fprintf (file, "%s:\n", binder_name);
26074 fprintf (file, "\tlea\t%s-LPC$%d(%%eax),%%eax\n", lazy_ptr_name, label);
26075 fprintf (file, "\tpushl\t%%eax\n");
26078 fprintf (file, "\tpushl\t$%s\n", lazy_ptr_name);
26080 fprintf (file, "\tjmp\tdyld_stub_binding_helper\n");
26082 switch_to_section (darwin_sections[machopic_lazy_symbol_ptr_section]);
26083 fprintf (file, "%s:\n", lazy_ptr_name);
26084 fprintf (file, "\t.indirect_symbol %s\n", symbol_name);
26085 fprintf (file, "\t.long %s\n", binder_name);
26089 darwin_x86_file_end (void)
26091 darwin_file_end ();
26094 #endif /* TARGET_MACHO */
26096 /* Order the registers for register allocator. */
26099 x86_order_regs_for_local_alloc (void)
26104 /* First allocate the local general purpose registers. */
26105 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
26106 if (GENERAL_REGNO_P (i) && call_used_regs[i])
26107 reg_alloc_order [pos++] = i;
26109 /* Global general purpose registers. */
26110 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
26111 if (GENERAL_REGNO_P (i) && !call_used_regs[i])
26112 reg_alloc_order [pos++] = i;
26114 /* x87 registers come first in case we are doing FP math
26116 if (!TARGET_SSE_MATH)
26117 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++)
26118 reg_alloc_order [pos++] = i;
26120 /* SSE registers. */
26121 for (i = FIRST_SSE_REG; i <= LAST_SSE_REG; i++)
26122 reg_alloc_order [pos++] = i;
26123 for (i = FIRST_REX_SSE_REG; i <= LAST_REX_SSE_REG; i++)
26124 reg_alloc_order [pos++] = i;
26126 /* x87 registers. */
26127 if (TARGET_SSE_MATH)
26128 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++)
26129 reg_alloc_order [pos++] = i;
26131 for (i = FIRST_MMX_REG; i <= LAST_MMX_REG; i++)
26132 reg_alloc_order [pos++] = i;
26134 /* Initialize the rest of array as we do not allocate some registers
26136 while (pos < FIRST_PSEUDO_REGISTER)
26137 reg_alloc_order [pos++] = 0;
26140 /* Handle a "ms_abi" or "sysv" attribute; arguments as in
26141 struct attribute_spec.handler. */
26143 ix86_handle_abi_attribute (tree *node, tree name,
26144 tree args ATTRIBUTE_UNUSED,
26145 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
26147 if (TREE_CODE (*node) != FUNCTION_TYPE
26148 && TREE_CODE (*node) != METHOD_TYPE
26149 && TREE_CODE (*node) != FIELD_DECL
26150 && TREE_CODE (*node) != TYPE_DECL)
26152 warning (OPT_Wattributes, "%qs attribute only applies to functions",
26153 IDENTIFIER_POINTER (name));
26154 *no_add_attrs = true;
26159 warning (OPT_Wattributes, "%qs attribute only available for 64-bit",
26160 IDENTIFIER_POINTER (name));
26161 *no_add_attrs = true;
26165 /* Can combine regparm with all attributes but fastcall. */
26166 if (is_attribute_p ("ms_abi", name))
26168 if (lookup_attribute ("sysv_abi", TYPE_ATTRIBUTES (*node)))
26170 error ("ms_abi and sysv_abi attributes are not compatible");
26175 else if (is_attribute_p ("sysv_abi", name))
26177 if (lookup_attribute ("ms_abi", TYPE_ATTRIBUTES (*node)))
26179 error ("ms_abi and sysv_abi attributes are not compatible");
26188 /* Handle a "ms_struct" or "gcc_struct" attribute; arguments as in
26189 struct attribute_spec.handler. */
26191 ix86_handle_struct_attribute (tree *node, tree name,
26192 tree args ATTRIBUTE_UNUSED,
26193 int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
26196 if (DECL_P (*node))
26198 if (TREE_CODE (*node) == TYPE_DECL)
26199 type = &TREE_TYPE (*node);
26204 if (!(type && (TREE_CODE (*type) == RECORD_TYPE
26205 || TREE_CODE (*type) == UNION_TYPE)))
26207 warning (OPT_Wattributes, "%qs attribute ignored",
26208 IDENTIFIER_POINTER (name));
26209 *no_add_attrs = true;
26212 else if ((is_attribute_p ("ms_struct", name)
26213 && lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (*type)))
26214 || ((is_attribute_p ("gcc_struct", name)
26215 && lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (*type)))))
26217 warning (OPT_Wattributes, "%qs incompatible attribute ignored",
26218 IDENTIFIER_POINTER (name));
26219 *no_add_attrs = true;
26226 ix86_ms_bitfield_layout_p (const_tree record_type)
26228 return (TARGET_MS_BITFIELD_LAYOUT &&
26229 !lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (record_type)))
26230 || lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (record_type));
26233 /* Returns an expression indicating where the this parameter is
26234 located on entry to the FUNCTION. */
26237 x86_this_parameter (tree function)
26239 tree type = TREE_TYPE (function);
26240 bool aggr = aggregate_value_p (TREE_TYPE (type), type) != 0;
26245 const int *parm_regs;
26247 if (ix86_function_type_abi (type) == MS_ABI)
26248 parm_regs = x86_64_ms_abi_int_parameter_registers;
26250 parm_regs = x86_64_int_parameter_registers;
26251 return gen_rtx_REG (DImode, parm_regs[aggr]);
26254 nregs = ix86_function_regparm (type, function);
26256 if (nregs > 0 && !stdarg_p (type))
26260 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type)))
26261 regno = aggr ? DX_REG : CX_REG;
26269 return gen_rtx_MEM (SImode,
26270 plus_constant (stack_pointer_rtx, 4));
26273 return gen_rtx_REG (SImode, regno);
26276 return gen_rtx_MEM (SImode, plus_constant (stack_pointer_rtx, aggr ? 8 : 4));
26279 /* Determine whether x86_output_mi_thunk can succeed. */
26282 x86_can_output_mi_thunk (const_tree thunk ATTRIBUTE_UNUSED,
26283 HOST_WIDE_INT delta ATTRIBUTE_UNUSED,
26284 HOST_WIDE_INT vcall_offset, const_tree function)
26286 /* 64-bit can handle anything. */
26290 /* For 32-bit, everything's fine if we have one free register. */
26291 if (ix86_function_regparm (TREE_TYPE (function), function) < 3)
26294 /* Need a free register for vcall_offset. */
26298 /* Need a free register for GOT references. */
26299 if (flag_pic && !(*targetm.binds_local_p) (function))
26302 /* Otherwise ok. */
26306 /* Output the assembler code for a thunk function. THUNK_DECL is the
26307 declaration for the thunk function itself, FUNCTION is the decl for
26308 the target function. DELTA is an immediate constant offset to be
26309 added to THIS. If VCALL_OFFSET is nonzero, the word at
26310 *(*this + vcall_offset) should be added to THIS. */
26313 x86_output_mi_thunk (FILE *file ATTRIBUTE_UNUSED,
26314 tree thunk ATTRIBUTE_UNUSED, HOST_WIDE_INT delta,
26315 HOST_WIDE_INT vcall_offset, tree function)
26318 rtx this_param = x86_this_parameter (function);
26321 /* If VCALL_OFFSET, we'll need THIS in a register. Might as well
26322 pull it in now and let DELTA benefit. */
26323 if (REG_P (this_param))
26324 this_reg = this_param;
26325 else if (vcall_offset)
26327 /* Put the this parameter into %eax. */
26328 xops[0] = this_param;
26329 xops[1] = this_reg = gen_rtx_REG (Pmode, AX_REG);
26330 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26333 this_reg = NULL_RTX;
26335 /* Adjust the this parameter by a fixed constant. */
26338 xops[0] = GEN_INT (delta);
26339 xops[1] = this_reg ? this_reg : this_param;
26342 if (!x86_64_general_operand (xops[0], DImode))
26344 tmp = gen_rtx_REG (DImode, R10_REG);
26346 output_asm_insn ("mov{q}\t{%1, %0|%0, %1}", xops);
26348 xops[1] = this_param;
26350 output_asm_insn ("add{q}\t{%0, %1|%1, %0}", xops);
26353 output_asm_insn ("add{l}\t{%0, %1|%1, %0}", xops);
26356 /* Adjust the this parameter by a value stored in the vtable. */
26360 tmp = gen_rtx_REG (DImode, R10_REG);
26363 int tmp_regno = CX_REG;
26364 if (lookup_attribute ("fastcall",
26365 TYPE_ATTRIBUTES (TREE_TYPE (function))))
26366 tmp_regno = AX_REG;
26367 tmp = gen_rtx_REG (SImode, tmp_regno);
26370 xops[0] = gen_rtx_MEM (Pmode, this_reg);
26372 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26374 /* Adjust the this parameter. */
26375 xops[0] = gen_rtx_MEM (Pmode, plus_constant (tmp, vcall_offset));
26376 if (TARGET_64BIT && !memory_operand (xops[0], Pmode))
26378 rtx tmp2 = gen_rtx_REG (DImode, R11_REG);
26379 xops[0] = GEN_INT (vcall_offset);
26381 output_asm_insn ("mov{q}\t{%0, %1|%1, %0}", xops);
26382 xops[0] = gen_rtx_MEM (Pmode, gen_rtx_PLUS (Pmode, tmp, tmp2));
26384 xops[1] = this_reg;
26385 output_asm_insn ("add%z1\t{%0, %1|%1, %0}", xops);
26388 /* If necessary, drop THIS back to its stack slot. */
26389 if (this_reg && this_reg != this_param)
26391 xops[0] = this_reg;
26392 xops[1] = this_param;
26393 output_asm_insn ("mov%z1\t{%0, %1|%1, %0}", xops);
26396 xops[0] = XEXP (DECL_RTL (function), 0);
26399 if (!flag_pic || (*targetm.binds_local_p) (function))
26400 output_asm_insn ("jmp\t%P0", xops);
26401 /* All thunks should be in the same object as their target,
26402 and thus binds_local_p should be true. */
26403 else if (TARGET_64BIT && cfun->machine->call_abi == MS_ABI)
26404 gcc_unreachable ();
26407 tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, xops[0]), UNSPEC_GOTPCREL);
26408 tmp = gen_rtx_CONST (Pmode, tmp);
26409 tmp = gen_rtx_MEM (QImode, tmp);
26411 output_asm_insn ("jmp\t%A0", xops);
26416 if (!flag_pic || (*targetm.binds_local_p) (function))
26417 output_asm_insn ("jmp\t%P0", xops);
26422 rtx sym_ref = XEXP (DECL_RTL (function), 0);
26423 tmp = (gen_rtx_SYMBOL_REF
26425 machopic_indirection_name (sym_ref, /*stub_p=*/true)));
26426 tmp = gen_rtx_MEM (QImode, tmp);
26428 output_asm_insn ("jmp\t%0", xops);
26431 #endif /* TARGET_MACHO */
26433 tmp = gen_rtx_REG (SImode, CX_REG);
26434 output_set_got (tmp, NULL_RTX);
26437 output_asm_insn ("mov{l}\t{%0@GOT(%1), %1|%1, %0@GOT[%1]}", xops);
26438 output_asm_insn ("jmp\t{*}%1", xops);
26444 x86_file_start (void)
26446 default_file_start ();
26448 darwin_file_start ();
26450 if (X86_FILE_START_VERSION_DIRECTIVE)
26451 fputs ("\t.version\t\"01.01\"\n", asm_out_file);
26452 if (X86_FILE_START_FLTUSED)
26453 fputs ("\t.global\t__fltused\n", asm_out_file);
26454 if (ix86_asm_dialect == ASM_INTEL)
26455 fputs ("\t.intel_syntax noprefix\n", asm_out_file);
26459 x86_field_alignment (tree field, int computed)
26461 enum machine_mode mode;
26462 tree type = TREE_TYPE (field);
26464 if (TARGET_64BIT || TARGET_ALIGN_DOUBLE)
26466 mode = TYPE_MODE (strip_array_types (type));
26467 if (mode == DFmode || mode == DCmode
26468 || GET_MODE_CLASS (mode) == MODE_INT
26469 || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT)
26470 return MIN (32, computed);
26474 /* Output assembler code to FILE to increment profiler label # LABELNO
26475 for profiling a function entry. */
26477 x86_function_profiler (FILE *file, int labelno ATTRIBUTE_UNUSED)
26481 #ifndef NO_PROFILE_COUNTERS
26482 fprintf (file, "\tleaq\t%sP%d@(%%rip),%%r11\n", LPREFIX, labelno);
26485 if (DEFAULT_ABI == SYSV_ABI && flag_pic)
26486 fprintf (file, "\tcall\t*%s@GOTPCREL(%%rip)\n", MCOUNT_NAME);
26488 fprintf (file, "\tcall\t%s\n", MCOUNT_NAME);
26492 #ifndef NO_PROFILE_COUNTERS
26493 fprintf (file, "\tleal\t%sP%d@GOTOFF(%%ebx),%%%s\n",
26494 LPREFIX, labelno, PROFILE_COUNT_REGISTER);
26496 fprintf (file, "\tcall\t*%s@GOT(%%ebx)\n", MCOUNT_NAME);
26500 #ifndef NO_PROFILE_COUNTERS
26501 fprintf (file, "\tmovl\t$%sP%d,%%%s\n", LPREFIX, labelno,
26502 PROFILE_COUNT_REGISTER);
26504 fprintf (file, "\tcall\t%s\n", MCOUNT_NAME);
26508 /* We don't have exact information about the insn sizes, but we may assume
26509 quite safely that we are informed about all 1 byte insns and memory
26510 address sizes. This is enough to eliminate unnecessary padding in
26514 min_insn_size (rtx insn)
26518 if (!INSN_P (insn) || !active_insn_p (insn))
26521 /* Discard alignments we've emit and jump instructions. */
26522 if (GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE
26523 && XINT (PATTERN (insn), 1) == UNSPECV_ALIGN)
26526 && (GET_CODE (PATTERN (insn)) == ADDR_VEC
26527 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC))
26530 /* Important case - calls are always 5 bytes.
