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
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"
51 #include "tree-gimple.h"
54 #include "tm-constrs.h"
57 static int x86_builtin_vectorization_cost (bool);
58 static rtx legitimize_dllimport_symbol (rtx, bool);
60 #ifndef CHECK_STACK_LIMIT
61 #define CHECK_STACK_LIMIT (-1)
64 /* Return index of given mode in mult and division cost tables. */
65 #define MODE_INDEX(mode) \
66 ((mode) == QImode ? 0 \
67 : (mode) == HImode ? 1 \
68 : (mode) == SImode ? 2 \
69 : (mode) == DImode ? 3 \
72 /* Processor costs (relative to an add) */
73 /* We assume COSTS_N_INSNS is defined as (N)*4 and an addition is 2 bytes. */
74 #define COSTS_N_BYTES(N) ((N) * 2)
76 #define DUMMY_STRINGOP_ALGS {libcall, {{-1, libcall}}}
79 struct processor_costs size_cost = { /* costs for tuning for size */
80 COSTS_N_BYTES (2), /* cost of an add instruction */
81 COSTS_N_BYTES (3), /* cost of a lea instruction */
82 COSTS_N_BYTES (2), /* variable shift costs */
83 COSTS_N_BYTES (3), /* constant shift costs */
84 {COSTS_N_BYTES (3), /* cost of starting multiply for QI */
85 COSTS_N_BYTES (3), /* HI */
86 COSTS_N_BYTES (3), /* SI */
87 COSTS_N_BYTES (3), /* DI */
88 COSTS_N_BYTES (5)}, /* other */
89 0, /* cost of multiply per each bit set */
90 {COSTS_N_BYTES (3), /* cost of a divide/mod for QI */
91 COSTS_N_BYTES (3), /* HI */
92 COSTS_N_BYTES (3), /* SI */
93 COSTS_N_BYTES (3), /* DI */
94 COSTS_N_BYTES (5)}, /* other */
95 COSTS_N_BYTES (3), /* cost of movsx */
96 COSTS_N_BYTES (3), /* cost of movzx */
99 2, /* cost for loading QImode using movzbl */
100 {2, 2, 2}, /* cost of loading integer registers
101 in QImode, HImode and SImode.
102 Relative to reg-reg move (2). */
103 {2, 2, 2}, /* cost of storing integer registers */
104 2, /* cost of reg,reg fld/fst */
105 {2, 2, 2}, /* cost of loading fp registers
106 in SFmode, DFmode and XFmode */
107 {2, 2, 2}, /* cost of storing fp registers
108 in SFmode, DFmode and XFmode */
109 3, /* cost of moving MMX register */
110 {3, 3}, /* cost of loading MMX registers
111 in SImode and DImode */
112 {3, 3}, /* cost of storing MMX registers
113 in SImode and DImode */
114 3, /* cost of moving SSE register */
115 {3, 3, 3}, /* cost of loading SSE registers
116 in SImode, DImode and TImode */
117 {3, 3, 3}, /* cost of storing SSE registers
118 in SImode, DImode and TImode */
119 3, /* MMX or SSE register to integer */
120 0, /* size of l1 cache */
121 0, /* size of l2 cache */
122 0, /* size of prefetch block */
123 0, /* number of parallel prefetches */
125 COSTS_N_BYTES (2), /* cost of FADD and FSUB insns. */
126 COSTS_N_BYTES (2), /* cost of FMUL instruction. */
127 COSTS_N_BYTES (2), /* cost of FDIV instruction. */
128 COSTS_N_BYTES (2), /* cost of FABS instruction. */
129 COSTS_N_BYTES (2), /* cost of FCHS instruction. */
130 COSTS_N_BYTES (2), /* cost of FSQRT instruction. */
131 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
132 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
133 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
134 {rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}}},
135 1, /* scalar_stmt_cost. */
136 1, /* scalar load_cost. */
137 1, /* scalar_store_cost. */
138 1, /* vec_stmt_cost. */
139 1, /* vec_to_scalar_cost. */
140 1, /* scalar_to_vec_cost. */
141 1, /* vec_align_load_cost. */
142 1, /* vec_unalign_load_cost. */
143 1, /* vec_store_cost. */
144 1, /* cond_taken_branch_cost. */
145 1, /* cond_not_taken_branch_cost. */
148 /* Processor costs (relative to an add) */
150 struct processor_costs i386_cost = { /* 386 specific costs */
151 COSTS_N_INSNS (1), /* cost of an add instruction */
152 COSTS_N_INSNS (1), /* cost of a lea instruction */
153 COSTS_N_INSNS (3), /* variable shift costs */
154 COSTS_N_INSNS (2), /* constant shift costs */
155 {COSTS_N_INSNS (6), /* cost of starting multiply for QI */
156 COSTS_N_INSNS (6), /* HI */
157 COSTS_N_INSNS (6), /* SI */
158 COSTS_N_INSNS (6), /* DI */
159 COSTS_N_INSNS (6)}, /* other */
160 COSTS_N_INSNS (1), /* cost of multiply per each bit set */
161 {COSTS_N_INSNS (23), /* cost of a divide/mod for QI */
162 COSTS_N_INSNS (23), /* HI */
163 COSTS_N_INSNS (23), /* SI */
164 COSTS_N_INSNS (23), /* DI */
165 COSTS_N_INSNS (23)}, /* other */
166 COSTS_N_INSNS (3), /* cost of movsx */
167 COSTS_N_INSNS (2), /* cost of movzx */
168 15, /* "large" insn */
170 4, /* cost for loading QImode using movzbl */
171 {2, 4, 2}, /* cost of loading integer registers
172 in QImode, HImode and SImode.
173 Relative to reg-reg move (2). */
174 {2, 4, 2}, /* cost of storing integer registers */
175 2, /* cost of reg,reg fld/fst */
176 {8, 8, 8}, /* cost of loading fp registers
177 in SFmode, DFmode and XFmode */
178 {8, 8, 8}, /* cost of storing fp registers
179 in SFmode, DFmode and XFmode */
180 2, /* cost of moving MMX register */
181 {4, 8}, /* cost of loading MMX registers
182 in SImode and DImode */
183 {4, 8}, /* cost of storing MMX registers
184 in SImode and DImode */
185 2, /* cost of moving SSE register */
186 {4, 8, 16}, /* cost of loading SSE registers
187 in SImode, DImode and TImode */
188 {4, 8, 16}, /* cost of storing SSE registers
189 in SImode, DImode and TImode */
190 3, /* MMX or SSE register to integer */
191 0, /* size of l1 cache */
192 0, /* size of l2 cache */
193 0, /* size of prefetch block */
194 0, /* number of parallel prefetches */
196 COSTS_N_INSNS (23), /* cost of FADD and FSUB insns. */
197 COSTS_N_INSNS (27), /* cost of FMUL instruction. */
198 COSTS_N_INSNS (88), /* cost of FDIV instruction. */
199 COSTS_N_INSNS (22), /* cost of FABS instruction. */
200 COSTS_N_INSNS (24), /* cost of FCHS instruction. */
201 COSTS_N_INSNS (122), /* cost of FSQRT instruction. */
202 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
203 DUMMY_STRINGOP_ALGS},
204 {{rep_prefix_1_byte, {{-1, rep_prefix_1_byte}}},
205 DUMMY_STRINGOP_ALGS},
206 1, /* scalar_stmt_cost. */
207 1, /* scalar load_cost. */
208 1, /* scalar_store_cost. */
209 1, /* vec_stmt_cost. */
210 1, /* vec_to_scalar_cost. */
211 1, /* scalar_to_vec_cost. */
212 1, /* vec_align_load_cost. */
213 2, /* vec_unalign_load_cost. */
214 1, /* vec_store_cost. */
215 3, /* cond_taken_branch_cost. */
216 1, /* cond_not_taken_branch_cost. */
220 struct processor_costs i486_cost = { /* 486 specific costs */
221 COSTS_N_INSNS (1), /* cost of an add instruction */
222 COSTS_N_INSNS (1), /* cost of a lea instruction */
223 COSTS_N_INSNS (3), /* variable shift costs */
224 COSTS_N_INSNS (2), /* constant shift costs */
225 {COSTS_N_INSNS (12), /* cost of starting multiply for QI */
226 COSTS_N_INSNS (12), /* HI */
227 COSTS_N_INSNS (12), /* SI */
228 COSTS_N_INSNS (12), /* DI */
229 COSTS_N_INSNS (12)}, /* other */
230 1, /* cost of multiply per each bit set */
231 {COSTS_N_INSNS (40), /* cost of a divide/mod for QI */
232 COSTS_N_INSNS (40), /* HI */
233 COSTS_N_INSNS (40), /* SI */
234 COSTS_N_INSNS (40), /* DI */
235 COSTS_N_INSNS (40)}, /* other */
236 COSTS_N_INSNS (3), /* cost of movsx */
237 COSTS_N_INSNS (2), /* cost of movzx */
238 15, /* "large" insn */
240 4, /* cost for loading QImode using movzbl */
241 {2, 4, 2}, /* cost of loading integer registers
242 in QImode, HImode and SImode.
243 Relative to reg-reg move (2). */
244 {2, 4, 2}, /* cost of storing integer registers */
245 2, /* cost of reg,reg fld/fst */
246 {8, 8, 8}, /* cost of loading fp registers
247 in SFmode, DFmode and XFmode */
248 {8, 8, 8}, /* cost of storing fp registers
249 in SFmode, DFmode and XFmode */
250 2, /* cost of moving MMX register */
251 {4, 8}, /* cost of loading MMX registers
252 in SImode and DImode */
253 {4, 8}, /* cost of storing MMX registers
254 in SImode and DImode */
255 2, /* cost of moving SSE register */
256 {4, 8, 16}, /* cost of loading SSE registers
257 in SImode, DImode and TImode */
258 {4, 8, 16}, /* cost of storing SSE registers
259 in SImode, DImode and TImode */
260 3, /* MMX or SSE register to integer */
261 4, /* size of l1 cache. 486 has 8kB cache
262 shared for code and data, so 4kB is
263 not really precise. */
264 4, /* size of l2 cache */
265 0, /* size of prefetch block */
266 0, /* number of parallel prefetches */
268 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
269 COSTS_N_INSNS (16), /* cost of FMUL instruction. */
270 COSTS_N_INSNS (73), /* cost of FDIV instruction. */
271 COSTS_N_INSNS (3), /* cost of FABS instruction. */
272 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
273 COSTS_N_INSNS (83), /* cost of FSQRT instruction. */
274 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
275 DUMMY_STRINGOP_ALGS},
276 {{rep_prefix_4_byte, {{-1, rep_prefix_4_byte}}},
277 DUMMY_STRINGOP_ALGS},
278 1, /* scalar_stmt_cost. */
279 1, /* scalar load_cost. */
280 1, /* scalar_store_cost. */
281 1, /* vec_stmt_cost. */
282 1, /* vec_to_scalar_cost. */
283 1, /* scalar_to_vec_cost. */
284 1, /* vec_align_load_cost. */
285 2, /* vec_unalign_load_cost. */
286 1, /* vec_store_cost. */
287 3, /* cond_taken_branch_cost. */
288 1, /* cond_not_taken_branch_cost. */
292 struct processor_costs pentium_cost = {
293 COSTS_N_INSNS (1), /* cost of an add instruction */
294 COSTS_N_INSNS (1), /* cost of a lea instruction */
295 COSTS_N_INSNS (4), /* variable shift costs */
296 COSTS_N_INSNS (1), /* constant shift costs */
297 {COSTS_N_INSNS (11), /* cost of starting multiply for QI */
298 COSTS_N_INSNS (11), /* HI */
299 COSTS_N_INSNS (11), /* SI */
300 COSTS_N_INSNS (11), /* DI */
301 COSTS_N_INSNS (11)}, /* other */
302 0, /* cost of multiply per each bit set */
303 {COSTS_N_INSNS (25), /* cost of a divide/mod for QI */
304 COSTS_N_INSNS (25), /* HI */
305 COSTS_N_INSNS (25), /* SI */
306 COSTS_N_INSNS (25), /* DI */
307 COSTS_N_INSNS (25)}, /* other */
308 COSTS_N_INSNS (3), /* cost of movsx */
309 COSTS_N_INSNS (2), /* cost of movzx */
310 8, /* "large" insn */
312 6, /* cost for loading QImode using movzbl */
313 {2, 4, 2}, /* cost of loading integer registers
314 in QImode, HImode and SImode.
315 Relative to reg-reg move (2). */
316 {2, 4, 2}, /* cost of storing integer registers */
317 2, /* cost of reg,reg fld/fst */
318 {2, 2, 6}, /* cost of loading fp registers
319 in SFmode, DFmode and XFmode */
320 {4, 4, 6}, /* cost of storing fp registers
321 in SFmode, DFmode and XFmode */
322 8, /* cost of moving MMX register */
323 {8, 8}, /* cost of loading MMX registers
324 in SImode and DImode */
325 {8, 8}, /* cost of storing MMX registers
326 in SImode and DImode */
327 2, /* cost of moving SSE register */
328 {4, 8, 16}, /* cost of loading SSE registers
329 in SImode, DImode and TImode */
330 {4, 8, 16}, /* cost of storing SSE registers
331 in SImode, DImode and TImode */
332 3, /* MMX or SSE register to integer */
333 8, /* size of l1 cache. */
334 8, /* size of l2 cache */
335 0, /* size of prefetch block */
336 0, /* number of parallel prefetches */
338 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
339 COSTS_N_INSNS (3), /* cost of FMUL instruction. */
340 COSTS_N_INSNS (39), /* cost of FDIV instruction. */
341 COSTS_N_INSNS (1), /* cost of FABS instruction. */
342 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
343 COSTS_N_INSNS (70), /* cost of FSQRT instruction. */
344 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
345 DUMMY_STRINGOP_ALGS},
346 {{libcall, {{-1, rep_prefix_4_byte}}},
347 DUMMY_STRINGOP_ALGS},
348 1, /* scalar_stmt_cost. */
349 1, /* scalar load_cost. */
350 1, /* scalar_store_cost. */
351 1, /* vec_stmt_cost. */
352 1, /* vec_to_scalar_cost. */
353 1, /* scalar_to_vec_cost. */
354 1, /* vec_align_load_cost. */
355 2, /* vec_unalign_load_cost. */
356 1, /* vec_store_cost. */
357 3, /* cond_taken_branch_cost. */
358 1, /* cond_not_taken_branch_cost. */
362 struct processor_costs pentiumpro_cost = {
363 COSTS_N_INSNS (1), /* cost of an add instruction */
364 COSTS_N_INSNS (1), /* cost of a lea instruction */
365 COSTS_N_INSNS (1), /* variable shift costs */
366 COSTS_N_INSNS (1), /* constant shift costs */
367 {COSTS_N_INSNS (4), /* cost of starting multiply for QI */
368 COSTS_N_INSNS (4), /* HI */
369 COSTS_N_INSNS (4), /* SI */
370 COSTS_N_INSNS (4), /* DI */
371 COSTS_N_INSNS (4)}, /* other */
372 0, /* cost of multiply per each bit set */
373 {COSTS_N_INSNS (17), /* cost of a divide/mod for QI */
374 COSTS_N_INSNS (17), /* HI */
375 COSTS_N_INSNS (17), /* SI */
376 COSTS_N_INSNS (17), /* DI */
377 COSTS_N_INSNS (17)}, /* other */
378 COSTS_N_INSNS (1), /* cost of movsx */
379 COSTS_N_INSNS (1), /* cost of movzx */
380 8, /* "large" insn */
382 2, /* cost for loading QImode using movzbl */
383 {4, 4, 4}, /* cost of loading integer registers
384 in QImode, HImode and SImode.
