1 /* Subroutines used for code generation on IA-32.
2 Copyright (C) 1988, 1992, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
3 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
30 #include "hard-reg-set.h"
32 #include "insn-config.h"
33 #include "conditions.h"
35 #include "insn-codes.h"
36 #include "insn-attr.h"
44 #include "basic-block.h"
47 #include "target-def.h"
48 #include "langhooks.h"
53 #include "tm-constrs.h"
57 static 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 ix86_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 storing fp registers
995 in SFmode, DFmode and XFmode */
996 2, /* cost of moving MMX register */
997 {6, 6}, /* cost of loading MMX registers
998 in SImode and DImode */
999 {4, 4}, /* cost of storing MMX registers
1000 in SImode and DImode */
1001 2, /* cost of moving SSE register */
1002 {6, 6, 6}, /* cost of loading SSE registers
1003 in SImode, DImode and TImode */
1004 {4, 4, 4}, /* cost of storing SSE registers
1005 in SImode, DImode and TImode */
1006 2, /* MMX or SSE register to integer */
1007 32, /* size of l1 cache. */
1008 2048, /* size of l2 cache. */
1009 128, /* size of prefetch block */
1010 8, /* number of parallel prefetches */
1011 3, /* Branch cost */
1012 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
1013 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
1014 COSTS_N_INSNS (32), /* cost of FDIV instruction. */
1015 COSTS_N_INSNS (1), /* cost of FABS instruction. */
1016 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
1017 COSTS_N_INSNS (58), /* cost of FSQRT instruction. */
1018 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1019 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1020 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1021 {{libcall, {{8, loop}, {15, unrolled_loop},
1022 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1023 {libcall, {{24, loop}, {32, unrolled_loop},
1024 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1025 1, /* scalar_stmt_cost. */
1026 1, /* scalar load_cost. */
1027 1, /* scalar_store_cost. */
1028 1, /* vec_stmt_cost. */
1029 1, /* vec_to_scalar_cost. */
1030 1, /* scalar_to_vec_cost. */
1031 1, /* vec_align_load_cost. */
1032 2, /* vec_unalign_load_cost. */
1033 1, /* vec_store_cost. */
1034 3, /* cond_taken_branch_cost. */
1035 1, /* cond_not_taken_branch_cost. */
1039 struct processor_costs atom_cost = {
1040 COSTS_N_INSNS (1), /* cost of an add instruction */
1041 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1042 COSTS_N_INSNS (1), /* variable shift costs */
1043 COSTS_N_INSNS (1), /* constant shift costs */
1044 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1045 COSTS_N_INSNS (4), /* HI */
1046 COSTS_N_INSNS (3), /* SI */
1047 COSTS_N_INSNS (4), /* DI */
1048 COSTS_N_INSNS (2)}, /* other */
1049 0, /* cost of multiply per each bit set */
1050 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1051 COSTS_N_INSNS (26), /* HI */
1052 COSTS_N_INSNS (42), /* SI */
1053 COSTS_N_INSNS (74), /* DI */
1054 COSTS_N_INSNS (74)}, /* other */
1055 COSTS_N_INSNS (1), /* cost of movsx */
1056 COSTS_N_INSNS (1), /* cost of movzx */
1057 8, /* "large" insn */
1058 17, /* MOVE_RATIO */
1059 2, /* cost for loading QImode using movzbl */
1060 {4, 4, 4}, /* cost of loading integer registers
1061 in QImode, HImode and SImode.
1062 Relative to reg-reg move (2). */
1063 {4, 4, 4}, /* cost of storing integer registers */
1064 4, /* cost of reg,reg fld/fst */
1065 {12, 12, 12}, /* cost of loading fp registers
1066 in SFmode, DFmode and XFmode */
1067 {6, 6, 8}, /* cost of storing fp registers
1068 in SFmode, DFmode and XFmode */
1069 2, /* cost of moving MMX register */
1070 {8, 8}, /* cost of loading MMX registers
1071 in SImode and DImode */
1072 {8, 8}, /* cost of storing MMX registers
1073 in SImode and DImode */
1074 2, /* cost of moving SSE register */
1075 {8, 8, 8}, /* cost of loading SSE registers
1076 in SImode, DImode and TImode */
1077 {8, 8, 8}, /* cost of storing SSE registers
1078 in SImode, DImode and TImode */
1079 5, /* MMX or SSE register to integer */
1080 32, /* size of l1 cache. */
1081 256, /* size of l2 cache. */
1082 64, /* size of prefetch block */
1083 6, /* number of parallel prefetches */
1084 3, /* Branch cost */
1085 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1086 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1087 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1088 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1089 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1090 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1091 {{libcall, {{11, loop}, {-1, rep_prefix_4_byte}}},
1092 {libcall, {{32, loop}, {64, rep_prefix_4_byte},
1093 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1094 {{libcall, {{8, loop}, {15, unrolled_loop},
1095 {2048, rep_prefix_4_byte}, {-1, libcall}}},