26531 It is common to have many calls in the row. */
26533 && symbolic_reference_mentioned_p (PATTERN (insn))
26534 && !SIBLING_CALL_P (insn))
26536 if (get_attr_length (insn) <= 1)
26539 /* For normal instructions we may rely on the sizes of addresses
26540 and the presence of symbol to require 4 bytes of encoding.
26541 This is not the case for jumps where references are PC relative. */
26542 if (!JUMP_P (insn))
26544 l = get_attr_length_address (insn);
26545 if (l < 4 && symbolic_reference_mentioned_p (PATTERN (insn)))
26554 /* AMD K8 core mispredicts jumps when there are more than 3 jumps in 16 byte
26558 ix86_avoid_jump_misspredicts (void)
26560 rtx insn, start = get_insns ();
26561 int nbytes = 0, njumps = 0;
26564 /* Look for all minimal intervals of instructions containing 4 jumps.
26565 The intervals are bounded by START and INSN. NBYTES is the total
26566 size of instructions in the interval including INSN and not including
26567 START. When the NBYTES is smaller than 16 bytes, it is possible
26568 that the end of START and INSN ends up in the same 16byte page.
26570 The smallest offset in the page INSN can start is the case where START
26571 ends on the offset 0. Offset of INSN is then NBYTES - sizeof (INSN).
26572 We add p2align to 16byte window with maxskip 17 - NBYTES + sizeof (INSN).
26574 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
26577 nbytes += min_insn_size (insn);
26579 fprintf(dump_file, "Insn %i estimated to %i bytes\n",
26580 INSN_UID (insn), min_insn_size (insn));
26582 && GET_CODE (PATTERN (insn)) != ADDR_VEC
26583 && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)
26591 start = NEXT_INSN (start);
26592 if ((JUMP_P (start)
26593 && GET_CODE (PATTERN (start)) != ADDR_VEC
26594 && GET_CODE (PATTERN (start)) != ADDR_DIFF_VEC)
26596 njumps--, isjump = 1;
26599 nbytes -= min_insn_size (start);
26601 gcc_assert (njumps >= 0);
26603 fprintf (dump_file, "Interval %i to %i has %i bytes\n",
26604 INSN_UID (start), INSN_UID (insn), nbytes);
26606 if (njumps == 3 && isjump && nbytes < 16)
26608 int padsize = 15 - nbytes + min_insn_size (insn);
26611 fprintf (dump_file, "Padding insn %i by %i bytes!\n",
26612 INSN_UID (insn), padsize);
26613 emit_insn_before (gen_align (GEN_INT (padsize)), insn);
26618 /* AMD Athlon works faster
26619 when RET is not destination of conditional jump or directly preceded
26620 by other jump instruction. We avoid the penalty by inserting NOP just
26621 before the RET instructions in such cases. */
26623 ix86_pad_returns (void)
26628 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
26630 basic_block bb = e->src;
26631 rtx ret = BB_END (bb);
26633 bool replace = false;
26635 if (!JUMP_P (ret) || GET_CODE (PATTERN (ret)) != RETURN
26636 || optimize_bb_for_size_p (bb))
26638 for (prev = PREV_INSN (ret); prev; prev = PREV_INSN (prev))
26639 if (active_insn_p (prev) || LABEL_P (prev))
26641 if (prev && LABEL_P (prev))
26646 FOR_EACH_EDGE (e, ei, bb->preds)
26647 if (EDGE_FREQUENCY (e) && e->src->index >= 0
26648 && !(e->flags & EDGE_FALLTHRU))
26653 prev = prev_active_insn (ret);
26655 && ((JUMP_P (prev) && any_condjump_p (prev))
26658 /* Empty functions get branch mispredict even when the jump destination
26659 is not visible to us. */
26660 if (!prev && cfun->function_frequency > FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
26665 emit_insn_before (gen_return_internal_long (), ret);
26671 /* Implement machine specific optimizations. We implement padding of returns
26672 for K8 CPUs and pass to avoid 4 jumps in the single 16 byte window. */
26676 if (TARGET_PAD_RETURNS && optimize
26677 && optimize_function_for_speed_p (cfun))
26678 ix86_pad_returns ();
26679 if (TARGET_FOUR_JUMP_LIMIT && optimize
26680 && optimize_function_for_speed_p (cfun))
26681 ix86_avoid_jump_misspredicts ();
26684 /* Return nonzero when QImode register that must be represented via REX prefix
26687 x86_extended_QIreg_mentioned_p (rtx insn)
26690 extract_insn_cached (insn);
26691 for (i = 0; i < recog_data.n_operands; i++)
26692 if (REG_P (recog_data.operand[i])
26693 && REGNO (recog_data.operand[i]) >= 4)
26698 /* Return nonzero when P points to register encoded via REX prefix.
26699 Called via for_each_rtx. */
26701 extended_reg_mentioned_1 (rtx *p, void *data ATTRIBUTE_UNUSED)
26703 unsigned int regno;
26706 regno = REGNO (*p);
26707 return REX_INT_REGNO_P (regno) || REX_SSE_REGNO_P (regno);
26710 /* Return true when INSN mentions register that must be encoded using REX
26713 x86_extended_reg_mentioned_p (rtx insn)
26715 return for_each_rtx (INSN_P (insn) ? &PATTERN (insn) : &insn,
26716 extended_reg_mentioned_1, NULL);
26719 /* Generate an unsigned DImode/SImode to FP conversion. This is the same code
26720 optabs would emit if we didn't have TFmode patterns. */
26723 x86_emit_floatuns (rtx operands[2])
26725 rtx neglab, donelab, i0, i1, f0, in, out;
26726 enum machine_mode mode, inmode;
26728 inmode = GET_MODE (operands[1]);
26729 gcc_assert (inmode == SImode || inmode == DImode);
26732 in = force_reg (inmode, operands[1]);
26733 mode = GET_MODE (out);
26734 neglab = gen_label_rtx ();
26735 donelab = gen_label_rtx ();
26736 f0 = gen_reg_rtx (mode);
26738 emit_cmp_and_jump_insns (in, const0_rtx, LT, const0_rtx, inmode, 0, neglab);
26740 expand_float (out, in, 0);
26742 emit_jump_insn (gen_jump (donelab));
26745 emit_label (neglab);
26747 i0 = expand_simple_binop (inmode, LSHIFTRT, in, const1_rtx, NULL,
26749 i1 = expand_simple_binop (inmode, AND, in, const1_rtx, NULL,
26751 i0 = expand_simple_binop (inmode, IOR, i0, i1, i0, 1, OPTAB_DIRECT);
26753 expand_float (f0, i0, 0);
26755 emit_insn (gen_rtx_SET (VOIDmode, out, gen_rtx_PLUS (mode, f0, f0)));
26757 emit_label (donelab);
26760 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
26761 with all elements equal to VAR. Return true if successful. */
26764 ix86_expand_vector_init_duplicate (bool mmx_ok, enum machine_mode mode,
26765 rtx target, rtx val)
26767 enum machine_mode hmode, smode, wsmode, wvmode;
26782 val = force_reg (GET_MODE_INNER (mode), val);
26783 x = gen_rtx_VEC_DUPLICATE (mode, val);
26784 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26790 if (TARGET_SSE || TARGET_3DNOW_A)
26792 val = gen_lowpart (SImode, val);
26793 x = gen_rtx_TRUNCATE (HImode, val);
26794 x = gen_rtx_VEC_DUPLICATE (mode, x);
26795 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26817 /* Extend HImode to SImode using a paradoxical SUBREG. */
26818 tmp1 = gen_reg_rtx (SImode);
26819 emit_move_insn (tmp1, gen_lowpart (SImode, val));
26820 /* Insert the SImode value as low element of V4SImode vector. */
26821 tmp2 = gen_reg_rtx (V4SImode);
26822 tmp1 = gen_rtx_VEC_MERGE (V4SImode,
26823 gen_rtx_VEC_DUPLICATE (V4SImode, tmp1),
26824 CONST0_RTX (V4SImode),
26826 emit_insn (gen_rtx_SET (VOIDmode, tmp2, tmp1));
26827 /* Cast the V4SImode vector back to a V8HImode vector. */
26828 tmp1 = gen_reg_rtx (V8HImode);
26829 emit_move_insn (tmp1, gen_lowpart (V8HImode, tmp2));
26830 /* Duplicate the low short through the whole low SImode word. */
26831 emit_insn (gen_sse2_punpcklwd (tmp1, tmp1, tmp1));
26832 /* Cast the V8HImode vector back to a V4SImode vector. */
26833 tmp2 = gen_reg_rtx (V4SImode);
26834 emit_move_insn (tmp2, gen_lowpart (V4SImode, tmp1));
26835 /* Replicate the low element of the V4SImode vector. */
26836 emit_insn (gen_sse2_pshufd (tmp2, tmp2, const0_rtx));
26837 /* Cast the V2SImode back to V8HImode, and store in target. */
26838 emit_move_insn (target, gen_lowpart (V8HImode, tmp2));
26849 /* Extend QImode to SImode using a paradoxical SUBREG. */
26850 tmp1 = gen_reg_rtx (SImode);
26851 emit_move_insn (tmp1, gen_lowpart (SImode, val));
26852 /* Insert the SImode value as low element of V4SImode vector. */
26853 tmp2 = gen_reg_rtx (V4SImode);
26854 tmp1 = gen_rtx_VEC_MERGE (V4SImode,
26855 gen_rtx_VEC_DUPLICATE (V4SImode, tmp1),
26856 CONST0_RTX (V4SImode),
26858 emit_insn (gen_rtx_SET (VOIDmode, tmp2, tmp1));
26859 /* Cast the V4SImode vector back to a V16QImode vector. */
26860 tmp1 = gen_reg_rtx (V16QImode);
26861 emit_move_insn (tmp1, gen_lowpart (V16QImode, tmp2));
26862 /* Duplicate the low byte through the whole low SImode word. */
26863 emit_insn (gen_sse2_punpcklbw (tmp1, tmp1, tmp1));
26864 emit_insn (gen_sse2_punpcklbw (tmp1, tmp1, tmp1));
26865 /* Cast the V16QImode vector back to a V4SImode vector. */
26866 tmp2 = gen_reg_rtx (V4SImode);
26867 emit_move_insn (tmp2, gen_lowpart (V4SImode, tmp1));
26868 /* Replicate the low element of the V4SImode vector. */
26869 emit_insn (gen_sse2_pshufd (tmp2, tmp2, const0_rtx));
26870 /* Cast the V2SImode back to V16QImode, and store in target. */
26871 emit_move_insn (target, gen_lowpart (V16QImode, tmp2));
26879 /* Replicate the value once into the next wider mode and recurse. */
26880 val = convert_modes (wsmode, smode, val, true);
26881 x = expand_simple_binop (wsmode, ASHIFT, val,
26882 GEN_INT (GET_MODE_BITSIZE (smode)),
26883 NULL_RTX, 1, OPTAB_LIB_WIDEN);
26884 val = expand_simple_binop (wsmode, IOR, val, x, x, 1, OPTAB_LIB_WIDEN);
26886 x = gen_reg_rtx (wvmode);
26887 if (!ix86_expand_vector_init_duplicate (mmx_ok, wvmode, x, val))
26888 gcc_unreachable ();
26889 emit_move_insn (target, gen_lowpart (mode, x));
26912 rtx tmp = gen_reg_rtx (hmode);
26913 ix86_expand_vector_init_duplicate (mmx_ok, hmode, tmp, val);
26914 emit_insn (gen_rtx_SET (VOIDmode, target,
26915 gen_rtx_VEC_CONCAT (mode, tmp, tmp)));
26924 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
26925 whose ONE_VAR element is VAR, and other elements are zero. Return true
26929 ix86_expand_vector_init_one_nonzero (bool mmx_ok, enum machine_mode mode,
26930 rtx target, rtx var, int one_var)
26932 enum machine_mode vsimode;
26935 bool use_vector_set = false;
26940 /* For SSE4.1, we normally use vector set. But if the second
26941 element is zero and inter-unit moves are OK, we use movq
26943 use_vector_set = (TARGET_64BIT
26945 && !(TARGET_INTER_UNIT_MOVES
26951 use_vector_set = TARGET_SSE4_1;
26954 use_vector_set = TARGET_SSE2;
26957 use_vector_set = TARGET_SSE || TARGET_3DNOW_A;
26965 use_vector_set = TARGET_AVX;
26971 if (use_vector_set)
26973 emit_insn (gen_rtx_SET (VOIDmode, target, CONST0_RTX (mode)));