385 Relative to reg-reg move (2). */
386 {2, 2, 2}, /* cost of storing integer registers */
387 2, /* cost of reg,reg fld/fst */
388 {2, 2, 6}, /* cost of loading fp registers
389 in SFmode, DFmode and XFmode */
390 {4, 4, 6}, /* cost of storing fp registers
391 in SFmode, DFmode and XFmode */
392 2, /* cost of moving MMX register */
393 {2, 2}, /* cost of loading MMX registers
394 in SImode and DImode */
395 {2, 2}, /* cost of storing MMX registers
396 in SImode and DImode */
397 2, /* cost of moving SSE register */
398 {2, 2, 8}, /* cost of loading SSE registers
399 in SImode, DImode and TImode */
400 {2, 2, 8}, /* cost of storing SSE registers
401 in SImode, DImode and TImode */
402 3, /* MMX or SSE register to integer */
403 8, /* size of l1 cache. */
404 256, /* size of l2 cache */
405 32, /* size of prefetch block */
406 6, /* number of parallel prefetches */
408 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
409 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
410 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
411 COSTS_N_INSNS (2), /* cost of FABS instruction. */
412 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
413 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
414 /* PentiumPro has optimized rep instructions for blocks aligned by 8 bytes (we ensure
415 the alignment). For small blocks inline loop is still a noticeable win, for bigger
416 blocks either rep movsl or rep movsb is way to go. Rep movsb has apparently
417 more expensive startup time in CPU, but after 4K the difference is down in the noise.
419 {{rep_prefix_4_byte, {{128, loop}, {1024, unrolled_loop},
420 {8192, rep_prefix_4_byte}, {-1, rep_prefix_1_byte}}},
421 DUMMY_STRINGOP_ALGS},
422 {{rep_prefix_4_byte, {{1024, unrolled_loop},
423 {8192, rep_prefix_4_byte}, {-1, libcall}}},
424 DUMMY_STRINGOP_ALGS},
425 1, /* scalar_stmt_cost. */
426 1, /* scalar load_cost. */
427 1, /* scalar_store_cost. */
428 1, /* vec_stmt_cost. */
429 1, /* vec_to_scalar_cost. */
430 1, /* scalar_to_vec_cost. */
431 1, /* vec_align_load_cost. */
432 2, /* vec_unalign_load_cost. */
433 1, /* vec_store_cost. */
434 3, /* cond_taken_branch_cost. */
435 1, /* cond_not_taken_branch_cost. */
439 struct processor_costs geode_cost = {
440 COSTS_N_INSNS (1), /* cost of an add instruction */
441 COSTS_N_INSNS (1), /* cost of a lea instruction */
442 COSTS_N_INSNS (2), /* variable shift costs */
443 COSTS_N_INSNS (1), /* constant shift costs */
444 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
445 COSTS_N_INSNS (4), /* HI */
446 COSTS_N_INSNS (7), /* SI */
447 COSTS_N_INSNS (7), /* DI */
448 COSTS_N_INSNS (7)}, /* other */
449 0, /* cost of multiply per each bit set */
450 {COSTS_N_INSNS (15), /* cost of a divide/mod for QI */
451 COSTS_N_INSNS (23), /* HI */
452 COSTS_N_INSNS (39), /* SI */
453 COSTS_N_INSNS (39), /* DI */
454 COSTS_N_INSNS (39)}, /* other */
455 COSTS_N_INSNS (1), /* cost of movsx */
456 COSTS_N_INSNS (1), /* cost of movzx */
457 8, /* "large" insn */
459 1, /* cost for loading QImode using movzbl */
460 {1, 1, 1}, /* cost of loading integer registers
461 in QImode, HImode and SImode.
462 Relative to reg-reg move (2). */
463 {1, 1, 1}, /* cost of storing integer registers */
464 1, /* cost of reg,reg fld/fst */
465 {1, 1, 1}, /* cost of loading fp registers
466 in SFmode, DFmode and XFmode */
467 {4, 6, 6}, /* cost of storing fp registers
468 in SFmode, DFmode and XFmode */
470 1, /* cost of moving MMX register */
471 {1, 1}, /* cost of loading MMX registers
472 in SImode and DImode */
473 {1, 1}, /* cost of storing MMX registers
474 in SImode and DImode */
475 1, /* cost of moving SSE register */
476 {1, 1, 1}, /* cost of loading SSE registers
477 in SImode, DImode and TImode */
478 {1, 1, 1}, /* cost of storing SSE registers
479 in SImode, DImode and TImode */
480 1, /* MMX or SSE register to integer */
481 64, /* size of l1 cache. */
482 128, /* size of l2 cache. */
483 32, /* size of prefetch block */
484 1, /* number of parallel prefetches */
486 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
487 COSTS_N_INSNS (11), /* cost of FMUL instruction. */
488 COSTS_N_INSNS (47), /* cost of FDIV instruction. */
489 COSTS_N_INSNS (1), /* cost of FABS instruction. */
490 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
491 COSTS_N_INSNS (54), /* cost of FSQRT instruction. */
492 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
493 DUMMY_STRINGOP_ALGS},
494 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
495 DUMMY_STRINGOP_ALGS},
496 1, /* scalar_stmt_cost. */
497 1, /* scalar load_cost. */
498 1, /* scalar_store_cost. */
499 1, /* vec_stmt_cost. */
500 1, /* vec_to_scalar_cost. */
501 1, /* scalar_to_vec_cost. */
502 1, /* vec_align_load_cost. */
503 2, /* vec_unalign_load_cost. */
504 1, /* vec_store_cost. */
505 3, /* cond_taken_branch_cost. */
506 1, /* cond_not_taken_branch_cost. */
510 struct processor_costs k6_cost = {
511 COSTS_N_INSNS (1), /* cost of an add instruction */
512 COSTS_N_INSNS (2), /* cost of a lea instruction */
513 COSTS_N_INSNS (1), /* variable shift costs */
514 COSTS_N_INSNS (1), /* constant shift costs */
515 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
516 COSTS_N_INSNS (3), /* HI */
517 COSTS_N_INSNS (3), /* SI */
518 COSTS_N_INSNS (3), /* DI */
519 COSTS_N_INSNS (3)}, /* other */
520 0, /* cost of multiply per each bit set */
521 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
522 COSTS_N_INSNS (18), /* HI */
523 COSTS_N_INSNS (18), /* SI */
524 COSTS_N_INSNS (18), /* DI */
525 COSTS_N_INSNS (18)}, /* other */
526 COSTS_N_INSNS (2), /* cost of movsx */
527 COSTS_N_INSNS (2), /* cost of movzx */
528 8, /* "large" insn */
530 3, /* cost for loading QImode using movzbl */
531 {4, 5, 4}, /* cost of loading integer registers
532 in QImode, HImode and SImode.
533 Relative to reg-reg move (2). */
534 {2, 3, 2}, /* cost of storing integer registers */
535 4, /* cost of reg,reg fld/fst */
536 {6, 6, 6}, /* cost of loading fp registers
537 in SFmode, DFmode and XFmode */
538 {4, 4, 4}, /* cost of storing fp registers
539 in SFmode, DFmode and XFmode */
540 2, /* cost of moving MMX register */
541 {2, 2}, /* cost of loading MMX registers
542 in SImode and DImode */
543 {2, 2}, /* cost of storing MMX registers
544 in SImode and DImode */
545 2, /* cost of moving SSE register */
546 {2, 2, 8}, /* cost of loading SSE registers
547 in SImode, DImode and TImode */
548 {2, 2, 8}, /* cost of storing SSE registers
549 in SImode, DImode and TImode */
550 6, /* MMX or SSE register to integer */
551 32, /* size of l1 cache. */
552 32, /* size of l2 cache. Some models
553 have integrated l2 cache, but
554 optimizing for k6 is not important
555 enough to worry about that. */
556 32, /* size of prefetch block */
557 1, /* number of parallel prefetches */
559 COSTS_N_INSNS (2), /* cost of FADD and FSUB insns. */
560 COSTS_N_INSNS (2), /* cost of FMUL instruction. */
561 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
562 COSTS_N_INSNS (2), /* cost of FABS instruction. */
563 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
564 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
565 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
566 DUMMY_STRINGOP_ALGS},
567 {{libcall, {{256, rep_prefix_4_byte}, {-1, libcall}}},
568 DUMMY_STRINGOP_ALGS},
569 1, /* scalar_stmt_cost. */
570 1, /* scalar load_cost. */
571 1, /* scalar_store_cost. */
572 1, /* vec_stmt_cost. */
573 1, /* vec_to_scalar_cost. */
574 1, /* scalar_to_vec_cost. */
575 1, /* vec_align_load_cost. */
576 2, /* vec_unalign_load_cost. */
577 1, /* vec_store_cost. */
578 3, /* cond_taken_branch_cost. */
579 1, /* cond_not_taken_branch_cost. */
583 struct processor_costs athlon_cost = {
584 COSTS_N_INSNS (1), /* cost of an add instruction */
585 COSTS_N_INSNS (2), /* cost of a lea instruction */
586 COSTS_N_INSNS (1), /* variable shift costs */
587 COSTS_N_INSNS (1), /* constant shift costs */
588 {COSTS_N_INSNS (5), /* cost of starting multiply for QI */
589 COSTS_N_INSNS (5), /* HI */
590 COSTS_N_INSNS (5), /* SI */
591 COSTS_N_INSNS (5), /* DI */
592 COSTS_N_INSNS (5)}, /* other */
593 0, /* cost of multiply per each bit set */
594 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
595 COSTS_N_INSNS (26), /* HI */
596 COSTS_N_INSNS (42), /* SI */
597 COSTS_N_INSNS (74), /* DI */
598 COSTS_N_INSNS (74)}, /* other */
599 COSTS_N_INSNS (1), /* cost of movsx */
600 COSTS_N_INSNS (1), /* cost of movzx */
601 8, /* "large" insn */
603 4, /* cost for loading QImode using movzbl */
604 {3, 4, 3}, /* cost of loading integer registers
605 in QImode, HImode and SImode.
606 Relative to reg-reg move (2). */
607 {3, 4, 3}, /* cost of storing integer registers */
608 4, /* cost of reg,reg fld/fst */
609 {4, 4, 12}, /* cost of loading fp registers
610 in SFmode, DFmode and XFmode */
611 {6, 6, 8}, /* cost of storing fp registers
612 in SFmode, DFmode and XFmode */
613 2, /* cost of moving MMX register */
614 {4, 4}, /* cost of loading MMX registers
615 in SImode and DImode */
616 {4, 4}, /* cost of storing MMX registers
617 in SImode and DImode */
618 2, /* cost of moving SSE register */
619 {4, 4, 6}, /* cost of loading SSE registers
620 in SImode, DImode and TImode */
621 {4, 4, 5}, /* cost of storing SSE registers
622 in SImode, DImode and TImode */
623 5, /* MMX or SSE register to integer */
624 64, /* size of l1 cache. */
625 256, /* size of l2 cache. */
626 64, /* size of prefetch block */
627 6, /* number of parallel prefetches */
629 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
630 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
631 COSTS_N_INSNS (24), /* cost of FDIV instruction. */
632 COSTS_N_INSNS (2), /* cost of FABS instruction. */
633 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
634 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
635 /* For some reason, Athlon deals better with REP prefix (relative to loops)
636 compared to K8. Alignment becomes important after 8 bytes for memcpy and
637 128 bytes for memset. */
638 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
639 DUMMY_STRINGOP_ALGS},
640 {{libcall, {{2048, rep_prefix_4_byte}, {-1, libcall}}},
641 DUMMY_STRINGOP_ALGS},
642 1, /* scalar_stmt_cost. */
643 1, /* scalar load_cost. */
644 1, /* scalar_store_cost. */
645 1, /* vec_stmt_cost. */
646 1, /* vec_to_scalar_cost. */
647 1, /* scalar_to_vec_cost. */
648 1, /* vec_align_load_cost. */
649 2, /* vec_unalign_load_cost. */
650 1, /* vec_store_cost. */
651 3, /* cond_taken_branch_cost. */
652 1, /* cond_not_taken_branch_cost. */
656 struct processor_costs k8_cost = {
657 COSTS_N_INSNS (1), /* cost of an add instruction */
658 COSTS_N_INSNS (2), /* cost of a lea instruction */
659 COSTS_N_INSNS (1), /* variable shift costs */
660 COSTS_N_INSNS (1), /* constant shift costs */
661 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
662 COSTS_N_INSNS (4), /* HI */
663 COSTS_N_INSNS (3), /* SI */
664 COSTS_N_INSNS (4), /* DI */
665 COSTS_N_INSNS (5)}, /* other */
666 0, /* cost of multiply per each bit set */
667 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
668 COSTS_N_INSNS (26), /* HI */
669 COSTS_N_INSNS (42), /* SI */
670 COSTS_N_INSNS (74), /* DI */
671 COSTS_N_INSNS (74)}, /* other */
672 COSTS_N_INSNS (1), /* cost of movsx */
673 COSTS_N_INSNS (1), /* cost of movzx */
674 8, /* "large" insn */
676 4, /* cost for loading QImode using movzbl */
677 {3, 4, 3}, /* cost of loading integer registers
678 in QImode, HImode and SImode.
679 Relative to reg-reg move (2). */
680 {3, 4, 3}, /* cost of storing integer registers */
681 4, /* cost of reg,reg fld/fst */
682 {4, 4, 12}, /* cost of loading fp registers
683 in SFmode, DFmode and XFmode */
684 {6, 6, 8}, /* cost of storing fp registers
685 in SFmode, DFmode and XFmode */
686 2, /* cost of moving MMX register */
687 {3, 3}, /* cost of loading MMX registers
688 in SImode and DImode */
689 {4, 4}, /* cost of storing MMX registers
690 in SImode and DImode */
691 2, /* cost of moving SSE register */
692 {4, 3, 6}, /* cost of loading SSE registers
693 in SImode, DImode and TImode */
694 {4, 4, 5}, /* cost of storing SSE registers
695 in SImode, DImode and TImode */
696 5, /* MMX or SSE register to integer */
697 64, /* size of l1 cache. */
698 512, /* size of l2 cache. */
699 64, /* size of prefetch block */
700 /* New AMD processors never drop prefetches; if they cannot be performed
701 immediately, they are queued. We set number of simultaneous prefetches
702 to a large constant to reflect this (it probably is not a good idea not
703 to limit number of prefetches at all, as their execution also takes some
705 100, /* number of parallel prefetches */
707 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
708 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
709 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
710 COSTS_N_INSNS (2), /* cost of FABS instruction. */
711 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
712 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
713 /* K8 has optimized REP instruction for medium sized blocks, but for very small
714 blocks it is better to use loop. For large blocks, libcall can do
715 nontemporary accesses and beat inline considerably. */
716 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
717 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
718 {{libcall, {{8, loop}, {24, unrolled_loop},
719 {2048, rep_prefix_4_byte}, {-1, libcall}}},
720 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
721 4, /* scalar_stmt_cost. */
722 2, /* scalar load_cost. */
723 2, /* scalar_store_cost. */
724 5, /* vec_stmt_cost. */
725 0, /* vec_to_scalar_cost. */
726 2, /* scalar_to_vec_cost. */
727 2, /* vec_align_load_cost. */
728 3, /* vec_unalign_load_cost. */
729 3, /* vec_store_cost. */
730 3, /* cond_taken_branch_cost. */
731 2, /* cond_not_taken_branch_cost. */
734 struct processor_costs amdfam10_cost = {
735 COSTS_N_INSNS (1), /* cost of an add instruction */
736 COSTS_N_INSNS (2), /* cost of a lea instruction */
737 COSTS_N_INSNS (1), /* variable shift costs */
738 COSTS_N_INSNS (1), /* constant shift costs */
739 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
740 COSTS_N_INSNS (4), /* HI */
741 COSTS_N_INSNS (3), /* SI */
742 COSTS_N_INSNS (4), /* DI */
743 COSTS_N_INSNS (5)}, /* other */
744 0, /* cost of multiply per each bit set */
745 {COSTS_N_INSNS (19), /* cost of a divide/mod for QI */
746 COSTS_N_INSNS (35), /* HI */
747 COSTS_N_INSNS (51), /* SI */
748 COSTS_N_INSNS (83), /* DI */
749 COSTS_N_INSNS (83)}, /* other */
750 COSTS_N_INSNS (1), /* cost of movsx */
751 COSTS_N_INSNS (1), /* cost of movzx */
752 8, /* "large" insn */
754 4, /* cost for loading QImode using movzbl */
755 {3, 4, 3}, /* cost of loading integer registers
756 in QImode, HImode and SImode.