1096 {libcall, {{24, loop}, {32, unrolled_loop},
1097 {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1098 1, /* scalar_stmt_cost. */
1099 1, /* scalar load_cost. */
1100 1, /* scalar_store_cost. */
1101 1, /* vec_stmt_cost. */
1102 1, /* vec_to_scalar_cost. */
1103 1, /* scalar_to_vec_cost. */
1104 1, /* vec_align_load_cost. */
1105 2, /* vec_unalign_load_cost. */
1106 1, /* vec_store_cost. */
1107 3, /* cond_taken_branch_cost. */
1108 1, /* cond_not_taken_branch_cost. */
1111 /* Generic64 should produce code tuned for Nocona and K8. */
1113 struct processor_costs generic64_cost = {
1114 COSTS_N_INSNS (1), /* cost of an add instruction */
1115 /* On all chips taken into consideration lea is 2 cycles and more. With
1116 this cost however our current implementation of synth_mult results in
1117 use of unnecessary temporary registers causing regression on several
1118 SPECfp benchmarks. */
1119 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1120 COSTS_N_INSNS (1), /* variable shift costs */
1121 COSTS_N_INSNS (1), /* constant shift costs */
1122 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1123 COSTS_N_INSNS (4), /* HI */
1124 COSTS_N_INSNS (3), /* SI */
1125 COSTS_N_INSNS (4), /* DI */
1126 COSTS_N_INSNS (2)}, /* other */
1127 0, /* cost of multiply per each bit set */
1128 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1129 COSTS_N_INSNS (26), /* HI */
1130 COSTS_N_INSNS (42), /* SI */
1131 COSTS_N_INSNS (74), /* DI */
1132 COSTS_N_INSNS (74)}, /* other */
1133 COSTS_N_INSNS (1), /* cost of movsx */
1134 COSTS_N_INSNS (1), /* cost of movzx */
1135 8, /* "large" insn */
1136 17, /* MOVE_RATIO */
1137 4, /* cost for loading QImode using movzbl */
1138 {4, 4, 4}, /* cost of loading integer registers
1139 in QImode, HImode and SImode.
1140 Relative to reg-reg move (2). */
1141 {4, 4, 4}, /* cost of storing integer registers */
1142 4, /* cost of reg,reg fld/fst */
1143 {12, 12, 12}, /* cost of loading fp registers
1144 in SFmode, DFmode and XFmode */
1145 {6, 6, 8}, /* cost of storing fp registers
1146 in SFmode, DFmode and XFmode */
1147 2, /* cost of moving MMX register */
1148 {8, 8}, /* cost of loading MMX registers
1149 in SImode and DImode */
1150 {8, 8}, /* cost of storing MMX registers
1151 in SImode and DImode */
1152 2, /* cost of moving SSE register */
1153 {8, 8, 8}, /* cost of loading SSE registers
1154 in SImode, DImode and TImode */
1155 {8, 8, 8}, /* cost of storing SSE registers
1156 in SImode, DImode and TImode */
1157 5, /* MMX or SSE register to integer */
1158 32, /* size of l1 cache. */
1159 512, /* size of l2 cache. */
1160 64, /* size of prefetch block */
1161 6, /* number of parallel prefetches */
1162 /* Benchmarks shows large regressions on K8 sixtrack benchmark when this value
1163 is increased to perhaps more appropriate value of 5. */
1164 3, /* Branch cost */
1165 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1166 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1167 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1168 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1169 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1170 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1171 {DUMMY_STRINGOP_ALGS,
1172 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1173 {DUMMY_STRINGOP_ALGS,
1174 {libcall, {{32, loop}, {8192, rep_prefix_8_byte}, {-1, libcall}}}},
1175 1, /* scalar_stmt_cost. */
1176 1, /* scalar load_cost. */
1177 1, /* scalar_store_cost. */
1178 1, /* vec_stmt_cost. */
1179 1, /* vec_to_scalar_cost. */
1180 1, /* scalar_to_vec_cost. */
1181 1, /* vec_align_load_cost. */
1182 2, /* vec_unalign_load_cost. */
1183 1, /* vec_store_cost. */
1184 3, /* cond_taken_branch_cost. */
1185 1, /* cond_not_taken_branch_cost. */
1188 /* Generic32 should produce code tuned for Athlon, PPro, Pentium4, Nocona and K8. */
1190 struct processor_costs generic32_cost = {
1191 COSTS_N_INSNS (1), /* cost of an add instruction */
1192 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
1193 COSTS_N_INSNS (1), /* variable shift costs */
1194 COSTS_N_INSNS (1), /* constant shift costs */
1195 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
1196 COSTS_N_INSNS (4), /* HI */
1197 COSTS_N_INSNS (3), /* SI */
1198 COSTS_N_INSNS (4), /* DI */
1199 COSTS_N_INSNS (2)}, /* other */
1200 0, /* cost of multiply per each bit set */
1201 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
1202 COSTS_N_INSNS (26), /* HI */
1203 COSTS_N_INSNS (42), /* SI */
1204 COSTS_N_INSNS (74), /* DI */
1205 COSTS_N_INSNS (74)}, /* other */
1206 COSTS_N_INSNS (1), /* cost of movsx */
1207 COSTS_N_INSNS (1), /* cost of movzx */
1208 8, /* "large" insn */
1209 17, /* MOVE_RATIO */
1210 4, /* cost for loading QImode using movzbl */
1211 {4, 4, 4}, /* cost of loading integer registers
1212 in QImode, HImode and SImode.