26974 var = force_reg (GET_MODE_INNER (mode), var);
26975 ix86_expand_vector_set (mmx_ok, target, var, one_var);
26991 var = force_reg (GET_MODE_INNER (mode), var);
26992 x = gen_rtx_VEC_CONCAT (mode, var, CONST0_RTX (GET_MODE_INNER (mode)));
26993 emit_insn (gen_rtx_SET (VOIDmode, target, x));
26998 if (!REG_P (target) || REGNO (target) < FIRST_PSEUDO_REGISTER)
26999 new_target = gen_reg_rtx (mode);
27001 new_target = target;
27002 var = force_reg (GET_MODE_INNER (mode), var);
27003 x = gen_rtx_VEC_DUPLICATE (mode, var);
27004 x = gen_rtx_VEC_MERGE (mode, x, CONST0_RTX (mode), const1_rtx);
27005 emit_insn (gen_rtx_SET (VOIDmode, new_target, x));
27008 /* We need to shuffle the value to the correct position, so
27009 create a new pseudo to store the intermediate result. */
27011 /* With SSE2, we can use the integer shuffle insns. */
27012 if (mode != V4SFmode && TARGET_SSE2)
27014 emit_insn (gen_sse2_pshufd_1 (new_target, new_target,
27016 GEN_INT (one_var == 1 ? 0 : 1),
27017 GEN_INT (one_var == 2 ? 0 : 1),
27018 GEN_INT (one_var == 3 ? 0 : 1)));
27019 if (target != new_target)
27020 emit_move_insn (target, new_target);
27024 /* Otherwise convert the intermediate result to V4SFmode and
27025 use the SSE1 shuffle instructions. */
27026 if (mode != V4SFmode)
27028 tmp = gen_reg_rtx (V4SFmode);
27029 emit_move_insn (tmp, gen_lowpart (V4SFmode, new_target));
27034 emit_insn (gen_sse_shufps_v4sf (tmp, tmp, tmp,
27036 GEN_INT (one_var == 1 ? 0 : 1),
27037 GEN_INT (one_var == 2 ? 0+4 : 1+4),
27038 GEN_INT (one_var == 3 ? 0+4 : 1+4)));
27040 if (mode != V4SFmode)
27041 emit_move_insn (target, gen_lowpart (V4SImode, tmp));
27042 else if (tmp != target)
27043 emit_move_insn (target, tmp);
27045 else if (target != new_target)
27046 emit_move_insn (target, new_target);
27051 vsimode = V4SImode;
27057 vsimode = V2SImode;
27063 /* Zero extend the variable element to SImode and recurse. */
27064 var = convert_modes (SImode, GET_MODE_INNER (mode), var, true);
27066 x = gen_reg_rtx (vsimode);
27067 if (!ix86_expand_vector_init_one_nonzero (mmx_ok, vsimode, x,
27069 gcc_unreachable ();
27071 emit_move_insn (target, gen_lowpart (mode, x));
27079 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
27080 consisting of the values in VALS. It is known that all elements
27081 except ONE_VAR are constants. Return true if successful. */
27084 ix86_expand_vector_init_one_var (bool mmx_ok, enum machine_mode mode,
27085 rtx target, rtx vals, int one_var)
27087 rtx var = XVECEXP (vals, 0, one_var);
27088 enum machine_mode wmode;
27091 const_vec = copy_rtx (vals);
27092 XVECEXP (const_vec, 0, one_var) = CONST0_RTX (GET_MODE_INNER (mode));
27093 const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (const_vec, 0));
27101 /* For the two element vectors, it's just as easy to use
27102 the general case. */
27126 /* There's no way to set one QImode entry easily. Combine
27127 the variable value with its adjacent constant value, and
27128 promote to an HImode set. */
27129 x = XVECEXP (vals, 0, one_var ^ 1);
27132 var = convert_modes (HImode, QImode, var, true);
27133 var = expand_simple_binop (HImode, ASHIFT, var, GEN_INT (8),
27134 NULL_RTX, 1, OPTAB_LIB_WIDEN);
27135 x = GEN_INT (INTVAL (x) & 0xff);
27139 var = convert_modes (HImode, QImode, var, true);
27140 x = gen_int_mode (INTVAL (x) << 8, HImode);
27142 if (x != const0_rtx)
27143 var = expand_simple_binop (HImode, IOR, var, x, var,
27144 1, OPTAB_LIB_WIDEN);
27146 x = gen_reg_rtx (wmode);
27147 emit_move_insn (x, gen_lowpart (wmode, const_vec));
27148 ix86_expand_vector_set (mmx_ok, x, var, one_var >> 1);
27150 emit_move_insn (target, gen_lowpart (mode, x));
27157 emit_move_insn (target, const_vec);
27158 ix86_expand_vector_set (mmx_ok, target, var, one_var);
27162 /* A subroutine of ix86_expand_vector_init_general. Use vector
27163 concatenate to handle the most general case: all values variable,
27164 and none identical. */
27167 ix86_expand_vector_init_concat (enum machine_mode mode,
27168 rtx target, rtx *ops, int n)
27170 enum machine_mode cmode, hmode = VOIDmode;
27171 rtx first[8], second[4];
27211 gcc_unreachable ();
27214 if (!register_operand (ops[1], cmode))
27215 ops[1] = force_reg (cmode, ops[1]);
27216 if (!register_operand (ops[0], cmode))
27217 ops[0] = force_reg (cmode, ops[0]);
27218 emit_insn (gen_rtx_SET (VOIDmode, target,
27219 gen_rtx_VEC_CONCAT (mode, ops[0],
27239 gcc_unreachable ();
27255 gcc_unreachable ();
27260 /* FIXME: We process inputs backward to help RA. PR 36222. */
27263 for (; i > 0; i -= 2, j--)
27265 first[j] = gen_reg_rtx (cmode);
27266 v = gen_rtvec (2, ops[i - 1], ops[i]);
27267 ix86_expand_vector_init (false, first[j],
27268 gen_rtx_PARALLEL (cmode, v));
27274 gcc_assert (hmode != VOIDmode);
27275 for (i = j = 0; i < n; i += 2, j++)
27277 second[j] = gen_reg_rtx (hmode);
27278 ix86_expand_vector_init_concat (hmode, second [j],
27282 ix86_expand_vector_init_concat (mode, target, second, n);
27285 ix86_expand_vector_init_concat (mode, target, first, n);
27289 gcc_unreachable ();
27293 /* A subroutine of ix86_expand_vector_init_general. Use vector
27294 interleave to handle the most general case: all values variable,
27295 and none identical. */
27298 ix86_expand_vector_init_interleave (enum machine_mode mode,
27299 rtx target, rtx *ops, int n)
27301 enum machine_mode first_imode, second_imode, third_imode, inner_mode;
27304 rtx (*gen_load_even) (rtx, rtx, rtx);
27305 rtx (*gen_interleave_first_low) (rtx, rtx, rtx);
27306 rtx (*gen_interleave_second_low) (rtx, rtx, rtx);
27311 gen_load_even = gen_vec_setv8hi;
27312 gen_interleave_first_low = gen_vec_interleave_lowv4si;
27313 gen_interleave_second_low = gen_vec_interleave_lowv2di;
27314 inner_mode = HImode;
27315 first_imode = V4SImode;
27316 second_imode = V2DImode;
27317 third_imode = VOIDmode;
27320 gen_load_even = gen_vec_setv16qi;
27321 gen_interleave_first_low = gen_vec_interleave_lowv8hi;
27322 gen_interleave_second_low = gen_vec_interleave_lowv4si;
27323 inner_mode = QImode;
27324 first_imode = V8HImode;
27325 second_imode = V4SImode;
27326 third_imode = V2DImode;
27329 gcc_unreachable ();
27332 for (i = 0; i < n; i++)
27334 /* Extend the odd elment to SImode using a paradoxical SUBREG. */
27335 op0 = gen_reg_rtx (SImode);
27336 emit_move_insn (op0, gen_lowpart (SImode, ops [i + i]));
27338 /* Insert the SImode value as low element of V4SImode vector. */
27339 op1 = gen_reg_rtx (V4SImode);
27340 op0 = gen_rtx_VEC_MERGE (V4SImode,
27341 gen_rtx_VEC_DUPLICATE (V4SImode,
27343 CONST0_RTX (V4SImode),
27345 emit_insn (gen_rtx_SET (VOIDmode, op1, op0));
27347 /* Cast the V4SImode vector back to a vector in orignal mode. */
27348 op0 = gen_reg_rtx (mode);
27349 emit_move_insn (op0, gen_lowpart (mode, op1));
27351 /* Load even elements into the second positon. */
27352 emit_insn ((*gen_load_even) (op0,
27353 force_reg (inner_mode,
27357 /* Cast vector to FIRST_IMODE vector. */
27358 ops[i] = gen_reg_rtx (first_imode);
27359 emit_move_insn (ops[i], gen_lowpart (first_imode, op0));
27362 /* Interleave low FIRST_IMODE vectors. */
27363 for (i = j = 0; i < n; i += 2, j++)
27365 op0 = gen_reg_rtx (first_imode);
27366 emit_insn ((*gen_interleave_first_low) (op0, ops[i], ops[i + 1]));
27368 /* Cast FIRST_IMODE vector to SECOND_IMODE vector. */
27369 ops[j] = gen_reg_rtx (second_imode);
27370 emit_move_insn (ops[j], gen_lowpart (second_imode, op0));
27373 /* Interleave low SECOND_IMODE vectors. */
27374 switch (second_imode)
27377 for (i = j = 0; i < n / 2; i += 2, j++)
27379 op0 = gen_reg_rtx (second_imode);
27380 emit_insn ((*gen_interleave_second_low) (op0, ops[i],
27383 /* Cast the SECOND_IMODE vector to the THIRD_IMODE
27385 ops[j] = gen_reg_rtx (third_imode);
27386 emit_move_insn (ops[j], gen_lowpart (third_imode, op0));
27388 second_imode = V2DImode;
27389 gen_interleave_second_low = gen_vec_interleave_lowv2di;
27393 op0 = gen_reg_rtx (second_imode);
27394 emit_insn ((*gen_interleave_second_low) (op0, ops[0],
27397 /* Cast the SECOND_IMODE vector back to a vector on original
27399 emit_insn (gen_rtx_SET (VOIDmode, target,
27400 gen_lowpart (mode, op0)));
27404 gcc_unreachable ();
27408 /* A subroutine of ix86_expand_vector_init. Handle the most general case:
27409 all values variable, and none identical. */
27412 ix86_expand_vector_init_general (bool mmx_ok, enum machine_mode mode,
27413 rtx target, rtx vals)
27415 rtx ops[32], op0, op1;
27416 enum machine_mode half_mode = VOIDmode;
27423 if (!mmx_ok && !TARGET_SSE)
27435 n = GET_MODE_NUNITS (mode);
27436 for (i = 0; i < n; i++)
27437 ops[i] = XVECEXP (vals, 0, i);
27438 ix86_expand_vector_init_concat (mode, target, ops, n);
27442 half_mode = V16QImode;
27446 half_mode = V8HImode;
27450 n = GET_MODE_NUNITS (mode);
27451 for (i = 0; i < n; i++)
27452 ops[i] = XVECEXP (vals, 0, i);
27453 op0 = gen_reg_rtx (half_mode);
27454 op1 = gen_reg_rtx (half_mode);
27455 ix86_expand_vector_init_interleave (half_mode, op0, ops,
27457 ix86_expand_vector_init_interleave (half_mode, op1,
27458 &ops [n >> 1], n >> 2);
27459 emit_insn (gen_rtx_SET (VOIDmode, target,
27460 gen_rtx_VEC_CONCAT (mode, op0, op1)));
27464 if (!TARGET_SSE4_1)
27472 /* Don't use ix86_expand_vector_init_interleave if we can't
27473 move from GPR to SSE register directly. */
27474 if (!TARGET_INTER_UNIT_MOVES)
27477 n = GET_MODE_NUNITS (mode);
27478 for (i = 0; i < n; i++)
27479 ops[i] = XVECEXP (vals, 0, i);
27480 ix86_expand_vector_init_interleave (mode, target, ops, n >> 1);
27488 gcc_unreachable ();
27492 int i, j, n_elts, n_words, n_elt_per_word;
27493 enum machine_mode inner_mode;
27494 rtx words[4], shift;
27496 inner_mode = GET_MODE_INNER (mode);
27497 n_elts = GET_MODE_NUNITS (mode);
27498 n_words = GET_MODE_SIZE (mode) / UNITS_PER_WORD;
27499 n_elt_per_word = n_elts / n_words;
27500 shift = GEN_INT (GET_MODE_BITSIZE (inner_mode));
27502 for (i = 0; i < n_words; ++i)
27504 rtx word = NULL_RTX;
27506 for (j = 0; j < n_elt_per_word; ++j)
27508 rtx elt = XVECEXP (vals, 0, (i+1)*n_elt_per_word - j - 1);
27509 elt = convert_modes (word_mode, inner_mode, elt, true);
27515 word = expand_simple_binop (word_mode, ASHIFT, word, shift,
27516 word, 1, OPTAB_LIB_WIDEN);