757 Relative to reg-reg move (2). */
758 {3, 4, 3}, /* cost of storing integer registers */
759 4, /* cost of reg,reg fld/fst */
760 {4, 4, 12}, /* cost of loading fp registers
761 in SFmode, DFmode and XFmode */
762 {6, 6, 8}, /* cost of storing fp registers
763 in SFmode, DFmode and XFmode */
764 2, /* cost of moving MMX register */
765 {3, 3}, /* cost of loading MMX registers
766 in SImode and DImode */
767 {4, 4}, /* cost of storing MMX registers
768 in SImode and DImode */
769 2, /* cost of moving SSE register */
770 {4, 4, 3}, /* cost of loading SSE registers
771 in SImode, DImode and TImode */
772 {4, 4, 5}, /* cost of storing SSE registers
773 in SImode, DImode and TImode */
774 3, /* MMX or SSE register to integer */
776 MOVD reg64, xmmreg Double FSTORE 4
777 MOVD reg32, xmmreg Double FSTORE 4
779 MOVD reg64, xmmreg Double FADD 3
781 MOVD reg32, xmmreg Double FADD 3
783 64, /* size of l1 cache. */
784 512, /* size of l2 cache. */
785 64, /* size of prefetch block */
786 /* New AMD processors never drop prefetches; if they cannot be performed
787 immediately, they are queued. We set number of simultaneous prefetches
788 to a large constant to reflect this (it probably is not a good idea not
789 to limit number of prefetches at all, as their execution also takes some
791 100, /* number of parallel prefetches */
793 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
794 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
795 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
796 COSTS_N_INSNS (2), /* cost of FABS instruction. */
797 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
798 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
800 /* AMDFAM10 has optimized REP instruction for medium sized blocks, but for
801 very small blocks it is better to use loop. For large blocks, libcall can
802 do nontemporary accesses and beat inline considerably. */
803 {{libcall, {{6, loop}, {14, unrolled_loop}, {-1, rep_prefix_4_byte}}},
804 {libcall, {{16, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
805 {{libcall, {{8, loop}, {24, unrolled_loop},
806 {2048, rep_prefix_4_byte}, {-1, libcall}}},
807 {libcall, {{48, unrolled_loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
808 4, /* scalar_stmt_cost. */
809 2, /* scalar load_cost. */
810 2, /* scalar_store_cost. */
811 6, /* vec_stmt_cost. */
812 0, /* vec_to_scalar_cost. */
813 2, /* scalar_to_vec_cost. */
814 2, /* vec_align_load_cost. */
815 2, /* vec_unalign_load_cost. */
816 2, /* vec_store_cost. */
817 2, /* cond_taken_branch_cost. */
818 1, /* cond_not_taken_branch_cost. */
822 struct processor_costs pentium4_cost = {
823 COSTS_N_INSNS (1), /* cost of an add instruction */
824 COSTS_N_INSNS (3), /* cost of a lea instruction */
825 COSTS_N_INSNS (4), /* variable shift costs */
826 COSTS_N_INSNS (4), /* constant shift costs */
827 {COSTS_N_INSNS (15), /* cost of starting multiply for QI */
828 COSTS_N_INSNS (15), /* HI */
829 COSTS_N_INSNS (15), /* SI */
830 COSTS_N_INSNS (15), /* DI */
831 COSTS_N_INSNS (15)}, /* other */
832 0, /* cost of multiply per each bit set */
833 {COSTS_N_INSNS (56), /* cost of a divide/mod for QI */
834 COSTS_N_INSNS (56), /* HI */
835 COSTS_N_INSNS (56), /* SI */
836 COSTS_N_INSNS (56), /* DI */
837 COSTS_N_INSNS (56)}, /* other */
838 COSTS_N_INSNS (1), /* cost of movsx */
839 COSTS_N_INSNS (1), /* cost of movzx */
840 16, /* "large" insn */
842 2, /* cost for loading QImode using movzbl */
843 {4, 5, 4}, /* cost of loading integer registers
844 in QImode, HImode and SImode.
845 Relative to reg-reg move (2). */
846 {2, 3, 2}, /* cost of storing integer registers */
847 2, /* cost of reg,reg fld/fst */
848 {2, 2, 6}, /* cost of loading fp registers
849 in SFmode, DFmode and XFmode */
850 {4, 4, 6}, /* cost of storing fp registers
851 in SFmode, DFmode and XFmode */
852 2, /* cost of moving MMX register */
853 {2, 2}, /* cost of loading MMX registers
854 in SImode and DImode */
855 {2, 2}, /* cost of storing MMX registers
856 in SImode and DImode */
857 12, /* cost of moving SSE register */
858 {12, 12, 12}, /* cost of loading SSE registers
859 in SImode, DImode and TImode */
860 {2, 2, 8}, /* cost of storing SSE registers
861 in SImode, DImode and TImode */
862 10, /* MMX or SSE register to integer */
863 8, /* size of l1 cache. */
864 256, /* size of l2 cache. */
865 64, /* size of prefetch block */
866 6, /* number of parallel prefetches */
868 COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */
869 COSTS_N_INSNS (7), /* cost of FMUL instruction. */
870 COSTS_N_INSNS (43), /* cost of FDIV instruction. */
871 COSTS_N_INSNS (2), /* cost of FABS instruction. */
872 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
873 COSTS_N_INSNS (43), /* cost of FSQRT instruction. */
874 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
875 DUMMY_STRINGOP_ALGS},
876 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
878 DUMMY_STRINGOP_ALGS},
879 1, /* scalar_stmt_cost. */
880 1, /* scalar load_cost. */
881 1, /* scalar_store_cost. */
882 1, /* vec_stmt_cost. */
883 1, /* vec_to_scalar_cost. */
884 1, /* scalar_to_vec_cost. */
885 1, /* vec_align_load_cost. */
886 2, /* vec_unalign_load_cost. */
887 1, /* vec_store_cost. */
888 3, /* cond_taken_branch_cost. */
889 1, /* cond_not_taken_branch_cost. */
893 struct processor_costs nocona_cost = {
894 COSTS_N_INSNS (1), /* cost of an add instruction */
895 COSTS_N_INSNS (1), /* cost of a lea instruction */
896 COSTS_N_INSNS (1), /* variable shift costs */
897 COSTS_N_INSNS (1), /* constant shift costs */
898 {COSTS_N_INSNS (10), /* cost of starting multiply for QI */
899 COSTS_N_INSNS (10), /* HI */
900 COSTS_N_INSNS (10), /* SI */
901 COSTS_N_INSNS (10), /* DI */
902 COSTS_N_INSNS (10)}, /* other */
903 0, /* cost of multiply per each bit set */
904 {COSTS_N_INSNS (66), /* cost of a divide/mod for QI */
905 COSTS_N_INSNS (66), /* HI */
906 COSTS_N_INSNS (66), /* SI */
907 COSTS_N_INSNS (66), /* DI */
908 COSTS_N_INSNS (66)}, /* other */
909 COSTS_N_INSNS (1), /* cost of movsx */
910 COSTS_N_INSNS (1), /* cost of movzx */
911 16, /* "large" insn */
913 4, /* cost for loading QImode using movzbl */
914 {4, 4, 4}, /* cost of loading integer registers
915 in QImode, HImode and SImode.
916 Relative to reg-reg move (2). */
917 {4, 4, 4}, /* cost of storing integer registers */
918 3, /* cost of reg,reg fld/fst */
919 {12, 12, 12}, /* cost of loading fp registers
920 in SFmode, DFmode and XFmode */
921 {4, 4, 4}, /* cost of storing fp registers
922 in SFmode, DFmode and XFmode */
923 6, /* cost of moving MMX register */
924 {12, 12}, /* cost of loading MMX registers
925 in SImode and DImode */
926 {12, 12}, /* cost of storing MMX registers
927 in SImode and DImode */
928 6, /* cost of moving SSE register */
929 {12, 12, 12}, /* cost of loading SSE registers
930 in SImode, DImode and TImode */
931 {12, 12, 12}, /* cost of storing SSE registers
932 in SImode, DImode and TImode */
933 8, /* MMX or SSE register to integer */
934 8, /* size of l1 cache. */
935 1024, /* size of l2 cache. */
936 128, /* size of prefetch block */
937 8, /* number of parallel prefetches */
939 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
940 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
941 COSTS_N_INSNS (40), /* cost of FDIV instruction. */
942 COSTS_N_INSNS (3), /* cost of FABS instruction. */
943 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
944 COSTS_N_INSNS (44), /* cost of FSQRT instruction. */
945 {{libcall, {{12, loop_1_byte}, {-1, rep_prefix_4_byte}}},
946 {libcall, {{32, loop}, {20000, rep_prefix_8_byte},
947 {100000, unrolled_loop}, {-1, libcall}}}},
948 {{libcall, {{6, loop_1_byte}, {48, loop}, {20480, rep_prefix_4_byte},
950 {libcall, {{24, loop}, {64, unrolled_loop},
951 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
952 1, /* scalar_stmt_cost. */
953 1, /* scalar load_cost. */
954 1, /* scalar_store_cost. */
955 1, /* vec_stmt_cost. */
956 1, /* vec_to_scalar_cost. */
957 1, /* scalar_to_vec_cost. */
958 1, /* vec_align_load_cost. */
959 2, /* vec_unalign_load_cost. */
960 1, /* vec_store_cost. */
961 3, /* cond_taken_branch_cost. */
962 1, /* cond_not_taken_branch_cost. */
966 struct processor_costs core2_cost = {
967 COSTS_N_INSNS (1), /* cost of an add instruction */
968 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
969 COSTS_N_INSNS (1), /* variable shift costs */
970 COSTS_N_INSNS (1), /* constant shift costs */
971 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
972 COSTS_N_INSNS (3), /* HI */
973 COSTS_N_INSNS (3), /* SI */
974 COSTS_N_INSNS (3), /* DI */
975 COSTS_N_INSNS (3)}, /* other */
976 0, /* cost of multiply per each bit set */
977 {COSTS_N_INSNS (22), /* cost of a divide/mod for QI */
978 COSTS_N_INSNS (22), /* HI */
979 COSTS_N_INSNS (22), /* SI */
980 COSTS_N_INSNS (22), /* DI */
981 COSTS_N_INSNS (22)}, /* other */
982 COSTS_N_INSNS (1), /* cost of movsx */
983 COSTS_N_INSNS (1), /* cost of movzx */
984 8, /* "large" insn */
986 2, /* cost for loading QImode using movzbl */
987 {6, 6, 6}, /* cost of loading integer registers
988 in QImode, HImode and SImode.
989 Relative to reg-reg move (2). */
990 {4, 4, 4}, /* cost of storing integer registers */
991 2, /* cost of reg,reg fld/fst */
992 {6, 6, 6}, /* cost of loading fp registers
993 in SFmode, DFmode and XFmode */
994 {4, 4, 4}, /* cost of loading integer registers */
995 2, /* cost of moving MMX register */
996 {6, 6}, /* cost of loading MMX registers
997 in SImode and DImode */
998 {4, 4}, /* cost of storing MMX registers
999 in SImode and DImode */
1000 2, /* cost of moving SSE register */
1001 {6, 6, 6}, /* cost of loading SSE registers
1002 in SImode, DImode and TImode */
1003 {4, 4, 4}, /* cost of storing SSE registers
1004 in SImode, DImode and TImode */
1005 2, /* MMX or SSE register to integer */
1006 32, /* size of l1 cache. */
1007 2048, /* size of l2 cache. */
1008 128, /* size of prefetch block */
1009 8, /* number of parallel prefetches */
1010 3, /* Branch cost */
1011 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
1012 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
1013 COSTS_N_INSNS (32), /* cost of FDIV instruction. */
1014 COSTS_N_INSNS (1), /* cost of FABS instruction. */
1015 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
1016 COSTS_N_INSNS (58), /* cost of FSQRT instruction. */
1017 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1018 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1019 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1020 {{libcall, {{8, loop}, {15, unrolled_loop},
1021 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1022 {libcall, {{24, loop}, {32, unrolled_loop},
1023 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1024 1, /* scalar_stmt_cost. */
1025 1, /* scalar load_cost. */
1026 1, /* scalar_store_cost. */
1027 1, /* vec_stmt_cost. */
1028 1, /* vec_to_scalar_cost. */
1029 1, /* scalar_to_vec_cost. */
1030 1, /* vec_align_load_cost. */
1031 2, /* vec_unalign_load_cost. */
1032 1, /* vec_store_cost. */
1033 3, /* cond_taken_branch_cost. */
1034 1, /* cond_not_taken_branch_cost. */
1037 /* Generic64 should produce code tuned for Nocona and K8. */
1039 struct processor_costs generic64_cost = {
1040 COSTS_N_INSNS (1), /* cost of an add instruction */
1041 /* On all chips taken into consideration lea is 2 cycles and more. With
1042 this cost however our current implementation of synth_mult results in
1043 use of unnecessary temporary registers causing regression on several
1044 SPECfp benchmarks. */
1045 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1046 COSTS_N_INSNS (1), /* variable shift costs */
1047 COSTS_N_INSNS (1), /* constant shift costs */
1048 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1049 COSTS_N_INSNS (4), /* HI */
1050 COSTS_N_INSNS (3), /* SI */
1051 COSTS_N_INSNS (4), /* DI */
1052 COSTS_N_INSNS (2)}, /* other */
1053 0, /* cost of multiply per each bit set */
1054 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1055 COSTS_N_INSNS (26), /* HI */
1056 COSTS_N_INSNS (42), /* SI */
1057 COSTS_N_INSNS (74), /* DI */
1058 COSTS_N_INSNS (74)}, /* other */
1059 COSTS_N_INSNS (1), /* cost of movsx */
1060 COSTS_N_INSNS (1), /* cost of movzx */
1061 8, /* "large" insn */
1062 17, /* MOVE_RATIO */
1063 4, /* cost for loading QImode using movzbl */
1064 {4, 4, 4}, /* cost of loading integer registers
1065 in QImode, HImode and SImode.