1213 Relative to reg-reg move (2). */
1214 {4, 4, 4}, /* cost of storing integer registers */
1215 4, /* cost of reg,reg fld/fst */
1216 {12, 12, 12}, /* cost of loading fp registers
1217 in SFmode, DFmode and XFmode */
1218 {6, 6, 8}, /* cost of storing fp registers
1219 in SFmode, DFmode and XFmode */
1220 2, /* cost of moving MMX register */
1221 {8, 8}, /* cost of loading MMX registers
1222 in SImode and DImode */
1223 {8, 8}, /* cost of storing MMX registers
1224 in SImode and DImode */
1225 2, /* cost of moving SSE register */
1226 {8, 8, 8}, /* cost of loading SSE registers
1227 in SImode, DImode and TImode */
1228 {8, 8, 8}, /* cost of storing SSE registers
1229 in SImode, DImode and TImode */
1230 5, /* MMX or SSE register to integer */
1231 32, /* size of l1 cache. */
1232 256, /* size of l2 cache. */
1233 64, /* size of prefetch block */
1234 6, /* number of parallel prefetches */
1235 3, /* Branch cost */
1236 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
1237 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
1238 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
1239 COSTS_N_INSNS (8), /* cost of FABS instruction. */
1240 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
1241 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
1242 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1243 DUMMY_STRINGOP_ALGS},
1244 {{libcall, {{32, loop}, {8192, rep_prefix_4_byte}, {-1, libcall}}},
1245 DUMMY_STRINGOP_ALGS},
1246 1, /* scalar_stmt_cost. */
1247 1, /* scalar load_cost. */
1248 1, /* scalar_store_cost. */
1249 1, /* vec_stmt_cost. */
1250 1, /* vec_to_scalar_cost. */
1251 1, /* scalar_to_vec_cost. */
1252 1, /* vec_align_load_cost. */
1253 2, /* vec_unalign_load_cost. */
1254 1, /* vec_store_cost. */
1255 3, /* cond_taken_branch_cost. */
1256 1, /* cond_not_taken_branch_cost. */
1259 const struct processor_costs *ix86_cost = &pentium_cost;
1261 /* Processor feature/optimization bitmasks. */
1262 #define m_386 (1<<PROCESSOR_I386)
1263 #define m_486 (1<<PROCESSOR_I486)
1264 #define m_PENT (1<<PROCESSOR_PENTIUM)
1265 #define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
1266 #define m_PENT4 (1<<PROCESSOR_PENTIUM4)
1267 #define m_NOCONA (1<<PROCESSOR_NOCONA)
1268 #define m_CORE2 (1<<PROCESSOR_CORE2)
1269 #define m_ATOM (1<<PROCESSOR_ATOM)
1271 #define m_GEODE (1<<PROCESSOR_GEODE)
1272 #define m_K6 (1<<PROCESSOR_K6)
1273 #define m_K6_GEODE (m_K6 | m_GEODE)
1274 #define m_K8 (1<<PROCESSOR_K8)
1275 #define m_ATHLON (1<<PROCESSOR_ATHLON)
1276 #define m_ATHLON_K8 (m_K8 | m_ATHLON)
1277 #define m_AMDFAM10 (1<<PROCESSOR_AMDFAM10)
1278 #define m_AMD_MULTIPLE (m_K8 | m_ATHLON | m_AMDFAM10)
1280 #define m_GENERIC32 (1<<PROCESSOR_GENERIC32)
1281 #define m_GENERIC64 (1<<PROCESSOR_GENERIC64)
1283 /* Generic instruction choice should be common subset of supported CPUs
1284 (PPro/PENT4/NOCONA/CORE2/Athlon/K8). */
1285 #define m_GENERIC (m_GENERIC32 | m_GENERIC64)
1287 /* Feature tests against the various tunings. */
1288 unsigned char ix86_tune_features[X86_TUNE_LAST];
1290 /* Feature tests against the various tunings used to create ix86_tune_features
1291 based on the processor mask. */
1292 static unsigned int initial_ix86_tune_features[X86_TUNE_LAST] = {
1293 /* X86_TUNE_USE_LEAVE: Leave does not affect Nocona SPEC2000 results
1294 negatively, so enabling for Generic64 seems like good code size
1295 tradeoff. We can't enable it for 32bit generic because it does not
1296 work well with PPro base chips. */
1297 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_CORE2 | m_GENERIC64,
1299 /* X86_TUNE_PUSH_MEMORY */
1300 m_386 | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4
1301 | m_NOCONA | m_CORE2 | m_GENERIC,
1303 /* X86_TUNE_ZERO_EXTEND_WITH_AND */
1306 /* X86_TUNE_UNROLL_STRLEN */
1307 m_486 | m_PENT | m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_K6
1308 | m_CORE2 | m_GENERIC,
1310 /* X86_TUNE_DEEP_BRANCH_PREDICTION */
1311 m_ATOM | m_PPRO | m_K6_GEODE | m_AMD_MULTIPLE | m_PENT4 | m_GENERIC,
1313 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
1314 on simulation result. But after P4 was made, no performance benefit
1315 was observed with branch hints. It also increases the code size.