27517 word = expand_simple_binop (word_mode, IOR, word, elt,
27518 word, 1, OPTAB_LIB_WIDEN);
27526 emit_move_insn (target, gen_lowpart (mode, words[0]));
27527 else if (n_words == 2)
27529 rtx tmp = gen_reg_rtx (mode);
27530 emit_clobber (tmp);
27531 emit_move_insn (gen_lowpart (word_mode, tmp), words[0]);
27532 emit_move_insn (gen_highpart (word_mode, tmp), words[1]);
27533 emit_move_insn (target, tmp);
27535 else if (n_words == 4)
27537 rtx tmp = gen_reg_rtx (V4SImode);
27538 gcc_assert (word_mode == SImode);
27539 vals = gen_rtx_PARALLEL (V4SImode, gen_rtvec_v (4, words));
27540 ix86_expand_vector_init_general (false, V4SImode, tmp, vals);
27541 emit_move_insn (target, gen_lowpart (mode, tmp));
27544 gcc_unreachable ();
27548 /* Initialize vector TARGET via VALS. Suppress the use of MMX
27549 instructions unless MMX_OK is true. */
27552 ix86_expand_vector_init (bool mmx_ok, rtx target, rtx vals)
27554 enum machine_mode mode = GET_MODE (target);
27555 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27556 int n_elts = GET_MODE_NUNITS (mode);
27557 int n_var = 0, one_var = -1;
27558 bool all_same = true, all_const_zero = true;
27562 for (i = 0; i < n_elts; ++i)
27564 x = XVECEXP (vals, 0, i);
27565 if (!(CONST_INT_P (x)
27566 || GET_CODE (x) == CONST_DOUBLE
27567 || GET_CODE (x) == CONST_FIXED))
27568 n_var++, one_var = i;
27569 else if (x != CONST0_RTX (inner_mode))
27570 all_const_zero = false;
27571 if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
27575 /* Constants are best loaded from the constant pool. */
27578 emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)));
27582 /* If all values are identical, broadcast the value. */
27584 && ix86_expand_vector_init_duplicate (mmx_ok, mode, target,
27585 XVECEXP (vals, 0, 0)))
27588 /* Values where only one field is non-constant are best loaded from
27589 the pool and overwritten via move later. */
27593 && ix86_expand_vector_init_one_nonzero (mmx_ok, mode, target,
27594 XVECEXP (vals, 0, one_var),
27598 if (ix86_expand_vector_init_one_var (mmx_ok, mode, target, vals, one_var))
27602 ix86_expand_vector_init_general (mmx_ok, mode, target, vals);
27606 ix86_expand_vector_set (bool mmx_ok, rtx target, rtx val, int elt)
27608 enum machine_mode mode = GET_MODE (target);
27609 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27610 enum machine_mode half_mode;
27611 bool use_vec_merge = false;
27613 static rtx (*gen_extract[6][2]) (rtx, rtx)
27615 { gen_vec_extract_lo_v32qi, gen_vec_extract_hi_v32qi },
27616 { gen_vec_extract_lo_v16hi, gen_vec_extract_hi_v16hi },
27617 { gen_vec_extract_lo_v8si, gen_vec_extract_hi_v8si },
27618 { gen_vec_extract_lo_v4di, gen_vec_extract_hi_v4di },
27619 { gen_vec_extract_lo_v8sf, gen_vec_extract_hi_v8sf },
27620 { gen_vec_extract_lo_v4df, gen_vec_extract_hi_v4df }
27622 static rtx (*gen_insert[6][2]) (rtx, rtx, rtx)
27624 { gen_vec_set_lo_v32qi, gen_vec_set_hi_v32qi },
27625 { gen_vec_set_lo_v16hi, gen_vec_set_hi_v16hi },
27626 { gen_vec_set_lo_v8si, gen_vec_set_hi_v8si },
27627 { gen_vec_set_lo_v4di, gen_vec_set_hi_v4di },
27628 { gen_vec_set_lo_v8sf, gen_vec_set_hi_v8sf },
27629 { gen_vec_set_lo_v4df, gen_vec_set_hi_v4df }
27639 tmp = gen_reg_rtx (GET_MODE_INNER (mode));
27640 ix86_expand_vector_extract (true, tmp, target, 1 - elt);
27642 tmp = gen_rtx_VEC_CONCAT (mode, tmp, val);
27644 tmp = gen_rtx_VEC_CONCAT (mode, val, tmp);
27645 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27651 use_vec_merge = TARGET_SSE4_1;
27659 /* For the two element vectors, we implement a VEC_CONCAT with
27660 the extraction of the other element. */
27662 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (1 - elt)));
27663 tmp = gen_rtx_VEC_SELECT (inner_mode, target, tmp);
27666 op0 = val, op1 = tmp;
27668 op0 = tmp, op1 = val;
27670 tmp = gen_rtx_VEC_CONCAT (mode, op0, op1);
27671 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27676 use_vec_merge = TARGET_SSE4_1;
27683 use_vec_merge = true;
27687 /* tmp = target = A B C D */
27688 tmp = copy_to_reg (target);
27689 /* target = A A B B */
27690 emit_insn (gen_sse_unpcklps (target, target, target));
27691 /* target = X A B B */
27692 ix86_expand_vector_set (false, target, val, 0);
27693 /* target = A X C D */
27694 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27695 GEN_INT (1), GEN_INT (0),
27696 GEN_INT (2+4), GEN_INT (3+4)));
27700 /* tmp = target = A B C D */
27701 tmp = copy_to_reg (target);
27702 /* tmp = X B C D */
27703 ix86_expand_vector_set (false, tmp, val, 0);
27704 /* target = A B X D */
27705 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27706 GEN_INT (0), GEN_INT (1),
27707 GEN_INT (0+4), GEN_INT (3+4)));
27711 /* tmp = target = A B C D */
27712 tmp = copy_to_reg (target);
27713 /* tmp = X B C D */
27714 ix86_expand_vector_set (false, tmp, val, 0);
27715 /* target = A B X D */
27716 emit_insn (gen_sse_shufps_v4sf (target, target, tmp,
27717 GEN_INT (0), GEN_INT (1),
27718 GEN_INT (2+4), GEN_INT (0+4)));
27722 gcc_unreachable ();
27727 use_vec_merge = TARGET_SSE4_1;
27731 /* Element 0 handled by vec_merge below. */
27734 use_vec_merge = true;
27740 /* With SSE2, use integer shuffles to swap element 0 and ELT,
27741 store into element 0, then shuffle them back. */
27745 order[0] = GEN_INT (elt);
27746 order[1] = const1_rtx;
27747 order[2] = const2_rtx;
27748 order[3] = GEN_INT (3);
27749 order[elt] = const0_rtx;
27751 emit_insn (gen_sse2_pshufd_1 (target, target, order[0],
27752 order[1], order[2], order[3]));
27754 ix86_expand_vector_set (false, target, val, 0);
27756 emit_insn (gen_sse2_pshufd_1 (target, target, order[0],
27757 order[1], order[2], order[3]));
27761 /* For SSE1, we have to reuse the V4SF code. */
27762 ix86_expand_vector_set (false, gen_lowpart (V4SFmode, target),
27763 gen_lowpart (SFmode, val), elt);
27768 use_vec_merge = TARGET_SSE2;
27771 use_vec_merge = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A);
27775 use_vec_merge = TARGET_SSE4_1;
27782 half_mode = V16QImode;
27788 half_mode = V8HImode;
27794 half_mode = V4SImode;
27800 half_mode = V2DImode;
27806 half_mode = V4SFmode;
27812 half_mode = V2DFmode;
27818 /* Compute offset. */
27822 gcc_assert (i <= 1);
27824 /* Extract the half. */
27825 tmp = gen_reg_rtx (half_mode);
27826 emit_insn ((*gen_extract[j][i]) (tmp, target));
27828 /* Put val in tmp at elt. */
27829 ix86_expand_vector_set (false, tmp, val, elt);
27832 emit_insn ((*gen_insert[j][i]) (target, target, tmp));
27841 tmp = gen_rtx_VEC_DUPLICATE (mode, val);
27842 tmp = gen_rtx_VEC_MERGE (mode, tmp, target, GEN_INT (1 << elt));
27843 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27847 rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), false);
27849 emit_move_insn (mem, target);
27851 tmp = adjust_address (mem, inner_mode, elt*GET_MODE_SIZE (inner_mode));
27852 emit_move_insn (tmp, val);
27854 emit_move_insn (target, mem);
27859 ix86_expand_vector_extract (bool mmx_ok, rtx target, rtx vec, int elt)
27861 enum machine_mode mode = GET_MODE (vec);
27862 enum machine_mode inner_mode = GET_MODE_INNER (mode);
27863 bool use_vec_extr = false;
27876 use_vec_extr = true;
27880 use_vec_extr = TARGET_SSE4_1;
27892 tmp = gen_reg_rtx (mode);
27893 emit_insn (gen_sse_shufps_v4sf (tmp, vec, vec,
27894 GEN_INT (elt), GEN_INT (elt),
27895 GEN_INT (elt+4), GEN_INT (elt+4)));
27899 tmp = gen_reg_rtx (mode);
27900 emit_insn (gen_sse_unpckhps (tmp, vec, vec));
27904 gcc_unreachable ();
27907 use_vec_extr = true;
27912 use_vec_extr = TARGET_SSE4_1;
27926 tmp = gen_reg_rtx (mode);
27927 emit_insn (gen_sse2_pshufd_1 (tmp, vec,
27928 GEN_INT (elt), GEN_INT (elt),
27929 GEN_INT (elt), GEN_INT (elt)));
27933 tmp = gen_reg_rtx (mode);
27934 emit_insn (gen_sse2_punpckhdq (tmp, vec, vec));
27938 gcc_unreachable ();
27941 use_vec_extr = true;
27946 /* For SSE1, we have to reuse the V4SF code. */
27947 ix86_expand_vector_extract (false, gen_lowpart (SFmode, target),
27948 gen_lowpart (V4SFmode, vec), elt);
27954 use_vec_extr = TARGET_SSE2;
27957 use_vec_extr = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A);
27961 use_vec_extr = TARGET_SSE4_1;
27965 /* ??? Could extract the appropriate HImode element and shift. */
27972 tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (elt)));
27973 tmp = gen_rtx_VEC_SELECT (inner_mode, vec, tmp);
27975 /* Let the rtl optimizers know about the zero extension performed. */
27976 if (inner_mode == QImode || inner_mode == HImode)
27978 tmp = gen_rtx_ZERO_EXTEND (SImode, tmp);
27979 target = gen_lowpart (SImode, target);
27982 emit_insn (gen_rtx_SET (VOIDmode, target, tmp));
27986 rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), false);
27988 emit_move_insn (mem, vec);
27990 tmp = adjust_address (mem, inner_mode, elt*GET_MODE_SIZE (inner_mode));
27991 emit_move_insn (target, tmp);
27995 /* Expand a vector reduction on V4SFmode for SSE1. FN is the binary
27996 pattern to reduce; DEST is the destination; IN is the input vector. */
27999 ix86_expand_reduc_v4sf (rtx (*fn) (rtx, rtx, rtx), rtx dest, rtx in)
28001 rtx tmp1, tmp2, tmp3;
28003 tmp1 = gen_reg_rtx (V4SFmode);
28004 tmp2 = gen_reg_rtx (V4SFmode);
28005 tmp3 = gen_reg_rtx (V4SFmode);
28007 emit_insn (gen_sse_movhlps (tmp1, in, in));
28008 emit_insn (fn (tmp2, tmp1, in));
28010 emit_insn (gen_sse_shufps_v4sf (tmp3, tmp2, tmp2,
28011 GEN_INT (1), GEN_INT (1),
28012 GEN_INT (1+4), GEN_INT (1+4)));
28013 emit_insn (fn (dest, tmp2, tmp3));
28016 /* Target hook for scalar_mode_supported_p. */
28018 ix86_scalar_mode_supported_p (enum machine_mode mode)
28020 if (DECIMAL_FLOAT_MODE_P (mode))
28022 else if (mode == TFmode)
28025 return default_scalar_mode_supported_p (mode);
28028 /* Implements target hook vector_mode_supported_p. */
28030 ix86_vector_mode_supported_p (enum machine_mode mode)
28032 if (TARGET_SSE && VALID_SSE_REG_MODE (mode))
28034 if (TARGET_SSE2 && VALID_SSE2_REG_MODE (mode))
28036 if (TARGET_AVX && VALID_AVX256_REG_MODE (mode))
28038 if (TARGET_MMX && VALID_MMX_REG_MODE (mode))
28040 if (TARGET_3DNOW && VALID_MMX_REG_MODE_3DNOW (mode))
28045 /* Target hook for c_mode_for_suffix. */
28046 static enum machine_mode
28047 ix86_c_mode_for_suffix (char suffix)
28057 /* Worker function for TARGET_MD_ASM_CLOBBERS.