1066 Relative to reg-reg move (2). */
1067 {4, 4, 4}, /* cost of storing integer registers */
1068 4, /* cost of reg,reg fld/fst */
1069 {12, 12, 12}, /* cost of loading fp registers
1070 in SFmode, DFmode and XFmode */
1071 {6, 6, 8}, /* cost of storing fp registers
1072 in SFmode, DFmode and XFmode */
1073 2, /* cost of moving MMX register */
1074 {8, 8}, /* cost of loading MMX registers
1075 in SImode and DImode */
1076 {8, 8}, /* cost of storing MMX registers
1077 in SImode and DImode */
1078 2, /* cost of moving SSE register */
1079 {8, 8, 8}, /* cost of loading SSE registers
1080 in SImode, DImode and TImode */
1081 {8, 8, 8}, /* cost of storing SSE registers
1082 in SImode, DImode and TImode */
1083 5, /* MMX or SSE register to integer */
1084 32, /* size of l1 cache. */
1085 512, /* size of l2 cache. */
1086 64, /* size of prefetch block */
1087 6, /* number of parallel prefetches */
1088 /* Benchmarks shows large regressions on K8 sixtrack benchmark when this value
1089 is increased to perhaps more appropriate value of 5. */
1090 3, /* Branch cost */
1091 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1092 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1093 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1094 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1095 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1096 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1097 {DUMMY_STRINGOP_ALGS,
1098 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1099 {DUMMY_STRINGOP_ALGS,
1100 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1101 1, /* scalar_stmt_cost. */
1102 1, /* scalar load_cost. */
1103 1, /* scalar_store_cost. */
1104 1, /* vec_stmt_cost. */
1105 1, /* vec_to_scalar_cost. */
1106 1, /* scalar_to_vec_cost. */
1107 1, /* vec_align_load_cost. */
1108 2, /* vec_unalign_load_cost. */
1109 1, /* vec_store_cost. */
1110 3, /* cond_taken_branch_cost. */
1111 1, /* cond_not_taken_branch_cost. */
1114 /* Generic32 should produce code tuned for Athlon, PPro, Pentium4, Nocona and K8. */
1116 struct processor_costs generic32_cost = {
1117 COSTS_N_INSNS (1), /* cost of an add instruction */
1118 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1119 COSTS_N_INSNS (1), /* variable shift costs */
1120 COSTS_N_INSNS (1), /* constant shift costs */
1121 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1122 COSTS_N_INSNS (4), /* HI */
1123 COSTS_N_INSNS (3), /* SI */
1124 COSTS_N_INSNS (4), /* DI */
1125 COSTS_N_INSNS (2)}, /* other */
1126 0, /* cost of multiply per each bit set */
1127 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1128 COSTS_N_INSNS (26), /* HI */
1129 COSTS_N_INSNS (42), /* SI */
1130 COSTS_N_INSNS (74), /* DI */
1131 COSTS_N_INSNS (74)}, /* other */
1132 COSTS_N_INSNS (1), /* cost of movsx */
1133 COSTS_N_INSNS (1), /* cost of movzx */
1134 8, /* "large" insn */
1135 17, /* MOVE_RATIO */
1136 4, /* cost for loading QImode using movzbl */
1137 {4, 4, 4}, /* cost of loading integer registers
1138 in QImode, HImode and SImode.
1139 Relative to reg-reg move (2). */
1140 {4, 4, 4}, /* cost of storing integer registers */
1141 4, /* cost of reg,reg fld/fst */
1142 {12, 12, 12}, /* cost of loading fp registers
1143 in SFmode, DFmode and XFmode */
1144 {6, 6, 8}, /* cost of storing fp registers
1145 in SFmode, DFmode and XFmode */
1146 2, /* cost of moving MMX register */
1147 {8, 8}, /* cost of loading MMX registers
1148 in SImode and DImode */
1149 {8, 8}, /* cost of storing MMX registers
1150 in SImode and DImode */
1151 2, /* cost of moving SSE register */
1152 {8, 8, 8}, /* cost of loading SSE registers
1153 in SImode, DImode and TImode */
1154 {8, 8, 8}, /* cost of storing SSE registers
1155 in SImode, DImode and TImode */
1156 5, /* MMX or SSE register to integer */
1157 32, /* size of l1 cache. */
1158 256, /* size of l2 cache. */
1159 64, /* size of prefetch block */
1160 6, /* number of parallel prefetches */
1161 3, /* Branch cost */
1162 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1163 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1164 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1165 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1166 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1167 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1168 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1169 DUMMY_STRINGOP_ALGS},
1170 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1171 DUMMY_STRINGOP_ALGS},
1172 1, /* scalar_stmt_cost. */
1173 1, /* scalar load_cost. */
1174 1, /* scalar_store_cost. */
1175 1, /* vec_stmt_cost. */
1176 1, /* vec_to_scalar_cost. */
1177 1, /* scalar_to_vec_cost. */
1178 1, /* vec_align_load_cost. */
1179 2, /* vec_unalign_load_cost. */
1180 1, /* vec_store_cost. */
1181 3, /* cond_taken_branch_cost. */
1182 1, /* cond_not_taken_branch_cost. */
1185 const struct processor_costs *ix86_cost = &pentium_cost;
1187 /* Processor feature/optimization bitmasks. */
1188 #define m_386 (1<<PROCESSOR_I386)
1189 #define m_486 (1<<PROCESSOR_I486)
1190 #define m_PENT (1<<PROCESSOR_PENTIUM)
1191 #define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
1192 #define m_PENT4 (1<<PROCESSOR_PENTIUM4)
1193 #define m_NOCONA (1<<PROCESSOR_NOCONA)
1194 #define m_CORE2 (1<<PROCESSOR_CORE2)
1196 #define m_GEODE (1<<PROCESSOR_GEODE)
1197 #define m_K6 (1<<PROCESSOR_K6)
1198 #define m_K6_GEODE (m_K6 | m_GEODE)
1199 #define m_K8 (1<<PROCESSOR_K8)
1200 #define m_ATHLON (1<<PROCESSOR_ATHLON)
1201 #define m_ATHLON_K8 (m_K8 | m_ATHLON)
1202 #define m_AMDFAM10 (1<<PROCESSOR_AMDFAM10)
1203 #define m_AMD_MULTIPLE (m_K8 | m_ATHLON | m_AMDFAM10)
1205 #define m_GENERIC32 (1<<PROCESSOR_GENERIC32)
1206 #define m_GENERIC64 (1<<PROCESSOR_GENERIC64)
1208 /* Generic instruction choice should be common subset of supported CPUs
1209 (PPro/PENT4/NOCONA/CORE2/Athlon/K8). */
1210 #define m_GENERIC (m_GENERIC32 | m_GENERIC64)
1212 /* Feature tests against the various tunings. */
1213 unsigned int ix86_tune_features[X86_TUNE_LAST] = {
1214 /* X86_TUNE_USE_LEAVE: Leave does not affect Nocona SPEC2000 results
1215 negatively, so enabling for Generic64 seems like good code size
1216 tradeoff. We can't enable it for 32bit generic because it does not
1217 work well with PPro base chips. */
1218 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_CORE2 | m_GENERIC64,
1220 /* X86_TUNE_PUSH_MEMORY */
1221 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4
1222 | m_NOCONA | m_CORE2 | m_GENERIC,
1224 /* X86_TUNE_ZERO_EXTEND_WITH_AND */
1227 /* X86_TUNE_USE_BIT_TEST */
1230 /* X86_TUNE_UNROLL_STRLEN */
1231 m_486 | m_PENT | m_PPRO | m_AMD_MULTIPLE | m_K6 | m_CORE2 | m_GENERIC,
1233 /* X86_TUNE_DEEP_BRANCH_PREDICTION */
1234 m_PPRO | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4 | m_GENERIC,
1236 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
1237 on simulation result. But after P4 was made, no performance benefit
1238 was observed with branch hints. It also increases the code size.
1239 As a result, icc never generates branch hints. */
1242 /* X86_TUNE_DOUBLE_WITH_ADD */
1245 /* X86_TUNE_USE_SAHF */
1246 m_PPRO | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_PENT4
1247 | m_NOCONA | m_CORE2 | m_GENERIC,
1249 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
1250 partial dependencies. */
1251 m_AMD_MULTIPLE | m_PPRO | m_PENT4 | m_NOCONA
1252 | m_CORE2 | m_GENERIC | m_GEODE /* m_386 | m_K6 */,
1254 /* X86_TUNE_PARTIAL_REG_STALL: We probably ought to watch for partial
1255 register stalls on Generic32 compilation setting as well. However
1256 in current implementation the partial register stalls are not eliminated
1257 very well - they can be introduced via subregs synthesized by combine
1258 and can happen in caller/callee saving sequences. Because this option
1259 pays back little on PPro based chips and is in conflict with partial reg
1260 dependencies used by Athlon/P4 based chips, it is better to leave it off
1261 for generic32 for now. */
1264 /* X86_TUNE_PARTIAL_FLAG_REG_STALL */
1265 m_CORE2 | m_GENERIC,
1267 /* X86_TUNE_USE_HIMODE_FIOP */
1268 m_386 | m_486 | m_K6_GEODE,
1270 /* X86_TUNE_USE_SIMODE_FIOP */
1271 ~(m_PPRO | m_AMD_MULTIPLE | m_PENT | m_CORE2 | m_GENERIC),
1273 /* X86_TUNE_USE_MOV0 */
1276 /* X86_TUNE_USE_CLTD */
1277 ~(m_PENT | m_K6 | m_CORE2 | m_GENERIC),
1279 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
1282 /* X86_TUNE_SPLIT_LONG_MOVES */
1285 /* X86_TUNE_READ_MODIFY_WRITE */
1288 /* X86_TUNE_READ_MODIFY */
1291 /* X86_TUNE_PROMOTE_QIMODE */
1292 m_K6_GEODE | m_PENT | m_386 | m_486 | m_AMD_MULTIPLE | m_CORE2
1293 | m_GENERIC /* | m_PENT4 ? */,
1295 /* X86_TUNE_FAST_PREFIX */
1296 ~(m_PENT | m_486 | m_386),
1298 /* X86_TUNE_SINGLE_STRINGOP */
1299 m_386 | m_PENT4 | m_NOCONA,
1301 /* X86_TUNE_QIMODE_MATH */
1304 /* X86_TUNE_HIMODE_MATH: On PPro this flag is meant to avoid partial
1305 register stalls. Just like X86_TUNE_PARTIAL_REG_STALL this option
1306 might be considered for Generic32 if our scheme for avoiding partial
1307 stalls was more effective. */
1310 /* X86_TUNE_PROMOTE_QI_REGS */
1313 /* X86_TUNE_PROMOTE_HI_REGS */
1316 /* X86_TUNE_ADD_ESP_4: Enable if add/sub is preferred over 1/2 push/pop. */
1317 m_AMD_MULTIPLE | m_K6_GEODE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1319 /* X86_TUNE_ADD_ESP_8 */
1320 m_AMD_MULTIPLE | m_PPRO | m_K6_GEODE | m_386
1321 | m_486 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1323 /* X86_TUNE_SUB_ESP_4 */
1324 m_AMD_MULTIPLE | m_PPRO | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1326 /* X86_TUNE_SUB_ESP_8 */
1327 m_AMD_MULTIPLE | m_PPRO | m_386 | m_486
1328 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1330 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
1331 for DFmode copies */
1332 ~(m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1333 | m_GENERIC | m_GEODE),
1335 /* X86_TUNE_PARTIAL_REG_DEPENDENCY */
1336 m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1338 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: In the Generic model we have a
1339 conflict here in between PPro/Pentium4 based chips that thread 128bit
1340 SSE registers as single units versus K8 based chips that divide SSE
1341 registers to two 64bit halves. This knob promotes all store destinations
1342 to be 128bit to allow register renaming on 128bit SSE units, but usually
1343 results in one extra microop on 64bit SSE units. Experimental results
1344 shows that disabling this option on P4 brings over 20% SPECfp regression,
1345 while enabling it on K8 brings roughly 2.4% regression that can be partly
1346 masked by careful scheduling of moves. */
1347 m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC | m_AMDFAM10,
1349 /* X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL */
1352 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
1353 are resolved on SSE register parts instead of whole registers, so we may
1354 maintain just lower part of scalar values in proper format leaving the
1355 upper part undefined. */
1358 /* X86_TUNE_SSE_TYPELESS_STORES */
1361 /* X86_TUNE_SSE_LOAD0_BY_PXOR */
1362 m_PPRO | m_PENT4 | m_NOCONA,
1364 /* X86_TUNE_MEMORY_MISMATCH_STALL */
1365 m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1367 /* X86_TUNE_PROLOGUE_USING_MOVE */
1368 m_ATHLON_K8 | m_PPRO | m_CORE2 | m_GENERIC,
1370 /* X86_TUNE_EPILOGUE_USING_MOVE */
1371 m_ATHLON_K8 | m_PPRO | m_CORE2 | m_GENERIC,
1373 /* X86_TUNE_SHIFT1 */
1376 /* X86_TUNE_USE_FFREEP */
1379 /* X86_TUNE_INTER_UNIT_MOVES */
1380 ~(m_AMD_MULTIPLE | m_GENERIC),
1382 /* X86_TUNE_INTER_UNIT_CONVERSIONS */
1385 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
1386 than 4 branch instructions in the 16 byte window. */
1387 m_PPRO | m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1389 /* X86_TUNE_SCHEDULE */
1390 m_PPRO | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT | m_CORE2 | m_GENERIC,
1392 /* X86_TUNE_USE_BT */
1395 /* X86_TUNE_USE_INCDEC */
1396 ~(m_PENT4 | m_NOCONA | m_GENERIC),
1398 /* X86_TUNE_PAD_RETURNS */
1399 m_AMD_MULTIPLE | m_CORE2 | m_GENERIC,
1401 /* X86_TUNE_EXT_80387_CONSTANTS */
1402 m_K6_GEODE | m_ATHLON_K8 | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC,
1404 /* X86_TUNE_SHORTEN_X87_SSE */
1407 /* X86_TUNE_AVOID_VECTOR_DECODE */
1410 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
1411 and SImode multiply, but 386 and 486 do HImode multiply faster. */
1414 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
1415 vector path on AMD machines. */
1416 m_K8 | m_GENERIC64 | m_AMDFAM10,
1418 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
1420 m_K8 | m_GENERIC64 | m_AMDFAM10,
1422 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
1426 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
1427 but one byte longer. */
1430 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
1431 operand that cannot be represented using a modRM byte. The XOR
1432 replacement is long decoded, so this split helps here as well. */
1435 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
1436 from integer to FP. */
1440 /* Feature tests against the various architecture variations. */
1441 unsigned int ix86_arch_features[X86_ARCH_LAST] = {
1442 /* X86_ARCH_CMOVE: Conditional move was added for pentiumpro. */
1443 ~(m_386 | m_486 | m_PENT | m_K6),
1445 /* X86_ARCH_CMPXCHG: Compare and exchange was added for 80486. */
1448 /* X86_ARCH_CMPXCHG8B: Compare and exchange 8 bytes was added for pentium. */
1451 /* X86_ARCH_XADD: Exchange and add was added for 80486. */
1454 /* X86_ARCH_BSWAP: Byteswap was added for 80486. */
1458 static const unsigned int x86_accumulate_outgoing_args
1459 = m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC;
1461 static const unsigned int x86_arch_always_fancy_math_387
1462 = m_PENT | m_PPRO | m_AMD_MULTIPLE | m_PENT4
1463 | m_NOCONA | m_CORE2 | m_GENERIC;
1465 static enum stringop_alg stringop_alg = no_stringop;
1467 /* In case the average insn count for single function invocation is
1468 lower than this constant, emit fast (but longer) prologue and
1470 #define FAST_PROLOGUE_INSN_COUNT 20
1472 /* Names for 8 (low), 8 (high), and 16-bit registers, respectively. */
1473 static const char *const qi_reg_name[] = QI_REGISTER_NAMES;
1474 static const char *const qi_high_reg_name[] = QI_HIGH_REGISTER_NAMES;
1475 static const char *const hi_reg_name[] = HI_REGISTER_NAMES;
1477 /* Array of the smallest class containing reg number REGNO, indexed by
1478 REGNO. Used by REGNO_REG_CLASS in i386.h. */
1480 enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER] =
1482 /* ax, dx, cx, bx */
1483 AREG, DREG, CREG, BREG,
1484 /* si, di, bp, sp */
1485 SIREG, DIREG, NON_Q_REGS, NON_Q_REGS,
1487 FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS,
1488 FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
1491 /* flags, fpsr, fpcr, frame */
1492 NO_REGS, NO_REGS, NO_REGS, NON_Q_REGS,
1494 SSE_FIRST_REG, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1497 MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS,
1500 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1501 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1502 /* SSE REX registers */
1503 SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1507 /* The "default" register map used in 32bit mode. */
1509 int const dbx_register_map[FIRST_PSEUDO_REGISTER] =
1511 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */
1512 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */
1513 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1514 21, 22, 23, 24, 25, 26, 27, 28, /* SSE */
1515 29, 30, 31, 32, 33, 34, 35, 36, /* MMX */
1516 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1517 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1520 static int const x86_64_int_parameter_registers[6] =
1522 5 /*RDI*/, 4 /*RSI*/, 1 /*RDX*/, 2 /*RCX*/,
1523 FIRST_REX_INT_REG /*R8 */, FIRST_REX_INT_REG + 1 /*R9 */
1526 static int const x86_64_ms_abi_int_parameter_registers[4] =
1528 2 /*RCX*/, 1 /*RDX*/,
1529 FIRST_REX_INT_REG /*R8 */, FIRST_REX_INT_REG + 1 /*R9 */
1532 static int const x86_64_int_return_registers[4] =
1534 0 /*RAX*/, 1 /*RDX*/, 5 /*RDI*/, 4 /*RSI*/
1537 /* The "default" register map used in 64bit mode. */
1538 int const dbx64_register_map[FIRST_PSEUDO_REGISTER] =
1540 0, 1, 2, 3, 4, 5, 6, 7, /* general regs */
1541 33, 34, 35, 36, 37, 38, 39, 40, /* fp regs */
1542 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1543 17, 18, 19, 20, 21, 22, 23, 24, /* SSE */
1544 41, 42, 43, 44, 45, 46, 47, 48, /* MMX */
1545 8,9,10,11,12,13,14,15, /* extended integer registers */
1546 25, 26, 27, 28, 29, 30, 31, 32, /* extended SSE registers */
1549 /* Define the register numbers to be used in Dwarf debugging information.