1316 As a result, icc never generates branch hints. */
1319 /* X86_TUNE_DOUBLE_WITH_ADD */
1322 /* X86_TUNE_USE_SAHF */
1323 m_ATOM | m_PPRO | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_PENT4
1324 | m_NOCONA | m_CORE2 | m_GENERIC,
1326 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
1327 partial dependencies. */
1328 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_PENT4 | m_NOCONA
1329 | m_CORE2 | m_GENERIC | m_GEODE /* m_386 | m_K6 */,
1331 /* X86_TUNE_PARTIAL_REG_STALL: We probably ought to watch for partial
1332 register stalls on Generic32 compilation setting as well. However
1333 in current implementation the partial register stalls are not eliminated
1334 very well - they can be introduced via subregs synthesized by combine
1335 and can happen in caller/callee saving sequences. Because this option
1336 pays back little on PPro based chips and is in conflict with partial reg
1337 dependencies used by Athlon/P4 based chips, it is better to leave it off
1338 for generic32 for now. */
1341 /* X86_TUNE_PARTIAL_FLAG_REG_STALL */
1342 m_CORE2 | m_GENERIC,
1344 /* X86_TUNE_USE_HIMODE_FIOP */
1345 m_386 | m_486 | m_K6_GEODE,
1347 /* X86_TUNE_USE_SIMODE_FIOP */
1348 ~(m_PPRO | m_AMD_MULTIPLE | m_PENT | m_ATOM | m_CORE2 | m_GENERIC),
1350 /* X86_TUNE_USE_MOV0 */
1353 /* X86_TUNE_USE_CLTD */
1354 ~(m_PENT | m_ATOM | m_K6 | m_CORE2 | m_GENERIC),
1356 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
1359 /* X86_TUNE_SPLIT_LONG_MOVES */
1362 /* X86_TUNE_READ_MODIFY_WRITE */
1365 /* X86_TUNE_READ_MODIFY */
1368 /* X86_TUNE_PROMOTE_QIMODE */
1369 m_K6_GEODE | m_PENT | m_ATOM | m_386 | m_486 | m_AMD_MULTIPLE
1370 | m_CORE2 | m_GENERIC /* | m_PENT4 ? */,
1372 /* X86_TUNE_FAST_PREFIX */
1373 ~(m_PENT | m_486 | m_386),
1375 /* X86_TUNE_SINGLE_STRINGOP */
1376 m_386 | m_PENT4 | m_NOCONA,
1378 /* X86_TUNE_QIMODE_MATH */
1381 /* X86_TUNE_HIMODE_MATH: On PPro this flag is meant to avoid partial
1382 register stalls. Just like X86_TUNE_PARTIAL_REG_STALL this option
1383 might be considered for Generic32 if our scheme for avoiding partial
1384 stalls was more effective. */
1387 /* X86_TUNE_PROMOTE_QI_REGS */
1390 /* X86_TUNE_PROMOTE_HI_REGS */
1393 /* X86_TUNE_ADD_ESP_4: Enable if add/sub is preferred over 1/2 push/pop. */
1394 m_ATOM | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT4 | m_NOCONA
1395 | m_CORE2 | m_GENERIC,
1397 /* X86_TUNE_ADD_ESP_8 */
1398 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_K6_GEODE | m_386
1399 | m_486 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1401 /* X86_TUNE_SUB_ESP_4 */
1402 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_PENT4 | m_NOCONA | m_CORE2
1405 /* X86_TUNE_SUB_ESP_8 */
1406 m_AMD_MULTIPLE | m_ATOM | m_PPRO | m_386 | m_486
1407 | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1409 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
1410 for DFmode copies */
1411 ~(m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1412 | m_GENERIC | m_GEODE),
1414 /* X86_TUNE_PARTIAL_REG_DEPENDENCY */
1415 m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1417 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: In the Generic model we have a
1418 conflict here in between PPro/Pentium4 based chips that thread 128bit
1419 SSE registers as single units versus K8 based chips that divide SSE
1420 registers to two 64bit halves. This knob promotes all store destinations
1421 to be 128bit to allow register renaming on 128bit SSE units, but usually
1422 results in one extra microop on 64bit SSE units. Experimental results
1423 shows that disabling this option on P4 brings over 20% SPECfp regression,
1424 while enabling it on K8 brings roughly 2.4% regression that can be partly
1425 masked by careful scheduling of moves. */
1426 m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2 | m_GENERIC
1429 /* X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL */
1432 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
1433 are resolved on SSE register parts instead of whole registers, so we may
1434 maintain just lower part of scalar values in proper format leaving the
1435 upper part undefined. */