28059 We do this in the new i386 backend to maintain source compatibility
28060 with the old cc0-based compiler. */
28063 ix86_md_asm_clobbers (tree outputs ATTRIBUTE_UNUSED,
28064 tree inputs ATTRIBUTE_UNUSED,
28067 clobbers = tree_cons (NULL_TREE, build_string (5, "flags"),
28069 clobbers = tree_cons (NULL_TREE, build_string (4, "fpsr"),
28074 /* Implements target vector targetm.asm.encode_section_info. This
28075 is not used by netware. */
28077 static void ATTRIBUTE_UNUSED
28078 ix86_encode_section_info (tree decl, rtx rtl, int first)
28080 default_encode_section_info (decl, rtl, first);
28082 if (TREE_CODE (decl) == VAR_DECL
28083 && (TREE_STATIC (decl) || DECL_EXTERNAL (decl))
28084 && ix86_in_large_data_p (decl))
28085 SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FAR_ADDR;
28088 /* Worker function for REVERSE_CONDITION. */
28091 ix86_reverse_condition (enum rtx_code code, enum machine_mode mode)
28093 return (mode != CCFPmode && mode != CCFPUmode
28094 ? reverse_condition (code)
28095 : reverse_condition_maybe_unordered (code));
28098 /* Output code to perform an x87 FP register move, from OPERANDS[1]
28102 output_387_reg_move (rtx insn, rtx *operands)
28104 if (REG_P (operands[0]))
28106 if (REG_P (operands[1])
28107 && find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
28109 if (REGNO (operands[0]) == FIRST_STACK_REG)
28110 return output_387_ffreep (operands, 0);
28111 return "fstp\t%y0";
28113 if (STACK_TOP_P (operands[0]))
28114 return "fld%z1\t%y1";
28117 else if (MEM_P (operands[0]))
28119 gcc_assert (REG_P (operands[1]));
28120 if (find_regno_note (insn, REG_DEAD, REGNO (operands[1])))
28121 return "fstp%z0\t%y0";
28124 /* There is no non-popping store to memory for XFmode.
28125 So if we need one, follow the store with a load. */
28126 if (GET_MODE (operands[0]) == XFmode)
28127 return "fstp%z0\t%y0\n\tfld%z0\t%y0";
28129 return "fst%z0\t%y0";
28136 /* Output code to perform a conditional jump to LABEL, if C2 flag in
28137 FP status register is set. */
28140 ix86_emit_fp_unordered_jump (rtx label)
28142 rtx reg = gen_reg_rtx (HImode);
28145 emit_insn (gen_x86_fnstsw_1 (reg));
28147 if (TARGET_SAHF && (TARGET_USE_SAHF || optimize_insn_for_size_p ()))
28149 emit_insn (gen_x86_sahf_1 (reg));
28151 temp = gen_rtx_REG (CCmode, FLAGS_REG);
28152 temp = gen_rtx_UNORDERED (VOIDmode, temp, const0_rtx);
28156 emit_insn (gen_testqi_ext_ccno_0 (reg, GEN_INT (0x04)));
28158 temp = gen_rtx_REG (CCNOmode, FLAGS_REG);
28159 temp = gen_rtx_NE (VOIDmode, temp, const0_rtx);
28162 temp = gen_rtx_IF_THEN_ELSE (VOIDmode, temp,
28163 gen_rtx_LABEL_REF (VOIDmode, label),
28165 temp = gen_rtx_SET (VOIDmode, pc_rtx, temp);
28167 emit_jump_insn (temp);
28168 predict_jump (REG_BR_PROB_BASE * 10 / 100);
28171 /* Output code to perform a log1p XFmode calculation. */
28173 void ix86_emit_i387_log1p (rtx op0, rtx op1)
28175 rtx label1 = gen_label_rtx ();
28176 rtx label2 = gen_label_rtx ();
28178 rtx tmp = gen_reg_rtx (XFmode);
28179 rtx tmp2 = gen_reg_rtx (XFmode);
28181 emit_insn (gen_absxf2 (tmp, op1));
28182 emit_insn (gen_cmpxf (tmp,
28183 CONST_DOUBLE_FROM_REAL_VALUE (
28184 REAL_VALUE_ATOF ("0.29289321881345247561810596348408353", XFmode),
28186 emit_jump_insn (gen_bge (label1));
28188 emit_move_insn (tmp2, standard_80387_constant_rtx (4)); /* fldln2 */
28189 emit_insn (gen_fyl2xp1xf3_i387 (op0, op1, tmp2));
28190 emit_jump (label2);
28192 emit_label (label1);
28193 emit_move_insn (tmp, CONST1_RTX (XFmode));
28194 emit_insn (gen_addxf3 (tmp, op1, tmp));
28195 emit_move_insn (tmp2, standard_80387_constant_rtx (4)); /* fldln2 */
28196 emit_insn (gen_fyl2xxf3_i387 (op0, tmp, tmp2));
28198 emit_label (label2);
28201 /* Output code to perform a Newton-Rhapson approximation of a single precision
28202 floating point divide [http://en.wikipedia.org/wiki/N-th_root_algorithm]. */
28204 void ix86_emit_swdivsf (rtx res, rtx a, rtx b, enum machine_mode mode)
28206 rtx x0, x1, e0, e1, two;
28208 x0 = gen_reg_rtx (mode);
28209 e0 = gen_reg_rtx (mode);
28210 e1 = gen_reg_rtx (mode);
28211 x1 = gen_reg_rtx (mode);
28213 two = CONST_DOUBLE_FROM_REAL_VALUE (dconst2, SFmode);
28215 if (VECTOR_MODE_P (mode))
28216 two = ix86_build_const_vector (SFmode, true, two);
28218 two = force_reg (mode, two);
28220 /* a / b = a * rcp(b) * (2.0 - b * rcp(b)) */
28222 /* x0 = rcp(b) estimate */
28223 emit_insn (gen_rtx_SET (VOIDmode, x0,
28224 gen_rtx_UNSPEC (mode, gen_rtvec (1, b),
28227 emit_insn (gen_rtx_SET (VOIDmode, e0,
28228 gen_rtx_MULT (mode, x0, b)));
28230 emit_insn (gen_rtx_SET (VOIDmode, e1,
28231 gen_rtx_MINUS (mode, two, e0)));
28233 emit_insn (gen_rtx_SET (VOIDmode, x1,
28234 gen_rtx_MULT (mode, x0, e1)));
28236 emit_insn (gen_rtx_SET (VOIDmode, res,
28237 gen_rtx_MULT (mode, a, x1)));
28240 /* Output code to perform a Newton-Rhapson approximation of a
28241 single precision floating point [reciprocal] square root. */
28243 void ix86_emit_swsqrtsf (rtx res, rtx a, enum machine_mode mode,
28246 rtx x0, e0, e1, e2, e3, mthree, mhalf;
28249 x0 = gen_reg_rtx (mode);
28250 e0 = gen_reg_rtx (mode);
28251 e1 = gen_reg_rtx (mode);
28252 e2 = gen_reg_rtx (mode);
28253 e3 = gen_reg_rtx (mode);
28255 real_from_integer (&r, VOIDmode, -3, -1, 0);
28256 mthree = CONST_DOUBLE_FROM_REAL_VALUE (r, SFmode);
28258 real_arithmetic (&r, NEGATE_EXPR, &dconsthalf, NULL);
28259 mhalf = CONST_DOUBLE_FROM_REAL_VALUE (r, SFmode);
28261 if (VECTOR_MODE_P (mode))
28263 mthree = ix86_build_const_vector (SFmode, true, mthree);
28264 mhalf = ix86_build_const_vector (SFmode, true, mhalf);
28267 /* sqrt(a) = -0.5 * a * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0)
28268 rsqrt(a) = -0.5 * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0) */
28270 /* x0 = rsqrt(a) estimate */
28271 emit_insn (gen_rtx_SET (VOIDmode, x0,
28272 gen_rtx_UNSPEC (mode, gen_rtvec (1, a),
28275 /* If (a == 0.0) Filter out infinity to prevent NaN for sqrt(0.0). */
28280 zero = gen_reg_rtx (mode);
28281 mask = gen_reg_rtx (mode);
28283 zero = force_reg (mode, CONST0_RTX(mode));
28284 emit_insn (gen_rtx_SET (VOIDmode, mask,
28285 gen_rtx_NE (mode, zero, a)));
28287 emit_insn (gen_rtx_SET (VOIDmode, x0,
28288 gen_rtx_AND (mode, x0, mask)));
28292 emit_insn (gen_rtx_SET (VOIDmode, e0,
28293 gen_rtx_MULT (mode, x0, a)));
28295 emit_insn (gen_rtx_SET (VOIDmode, e1,
28296 gen_rtx_MULT (mode, e0, x0)));
28299 mthree = force_reg (mode, mthree);
28300 emit_insn (gen_rtx_SET (VOIDmode, e2,
28301 gen_rtx_PLUS (mode, e1, mthree)));
28303 mhalf = force_reg (mode, mhalf);
28305 /* e3 = -.5 * x0 */
28306 emit_insn (gen_rtx_SET (VOIDmode, e3,
28307 gen_rtx_MULT (mode, x0, mhalf)));
28309 /* e3 = -.5 * e0 */
28310 emit_insn (gen_rtx_SET (VOIDmode, e3,
28311 gen_rtx_MULT (mode, e0, mhalf)));
28312 /* ret = e2 * e3 */
28313 emit_insn (gen_rtx_SET (VOIDmode, res,
28314 gen_rtx_MULT (mode, e2, e3)));
28317 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
28319 static void ATTRIBUTE_UNUSED
28320 i386_solaris_elf_named_section (const char *name, unsigned int flags,
28323 /* With Binutils 2.15, the "@unwind" marker must be specified on
28324 every occurrence of the ".eh_frame" section, not just the first
28327 && strcmp (name, ".eh_frame") == 0)
28329 fprintf (asm_out_file, "\t.section\t%s,\"%s\",@unwind\n", name,
28330 flags & SECTION_WRITE ? "aw" : "a");
28333 default_elf_asm_named_section (name, flags, decl);
28336 /* Return the mangling of TYPE if it is an extended fundamental type. */
28338 static const char *
28339 ix86_mangle_type (const_tree type)
28341 type = TYPE_MAIN_VARIANT (type);
28343 if (TREE_CODE (type) != VOID_TYPE && TREE_CODE (type) != BOOLEAN_TYPE
28344 && TREE_CODE (type) != INTEGER_TYPE && TREE_CODE (type) != REAL_TYPE)
28347 switch (TYPE_MODE (type))
28350 /* __float128 is "g". */
28353 /* "long double" or __float80 is "e". */
28360 /* For 32-bit code we can save PIC register setup by using
28361 __stack_chk_fail_local hidden function instead of calling
28362 __stack_chk_fail directly. 64-bit code doesn't need to setup any PIC
28363 register, so it is better to call __stack_chk_fail directly. */
28366 ix86_stack_protect_fail (void)
28368 return TARGET_64BIT
28369 ? default_external_stack_protect_fail ()
28370 : default_hidden_stack_protect_fail ();
28373 /* Select a format to encode pointers in exception handling data. CODE
28374 is 0 for data, 1 for code labels, 2 for function pointers. GLOBAL is
28375 true if the symbol may be affected by dynamic relocations.
28377 ??? All x86 object file formats are capable of representing this.