1550 The SVR4 reference port C compiler uses the following register numbers
1551 in its Dwarf output code:
1552 0 for %eax (gcc regno = 0)
1553 1 for %ecx (gcc regno = 2)
1554 2 for %edx (gcc regno = 1)
1555 3 for %ebx (gcc regno = 3)
1556 4 for %esp (gcc regno = 7)
1557 5 for %ebp (gcc regno = 6)
1558 6 for %esi (gcc regno = 4)
1559 7 for %edi (gcc regno = 5)
1560 The following three DWARF register numbers are never generated by
1561 the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4
1562 believes these numbers have these meanings.
1563 8 for %eip (no gcc equivalent)
1564 9 for %eflags (gcc regno = 17)
1565 10 for %trapno (no gcc equivalent)
1566 It is not at all clear how we should number the FP stack registers
1567 for the x86 architecture. If the version of SDB on x86/svr4 were
1568 a bit less brain dead with respect to floating-point then we would
1569 have a precedent to follow with respect to DWARF register numbers
1570 for x86 FP registers, but the SDB on x86/svr4 is so completely
1571 broken with respect to FP registers that it is hardly worth thinking
1572 of it as something to strive for compatibility with.
1573 The version of x86/svr4 SDB I have at the moment does (partially)
1574 seem to believe that DWARF register number 11 is associated with
1575 the x86 register %st(0), but that's about all. Higher DWARF
1576 register numbers don't seem to be associated with anything in
1577 particular, and even for DWARF regno 11, SDB only seems to under-
1578 stand that it should say that a variable lives in %st(0) (when
1579 asked via an `=' command) if we said it was in DWARF regno 11,
1580 but SDB still prints garbage when asked for the value of the
1581 variable in question (via a `/' command).
1582 (Also note that the labels SDB prints for various FP stack regs
1583 when doing an `x' command are all wrong.)
1584 Note that these problems generally don't affect the native SVR4
1585 C compiler because it doesn't allow the use of -O with -g and
1586 because when it is *not* optimizing, it allocates a memory
1587 location for each floating-point variable, and the memory
1588 location is what gets described in the DWARF AT_location
1589 attribute for the variable in question.
1590 Regardless of the severe mental illness of the x86/svr4 SDB, we
1591 do something sensible here and we use the following DWARF
1592 register numbers. Note that these are all stack-top-relative
1594 11 for %st(0) (gcc regno = 8)
1595 12 for %st(1) (gcc regno = 9)
1596 13 for %st(2) (gcc regno = 10)
1597 14 for %st(3) (gcc regno = 11)
1598 15 for %st(4) (gcc regno = 12)
1599 16 for %st(5) (gcc regno = 13)
1600 17 for %st(6) (gcc regno = 14)
1601 18 for %st(7) (gcc regno = 15)
1603 int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER] =
1605 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */
1606 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */
1607 -1, 9, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1608 21, 22, 23, 24, 25, 26, 27, 28, /* SSE registers */
1609 29, 30, 31, 32, 33, 34, 35, 36, /* MMX registers */
1610 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1611 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1614 /* Test and compare insns in i386.md store the information needed to
1615 generate branch and scc insns here. */
1617 rtx ix86_compare_op0 = NULL_RTX;
1618 rtx ix86_compare_op1 = NULL_RTX;
1619 rtx ix86_compare_emitted = NULL_RTX;
1621 /* Size of the register save area. */
1622 #define X86_64_VARARGS_SIZE (X86_64_REGPARM_MAX * UNITS_PER_WORD + X86_64_SSE_REGPARM_MAX * 16)
1624 /* Define the structure for the machine field in struct function. */
1626 struct stack_local_entry GTY(())
1628 unsigned short mode;
1631 struct stack_local_entry *next;
1634 /* Structure describing stack frame layout.
1635 Stack grows downward:
1641 saved frame pointer if frame_pointer_needed
1642 <- HARD_FRAME_POINTER
1647 [va_arg registers] (
1648 > to_allocate <- FRAME_POINTER
1658 HOST_WIDE_INT frame;
1660 int outgoing_arguments_size;
1663 HOST_WIDE_INT to_allocate;
1664 /* The offsets relative to ARG_POINTER. */
1665 HOST_WIDE_INT frame_pointer_offset;
1666 HOST_WIDE_INT hard_frame_pointer_offset;
1667 HOST_WIDE_INT stack_pointer_offset;
1669 /* When save_regs_using_mov is set, emit prologue using
1670 move instead of push instructions. */
1671 bool save_regs_using_mov;
1674 /* Code model option. */
1675 enum cmodel ix86_cmodel;
1677 enum asm_dialect ix86_asm_dialect = ASM_ATT;
1679 enum tls_dialect ix86_tls_dialect = TLS_DIALECT_GNU;
1681 /* Which unit we are generating floating point math for. */
1682 enum fpmath_unit ix86_fpmath;
1684 /* Which cpu are we scheduling for. */
1685 enum processor_type ix86_tune;
1687 /* Which instruction set architecture to use. */
1688 enum processor_type ix86_arch;
1690 /* true if sse prefetch instruction is not NOOP. */
1691 int x86_prefetch_sse;
1693 /* ix86_regparm_string as a number */
1694 static int ix86_regparm;
1696 /* -mstackrealign option */
1697 extern int ix86_force_align_arg_pointer;
1698 static const char ix86_force_align_arg_pointer_string[] = "force_align_arg_pointer";
1700 /* Preferred alignment for stack boundary in bits. */
1701 unsigned int ix86_preferred_stack_boundary;
1703 /* Values 1-5: see jump.c */
1704 int ix86_branch_cost;
1706 /* Variables which are this size or smaller are put in the data/bss
1707 or ldata/lbss sections. */
1709 int ix86_section_threshold = 65536;
1711 /* Prefix built by ASM_GENERATE_INTERNAL_LABEL. */
1712 char internal_label_prefix[16];
1713 int internal_label_prefix_len;
1715 /* Fence to use after loop using movnt. */
1718 /* Register class used for passing given 64bit part of the argument.
1719 These represent classes as documented by the PS ABI, with the exception
1720 of SSESF, SSEDF classes, that are basically SSE class, just gcc will
1721 use SF or DFmode move instead of DImode to avoid reformatting penalties.
1723 Similarly we play games with INTEGERSI_CLASS to use cheaper SImode moves
1724 whenever possible (upper half does contain padding). */
1725 enum x86_64_reg_class
1728 X86_64_INTEGER_CLASS,
1729 X86_64_INTEGERSI_CLASS,
1736 X86_64_COMPLEX_X87_CLASS,
1739 static const char * const x86_64_reg_class_name[] =
1741 "no", "integer", "integerSI", "sse", "sseSF", "sseDF",
1742 "sseup", "x87", "x87up", "cplx87", "no"
1745 #define MAX_CLASSES 4
1747 /* Table of constants used by fldpi, fldln2, etc.... */
1748 static REAL_VALUE_TYPE ext_80387_constants_table [5];
1749 static bool ext_80387_constants_init = 0;
1752 static struct machine_function * ix86_init_machine_status (void);
1753 static rtx ix86_function_value (const_tree, const_tree, bool);
1754 static int ix86_function_regparm (const_tree, const_tree);
1755 static void ix86_compute_frame_layout (struct ix86_frame *);
1756 static bool ix86_expand_vector_init_one_nonzero (bool, enum machine_mode,
1760 /* The svr4 ABI for the i386 says that records and unions are returned
1762 #ifndef DEFAULT_PCC_STRUCT_RETURN
1763 #define DEFAULT_PCC_STRUCT_RETURN 1
1766 /* Bit flags that specify the ISA we are compiling for. */
1767 int ix86_isa_flags = TARGET_64BIT_DEFAULT | TARGET_SUBTARGET_ISA_DEFAULT;
1769 /* A mask of ix86_isa_flags that includes bit X if X
1770 was set or cleared on the command line. */
1771 static int ix86_isa_flags_explicit;
1773 /* Define a set of ISAs which are available when a given ISA is
1774 enabled. MMX and SSE ISAs are handled separately. */
1776 #define OPTION_MASK_ISA_MMX_SET OPTION_MASK_ISA_MMX
1777 #define OPTION_MASK_ISA_3DNOW_SET \
1778 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_MMX_SET)
1780 #define OPTION_MASK_ISA_SSE_SET OPTION_MASK_ISA_SSE
1781 #define OPTION_MASK_ISA_SSE2_SET \
1782 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE_SET)
1783 #define OPTION_MASK_ISA_SSE3_SET \
1784 (OPTION_MASK_ISA_SSE3 | OPTION_MASK_ISA_SSE2_SET)
1785 #define OPTION_MASK_ISA_SSSE3_SET \
1786 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE3_SET)
1787 #define OPTION_MASK_ISA_SSE4_1_SET \
1788 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSSE3_SET)
1789 #define OPTION_MASK_ISA_SSE4_2_SET \
1790 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_SSE4_1_SET)
1792 /* SSE4 includes both SSE4.1 and SSE4.2. -msse4 should be the same
1794 #define OPTION_MASK_ISA_SSE4_SET OPTION_MASK_ISA_SSE4_2_SET
1796 #define OPTION_MASK_ISA_SSE4A_SET \
1797 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE3_SET)
1798 #define OPTION_MASK_ISA_SSE5_SET \
1799 (OPTION_MASK_ISA_SSE5 | OPTION_MASK_ISA_SSE4A_SET)
1801 /* Define a set of ISAs which aren't available when a given ISA is
1802 disabled. MMX and SSE ISAs are handled separately. */
1804 #define OPTION_MASK_ISA_MMX_UNSET \
1805 (OPTION_MASK_ISA_MMX | OPTION_MASK_ISA_3DNOW_UNSET)
1806 #define OPTION_MASK_ISA_3DNOW_UNSET \
1807 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_3DNOW_A_UNSET)
1808 #define OPTION_MASK_ISA_3DNOW_A_UNSET OPTION_MASK_ISA_3DNOW_A
1810 #define OPTION_MASK_ISA_SSE_UNSET \
1811 (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_SSE2_UNSET)
1812 #define OPTION_MASK_ISA_SSE2_UNSET \
1813 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE3_UNSET)
1814 #define OPTION_MASK_ISA_SSE3_UNSET \
1815 (OPTION_MASK_ISA_SSE3 \
1816 | OPTION_MASK_ISA_SSSE3_UNSET \
1817 | OPTION_MASK_ISA_SSE4A_UNSET )
1818 #define OPTION_MASK_ISA_SSSE3_UNSET \
1819 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE4_1_UNSET)
1820 #define OPTION_MASK_ISA_SSE4_1_UNSET \
1821 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSE4_2_UNSET)
1822 #define OPTION_MASK_ISA_SSE4_2_UNSET OPTION_MASK_ISA_SSE4_2
1824 /* SSE4 includes both SSE4.1 and SSE4.2. -mno-sse4 should the same
1826 #define OPTION_MASK_ISA_SSE4_UNSET OPTION_MASK_ISA_SSE4_1_UNSET
1828 #define OPTION_MASK_ISA_SSE4A_UNSET \
1829 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE5_UNSET)
1831 #define OPTION_MASK_ISA_SSE5_UNSET OPTION_MASK_ISA_SSE5
1833 /* Vectorization library interface and handlers. */
1834 tree (*ix86_veclib_handler)(enum built_in_function, tree, tree) = NULL;
1835 static tree ix86_veclibabi_svml (enum built_in_function, tree, tree);
1836 static tree ix86_veclibabi_acml (enum built_in_function, tree, tree);
1838 /* Implement TARGET_HANDLE_OPTION. */
1841 ix86_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, int value)
1848 ix86_isa_flags |= OPTION_MASK_ISA_MMX_SET;
1849 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_SET;
1853 ix86_isa_flags &= ~OPTION_MASK_ISA_MMX_UNSET;
1854 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_UNSET;
1861 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_SET;
1862 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_SET;
1866 ix86_isa_flags &= ~OPTION_MASK_ISA_3DNOW_UNSET;
1867 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_UNSET;
1877 ix86_isa_flags |= OPTION_MASK_ISA_SSE_SET;
1878 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_SET;
1882 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE_UNSET;
1883 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_UNSET;
1890 ix86_isa_flags |= OPTION_MASK_ISA_SSE2_SET;
1891 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_SET;
1895 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE2_UNSET;
1896 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_UNSET;
1903 ix86_isa_flags |= OPTION_MASK_ISA_SSE3_SET;
1904 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_SET;
1908 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE3_UNSET;
1909 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_UNSET;
1916 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3_SET;
1917 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_SET;
1921 ix86_isa_flags &= ~OPTION_MASK_ISA_SSSE3_UNSET;
1922 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_UNSET;
1929 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1_SET;
1930 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_SET;
1934 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_1_UNSET;
1935 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_UNSET;
1942 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2_SET;
1943 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_SET;
1947 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_2_UNSET;
1948 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_UNSET;
1953 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
1954 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
1958 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
1959 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
1965 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A_SET;
1966 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_SET;
1970 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4A_UNSET;
1971 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_UNSET;
1978 ix86_isa_flags |= OPTION_MASK_ISA_SSE5_SET;
1979 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_SET;
1983 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE5_UNSET;
1984 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_UNSET;
1993 /* Sometimes certain combinations of command options do not make
1994 sense on a particular target machine. You can define a macro
1995 `OVERRIDE_OPTIONS' to take account of this. This macro, if
1996 defined, is executed once just after all the command options have
1999 Don't use this macro to turn on various extra optimizations for
2000 `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */
2003 override_options (void)
2006 int ix86_tune_defaulted = 0;
2007 int ix86_arch_specified = 0;
2008 unsigned int ix86_arch_mask, ix86_tune_mask;
2010 /* Comes from final.c -- no real reason to change it. */