1438 /* X86_TUNE_SSE_TYPELESS_STORES */
1441 /* X86_TUNE_SSE_LOAD0_BY_PXOR */
1442 m_PPRO | m_PENT4 | m_NOCONA,
1444 /* X86_TUNE_MEMORY_MISMATCH_STALL */
1445 m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_CORE2 | m_GENERIC,
1447 /* X86_TUNE_PROLOGUE_USING_MOVE */
1448 m_ATHLON_K8 | m_ATOM | m_PPRO | m_CORE2 | m_GENERIC,
1450 /* X86_TUNE_EPILOGUE_USING_MOVE */
1451 m_ATHLON_K8 | m_ATOM | m_PPRO | m_CORE2 | m_GENERIC,
1453 /* X86_TUNE_SHIFT1 */
1456 /* X86_TUNE_USE_FFREEP */
1459 /* X86_TUNE_INTER_UNIT_MOVES */
1460 ~(m_AMD_MULTIPLE | m_ATOM | m_GENERIC),
1462 /* X86_TUNE_INTER_UNIT_CONVERSIONS */
1465 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
1466 than 4 branch instructions in the 16 byte window. */
1467 m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_PENT4 | m_NOCONA | m_CORE2
1470 /* X86_TUNE_SCHEDULE */
1471 m_PPRO | m_AMD_MULTIPLE | m_K6_GEODE | m_PENT | m_ATOM | m_CORE2
1474 /* X86_TUNE_USE_BT */
1475 m_AMD_MULTIPLE | m_ATOM | m_CORE2 | m_GENERIC,
1477 /* X86_TUNE_USE_INCDEC */
1478 ~(m_PENT4 | m_NOCONA | m_GENERIC | m_ATOM),
1480 /* X86_TUNE_PAD_RETURNS */
1481 m_AMD_MULTIPLE | m_CORE2 | m_GENERIC,
1483 /* X86_TUNE_EXT_80387_CONSTANTS */
1484 m_K6_GEODE | m_ATHLON_K8 | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO
1485 | m_CORE2 | m_GENERIC,
1487 /* X86_TUNE_SHORTEN_X87_SSE */
1490 /* X86_TUNE_AVOID_VECTOR_DECODE */
1493 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
1494 and SImode multiply, but 386 and 486 do HImode multiply faster. */
1497 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
1498 vector path on AMD machines. */
1499 m_K8 | m_GENERIC64 | m_AMDFAM10,
1501 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
1503 m_K8 | m_GENERIC64 | m_AMDFAM10,
1505 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
1509 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
1510 but one byte longer. */
1513 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
1514 operand that cannot be represented using a modRM byte. The XOR
1515 replacement is long decoded, so this split helps here as well. */
1518 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
1520 m_AMDFAM10 | m_GENERIC,
1522 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
1523 from integer to FP. */
1526 /* X86_TUNE_FUSE_CMP_AND_BRANCH: Fuse a compare or test instruction
1527 with a subsequent conditional jump instruction into a single
1528 compare-and-branch uop. */
1531 /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
1532 will impact LEA instruction selection. */
1536 /* Feature tests against the various architecture variations. */
1537 unsigned char ix86_arch_features[X86_ARCH_LAST];
1539 /* Feature tests against the various architecture variations, used to create
1540 ix86_arch_features based on the processor mask. */
1541 static unsigned int initial_ix86_arch_features[X86_ARCH_LAST] = {
1542 /* X86_ARCH_CMOVE: Conditional move was added for pentiumpro. */
1543 ~(m_386 | m_486 | m_PENT | m_K6),
1545 /* X86_ARCH_CMPXCHG: Compare and exchange was added for 80486. */
1548 /* X86_ARCH_CMPXCHG8B: Compare and exchange 8 bytes was added for pentium. */
1551 /* X86_ARCH_XADD: Exchange and add was added for 80486. */
1554 /* X86_ARCH_BSWAP: Byteswap was added for 80486. */
1558 static const unsigned int x86_accumulate_outgoing_args
1559 = m_AMD_MULTIPLE | m_ATOM | m_PENT4 | m_NOCONA | m_PPRO | m_CORE2
1562 static const unsigned int x86_arch_always_fancy_math_387
1563 = m_PENT | m_ATOM | m_PPRO | m_AMD_MULTIPLE | m_PENT4
1564 | m_NOCONA | m_CORE2 | m_GENERIC;
1566 static enum stringop_alg stringop_alg = no_stringop;
1568 /* In case the average insn count for single function invocation is
1569 lower than this constant, emit fast (but longer) prologue and
1571 #define FAST_PROLOGUE_INSN_COUNT 20
1573 /* Names for 8 (low), 8 (high), and 16-bit registers, respectively. */
1574 static const char *const qi_reg_name[] = QI_REGISTER_NAMES;
1575 static const char *const qi_high_reg_name[] = QI_HIGH_REGISTER_NAMES;
1576 static const char *const hi_reg_name[] = HI_REGISTER_NAMES;
1578 /* Array of the smallest class containing reg number REGNO, indexed by