28378 After all, the relocation needed is the same as for the call insn.
28379 Whether or not a particular assembler allows us to enter such, I
28380 guess we'll have to see. */
28382 asm_preferred_eh_data_format (int code, int global)
28386 int type = DW_EH_PE_sdata8;
28388 || ix86_cmodel == CM_SMALL_PIC
28389 || (ix86_cmodel == CM_MEDIUM_PIC && (global || code)))
28390 type = DW_EH_PE_sdata4;
28391 return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type;
28393 if (ix86_cmodel == CM_SMALL
28394 || (ix86_cmodel == CM_MEDIUM && code))
28395 return DW_EH_PE_udata4;
28396 return DW_EH_PE_absptr;
28399 /* Expand copysign from SIGN to the positive value ABS_VALUE
28400 storing in RESULT. If MASK is non-null, it shall be a mask to mask out
28403 ix86_sse_copysign_to_positive (rtx result, rtx abs_value, rtx sign, rtx mask)
28405 enum machine_mode mode = GET_MODE (sign);
28406 rtx sgn = gen_reg_rtx (mode);
28407 if (mask == NULL_RTX)
28409 mask = ix86_build_signbit_mask (mode, VECTOR_MODE_P (mode), false);
28410 if (!VECTOR_MODE_P (mode))
28412 /* We need to generate a scalar mode mask in this case. */
28413 rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx));
28414 tmp = gen_rtx_VEC_SELECT (mode, mask, tmp);
28415 mask = gen_reg_rtx (mode);
28416 emit_insn (gen_rtx_SET (VOIDmode, mask, tmp));
28420 mask = gen_rtx_NOT (mode, mask);
28421 emit_insn (gen_rtx_SET (VOIDmode, sgn,
28422 gen_rtx_AND (mode, mask, sign)));
28423 emit_insn (gen_rtx_SET (VOIDmode, result,
28424 gen_rtx_IOR (mode, abs_value, sgn)));
28427 /* Expand fabs (OP0) and return a new rtx that holds the result. The
28428 mask for masking out the sign-bit is stored in *SMASK, if that is
28431 ix86_expand_sse_fabs (rtx op0, rtx *smask)
28433 enum machine_mode mode = GET_MODE (op0);
28436 xa = gen_reg_rtx (mode);
28437 mask = ix86_build_signbit_mask (mode, VECTOR_MODE_P (mode), true);
28438 if (!VECTOR_MODE_P (mode))
28440 /* We need to generate a scalar mode mask in this case. */
28441 rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx));
28442 tmp = gen_rtx_VEC_SELECT (mode, mask, tmp);
28443 mask = gen_reg_rtx (mode);
28444 emit_insn (gen_rtx_SET (VOIDmode, mask, tmp));
28446 emit_insn (gen_rtx_SET (VOIDmode, xa,
28447 gen_rtx_AND (mode, op0, mask)));
28455 /* Expands a comparison of OP0 with OP1 using comparison code CODE,
28456 swapping the operands if SWAP_OPERANDS is true. The expanded
28457 code is a forward jump to a newly created label in case the
28458 comparison is true. The generated label rtx is returned. */
28460 ix86_expand_sse_compare_and_jump (enum rtx_code code, rtx op0, rtx op1,
28461 bool swap_operands)
28472 label = gen_label_rtx ();
28473 tmp = gen_rtx_REG (CCFPUmode, FLAGS_REG);
28474 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28475 gen_rtx_COMPARE (CCFPUmode, op0, op1)));
28476 tmp = gen_rtx_fmt_ee (code, VOIDmode, tmp, const0_rtx);
28477 tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
28478 gen_rtx_LABEL_REF (VOIDmode, label), pc_rtx);
28479 tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
28480 JUMP_LABEL (tmp) = label;
28485 /* Expand a mask generating SSE comparison instruction comparing OP0 with OP1
28486 using comparison code CODE. Operands are swapped for the comparison if
28487 SWAP_OPERANDS is true. Returns a rtx for the generated mask. */
28489 ix86_expand_sse_compare_mask (enum rtx_code code, rtx op0, rtx op1,
28490 bool swap_operands)
28492 enum machine_mode mode = GET_MODE (op0);
28493 rtx mask = gen_reg_rtx (mode);
28502 if (mode == DFmode)
28503 emit_insn (gen_sse2_maskcmpdf3 (mask, op0, op1,
28504 gen_rtx_fmt_ee (code, mode, op0, op1)));
28506 emit_insn (gen_sse_maskcmpsf3 (mask, op0, op1,
28507 gen_rtx_fmt_ee (code, mode, op0, op1)));
28512 /* Generate and return a rtx of mode MODE for 2**n where n is the number
28513 of bits of the mantissa of MODE, which must be one of DFmode or SFmode. */
28515 ix86_gen_TWO52 (enum machine_mode mode)
28517 REAL_VALUE_TYPE TWO52r;
28520 real_ldexp (&TWO52r, &dconst1, mode == DFmode ? 52 : 23);
28521 TWO52 = const_double_from_real_value (TWO52r, mode);
28522 TWO52 = force_reg (mode, TWO52);
28527 /* Expand SSE sequence for computing lround from OP1 storing
28530 ix86_expand_lround (rtx op0, rtx op1)
28532 /* C code for the stuff we're doing below:
28533 tmp = op1 + copysign (nextafter (0.5, 0.0), op1)
28536 enum machine_mode mode = GET_MODE (op1);
28537 const struct real_format *fmt;
28538 REAL_VALUE_TYPE pred_half, half_minus_pred_half;
28541 /* load nextafter (0.5, 0.0) */
28542 fmt = REAL_MODE_FORMAT (mode);
28543 real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode);
28544 REAL_ARITHMETIC (pred_half, MINUS_EXPR, dconsthalf, half_minus_pred_half);
28546 /* adj = copysign (0.5, op1) */
28547 adj = force_reg (mode, const_double_from_real_value (pred_half, mode));
28548 ix86_sse_copysign_to_positive (adj, adj, force_reg (mode, op1), NULL_RTX);
28550 /* adj = op1 + adj */
28551 adj = expand_simple_binop (mode, PLUS, adj, op1, NULL_RTX, 0, OPTAB_DIRECT);
28553 /* op0 = (imode)adj */
28554 expand_fix (op0, adj, 0);
28557 /* Expand SSE2 sequence for computing lround from OPERAND1 storing
28560 ix86_expand_lfloorceil (rtx op0, rtx op1, bool do_floor)
28562 /* C code for the stuff we're doing below (for do_floor):
28564 xi -= (double)xi > op1 ? 1 : 0;
28567 enum machine_mode fmode = GET_MODE (op1);
28568 enum machine_mode imode = GET_MODE (op0);
28569 rtx ireg, freg, label, tmp;
28571 /* reg = (long)op1 */
28572 ireg = gen_reg_rtx (imode);
28573 expand_fix (ireg, op1, 0);
28575 /* freg = (double)reg */
28576 freg = gen_reg_rtx (fmode);
28577 expand_float (freg, ireg, 0);
28579 /* ireg = (freg > op1) ? ireg - 1 : ireg */
28580 label = ix86_expand_sse_compare_and_jump (UNLE,
28581 freg, op1, !do_floor);
28582 tmp = expand_simple_binop (imode, do_floor ? MINUS : PLUS,
28583 ireg, const1_rtx, NULL_RTX, 0, OPTAB_DIRECT);
28584 emit_move_insn (ireg, tmp);
28586 emit_label (label);
28587 LABEL_NUSES (label) = 1;
28589 emit_move_insn (op0, ireg);
28592 /* Expand rint (IEEE round to nearest) rounding OPERAND1 and storing the
28593 result in OPERAND0. */
28595 ix86_expand_rint (rtx operand0, rtx operand1)
28597 /* C code for the stuff we're doing below:
28598 xa = fabs (operand1);
28599 if (!isless (xa, 2**52))
28601 xa = xa + 2**52 - 2**52;
28602 return copysign (xa, operand1);
28604 enum machine_mode mode = GET_MODE (operand0);
28605 rtx res, xa, label, TWO52, mask;
28607 res = gen_reg_rtx (mode);
28608 emit_move_insn (res, operand1);
28610 /* xa = abs (operand1) */
28611 xa = ix86_expand_sse_fabs (res, &mask);
28613 /* if (!isless (xa, TWO52)) goto label; */
28614 TWO52 = ix86_gen_TWO52 (mode);
28615 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28617 xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28618 xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT);
28620 ix86_sse_copysign_to_positive (res, xa, res, mask);
28622 emit_label (label);
28623 LABEL_NUSES (label) = 1;
28625 emit_move_insn (operand0, res);
28628 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
28631 ix86_expand_floorceildf_32 (rtx operand0, rtx operand1, bool do_floor)
28633 /* C code for the stuff we expand below.
28634 double xa = fabs (x), x2;
28635 if (!isless (xa, TWO52))
28637 xa = xa + TWO52 - TWO52;
28638 x2 = copysign (xa, x);
28647 enum machine_mode mode = GET_MODE (operand0);
28648 rtx xa, TWO52, tmp, label, one, res, mask;
28650 TWO52 = ix86_gen_TWO52 (mode);
28652 /* Temporary for holding the result, initialized to the input
28653 operand to ease control flow. */
28654 res = gen_reg_rtx (mode);
28655 emit_move_insn (res, operand1);
28657 /* xa = abs (operand1) */
28658 xa = ix86_expand_sse_fabs (res, &mask);
28660 /* if (!isless (xa, TWO52)) goto label; */
28661 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28663 /* xa = xa + TWO52 - TWO52; */
28664 xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28665 xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT);
28667 /* xa = copysign (xa, operand1) */
28668 ix86_sse_copysign_to_positive (xa, xa, res, mask);
28670 /* generate 1.0 or -1.0 */
28671 one = force_reg (mode,
28672 const_double_from_real_value (do_floor
28673 ? dconst1 : dconstm1, mode));
28675 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
28676 tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor);
28677 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28678 gen_rtx_AND (mode, one, tmp)));
28679 /* We always need to subtract here to preserve signed zero. */
28680 tmp = expand_simple_binop (mode, MINUS,
28681 xa, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28682 emit_move_insn (res, tmp);
28684 emit_label (label);
28685 LABEL_NUSES (label) = 1;
28687 emit_move_insn (operand0, res);
28690 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
28693 ix86_expand_floorceil (rtx operand0, rtx operand1, bool do_floor)
28695 /* C code for the stuff we expand below.
28696 double xa = fabs (x), x2;
28697 if (!isless (xa, TWO52))
28699 x2 = (double)(long)x;
28706 if (HONOR_SIGNED_ZEROS (mode))
28707 return copysign (x2, x);
28710 enum machine_mode mode = GET_MODE (operand0);
28711 rtx xa, xi, TWO52, tmp, label, one, res, mask;
28713 TWO52 = ix86_gen_TWO52 (mode);
28715 /* Temporary for holding the result, initialized to the input
28716 operand to ease control flow. */
28717 res = gen_reg_rtx (mode);
28718 emit_move_insn (res, operand1);
28720 /* xa = abs (operand1) */
28721 xa = ix86_expand_sse_fabs (res, &mask);
28723 /* if (!isless (xa, TWO52)) goto label; */
28724 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28726 /* xa = (double)(long)x */
28727 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28728 expand_fix (xi, res, 0);
28729 expand_float (xa, xi, 0);
28732 one = force_reg (mode, const_double_from_real_value (dconst1, mode));
28734 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
28735 tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor);
28736 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28737 gen_rtx_AND (mode, one, tmp)));
28738 tmp = expand_simple_binop (mode, do_floor ? MINUS : PLUS,
28739 xa, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28740 emit_move_insn (res, tmp);
28742 if (HONOR_SIGNED_ZEROS (mode))
28743 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), mask);
28745 emit_label (label);
28746 LABEL_NUSES (label) = 1;
28748 emit_move_insn (operand0, res);
28751 /* Expand SSE sequence for computing round from OPERAND1 storing
28752 into OPERAND0. Sequence that works without relying on DImode truncation
28753 via cvttsd2siq that is only available on 64bit targets. */
28755 ix86_expand_rounddf_32 (rtx operand0, rtx operand1)
28757 /* C code for the stuff we expand below.
28758 double xa = fabs (x), xa2, x2;
28759 if (!isless (xa, TWO52))
28761 Using the absolute value and copying back sign makes
28762 -0.0 -> -0.0 correct.
28763 xa2 = xa + TWO52 - TWO52;
28768 else if (dxa > 0.5)
28770 x2 = copysign (xa2, x);
28773 enum machine_mode mode = GET_MODE (operand0);
28774 rtx xa, xa2, dxa, TWO52, tmp, label, half, mhalf, one, res, mask;
28776 TWO52 = ix86_gen_TWO52 (mode);
28778 /* Temporary for holding the result, initialized to the input
28779 operand to ease control flow. */
28780 res = gen_reg_rtx (mode);
28781 emit_move_insn (res, operand1);
28783 /* xa = abs (operand1) */
28784 xa = ix86_expand_sse_fabs (res, &mask);
28786 /* if (!isless (xa, TWO52)) goto label; */
28787 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28789 /* xa2 = xa + TWO52 - TWO52; */
28790 xa2 = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28791 xa2 = expand_simple_binop (mode, MINUS, xa2, TWO52, xa2, 0, OPTAB_DIRECT);
28793 /* dxa = xa2 - xa; */
28794 dxa = expand_simple_binop (mode, MINUS, xa2, xa, NULL_RTX, 0, OPTAB_DIRECT);
28796 /* generate 0.5, 1.0 and -0.5 */
28797 half = force_reg (mode, const_double_from_real_value (dconsthalf, mode));
28798 one = expand_simple_binop (mode, PLUS, half, half, NULL_RTX, 0, OPTAB_DIRECT);
28799 mhalf = expand_simple_binop (mode, MINUS, half, one, NULL_RTX,
28803 tmp = gen_reg_rtx (mode);
28804 /* xa2 = xa2 - (dxa > 0.5 ? 1 : 0) */
28805 tmp = ix86_expand_sse_compare_mask (UNGT, dxa, half, false);
28806 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28807 gen_rtx_AND (mode, one, tmp)));
28808 xa2 = expand_simple_binop (mode, MINUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28809 /* xa2 = xa2 + (dxa <= -0.5 ? 1 : 0) */
28810 tmp = ix86_expand_sse_compare_mask (UNGE, mhalf, dxa, false);
28811 emit_insn (gen_rtx_SET (VOIDmode, tmp,
28812 gen_rtx_AND (mode, one, tmp)));
28813 xa2 = expand_simple_binop (mode, PLUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT);
28815 /* res = copysign (xa2, operand1) */
28816 ix86_sse_copysign_to_positive (res, xa2, force_reg (mode, operand1), mask);
28818 emit_label (label);
28819 LABEL_NUSES (label) = 1;
28821 emit_move_insn (operand0, res);
28824 /* Expand SSE sequence for computing trunc from OPERAND1 storing
28827 ix86_expand_trunc (rtx operand0, rtx operand1)
28829 /* C code for SSE variant we expand below.
28830 double xa = fabs (x), x2;
28831 if (!isless (xa, TWO52))
28833 x2 = (double)(long)x;
28834 if (HONOR_SIGNED_ZEROS (mode))
28835 return copysign (x2, x);
28838 enum machine_mode mode = GET_MODE (operand0);
28839 rtx xa, xi, TWO52, label, res, mask;
28841 TWO52 = ix86_gen_TWO52 (mode);
28843 /* Temporary for holding the result, initialized to the input
28844 operand to ease control flow. */
28845 res = gen_reg_rtx (mode);
28846 emit_move_insn (res, operand1);
28848 /* xa = abs (operand1) */
28849 xa = ix86_expand_sse_fabs (res, &mask);
28851 /* if (!isless (xa, TWO52)) goto label; */
28852 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28854 /* x = (double)(long)x */
28855 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28856 expand_fix (xi, res, 0);
28857 expand_float (res, xi, 0);
28859 if (HONOR_SIGNED_ZEROS (mode))
28860 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), mask);
28862 emit_label (label);
28863 LABEL_NUSES (label) = 1;
28865 emit_move_insn (operand0, res);
28868 /* Expand SSE sequence for computing trunc from OPERAND1 storing
28871 ix86_expand_truncdf_32 (rtx operand0, rtx operand1)
28873 enum machine_mode mode = GET_MODE (operand0);
28874 rtx xa, mask, TWO52, label, one, res, smask, tmp;
28876 /* C code for SSE variant we expand below.