2011 #define MAX_CODE_ALIGN 16
2015 const struct processor_costs *cost; /* Processor costs */
2016 const int align_loop; /* Default alignments. */
2017 const int align_loop_max_skip;
2018 const int align_jump;
2019 const int align_jump_max_skip;
2020 const int align_func;
2022 const processor_target_table[PROCESSOR_max] =
2024 {&i386_cost, 4, 3, 4, 3, 4},
2025 {&i486_cost, 16, 15, 16, 15, 16},
2026 {&pentium_cost, 16, 7, 16, 7, 16},
2027 {&pentiumpro_cost, 16, 15, 16, 10, 16},
2028 {&geode_cost, 0, 0, 0, 0, 0},
2029 {&k6_cost, 32, 7, 32, 7, 32},
2030 {&athlon_cost, 16, 7, 16, 7, 16},
2031 {&pentium4_cost, 0, 0, 0, 0, 0},
2032 {&k8_cost, 16, 7, 16, 7, 16},
2033 {&nocona_cost, 0, 0, 0, 0, 0},
2034 {&core2_cost, 16, 10, 16, 10, 16},
2035 {&generic32_cost, 16, 7, 16, 7, 16},
2036 {&generic64_cost, 16, 10, 16, 10, 16},
2037 {&amdfam10_cost, 32, 24, 32, 7, 32}
2040 static const char *const cpu_names[TARGET_CPU_DEFAULT_max] =
2071 PTA_PREFETCH_SSE = 1 << 4,
2073 PTA_3DNOW_A = 1 << 6,
2077 PTA_POPCNT = 1 << 10,
2079 PTA_SSE4A = 1 << 12,
2080 PTA_NO_SAHF = 1 << 13,
2081 PTA_SSE4_1 = 1 << 14,
2082 PTA_SSE4_2 = 1 << 15,
2085 PTA_PCLMUL = 1 << 18
2090 const char *const name; /* processor name or nickname. */
2091 const enum processor_type processor;
2092 const unsigned /*enum pta_flags*/ flags;
2094 const processor_alias_table[] =
2096 {"i386", PROCESSOR_I386, 0},
2097 {"i486", PROCESSOR_I486, 0},
2098 {"i586", PROCESSOR_PENTIUM, 0},
2099 {"pentium", PROCESSOR_PENTIUM, 0},
2100 {"pentium-mmx", PROCESSOR_PENTIUM, PTA_MMX},
2101 {"winchip-c6", PROCESSOR_I486, PTA_MMX},
2102 {"winchip2", PROCESSOR_I486, PTA_MMX | PTA_3DNOW},
2103 {"c3", PROCESSOR_I486, PTA_MMX | PTA_3DNOW},
2104 {"c3-2", PROCESSOR_PENTIUMPRO, PTA_MMX | PTA_SSE},
2105 {"i686", PROCESSOR_PENTIUMPRO, 0},
2106 {"pentiumpro", PROCESSOR_PENTIUMPRO, 0},
2107 {"pentium2", PROCESSOR_PENTIUMPRO, PTA_MMX},
2108 {"pentium3", PROCESSOR_PENTIUMPRO, PTA_MMX | PTA_SSE},
2109 {"pentium3m", PROCESSOR_PENTIUMPRO, PTA_MMX | PTA_SSE},
2110 {"pentium-m", PROCESSOR_PENTIUMPRO, PTA_MMX | PTA_SSE | PTA_SSE2},
2111 {"pentium4", PROCESSOR_PENTIUM4, PTA_MMX |PTA_SSE | PTA_SSE2},
2112 {"pentium4m", PROCESSOR_PENTIUM4, PTA_MMX | PTA_SSE | PTA_SSE2},
2113 {"prescott", PROCESSOR_NOCONA, PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3},
2114 {"nocona", PROCESSOR_NOCONA, (PTA_64BIT
2115 | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2116 | PTA_CX16 | PTA_NO_SAHF)},
2117 {"core2", PROCESSOR_CORE2, (PTA_64BIT
2118 | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2121 {"geode", PROCESSOR_GEODE, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2122 |PTA_PREFETCH_SSE)},
2123 {"k6", PROCESSOR_K6, PTA_MMX},
2124 {"k6-2", PROCESSOR_K6, PTA_MMX | PTA_3DNOW},
2125 {"k6-3", PROCESSOR_K6, PTA_MMX | PTA_3DNOW},
2126 {"athlon", PROCESSOR_ATHLON, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2127 | PTA_PREFETCH_SSE)},
2128 {"athlon-tbird", PROCESSOR_ATHLON, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2129 | PTA_PREFETCH_SSE)},
2130 {"athlon-4", PROCESSOR_ATHLON, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2132 {"athlon-xp", PROCESSOR_ATHLON, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2134 {"athlon-mp", PROCESSOR_ATHLON, (PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2136 {"x86-64", PROCESSOR_K8, (PTA_64BIT
2137 | PTA_MMX | PTA_SSE | PTA_SSE2
2139 {"k8", PROCESSOR_K8, (PTA_64BIT
2140 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2141 | PTA_SSE | PTA_SSE2
2143 {"k8-sse3", PROCESSOR_K8, (PTA_64BIT
2144 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2145 | PTA_SSE | PTA_SSE2 | PTA_SSE3
2147 {"opteron", PROCESSOR_K8, (PTA_64BIT
2148 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2149 | PTA_SSE | PTA_SSE2
2151 {"opteron-sse3", PROCESSOR_K8, (PTA_64BIT
2152 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2153 | PTA_SSE | PTA_SSE2 | PTA_SSE3
2155 {"athlon64", PROCESSOR_K8, (PTA_64BIT
2156 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2157 | PTA_SSE | PTA_SSE2
2159 {"athlon64-sse3", PROCESSOR_K8, (PTA_64BIT
2160 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2161 | PTA_SSE | PTA_SSE2 | PTA_SSE3
2163 {"athlon-fx", PROCESSOR_K8, (PTA_64BIT
2164 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2165 | PTA_SSE | PTA_SSE2
2167 {"amdfam10", PROCESSOR_AMDFAM10, (PTA_64BIT
2168 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2169 | PTA_SSE | PTA_SSE2 | PTA_SSE3
2171 | PTA_CX16 | PTA_ABM)},
2172 {"barcelona", PROCESSOR_AMDFAM10, (PTA_64BIT
2173 | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A
2174 | PTA_SSE | PTA_SSE2 | PTA_SSE3
2176 | PTA_CX16 | PTA_ABM)},
2177 {"generic32", PROCESSOR_GENERIC32, 0 /* flags are only used for -march switch. */ },
2178 {"generic64", PROCESSOR_GENERIC64, PTA_64BIT /* flags are only used for -march switch. */ },
2181 int const pta_size = ARRAY_SIZE (processor_alias_table);
2183 #ifdef SUBTARGET_OVERRIDE_OPTIONS
2184 SUBTARGET_OVERRIDE_OPTIONS;
2187 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
2188 SUBSUBTARGET_OVERRIDE_OPTIONS;
2191 /* -fPIC is the default for x86_64. */
2192 if (TARGET_MACHO && TARGET_64BIT)
2195 /* Set the default values for switches whose default depends on TARGET_64BIT
2196 in case they weren't overwritten by command line options. */
2199 /* Mach-O doesn't support omitting the frame pointer for now. */
2200 if (flag_omit_frame_pointer == 2)
2201 flag_omit_frame_pointer = (TARGET_MACHO ? 0 : 1);
2202 if (flag_asynchronous_unwind_tables == 2)
2203 flag_asynchronous_unwind_tables = 1;
2204 if (flag_pcc_struct_return == 2)
2205 flag_pcc_struct_return = 0;
2209 if (flag_omit_frame_pointer == 2)
2210 flag_omit_frame_pointer = 0;
2211 if (flag_asynchronous_unwind_tables == 2)
2212 flag_asynchronous_unwind_tables = 0;
2213 if (flag_pcc_struct_return == 2)
2214 flag_pcc_struct_return = DEFAULT_PCC_STRUCT_RETURN;
2217 /* Need to check -mtune=generic first. */
2218 if (ix86_tune_string)
2220 if (!strcmp (ix86_tune_string, "generic")
2221 || !strcmp (ix86_tune_string, "i686")
2222 /* As special support for cross compilers we read -mtune=native
2223 as -mtune=generic. With native compilers we won't see the
2224 -mtune=native, as it was changed by the driver. */
2225 || !strcmp (ix86_tune_string, "native"))
2228 ix86_tune_string = "generic64";
2230 ix86_tune_string = "generic32";
2232 else if (!strncmp (ix86_tune_string, "generic", 7))
2233 error ("bad value (%s) for -mtune= switch", ix86_tune_string);
2237 if (ix86_arch_string)
2238 ix86_tune_string = ix86_arch_string;
2239 if (!ix86_tune_string)
2241 ix86_tune_string = cpu_names[TARGET_CPU_DEFAULT];
2242 ix86_tune_defaulted = 1;
2245 /* ix86_tune_string is set to ix86_arch_string or defaulted. We
2246 need to use a sensible tune option. */
2247 if (!strcmp (ix86_tune_string, "generic")
2248 || !strcmp (ix86_tune_string, "x86-64")
2249 || !strcmp (ix86_tune_string, "i686"))
2252 ix86_tune_string = "generic64";
2254 ix86_tune_string = "generic32";
2257 if (ix86_stringop_string)
2259 if (!strcmp (ix86_stringop_string, "rep_byte"))
2260 stringop_alg = rep_prefix_1_byte;
2261 else if (!strcmp (ix86_stringop_string, "libcall"))
2262 stringop_alg = libcall;
2263 else if (!strcmp (ix86_stringop_string, "rep_4byte"))
2264 stringop_alg = rep_prefix_4_byte;
2265 else if (!strcmp (ix86_stringop_string, "rep_8byte"))
2266 stringop_alg = rep_prefix_8_byte;
2267 else if (!strcmp (ix86_stringop_string, "byte_loop"))
2268 stringop_alg = loop_1_byte;
2269 else if (!strcmp (ix86_stringop_string, "loop"))
2270 stringop_alg = loop;
2271 else if (!strcmp (ix86_stringop_string, "unrolled_loop"))
2272 stringop_alg = unrolled_loop;
2274 error ("bad value (%s) for -mstringop-strategy= switch", ix86_stringop_string);
2276 if (!strcmp (ix86_tune_string, "x86-64"))
2277 warning (OPT_Wdeprecated, "-mtune=x86-64 is deprecated. Use -mtune=k8 or "
2278 "-mtune=generic instead as appropriate.");
2280 if (!ix86_arch_string)
2281 ix86_arch_string = TARGET_64BIT ? "x86-64" : "i386";
2283 ix86_arch_specified = 1;
2285 if (!strcmp (ix86_arch_string, "generic"))
2286 error ("generic CPU can be used only for -mtune= switch");
2287 if (!strncmp (ix86_arch_string, "generic", 7))
2288 error ("bad value (%s) for -march= switch", ix86_arch_string);
2290 if (ix86_cmodel_string != 0)
2292 if (!strcmp (ix86_cmodel_string, "small"))
2293 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2294 else if (!strcmp (ix86_cmodel_string, "medium"))
2295 ix86_cmodel = flag_pic ? CM_MEDIUM_PIC : CM_MEDIUM;
2296 else if (!strcmp (ix86_cmodel_string, "large"))
2297 ix86_cmodel = flag_pic ? CM_LARGE_PIC : CM_LARGE;
2299 error ("code model %s does not support PIC mode", ix86_cmodel_string);
2300 else if (!strcmp (ix86_cmodel_string, "32"))
2301 ix86_cmodel = CM_32;
2302 else if (!strcmp (ix86_cmodel_string, "kernel") && !flag_pic)
2303 ix86_cmodel = CM_KERNEL;
2305 error ("bad value (%s) for -mcmodel= switch", ix86_cmodel_string);
2309 /* For TARGET_64BIT and MS_ABI, force pic on, in order to enable the
2310 use of rip-relative addressing. This eliminates fixups that
2311 would otherwise be needed if this object is to be placed in a
2312 DLL, and is essentially just as efficient as direct addressing. */
2313 if (TARGET_64BIT && DEFAULT_ABI == MS_ABI)
2314 ix86_cmodel = CM_SMALL_PIC, flag_pic = 1;
2315 else if (TARGET_64BIT)
2316 ix86_cmodel = flag_pic ? CM_SMALL_PIC : CM_SMALL;
2318 ix86_cmodel = CM_32;
2320 if (ix86_asm_string != 0)
2323 && !strcmp (ix86_asm_string, "intel"))
2324 ix86_asm_dialect = ASM_INTEL;
2325 else if (!strcmp (ix86_asm_string, "att"))
2326 ix86_asm_dialect = ASM_ATT;
2328 error ("bad value (%s) for -masm= switch", ix86_asm_string);
2330 if ((TARGET_64BIT == 0) != (ix86_cmodel == CM_32))
2331 error ("code model %qs not supported in the %s bit mode",
2332 ix86_cmodel_string, TARGET_64BIT ? "64" : "32");
2333 if ((TARGET_64BIT != 0) != ((ix86_isa_flags & OPTION_MASK_ISA_64BIT) != 0))
2334 sorry ("%i-bit mode not compiled in",
2335 (ix86_isa_flags & OPTION_MASK_ISA_64BIT) ? 64 : 32);
2337 for (i = 0; i < pta_size; i++)
2338 if (! strcmp (ix86_arch_string, processor_alias_table[i].name))
2340 ix86_arch = processor_alias_table[i].processor;
2341 /* Default cpu tuning to the architecture. */
2342 ix86_tune = ix86_arch;
2344 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
2345 error ("CPU you selected does not support x86-64 "
2348 if (processor_alias_table[i].flags & PTA_MMX
2349 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_MMX))
2350 ix86_isa_flags |= OPTION_MASK_ISA_MMX;
2351 if (processor_alias_table[i].flags & PTA_3DNOW
2352 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW))
2353 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW;
2354 if (processor_alias_table[i].flags & PTA_3DNOW_A
2355 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_3DNOW_A))
2356 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_A;
2357 if (processor_alias_table[i].flags & PTA_SSE
2358 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE))
2359 ix86_isa_flags |= OPTION_MASK_ISA_SSE;
2360 if (processor_alias_table[i].flags & PTA_SSE2
2361 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE2))
2362 ix86_isa_flags |= OPTION_MASK_ISA_SSE2;
2363 if (processor_alias_table[i].flags & PTA_SSE3
2364 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE3))
2365 ix86_isa_flags |= OPTION_MASK_ISA_SSE3;
2366 if (processor_alias_table[i].flags & PTA_SSSE3
2367 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSSE3))
2368 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3;
2369 if (processor_alias_table[i].flags & PTA_SSE4_1
2370 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_1))
2371 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1;
2372 if (processor_alias_table[i].flags & PTA_SSE4_2
2373 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4_2))
2374 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2;
2375 if (processor_alias_table[i].flags & PTA_SSE4A
2376 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE4A))
2377 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A;
2378 if (processor_alias_table[i].flags & PTA_SSE5
2379 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE5))
2380 ix86_isa_flags |= OPTION_MASK_ISA_SSE5;
2382 if (processor_alias_table[i].flags & PTA_ABM)
2384 if (processor_alias_table[i].flags & PTA_CX16)
2385 x86_cmpxchg16b = true;
2386 if (processor_alias_table[i].flags & (PTA_POPCNT | PTA_ABM))
2388 if (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE))
2389 x86_prefetch_sse = true;
2390 if (!(TARGET_64BIT && (processor_alias_table[i].flags & PTA_NO_SAHF)))
2392 if (processor_alias_table[i].flags & PTA_AES)
2394 if (processor_alias_table[i].flags & PTA_PCLMUL)
2401 error ("bad value (%s) for -march= switch", ix86_arch_string);
2403 ix86_arch_mask = 1u << ix86_arch;
2404 for (i = 0; i < X86_ARCH_LAST; ++i)
2405 ix86_arch_features[i] &= ix86_arch_mask;
2407 for (i = 0; i < pta_size; i++)
2408 if (! strcmp (ix86_tune_string, processor_alias_table[i].name))
2410 ix86_tune = processor_alias_table[i].processor;
2411 if (TARGET_64BIT && !(processor_alias_table[i].flags & PTA_64BIT))
2413 if (ix86_tune_defaulted)
2415 ix86_tune_string = "x86-64";
2416 for (i = 0; i < pta_size; i++)
2417 if (! strcmp (ix86_tune_string,
2418 processor_alias_table[i].name))
2420 ix86_tune = processor_alias_table[i].processor;
2423 error ("CPU you selected does not support x86-64 "
2426 /* Intel CPUs have always interpreted SSE prefetch instructions as
2427 NOPs; so, we can enable SSE prefetch instructions even when
2428 -mtune (rather than -march) points us to a processor that has them.