1579 REGNO. Used by REGNO_REG_CLASS in i386.h. */
1581 enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER] =
1583 /* ax, dx, cx, bx */
1584 AREG, DREG, CREG, BREG,
1585 /* si, di, bp, sp */
1586 SIREG, DIREG, NON_Q_REGS, NON_Q_REGS,
1588 FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS,
1589 FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS,
1592 /* flags, fpsr, fpcr, frame */
1593 NO_REGS, NO_REGS, NO_REGS, NON_Q_REGS,
1595 SSE_FIRST_REG, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1598 MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS, MMX_REGS,
1601 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1602 NON_Q_REGS, NON_Q_REGS, NON_Q_REGS, NON_Q_REGS,
1603 /* SSE REX registers */
1604 SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS, SSE_REGS,
1608 /* The "default" register map used in 32bit mode. */
1610 int const dbx_register_map[FIRST_PSEUDO_REGISTER] =
1612 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */
1613 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */
1614 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1615 21, 22, 23, 24, 25, 26, 27, 28, /* SSE */
1616 29, 30, 31, 32, 33, 34, 35, 36, /* MMX */
1617 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1618 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1621 /* The "default" register map used in 64bit mode. */
1623 int const dbx64_register_map[FIRST_PSEUDO_REGISTER] =
1625 0, 1, 2, 3, 4, 5, 6, 7, /* general regs */
1626 33, 34, 35, 36, 37, 38, 39, 40, /* fp regs */
1627 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1628 17, 18, 19, 20, 21, 22, 23, 24, /* SSE */
1629 41, 42, 43, 44, 45, 46, 47, 48, /* MMX */
1630 8,9,10,11,12,13,14,15, /* extended integer registers */
1631 25, 26, 27, 28, 29, 30, 31, 32, /* extended SSE registers */
1634 /* Define the register numbers to be used in Dwarf debugging information.
1635 The SVR4 reference port C compiler uses the following register numbers
1636 in its Dwarf output code:
1637 0 for %eax (gcc regno = 0)
1638 1 for %ecx (gcc regno = 2)
1639 2 for %edx (gcc regno = 1)
1640 3 for %ebx (gcc regno = 3)
1641 4 for %esp (gcc regno = 7)
1642 5 for %ebp (gcc regno = 6)
1643 6 for %esi (gcc regno = 4)
1644 7 for %edi (gcc regno = 5)
1645 The following three DWARF register numbers are never generated by
1646 the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4
1647 believes these numbers have these meanings.
1648 8 for %eip (no gcc equivalent)
1649 9 for %eflags (gcc regno = 17)
1650 10 for %trapno (no gcc equivalent)
1651 It is not at all clear how we should number the FP stack registers
1652 for the x86 architecture. If the version of SDB on x86/svr4 were
1653 a bit less brain dead with respect to floating-point then we would
1654 have a precedent to follow with respect to DWARF register numbers
1655 for x86 FP registers, but the SDB on x86/svr4 is so completely
1656 broken with respect to FP registers that it is hardly worth thinking
1657 of it as something to strive for compatibility with.
1658 The version of x86/svr4 SDB I have at the moment does (partially)
1659 seem to believe that DWARF register number 11 is associated with
1660 the x86 register %st(0), but that's about all. Higher DWARF
1661 register numbers don't seem to be associated with anything in
1662 particular, and even for DWARF regno 11, SDB only seems to under-
1663 stand that it should say that a variable lives in %st(0) (when
1664 asked via an `=' command) if we said it was in DWARF regno 11,
1665 but SDB still prints garbage when asked for the value of the
1666 variable in question (via a `/' command).
1667 (Also note that the labels SDB prints for various FP stack regs
1668 when doing an `x' command are all wrong.)
1669 Note that these problems generally don't affect the native SVR4
1670 C compiler because it doesn't allow the use of -O with -g and
1671 because when it is *not* optimizing, it allocates a memory
1672 location for each floating-point variable, and the memory
1673 location is what gets described in the DWARF AT_location
1674 attribute for the variable in question.