28877 double xa = fabs (x), x2;
28878 if (!isless (xa, TWO52))
28880 xa2 = xa + TWO52 - TWO52;
28884 x2 = copysign (xa2, x);
28888 TWO52 = ix86_gen_TWO52 (mode);
28890 /* Temporary for holding the result, initialized to the input
28891 operand to ease control flow. */
28892 res = gen_reg_rtx (mode);
28893 emit_move_insn (res, operand1);
28895 /* xa = abs (operand1) */
28896 xa = ix86_expand_sse_fabs (res, &smask);
28898 /* if (!isless (xa, TWO52)) goto label; */
28899 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28901 /* res = xa + TWO52 - TWO52; */
28902 tmp = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT);
28903 tmp = expand_simple_binop (mode, MINUS, tmp, TWO52, tmp, 0, OPTAB_DIRECT);
28904 emit_move_insn (res, tmp);
28907 one = force_reg (mode, const_double_from_real_value (dconst1, mode));
28909 /* Compensate: res = xa2 - (res > xa ? 1 : 0) */
28910 mask = ix86_expand_sse_compare_mask (UNGT, res, xa, false);
28911 emit_insn (gen_rtx_SET (VOIDmode, mask,
28912 gen_rtx_AND (mode, mask, one)));
28913 tmp = expand_simple_binop (mode, MINUS,
28914 res, mask, NULL_RTX, 0, OPTAB_DIRECT);
28915 emit_move_insn (res, tmp);
28917 /* res = copysign (res, operand1) */
28918 ix86_sse_copysign_to_positive (res, res, force_reg (mode, operand1), smask);
28920 emit_label (label);
28921 LABEL_NUSES (label) = 1;
28923 emit_move_insn (operand0, res);
28926 /* Expand SSE sequence for computing round from OPERAND1 storing
28929 ix86_expand_round (rtx operand0, rtx operand1)
28931 /* C code for the stuff we're doing below:
28932 double xa = fabs (x);
28933 if (!isless (xa, TWO52))
28935 xa = (double)(long)(xa + nextafter (0.5, 0.0));
28936 return copysign (xa, x);
28938 enum machine_mode mode = GET_MODE (operand0);
28939 rtx res, TWO52, xa, label, xi, half, mask;
28940 const struct real_format *fmt;
28941 REAL_VALUE_TYPE pred_half, half_minus_pred_half;
28943 /* Temporary for holding the result, initialized to the input
28944 operand to ease control flow. */
28945 res = gen_reg_rtx (mode);
28946 emit_move_insn (res, operand1);
28948 TWO52 = ix86_gen_TWO52 (mode);
28949 xa = ix86_expand_sse_fabs (res, &mask);
28950 label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false);
28952 /* load nextafter (0.5, 0.0) */
28953 fmt = REAL_MODE_FORMAT (mode);
28954 real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode);
28955 REAL_ARITHMETIC (pred_half, MINUS_EXPR, dconsthalf, half_minus_pred_half);
28957 /* xa = xa + 0.5 */
28958 half = force_reg (mode, const_double_from_real_value (pred_half, mode));
28959 xa = expand_simple_binop (mode, PLUS, xa, half, NULL_RTX, 0, OPTAB_DIRECT);
28961 /* xa = (double)(int64_t)xa */
28962 xi = gen_reg_rtx (mode == DFmode ? DImode : SImode);
28963 expand_fix (xi, xa, 0);
28964 expand_float (xa, xi, 0);
28966 /* res = copysign (xa, operand1) */
28967 ix86_sse_copysign_to_positive (res, xa, force_reg (mode, operand1), mask);
28969 emit_label (label);
28970 LABEL_NUSES (label) = 1;
28972 emit_move_insn (operand0, res);
28976 /* Validate whether a SSE5 instruction is valid or not.
28977 OPERANDS is the array of operands.
28978 NUM is the number of operands.
28979 USES_OC0 is true if the instruction uses OC0 and provides 4 variants.
28980 NUM_MEMORY is the maximum number of memory operands to accept.
28981 when COMMUTATIVE is set, operand 1 and 2 can be swapped. */
28984 ix86_sse5_valid_op_p (rtx operands[], rtx insn ATTRIBUTE_UNUSED, int num,
28985 bool uses_oc0, int num_memory, bool commutative)
28991 /* Count the number of memory arguments */
28994 for (i = 0; i < num; i++)
28996 enum machine_mode mode = GET_MODE (operands[i]);
28997 if (register_operand (operands[i], mode))
29000 else if (memory_operand (operands[i], mode))
29002 mem_mask |= (1 << i);
29008 rtx pattern = PATTERN (insn);
29010 /* allow 0 for pcmov */
29011 if (GET_CODE (pattern) != SET
29012 || GET_CODE (SET_SRC (pattern)) != IF_THEN_ELSE
29014 || operands[i] != CONST0_RTX (mode))
29019 /* Special case pmacsdq{l,h} where we allow the 3rd argument to be
29020 a memory operation. */
29021 if (num_memory < 0)
29023 num_memory = -num_memory;
29024 if ((mem_mask & (1 << (num-1))) != 0)
29026 mem_mask &= ~(1 << (num-1));
29031 /* If there were no memory operations, allow the insn */
29035 /* Do not allow the destination register to be a memory operand. */
29036 else if (mem_mask & (1 << 0))
29039 /* If there are too many memory operations, disallow the instruction. While
29040 the hardware only allows 1 memory reference, before register allocation
29041 for some insns, we allow two memory operations sometimes in order to allow
29042 code like the following to be optimized:
29044 float fmadd (float *a, float *b, float *c) { return (*a * *b) + *c; }
29046 or similar cases that are vectorized into using the fmaddss
29048 else if (mem_count > num_memory)
29051 /* Don't allow more than one memory operation if not optimizing. */
29052 else if (mem_count > 1 && !optimize)
29055 else if (num == 4 && mem_count == 1)
29057 /* formats (destination is the first argument), example fmaddss:
29058 xmm1, xmm1, xmm2, xmm3/mem
29059 xmm1, xmm1, xmm2/mem, xmm3
29060 xmm1, xmm2, xmm3/mem, xmm1
29061 xmm1, xmm2/mem, xmm3, xmm1 */
29063 return ((mem_mask == (1 << 1))
29064 || (mem_mask == (1 << 2))
29065 || (mem_mask == (1 << 3)));
29067 /* format, example pmacsdd:
29068 xmm1, xmm2, xmm3/mem, xmm1 */
29070 return (mem_mask == (1 << 2) || mem_mask == (1 << 1));
29072 return (mem_mask == (1 << 2));
29075 else if (num == 4 && num_memory == 2)
29077 /* If there are two memory operations, we can load one of the memory ops
29078 into the destination register. This is for optimizing the
29079 multiply/add ops, which the combiner has optimized both the multiply
29080 and the add insns to have a memory operation. We have to be careful
29081 that the destination doesn't overlap with the inputs. */
29082 rtx op0 = operands[0];
29084 if (reg_mentioned_p (op0, operands[1])
29085 || reg_mentioned_p (op0, operands[2])
29086 || reg_mentioned_p (op0, operands[3]))
29089 /* formats (destination is the first argument), example fmaddss:
29090 xmm1, xmm1, xmm2, xmm3/mem
29091 xmm1, xmm1, xmm2/mem, xmm3
29092 xmm1, xmm2, xmm3/mem, xmm1
29093 xmm1, xmm2/mem, xmm3, xmm1
29095 For the oc0 case, we will load either operands[1] or operands[3] into
29096 operands[0], so any combination of 2 memory operands is ok. */
29100 /* format, example pmacsdd:
29101 xmm1, xmm2, xmm3/mem, xmm1
29103 For the integer multiply/add instructions be more restrictive and
29104 require operands[2] and operands[3] to be the memory operands. */
29106 return (mem_mask == ((1 << 1) | (1 << 3)) || ((1 << 2) | (1 << 3)));
29108 return (mem_mask == ((1 << 2) | (1 << 3)));
29111 else if (num == 3 && num_memory == 1)
29113 /* formats, example protb:
29114 xmm1, xmm2, xmm3/mem
29115 xmm1, xmm2/mem, xmm3 */
29117 return ((mem_mask == (1 << 1)) || (mem_mask == (1 << 2)));
29119 /* format, example comeq:
29120 xmm1, xmm2, xmm3/mem */
29122 return (mem_mask == (1 << 2));
29126 gcc_unreachable ();
29132 /* Fixup an SSE5 instruction that has 2 memory input references into a form the
29133 hardware will allow by using the destination register to load one of the
29134 memory operations. Presently this is used by the multiply/add routines to
29135 allow 2 memory references. */
29138 ix86_expand_sse5_multiple_memory (rtx operands[],
29140 enum machine_mode mode)
29142 rtx op0 = operands[0];
29144 || memory_operand (op0, mode)
29145 || reg_mentioned_p (op0, operands[1])
29146 || reg_mentioned_p (op0, operands[2])
29147 || reg_mentioned_p (op0, operands[3]))
29148 gcc_unreachable ();
29150 /* For 2 memory operands, pick either operands[1] or operands[3] to move into
29151 the destination register. */
29152 if (memory_operand (operands[1], mode))
29154 emit_move_insn (op0, operands[1]);
29157 else if (memory_operand (operands[3], mode))
29159 emit_move_insn (op0, operands[3]);
29163 gcc_unreachable ();
29169 /* Table of valid machine attributes. */
29170 static const struct attribute_spec ix86_attribute_table[] =
29172 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
29173 /* Stdcall attribute says callee is responsible for popping arguments
29174 if they are not variable. */
29175 { "stdcall", 0, 0, false, true, true, ix86_handle_cconv_attribute },
29176 /* Fastcall attribute says callee is responsible for popping arguments
29177 if they are not variable. */
29178 { "fastcall", 0, 0, false, true, true, ix86_handle_cconv_attribute },
29179 /* Cdecl attribute says the callee is a normal C declaration */
29180 { "cdecl", 0, 0, false, true, true, ix86_handle_cconv_attribute },
29181 /* Regparm attribute specifies how many integer arguments are to be
29182 passed in registers. */
29183 { "regparm", 1, 1, false, true, true, ix86_handle_cconv_attribute },
29184 /* Sseregparm attribute says we are using x86_64 calling conventions
29185 for FP arguments. */
29186 { "sseregparm", 0, 0, false, true, true, ix86_handle_cconv_attribute },
29187 /* force_align_arg_pointer says this function realigns the stack at entry. */
29188 { (const char *)&ix86_force_align_arg_pointer_string, 0, 0,
29189 false, true, true, ix86_handle_cconv_attribute },
29190 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
29191 { "dllimport", 0, 0, false, false, false, handle_dll_attribute },
29192 { "dllexport", 0, 0, false, false, false, handle_dll_attribute },
29193 { "shared", 0, 0, true, false, false, ix86_handle_shared_attribute },
29195 { "ms_struct", 0, 0, false, false, false, ix86_handle_struct_attribute },
29196 { "gcc_struct", 0, 0, false, false, false, ix86_handle_struct_attribute },
29197 #ifdef SUBTARGET_ATTRIBUTE_TABLE
29198 SUBTARGET_ATTRIBUTE_TABLE,
29200 /* ms_abi and sysv_abi calling convention function attributes. */
29201 { "ms_abi", 0, 0, false, true, true, ix86_handle_abi_attribute },
29202 { "sysv_abi", 0, 0, false, true, true, ix86_handle_abi_attribute },
29204 { NULL, 0, 0, false, false, false, NULL }
29207 /* Implement targetm.vectorize.builtin_vectorization_cost. */
29209 x86_builtin_vectorization_cost (bool runtime_test)
29211 /* If the branch of the runtime test is taken - i.e. - the vectorized
29212 version is skipped - this incurs a misprediction cost (because the
29213 vectorized version is expected to be the fall-through). So we subtract
29214 the latency of a mispredicted branch from the costs that are incured
29215 when the vectorized version is executed.
29217 TODO: The values in individual target tables have to be tuned or new
29218 fields may be needed. For eg. on K8, the default branch path is the
29219 not-taken path. If the taken path is predicted correctly, the minimum
29220 penalty of going down the taken-path is 1 cycle. If the taken-path is
29221 not predicted correctly, then the minimum penalty is 10 cycles. */
29225 return (-(ix86_cost->cond_taken_branch_cost));
29231 /* This function returns the calling abi specific va_list type node.
29232 It returns the FNDECL specific va_list type. */
29235 ix86_fn_abi_va_list (tree fndecl)
29240 return va_list_type_node;
29241 gcc_assert (fndecl != NULL_TREE);
29242 abi = ix86_function_abi ((const_tree) fndecl);
29245 return ms_va_list_type_node;
29247 return sysv_va_list_type_node;
29250 /* Returns the canonical va_list type specified by TYPE. If there
29251 is no valid TYPE provided, it return NULL_TREE. */
29254 ix86_canonical_va_list_type (tree type)
29258 /* Resolve references and pointers to va_list type. */
29259 if (INDIRECT_REF_P (type))
29260 type = TREE_TYPE (type);
29261 else if (POINTER_TYPE_P (type) && POINTER_TYPE_P (TREE_TYPE(type)))
29262 type = TREE_TYPE (type);
29266 wtype = va_list_type_node;
29267 gcc_assert (wtype != NULL_TREE);
29269 if (TREE_CODE (wtype) == ARRAY_TYPE)
29271 /* If va_list is an array type, the argument may have decayed
29272 to a pointer type, e.g. by being passed to another function.