2429 However, the VIA C3 gives a SIGILL, so we only do that for i686 and
2430 higher processors. */
2432 && (processor_alias_table[i].flags & (PTA_PREFETCH_SSE | PTA_SSE)))
2433 x86_prefetch_sse = true;
2437 error ("bad value (%s) for -mtune= switch", ix86_tune_string);
2439 /* Enable SSE2 if AES or PCLMUL is enabled. */
2440 if ((x86_aes || x86_pclmul)
2441 && !(ix86_isa_flags_explicit & OPTION_MASK_ISA_SSE2))
2443 ix86_isa_flags |= OPTION_MASK_ISA_SSE2_SET;
2444 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_SET;
2447 ix86_tune_mask = 1u << ix86_tune;
2448 for (i = 0; i < X86_TUNE_LAST; ++i)
2449 ix86_tune_features[i] &= ix86_tune_mask;
2452 ix86_cost = &size_cost;
2454 ix86_cost = processor_target_table[ix86_tune].cost;
2456 /* Arrange to set up i386_stack_locals for all functions. */
2457 init_machine_status = ix86_init_machine_status;
2459 /* Validate -mregparm= value. */
2460 if (ix86_regparm_string)
2463 warning (0, "-mregparm is ignored in 64-bit mode");
2464 i = atoi (ix86_regparm_string);
2465 if (i < 0 || i > REGPARM_MAX)
2466 error ("-mregparm=%d is not between 0 and %d", i, REGPARM_MAX);
2471 ix86_regparm = REGPARM_MAX;
2473 /* If the user has provided any of the -malign-* options,
2474 warn and use that value only if -falign-* is not set.
2475 Remove this code in GCC 3.2 or later. */
2476 if (ix86_align_loops_string)
2478 warning (0, "-malign-loops is obsolete, use -falign-loops");
2479 if (align_loops == 0)
2481 i = atoi (ix86_align_loops_string);
2482 if (i < 0 || i > MAX_CODE_ALIGN)
2483 error ("-malign-loops=%d is not between 0 and %d", i, MAX_CODE_ALIGN);
2485 align_loops = 1 << i;
2489 if (ix86_align_jumps_string)
2491 warning (0, "-malign-jumps is obsolete, use -falign-jumps");
2492 if (align_jumps == 0)
2494 i = atoi (ix86_align_jumps_string);
2495 if (i < 0 || i > MAX_CODE_ALIGN)
2496 error ("-malign-loops=%d is not between 0 and %d", i, MAX_CODE_ALIGN);
2498 align_jumps = 1 << i;
2502 if (ix86_align_funcs_string)
2504 warning (0, "-malign-functions is obsolete, use -falign-functions");
2505 if (align_functions == 0)
2507 i = atoi (ix86_align_funcs_string);
2508 if (i < 0 || i > MAX_CODE_ALIGN)
2509 error ("-malign-loops=%d is not between 0 and %d", i, MAX_CODE_ALIGN);
2511 align_functions = 1 << i;
2515 /* Default align_* from the processor table. */
2516 if (align_loops == 0)
2518 align_loops = processor_target_table[ix86_tune].align_loop;
2519 align_loops_max_skip = processor_target_table[ix86_tune].align_loop_max_skip;
2521 if (align_jumps == 0)
2523 align_jumps = processor_target_table[ix86_tune].align_jump;
2524 align_jumps_max_skip = processor_target_table[ix86_tune].align_jump_max_skip;
2526 if (align_functions == 0)
2528 align_functions = processor_target_table[ix86_tune].align_func;
2531 /* Validate -mbranch-cost= value, or provide default. */
2532 ix86_branch_cost = ix86_cost->branch_cost;
2533 if (ix86_branch_cost_string)
2535 i = atoi (ix86_branch_cost_string);
2537 error ("-mbranch-cost=%d is not between 0 and 5", i);
2539 ix86_branch_cost = i;
2541 if (ix86_section_threshold_string)
2543 i = atoi (ix86_section_threshold_string);
2545 error ("-mlarge-data-threshold=%d is negative", i);
2547 ix86_section_threshold = i;
2550 if (ix86_tls_dialect_string)
2552 if (strcmp (ix86_tls_dialect_string, "gnu") == 0)
2553 ix86_tls_dialect = TLS_DIALECT_GNU;
2554 else if (strcmp (ix86_tls_dialect_string, "gnu2") == 0)
2555 ix86_tls_dialect = TLS_DIALECT_GNU2;
2556 else if (strcmp (ix86_tls_dialect_string, "sun") == 0)
2557 ix86_tls_dialect = TLS_DIALECT_SUN;
2559 error ("bad value (%s) for -mtls-dialect= switch",
2560 ix86_tls_dialect_string);
2563 if (ix87_precision_string)
2565 i = atoi (ix87_precision_string);
2566 if (i != 32 && i != 64 && i != 80)
2567 error ("pc%d is not valid precision setting (32, 64 or 80)", i);
2572 target_flags |= TARGET_SUBTARGET64_DEFAULT & ~target_flags_explicit;
2574 /* Enable by default the SSE and MMX builtins. Do allow the user to
2575 explicitly disable any of these. In particular, disabling SSE and
2576 MMX for kernel code is extremely useful. */
2577 if (!ix86_arch_specified)
2579 |= ((OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_MMX
2580 | TARGET_SUBTARGET64_ISA_DEFAULT) & ~ix86_isa_flags_explicit);
2583 warning (0, "-mrtd is ignored in 64bit mode");
2587 target_flags |= TARGET_SUBTARGET32_DEFAULT & ~target_flags_explicit;
2589 if (!ix86_arch_specified)
2591 |= TARGET_SUBTARGET32_ISA_DEFAULT & ~ix86_isa_flags_explicit;
2593 /* i386 ABI does not specify red zone. It still makes sense to use it
2594 when programmer takes care to stack from being destroyed. */
2595 if (!(target_flags_explicit & MASK_NO_RED_ZONE))
2596 target_flags |= MASK_NO_RED_ZONE;
2599 /* Keep nonleaf frame pointers. */
2600 if (flag_omit_frame_pointer)
2601 target_flags &= ~MASK_OMIT_LEAF_FRAME_POINTER;
2602 else if (TARGET_OMIT_LEAF_FRAME_POINTER)
2603 flag_omit_frame_pointer = 1;
2605 /* If we're doing fast math, we don't care about comparison order
2606 wrt NaNs. This lets us use a shorter comparison sequence. */
2607 if (flag_finite_math_only)
2608 target_flags &= ~MASK_IEEE_FP;
2610 /* If the architecture always has an FPU, turn off NO_FANCY_MATH_387,
2611 since the insns won't need emulation. */
2612 if (x86_arch_always_fancy_math_387 & ix86_arch_mask)
2613 target_flags &= ~MASK_NO_FANCY_MATH_387;
2615 /* Likewise, if the target doesn't have a 387, or we've specified
2616 software floating point, don't use 387 inline intrinsics. */
2618 target_flags |= MASK_NO_FANCY_MATH_387;
2620 /* Turn on MMX builtins for -msse. */
2623 ix86_isa_flags |= OPTION_MASK_ISA_MMX & ~ix86_isa_flags_explicit;
2624 x86_prefetch_sse = true;
2627 /* Turn on popcnt instruction for -msse4.2 or -mabm. */
2628 if (TARGET_SSE4_2 || TARGET_ABM)
2631 /* Validate -mpreferred-stack-boundary= value, or provide default.
2632 The default of 128 bits is for Pentium III's SSE __m128. We can't
2633 change it because of optimize_size. Otherwise, we can't mix object
2634 files compiled with -Os and -On. */
2635 ix86_preferred_stack_boundary = 128;
2636 if (ix86_preferred_stack_boundary_string)
2638 i = atoi (ix86_preferred_stack_boundary_string);
2639 if (i < (TARGET_64BIT ? 4 : 2) || i > 12)
2640 error ("-mpreferred-stack-boundary=%d is not between %d and 12", i,
2641 TARGET_64BIT ? 4 : 2);
2643 ix86_preferred_stack_boundary = (1 << i) * BITS_PER_UNIT;
2646 /* Accept -msseregparm only if at least SSE support is enabled. */
2647 if (TARGET_SSEREGPARM
2649 error ("-msseregparm used without SSE enabled");
2651 ix86_fpmath = TARGET_FPMATH_DEFAULT;
2652 if (ix86_fpmath_string != 0)
2654 if (! strcmp (ix86_fpmath_string, "387"))
2655 ix86_fpmath = FPMATH_387;
2656 else if (! strcmp (ix86_fpmath_string, "sse"))
2660 warning (0, "SSE instruction set disabled, using 387 arithmetics");
2661 ix86_fpmath = FPMATH_387;
2664 ix86_fpmath = FPMATH_SSE;
2666 else if (! strcmp (ix86_fpmath_string, "387,sse")
2667 || ! strcmp (ix86_fpmath_string, "sse,387"))
2671 warning (0, "SSE instruction set disabled, using 387 arithmetics");
2672 ix86_fpmath = FPMATH_387;
2674 else if (!TARGET_80387)
2676 warning (0, "387 instruction set disabled, using SSE arithmetics");
2677 ix86_fpmath = FPMATH_SSE;
2680 ix86_fpmath = (enum fpmath_unit) (FPMATH_SSE | FPMATH_387);
2683 error ("bad value (%s) for -mfpmath= switch", ix86_fpmath_string);
2686 /* If the i387 is disabled, then do not return values in it. */
2688 target_flags &= ~MASK_FLOAT_RETURNS;
2690 /* Use external vectorized library in vectorizing intrinsics. */
2691 if (ix86_veclibabi_string)
2693 if (strcmp (ix86_veclibabi_string, "svml") == 0)
2694 ix86_veclib_handler = ix86_veclibabi_svml;
2695 else if (strcmp (ix86_veclibabi_string, "acml") == 0)
2696 ix86_veclib_handler = ix86_veclibabi_acml;
2698 error ("unknown vectorization library ABI type (%s) for "
2699 "-mveclibabi= switch", ix86_veclibabi_string);
2702 if ((x86_accumulate_outgoing_args & ix86_tune_mask)
2703 && !(target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
2705 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
2707 /* ??? Unwind info is not correct around the CFG unless either a frame
2708 pointer is present or M_A_O_A is set. Fixing this requires rewriting
2709 unwind info generation to be aware of the CFG and propagating states
2711 if ((flag_unwind_tables || flag_asynchronous_unwind_tables
2712 || flag_exceptions || flag_non_call_exceptions)
2713 && flag_omit_frame_pointer
2714 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
2716 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
2717 warning (0, "unwind tables currently require either a frame pointer "
2718 "or -maccumulate-outgoing-args for correctness");
2719 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
2722 /* If stack probes are required, the space used for large function
2723 arguments on the stack must also be probed, so enable
2724 -maccumulate-outgoing-args so this happens in the prologue. */
2725 if (TARGET_STACK_PROBE
2726 && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS))
2728 if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS)
2729 warning (0, "stack probing requires -maccumulate-outgoing-args "
2731 target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS;
2734 /* For sane SSE instruction set generation we need fcomi instruction.
2735 It is safe to enable all CMOVE instructions. */
2739 /* Figure out what ASM_GENERATE_INTERNAL_LABEL builds as a prefix. */
2742 ASM_GENERATE_INTERNAL_LABEL (internal_label_prefix, "LX", 0);
2743 p = strchr (internal_label_prefix, 'X');
2744 internal_label_prefix_len = p - internal_label_prefix;
2748 /* When scheduling description is not available, disable scheduler pass
2749 so it won't slow down the compilation and make x87 code slower. */
2750 if (!TARGET_SCHEDULE)
2751 flag_schedule_insns_after_reload = flag_schedule_insns = 0;
2753 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES))
2754 set_param_value ("simultaneous-prefetches",
2755 ix86_cost->simultaneous_prefetches);
2756 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE))
2757 set_param_value ("l1-cache-line-size", ix86_cost->prefetch_block);
2758 if (!PARAM_SET_P (PARAM_L1_CACHE_SIZE))
2759 set_param_value ("l1-cache-size", ix86_cost->l1_cache_size);
2760 if (!PARAM_SET_P (PARAM_L2_CACHE_SIZE))
2761 set_param_value ("l2-cache-size", ix86_cost->l2_cache_size);
2763 /* If using typedef char *va_list, signal that __builtin_va_start (&ap, 0)
2764 can be optimized to ap = __builtin_next_arg (0).