1675 Regardless of the severe mental illness of the x86/svr4 SDB, we
1676 do something sensible here and we use the following DWARF
1677 register numbers. Note that these are all stack-top-relative
1679 11 for %st(0) (gcc regno = 8)
1680 12 for %st(1) (gcc regno = 9)
1681 13 for %st(2) (gcc regno = 10)
1682 14 for %st(3) (gcc regno = 11)
1683 15 for %st(4) (gcc regno = 12)
1684 16 for %st(5) (gcc regno = 13)
1685 17 for %st(6) (gcc regno = 14)
1686 18 for %st(7) (gcc regno = 15)
1688 int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER] =
1690 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */
1691 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */
1692 -1, 9, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1693 21, 22, 23, 24, 25, 26, 27, 28, /* SSE registers */
1694 29, 30, 31, 32, 33, 34, 35, 36, /* MMX registers */
1695 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1696 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1699 /* Test and compare insns in i386.md store the information needed to
1700 generate branch and scc insns here. */
1702 rtx ix86_compare_op0 = NULL_RTX;
1703 rtx ix86_compare_op1 = NULL_RTX;
1705 /* Define parameter passing and return registers. */
1707 static int const x86_64_int_parameter_registers[6] =
1709 DI_REG, SI_REG, DX_REG, CX_REG, R8_REG, R9_REG
1712 static int const x86_64_ms_abi_int_parameter_registers[4] =
1714 CX_REG, DX_REG, R8_REG, R9_REG
1717 static int const x86_64_int_return_registers[4] =
1719 AX_REG, DX_REG, DI_REG, SI_REG
1722 /* Define the structure for the machine field in struct function. */
1724 struct GTY(()) stack_local_entry {
1725 unsigned short mode;
1728 struct stack_local_entry *next;
1731 /* Structure describing stack frame layout.
1732 Stack grows downward:
1738 saved frame pointer if frame_pointer_needed
1739 <- HARD_FRAME_POINTER
1748 [va_arg registers] (
1749 > to_allocate <- FRAME_POINTER
1761 HOST_WIDE_INT frame;
1763 int outgoing_arguments_size;
1766 HOST_WIDE_INT to_allocate;
1767 /* The offsets relative to ARG_POINTER. */
1768 HOST_WIDE_INT frame_pointer_offset;
1769 HOST_WIDE_INT hard_frame_pointer_offset;
1770 HOST_WIDE_INT stack_pointer_offset;
1772 /* When save_regs_using_mov is set, emit prologue using
1773 move instead of push instructions. */
1774 bool save_regs_using_mov;
1777 /* Code model option. */
1778 enum cmodel ix86_cmodel;
1780 enum asm_dialect ix86_asm_dialect = ASM_ATT;
1782 enum tls_dialect ix86_tls_dialect = TLS_DIALECT_GNU;
1784 /* Which unit we are generating floating point math for. */
1785 enum fpmath_unit ix86_fpmath;
1787 /* Which cpu are we scheduling for. */
1788 enum attr_cpu ix86_schedule;
1790 /* Which cpu are we optimizing for. */
1791 enum processor_type ix86_tune;
1793 /* Which instruction set architecture to use. */
1794 enum processor_type ix86_arch;
1796 /* true if sse prefetch instruction is not NOOP. */
1797 int x86_prefetch_sse;
1799 /* ix86_regparm_string as a number */
1800 static int ix86_regparm;
1802 /* -mstackrealign option */
1803 extern int ix86_force_align_arg_pointer;
1804 static const char ix86_force_align_arg_pointer_string[]
1805 = "force_align_arg_pointer";
1807 static rtx (*ix86_gen_leave) (void);
1808 static rtx (*ix86_gen_pop1) (rtx);
1809 static rtx (*ix86_gen_add3) (rtx, rtx, rtx);
1810 static rtx (*ix86_gen_sub3) (rtx, rtx, rtx);
1811 static rtx (*ix86_gen_sub3_carry) (rtx, rtx, rtx, rtx);
1812 static rtx (*ix86_gen_one_cmpl2) (rtx, rtx);
1813 static rtx (*ix86_gen_monitor) (rtx, rtx, rtx);
1814 static rtx (*ix86_gen_andsp) (rtx, rtx, rtx);
1816 /* Preferred alignment for stack boundary in bits. */
1817 unsigned int ix86_preferred_stack_boundary;
1819 /* Alignment for incoming stack boundary in bits specified at
1821 static unsigned int ix86_user_incoming_stack_boundary;
1823 /* Default alignment for incoming stack boundary in bits. */
1824 static unsigned int ix86_default_incoming_stack_boundary;
1826 /* Alignment for incoming stack boundary in bits. */
1827 unsigned int ix86_incoming_stack_boundary;
1829 /* The abi used by target. */
1830 enum calling_abi ix86_abi;
1832 /* Values 1-5: see jump.c */
1833 int ix86_branch_cost;
1835 /* Calling abi specific va_list type nodes. */
1836 static GTY(()) tree sysv_va_list_type_node;
1837 static GTY(()) tree ms_va_list_type_node;
1839 /* Variables which are this size or smaller are put in the data/bss
1840 or ldata/lbss sections. */
1842 int ix86_section_threshold = 65536;
1844 /* Prefix built by ASM_GENERATE_INTERNAL_LABEL. */
1845 char internal_label_prefix[16];
1846 int internal_label_prefix_len;
1848 /* Fence to use after loop using movnt. */
1851 /* Register class used for passing given 64bit part of the argument.
1852 These represent classes as documented by the PS ABI, with the exception
1853 of SSESF, SSEDF classes, that are basically SSE class, just gcc will
1854 use SF or DFmode move instead of DImode to avoid reformatting penalties.