29273 In that case, unwrap both types so that we can compare the
29274 underlying records. */
29275 if (TREE_CODE (htype) == ARRAY_TYPE
29276 || POINTER_TYPE_P (htype))
29278 wtype = TREE_TYPE (wtype);
29279 htype = TREE_TYPE (htype);
29282 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
29283 return va_list_type_node;
29284 wtype = sysv_va_list_type_node;
29285 gcc_assert (wtype != NULL_TREE);
29287 if (TREE_CODE (wtype) == ARRAY_TYPE)
29289 /* If va_list is an array type, the argument may have decayed
29290 to a pointer type, e.g. by being passed to another function.
29291 In that case, unwrap both types so that we can compare the
29292 underlying records. */
29293 if (TREE_CODE (htype) == ARRAY_TYPE
29294 || POINTER_TYPE_P (htype))
29296 wtype = TREE_TYPE (wtype);
29297 htype = TREE_TYPE (htype);
29300 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
29301 return sysv_va_list_type_node;
29302 wtype = ms_va_list_type_node;
29303 gcc_assert (wtype != NULL_TREE);
29305 if (TREE_CODE (wtype) == ARRAY_TYPE)
29307 /* If va_list is an array type, the argument may have decayed
29308 to a pointer type, e.g. by being passed to another function.
29309 In that case, unwrap both types so that we can compare the
29310 underlying records. */
29311 if (TREE_CODE (htype) == ARRAY_TYPE
29312 || POINTER_TYPE_P (htype))
29314 wtype = TREE_TYPE (wtype);
29315 htype = TREE_TYPE (htype);
29318 if (TYPE_MAIN_VARIANT (wtype) == TYPE_MAIN_VARIANT (htype))
29319 return ms_va_list_type_node;
29322 return std_canonical_va_list_type (type);
29325 /* Iterate through the target-specific builtin types for va_list.
29326 IDX denotes the iterator, *PTREE is set to the result type of
29327 the va_list builtin, and *PNAME to its internal type.
29328 Returns zero if there is no element for this index, otherwise
29329 IDX should be increased upon the next call.
29330 Note, do not iterate a base builtin's name like __builtin_va_list.
29331 Used from c_common_nodes_and_builtins. */
29334 ix86_enum_va_list (int idx, const char **pname, tree *ptree)
29340 *ptree = ms_va_list_type_node;
29341 *pname = "__builtin_ms_va_list";
29344 *ptree = sysv_va_list_type_node;
29345 *pname = "__builtin_sysv_va_list";
29353 /* Initialize the GCC target structure. */
29354 #undef TARGET_RETURN_IN_MEMORY
29355 #define TARGET_RETURN_IN_MEMORY ix86_return_in_memory
29357 #undef TARGET_ATTRIBUTE_TABLE
29358 #define TARGET_ATTRIBUTE_TABLE ix86_attribute_table
29359 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
29360 # undef TARGET_MERGE_DECL_ATTRIBUTES
29361 # define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes
29364 #undef TARGET_COMP_TYPE_ATTRIBUTES
29365 #define TARGET_COMP_TYPE_ATTRIBUTES ix86_comp_type_attributes
29367 #undef TARGET_INIT_BUILTINS
29368 #define TARGET_INIT_BUILTINS ix86_init_builtins
29369 #undef TARGET_EXPAND_BUILTIN
29370 #define TARGET_EXPAND_BUILTIN ix86_expand_builtin
29372 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
29373 #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
29374 ix86_builtin_vectorized_function
29376 #undef TARGET_VECTORIZE_BUILTIN_CONVERSION
29377 #define TARGET_VECTORIZE_BUILTIN_CONVERSION ix86_vectorize_builtin_conversion
29379 #undef TARGET_BUILTIN_RECIPROCAL
29380 #define TARGET_BUILTIN_RECIPROCAL ix86_builtin_reciprocal
29382 #undef TARGET_ASM_FUNCTION_EPILOGUE
29383 #define TARGET_ASM_FUNCTION_EPILOGUE ix86_output_function_epilogue
29385 #undef TARGET_ENCODE_SECTION_INFO
29386 #ifndef SUBTARGET_ENCODE_SECTION_INFO
29387 #define TARGET_ENCODE_SECTION_INFO ix86_encode_section_info
29389 #define TARGET_ENCODE_SECTION_INFO SUBTARGET_ENCODE_SECTION_INFO
29392 #undef TARGET_ASM_OPEN_PAREN
29393 #define TARGET_ASM_OPEN_PAREN ""
29394 #undef TARGET_ASM_CLOSE_PAREN
29395 #define TARGET_ASM_CLOSE_PAREN ""
29397 #undef TARGET_ASM_ALIGNED_HI_OP
29398 #define TARGET_ASM_ALIGNED_HI_OP ASM_SHORT
29399 #undef TARGET_ASM_ALIGNED_SI_OP
29400 #define TARGET_ASM_ALIGNED_SI_OP ASM_LONG
29402 #undef TARGET_ASM_ALIGNED_DI_OP
29403 #define TARGET_ASM_ALIGNED_DI_OP ASM_QUAD
29406 #undef TARGET_ASM_UNALIGNED_HI_OP
29407 #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP
29408 #undef TARGET_ASM_UNALIGNED_SI_OP
29409 #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP
29410 #undef TARGET_ASM_UNALIGNED_DI_OP
29411 #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP
29413 #undef TARGET_SCHED_ADJUST_COST
29414 #define TARGET_SCHED_ADJUST_COST ix86_adjust_cost
29415 #undef TARGET_SCHED_ISSUE_RATE
29416 #define TARGET_SCHED_ISSUE_RATE ix86_issue_rate
29417 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
29418 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
29419 ia32_multipass_dfa_lookahead
29421 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
29422 #define TARGET_FUNCTION_OK_FOR_SIBCALL ix86_function_ok_for_sibcall
29425 #undef TARGET_HAVE_TLS
29426 #define TARGET_HAVE_TLS true
29428 #undef TARGET_CANNOT_FORCE_CONST_MEM
29429 #define TARGET_CANNOT_FORCE_CONST_MEM ix86_cannot_force_const_mem
29430 #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
29431 #define TARGET_USE_BLOCKS_FOR_CONSTANT_P hook_bool_mode_const_rtx_true
29433 #undef TARGET_DELEGITIMIZE_ADDRESS
29434 #define TARGET_DELEGITIMIZE_ADDRESS ix86_delegitimize_address
29436 #undef TARGET_MS_BITFIELD_LAYOUT_P
29437 #define TARGET_MS_BITFIELD_LAYOUT_P ix86_ms_bitfield_layout_p
29440 #undef TARGET_BINDS_LOCAL_P
29441 #define TARGET_BINDS_LOCAL_P darwin_binds_local_p
29443 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
29444 #undef TARGET_BINDS_LOCAL_P
29445 #define TARGET_BINDS_LOCAL_P i386_pe_binds_local_p
29448 #undef TARGET_ASM_OUTPUT_MI_THUNK
29449 #define TARGET_ASM_OUTPUT_MI_THUNK x86_output_mi_thunk
29450 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
29451 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK x86_can_output_mi_thunk
29453 #undef TARGET_ASM_FILE_START
29454 #define TARGET_ASM_FILE_START x86_file_start
29456 #undef TARGET_DEFAULT_TARGET_FLAGS
29457 #define TARGET_DEFAULT_TARGET_FLAGS \
29459 | TARGET_SUBTARGET_DEFAULT \
29460 | TARGET_TLS_DIRECT_SEG_REFS_DEFAULT)
29462 #undef TARGET_HANDLE_OPTION
29463 #define TARGET_HANDLE_OPTION ix86_handle_option
29465 #undef TARGET_RTX_COSTS
29466 #define TARGET_RTX_COSTS ix86_rtx_costs
29467 #undef TARGET_ADDRESS_COST
29468 #define TARGET_ADDRESS_COST ix86_address_cost
29470 #undef TARGET_FIXED_CONDITION_CODE_REGS
29471 #define TARGET_FIXED_CONDITION_CODE_REGS ix86_fixed_condition_code_regs
29472 #undef TARGET_CC_MODES_COMPATIBLE
29473 #define TARGET_CC_MODES_COMPATIBLE ix86_cc_modes_compatible
29475 #undef TARGET_MACHINE_DEPENDENT_REORG
29476 #define TARGET_MACHINE_DEPENDENT_REORG ix86_reorg
29478 #undef TARGET_BUILD_BUILTIN_VA_LIST
29479 #define TARGET_BUILD_BUILTIN_VA_LIST ix86_build_builtin_va_list
29481 #undef TARGET_FN_ABI_VA_LIST
29482 #define TARGET_FN_ABI_VA_LIST ix86_fn_abi_va_list
29484 #undef TARGET_CANONICAL_VA_LIST_TYPE
29485 #define TARGET_CANONICAL_VA_LIST_TYPE ix86_canonical_va_list_type
29487 #undef TARGET_EXPAND_BUILTIN_VA_START
29488 #define TARGET_EXPAND_BUILTIN_VA_START ix86_va_start
29490 #undef TARGET_MD_ASM_CLOBBERS
29491 #define TARGET_MD_ASM_CLOBBERS ix86_md_asm_clobbers
29493 #undef TARGET_PROMOTE_PROTOTYPES
29494 #define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
29495 #undef TARGET_STRUCT_VALUE_RTX
29496 #define TARGET_STRUCT_VALUE_RTX ix86_struct_value_rtx
29497 #undef TARGET_SETUP_INCOMING_VARARGS
29498 #define TARGET_SETUP_INCOMING_VARARGS ix86_setup_incoming_varargs
29499 #undef TARGET_MUST_PASS_IN_STACK
29500 #define TARGET_MUST_PASS_IN_STACK ix86_must_pass_in_stack
29501 #undef TARGET_PASS_BY_REFERENCE
29502 #define TARGET_PASS_BY_REFERENCE ix86_pass_by_reference
29503 #undef TARGET_INTERNAL_ARG_POINTER
29504 #define TARGET_INTERNAL_ARG_POINTER ix86_internal_arg_pointer
29505 #undef TARGET_UPDATE_STACK_BOUNDARY
29506 #define TARGET_UPDATE_STACK_BOUNDARY ix86_update_stack_boundary
29507 #undef TARGET_GET_DRAP_RTX
29508 #define TARGET_GET_DRAP_RTX ix86_get_drap_rtx
29509 #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC
29510 #define TARGET_DWARF_HANDLE_FRAME_UNSPEC ix86_dwarf_handle_frame_unspec
29511 #undef TARGET_STRICT_ARGUMENT_NAMING
29512 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
29514 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
29515 #define TARGET_GIMPLIFY_VA_ARG_EXPR ix86_gimplify_va_arg
29517 #undef TARGET_SCALAR_MODE_SUPPORTED_P
29518 #define TARGET_SCALAR_MODE_SUPPORTED_P ix86_scalar_mode_supported_p
29520 #undef TARGET_VECTOR_MODE_SUPPORTED_P
29521 #define TARGET_VECTOR_MODE_SUPPORTED_P ix86_vector_mode_supported_p
29523 #undef TARGET_C_MODE_FOR_SUFFIX
29524 #define TARGET_C_MODE_FOR_SUFFIX ix86_c_mode_for_suffix
29527 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
29528 #define TARGET_ASM_OUTPUT_DWARF_DTPREL i386_output_dwarf_dtprel
29531 #ifdef SUBTARGET_INSERT_ATTRIBUTES
29532 #undef TARGET_INSERT_ATTRIBUTES
29533 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
29536 #undef TARGET_MANGLE_TYPE
29537 #define TARGET_MANGLE_TYPE ix86_mangle_type
29539 #undef TARGET_STACK_PROTECT_FAIL
29540 #define TARGET_STACK_PROTECT_FAIL ix86_stack_protect_fail
29542 #undef TARGET_FUNCTION_VALUE
29543 #define TARGET_FUNCTION_VALUE ix86_function_value
29545 #undef TARGET_SECONDARY_RELOAD
29546 #define TARGET_SECONDARY_RELOAD ix86_secondary_reload
29548 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST
29549 #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST x86_builtin_vectorization_cost
29551 #undef TARGET_SET_CURRENT_FUNCTION
29552 #define TARGET_SET_CURRENT_FUNCTION ix86_set_current_function
29554 #undef TARGET_OPTION_VALID_ATTRIBUTE_P
29555 #define TARGET_OPTION_VALID_ATTRIBUTE_P ix86_valid_target_attribute_p
29557 #undef TARGET_OPTION_SAVE
29558 #define TARGET_OPTION_SAVE ix86_function_specific_save
29560 #undef TARGET_OPTION_RESTORE
29561 #define TARGET_OPTION_RESTORE ix86_function_specific_restore
29563 #undef TARGET_OPTION_PRINT
29564 #define TARGET_OPTION_PRINT ix86_function_specific_print
29566 #undef TARGET_OPTION_CAN_INLINE_P
29567 #define TARGET_OPTION_CAN_INLINE_P ix86_can_inline_p
29569 #undef TARGET_EXPAND_TO_RTL_HOOK
29570 #define TARGET_EXPAND_TO_RTL_HOOK ix86_maybe_switch_abi
29572 struct gcc_target targetm = TARGET_INITIALIZER;
29574 #include "gt-i386.h"