2765 For abi switching it should be corrected. */
2766 if (!TARGET_64BIT || DEFAULT_ABI == MS_ABI)
2767 targetm.expand_builtin_va_start = NULL;
2770 /* Use -mcld by default for 32-bit code if configured with --enable-cld. */
2772 target_flags |= MASK_CLD & ~target_flags_explicit;
2776 /* Return true if this goes in large data/bss. */
2779 ix86_in_large_data_p (tree exp)
2781 if (ix86_cmodel != CM_MEDIUM && ix86_cmodel != CM_MEDIUM_PIC)
2784 /* Functions are never large data. */
2785 if (TREE_CODE (exp) == FUNCTION_DECL)
2788 if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
2790 const char *section = TREE_STRING_POINTER (DECL_SECTION_NAME (exp));
2791 if (strcmp (section, ".ldata") == 0
2792 || strcmp (section, ".lbss") == 0)
2798 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (exp));
2800 /* If this is an incomplete type with size 0, then we can't put it
2801 in data because it might be too big when completed. */
2802 if (!size || size > ix86_section_threshold)
2809 /* Switch to the appropriate section for output of DECL.
2810 DECL is either a `VAR_DECL' node or a constant of some sort.
2811 RELOC indicates whether forming the initial value of DECL requires
2812 link-time relocations. */
2814 static section * x86_64_elf_select_section (tree, int, unsigned HOST_WIDE_INT)
2818 x86_64_elf_select_section (tree decl, int reloc,
2819 unsigned HOST_WIDE_INT align)
2821 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
2822 && ix86_in_large_data_p (decl))
2824 const char *sname = NULL;
2825 unsigned int flags = SECTION_WRITE;
2826 switch (categorize_decl_for_section (decl, reloc))
2831 case SECCAT_DATA_REL:
2832 sname = ".ldata.rel";
2834 case SECCAT_DATA_REL_LOCAL:
2835 sname = ".ldata.rel.local";
2837 case SECCAT_DATA_REL_RO:
2838 sname = ".ldata.rel.ro";
2840 case SECCAT_DATA_REL_RO_LOCAL:
2841 sname = ".ldata.rel.ro.local";
2845 flags |= SECTION_BSS;
2848 case SECCAT_RODATA_MERGE_STR:
2849 case SECCAT_RODATA_MERGE_STR_INIT:
2850 case SECCAT_RODATA_MERGE_CONST:
2854 case SECCAT_SRODATA:
2861 /* We don't split these for medium model. Place them into
2862 default sections and hope for best. */
2864 case SECCAT_EMUTLS_VAR:
2865 case SECCAT_EMUTLS_TMPL:
2870 /* We might get called with string constants, but get_named_section
2871 doesn't like them as they are not DECLs. Also, we need to set
2872 flags in that case. */
2874 return get_section (sname, flags, NULL);
2875 return get_named_section (decl, sname, reloc);
2878 return default_elf_select_section (decl, reloc, align);
2881 /* Build up a unique section name, expressed as a
2882 STRING_CST node, and assign it to DECL_SECTION_NAME (decl).
2883 RELOC indicates whether the initial value of EXP requires
2884 link-time relocations. */
2886 static void ATTRIBUTE_UNUSED
2887 x86_64_elf_unique_section (tree decl, int reloc)
2889 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
2890 && ix86_in_large_data_p (decl))
2892 const char *prefix = NULL;
2893 /* We only need to use .gnu.linkonce if we don't have COMDAT groups. */
2894 bool one_only = DECL_ONE_ONLY (decl) && !HAVE_COMDAT_GROUP;
2896 switch (categorize_decl_for_section (decl, reloc))
2899 case SECCAT_DATA_REL:
2900 case SECCAT_DATA_REL_LOCAL:
2901 case SECCAT_DATA_REL_RO:
2902 case SECCAT_DATA_REL_RO_LOCAL:
2903 prefix = one_only ? ".ld" : ".ldata";
2906 prefix = one_only ? ".lb" : ".lbss";
2909 case SECCAT_RODATA_MERGE_STR:
2910 case SECCAT_RODATA_MERGE_STR_INIT:
2911 case SECCAT_RODATA_MERGE_CONST:
2912 prefix = one_only ? ".lr" : ".lrodata";
2914 case SECCAT_SRODATA:
2921 /* We don't split these for medium model. Place them into
2922 default sections and hope for best. */
2924 case SECCAT_EMUTLS_VAR:
2925 prefix = targetm.emutls.var_section;
2927 case SECCAT_EMUTLS_TMPL:
2928 prefix = targetm.emutls.tmpl_section;
2933 const char *name, *linkonce;
2936 name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl));
2937 name = targetm.strip_name_encoding (name);
2939 /* If we're using one_only, then there needs to be a .gnu.linkonce
2940 prefix to the section name. */
2941 linkonce = one_only ? ".gnu.linkonce" : "";
2943 string = ACONCAT ((linkonce, prefix, ".", name, NULL));
2945 DECL_SECTION_NAME (decl) = build_string (strlen (string), string);
2949 default_unique_section (decl, reloc);
2952 #ifdef COMMON_ASM_OP
2953 /* This says how to output assembler code to declare an
2954 uninitialized external linkage data object.
2956 For medium model x86-64 we need to use .largecomm opcode for
2959 x86_elf_aligned_common (FILE *file,
2960 const char *name, unsigned HOST_WIDE_INT size,
2963 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
2964 && size > (unsigned int)ix86_section_threshold)
2965 fprintf (file, ".largecomm\t");
2967 fprintf (file, "%s", COMMON_ASM_OP);
2968 assemble_name (file, name);
2969 fprintf (file, ","HOST_WIDE_INT_PRINT_UNSIGNED",%u\n",
2970 size, align / BITS_PER_UNIT);
2974 /* Utility function for targets to use in implementing
2975 ASM_OUTPUT_ALIGNED_BSS. */
2978 x86_output_aligned_bss (FILE *file, tree decl ATTRIBUTE_UNUSED,
2979 const char *name, unsigned HOST_WIDE_INT size,
2982 if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
2983 && size > (unsigned int)ix86_section_threshold)
2984 switch_to_section (get_named_section (decl, ".lbss", 0));
2986 switch_to_section (bss_section);
2987 ASM_OUTPUT_ALIGN (file, floor_log2 (align / BITS_PER_UNIT));
2988 #ifdef ASM_DECLARE_OBJECT_NAME
2989 last_assemble_variable_decl = decl;
2990 ASM_DECLARE_OBJECT_NAME (file, name, decl);
2992 /* Standard thing is just output label for the object. */
2993 ASM_OUTPUT_LABEL (file, name);
2994 #endif /* ASM_DECLARE_OBJECT_NAME */
2995 ASM_OUTPUT_SKIP (file, size ? size : 1);
2999 optimization_options (int level, int size ATTRIBUTE_UNUSED)
3001 /* For -O2 and beyond, turn off -fschedule-insns by default. It tends to
3002 make the problem with not enough registers even worse. */
3003 #ifdef INSN_SCHEDULING
3005 flag_schedule_insns = 0;
3009 /* The Darwin libraries never set errno, so we might as well
3010 avoid calling them when that's the only reason we would. */
3011 flag_errno_math = 0;
3013 /* The default values of these switches depend on the TARGET_64BIT
3014 that is not known at this moment. Mark these values with 2 and
3015 let user the to override these. In case there is no command line option
3016 specifying them, we will set the defaults in override_options. */
3018 flag_omit_frame_pointer = 2;
3019 flag_pcc_struct_return = 2;
3020 flag_asynchronous_unwind_tables = 2;
3021 flag_vect_cost_model = 1;
3022 #ifdef SUBTARGET_OPTIMIZATION_OPTIONS
3023 SUBTARGET_OPTIMIZATION_OPTIONS;
3027 /* Decide whether we can make a sibling call to a function. DECL is the
3028 declaration of the function being targeted by the call and EXP is the
3029 CALL_EXPR representing the call. */
3032 ix86_function_ok_for_sibcall (tree decl, tree exp)
3037 /* If we are generating position-independent code, we cannot sibcall
3038 optimize any indirect call, or a direct call to a global function,
3039 as the PLT requires %ebx be live. */
3040 if (!TARGET_64BIT && flag_pic && (!decl || !targetm.binds_local_p (decl)))
3047 func = TREE_TYPE (CALL_EXPR_FN (exp));
3048 if (POINTER_TYPE_P (func))
3049 func = TREE_TYPE (func);
3052 /* Check that the return value locations are the same. Like
3053 if we are returning floats on the 80387 register stack, we cannot
3054 make a sibcall from a function that doesn't return a float to a
3055 function that does or, conversely, from a function that does return
3056 a float to a function that doesn't; the necessary stack adjustment
3057 would not be executed. This is also the place we notice
3058 differences in the return value ABI. Note that it is ok for one
3059 of the functions to have void return type as long as the return
3060 value of the other is passed in a register. */
3061 a = ix86_function_value (TREE_TYPE (exp), func, false);
3062 b = ix86_function_value (TREE_TYPE (DECL_RESULT (cfun->decl)),
3064 if (STACK_REG_P (a) || STACK_REG_P (b))
3066 if (!rtx_equal_p (a, b))
3069 else if (VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun->decl))))
3071 else if (!rtx_equal_p (a, b))
3074 /* If this call is indirect, we'll need to be able to use a call-clobbered
3075 register for the address of the target function. Make sure that all
3076 such registers are not used for passing parameters. */
3077 if (!decl && !TARGET_64BIT)
3081 /* We're looking at the CALL_EXPR, we need the type of the function. */
3082 type = CALL_EXPR_FN (exp); /* pointer expression */
3083 type = TREE_TYPE (type); /* pointer type */
3084 type = TREE_TYPE (type); /* function type */
3086 if (ix86_function_regparm (type, NULL) >= 3)
3088 /* ??? Need to count the actual number of registers to be used,
3089 not the possible number of registers. Fix later. */
3094 /* Dllimport'd functions are also called indirectly. */
3095 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
3096 && decl && DECL_DLLIMPORT_P (decl)
3097 && ix86_function_regparm (TREE_TYPE (decl), NULL) >= 3)
3100 /* If we forced aligned the stack, then sibcalling would unalign the
3101 stack, which may break the called function. */
3102 if (cfun->machine->force_align_arg_pointer)
3105 /* Otherwise okay. That also includes certain types of indirect calls. */
3109 /* Handle "cdecl", "stdcall", "fastcall", "regparm" and "sseregparm"
3110 calling convention attributes;
3111 arguments as in struct attribute_spec.handler. */
3114 ix86_handle_cconv_attribute (tree *node, tree name,
3116 int flags ATTRIBUTE_UNUSED,
3119 if (TREE_CODE (*node) != FUNCTION_TYPE
3120 && TREE_CODE (*node) != METHOD_TYPE
3121 && TREE_CODE (*node) != FIELD_DECL
3122 && TREE_CODE (*node) != TYPE_DECL)
3124 warning (OPT_Wattributes, "%qs attribute only applies to functions",
3125 IDENTIFIER_POINTER (name));
3126 *no_add_attrs = true;
3130 /* Can combine regparm with all attributes but fastcall. */
3131 if (is_attribute_p ("regparm", name))
3135 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
3137 error ("fastcall and regparm attributes are not compatible");
3140 cst = TREE_VALUE (args);
3141 if (TREE_CODE (cst) != INTEGER_CST)
3143 warning (OPT_Wattributes,
3144 "%qs attribute requires an integer constant argument",
3145 IDENTIFIER_POINTER (name));
3146 *no_add_attrs = true;
3148 else if (compare_tree_int (cst, REGPARM_MAX) > 0)
3150 warning (OPT_Wattributes, "argument to %qs attribute larger than %d",
3151 IDENTIFIER_POINTER (name), REGPARM_MAX);
3152 *no_add_attrs = true;
3156 && lookup_attribute (ix86_force_align_arg_pointer_string,
3157 TYPE_ATTRIBUTES (*node))
3158 && compare_tree_int (cst, REGPARM_MAX-1))
3160 error ("%s functions limited to %d register parameters",
3161 ix86_force_align_arg_pointer_string, REGPARM_MAX-1);
3169 /* Do not warn when emulating the MS ABI. */
3170 if (TREE_CODE (*node) != FUNCTION_TYPE || !ix86_function_type_abi (*node))
3171 warning (OPT_Wattributes, "%qs attribute ignored",
3172 IDENTIFIER_POINTER (name));
3173 *no_add_attrs = true;
3177 /* Can combine fastcall with stdcall (redundant) and sseregparm. */
3178 if (is_attribute_p ("fastcall", name))
3180 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
3182 error ("fastcall and cdecl attributes are not compatible");
3184 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
3186 error ("fastcall and stdcall attributes are not compatible");
3188 if (lookup_attribute ("regparm", TYPE_ATTRIBUTES (*node)))
3190 error ("fastcall and regparm attributes are not compatible");
3194 /* Can combine stdcall with fastcall (redundant), regparm and
3196 else if (is_attribute_p ("stdcall", name))
3198 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node)))
3200 error ("stdcall and cdecl attributes are not compatible");
3202 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
3204 error ("stdcall and fastcall attributes are not compatible");
3208 /* Can combine cdecl with regparm and sseregparm. */
3209 else if (is_attribute_p ("cdecl", name))
3211 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node)))
3213 error ("stdcall and cdecl attributes are not compatible");
3215 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node)))
3217 error ("fastcall and cdecl attributes are not compatible");
3221 /* Can combine sseregparm with all attributes. */
3226 /* Return 0 if the attributes for two types are incompatible, 1 if they
3227 are compatible, and 2 if they are nearly compatible (which causes a
3228 warning to be generated). */
3231 ix86_comp_type_attributes (const_tree type1, const_tree type2)
3233 /* Check for mismatch of non-default calling convention. */
3234 const char *const rtdstr = TARGET_RTD ? "cdecl" : "stdcall";
3236 if (TREE_CODE (type1) != FUNCTION_TYPE
3237 && TREE_CODE (type1) != METHOD_TYPE)
3240 /* Check for mismatched fastcall/regparm types. */
3241 if ((!lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type1))
3242 != !lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type2)))
3243 || (ix86_function_regparm (type1, NULL)
3244 != ix86_function_regparm (type2, NULL)))
3247 /* Check for mismatched sseregparm types. */
3248 if (!lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type1))
3249 != !lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type2)))
3252 /* Check for mismatched return types (cdecl vs stdcall). */
3253 if (!lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type1))
3254 != !lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type2)))
3260 /* Return the regparm value for a function with the indicated TYPE and DECL.
3261 DECL may be NULL when calling function indirectly
3262 or considering a libcall. */
3265 ix86_function_regparm (const_tree type, const_tree decl)
3268 int regparm = ix86_regparm;
3270 static bool error_issued;
3274 if (ix86_function_type_abi (type) == DEFAULT_ABI)
3276 return DEFAULT_ABI != SYSV_ABI ? X86_64_REGPARM_MAX : X64_REGPARM_MAX;
3279 attr = lookup_attribute ("regparm", TYPE_ATTRIBUTES (type));
3283 = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attr)));
3285 if (decl && TREE_CODE (decl) == FUNCTION_DECL)
3287 /* We can't use regparm(3) for nested functions because
3288 these pass static chain pointer in %ecx register. */
3289 if (!error_issued && regparm == 3
3290 && decl_function_context (decl)
3291 && !DECL_NO_STATIC_CHAIN (decl))
3293 error ("nested functions are limited to 2 register parameters");
3294 error_issued = true;
3302 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type)))
3305 /* Use register calling convention for local functions when possible. */
3306 if (decl && TREE_CODE (decl) == FUNCTION_DECL
3307 && flag_unit_at_a_time && !profile_flag)
OSDN Git Service