1856 Similarly we play games with INTEGERSI_CLASS to use cheaper SImode moves
1857 whenever possible (upper half does contain padding). */
1858 enum x86_64_reg_class
1861 X86_64_INTEGER_CLASS,
1862 X86_64_INTEGERSI_CLASS,
1869 X86_64_COMPLEX_X87_CLASS,
1873 #define MAX_CLASSES 4
1875 /* Table of constants used by fldpi, fldln2, etc.... */
1876 static REAL_VALUE_TYPE ext_80387_constants_table [5];
1877 static bool ext_80387_constants_init = 0;
1880 static struct machine_function * ix86_init_machine_status (void);
1881 static rtx ix86_function_value (const_tree, const_tree, bool);
1882 static int ix86_function_regparm (const_tree, const_tree);
1883 static void ix86_compute_frame_layout (struct ix86_frame *);
1884 static bool ix86_expand_vector_init_one_nonzero (bool, enum machine_mode,
1886 static void ix86_add_new_builtins (int);
1888 enum ix86_function_specific_strings
1890 IX86_FUNCTION_SPECIFIC_ARCH,
1891 IX86_FUNCTION_SPECIFIC_TUNE,
1892 IX86_FUNCTION_SPECIFIC_FPMATH,
1893 IX86_FUNCTION_SPECIFIC_MAX
1896 static char *ix86_target_string (int, int, const char *, const char *,
1897 const char *, bool);
1898 static void ix86_debug_options (void) ATTRIBUTE_UNUSED;
1899 static void ix86_function_specific_save (struct cl_target_option *);
1900 static void ix86_function_specific_restore (struct cl_target_option *);
1901 static void ix86_function_specific_print (FILE *, int,
1902 struct cl_target_option *);
1903 static bool ix86_valid_target_attribute_p (tree, tree, tree, int);
1904 static bool ix86_valid_target_attribute_inner_p (tree, char *[]);
1905 static bool ix86_can_inline_p (tree, tree);
1906 static void ix86_set_current_function (tree);
1908 static enum calling_abi ix86_function_abi (const_tree);
1911 /* The svr4 ABI for the i386 says that records and unions are returned
1913 #ifndef DEFAULT_PCC_STRUCT_RETURN
1914 #define DEFAULT_PCC_STRUCT_RETURN 1
1917 /* Whether -mtune= or -march= were specified */
1918 static int ix86_tune_defaulted;
1919 static int ix86_arch_specified;
1921 /* Bit flags that specify the ISA we are compiling for. */
1922 int ix86_isa_flags = TARGET_64BIT_DEFAULT | TARGET_SUBTARGET_ISA_DEFAULT;
1924 /* A mask of ix86_isa_flags that includes bit X if X
1925 was set or cleared on the command line. */
1926 static int ix86_isa_flags_explicit;
1928 /* Define a set of ISAs which are available when a given ISA is
1929 enabled. MMX and SSE ISAs are handled separately. */
1931 #define OPTION_MASK_ISA_MMX_SET OPTION_MASK_ISA_MMX
1932 #define OPTION_MASK_ISA_3DNOW_SET \
1933 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_MMX_SET)
1935 #define OPTION_MASK_ISA_SSE_SET OPTION_MASK_ISA_SSE
1936 #define OPTION_MASK_ISA_SSE2_SET \
1937 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE_SET)
1938 #define OPTION_MASK_ISA_SSE3_SET \
1939 (OPTION_MASK_ISA_SSE3 | OPTION_MASK_ISA_SSE2_SET)
1940 #define OPTION_MASK_ISA_SSSE3_SET \
1941 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE3_SET)
1942 #define OPTION_MASK_ISA_SSE4_1_SET \
1943 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSSE3_SET)
1944 #define OPTION_MASK_ISA_SSE4_2_SET \
1945 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_SSE4_1_SET)
1946 #define OPTION_MASK_ISA_AVX_SET \
1947 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_SSE4_2_SET)
1948 #define OPTION_MASK_ISA_FMA_SET \
1949 (OPTION_MASK_ISA_FMA | OPTION_MASK_ISA_AVX_SET)
1951 /* SSE4 includes both SSE4.1 and SSE4.2. -msse4 should be the same
1953 #define OPTION_MASK_ISA_SSE4_SET OPTION_MASK_ISA_SSE4_2_SET
1955 #define OPTION_MASK_ISA_SSE4A_SET \
1956 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE3_SET)
1957 #define OPTION_MASK_ISA_SSE5_SET \
1958 (OPTION_MASK_ISA_SSE5 | OPTION_MASK_ISA_SSE4A_SET)
1960 /* AES and PCLMUL need SSE2 because they use xmm registers */
1961 #define OPTION_MASK_ISA_AES_SET \
1962 (OPTION_MASK_ISA_AES | OPTION_MASK_ISA_SSE2_SET)
1963 #define OPTION_MASK_ISA_PCLMUL_SET \
1964 (OPTION_MASK_ISA_PCLMUL | OPTION_MASK_ISA_SSE2_SET)
1966 #define OPTION_MASK_ISA_ABM_SET \
1967 (OPTION_MASK_ISA_ABM | OPTION_MASK_ISA_POPCNT)
1969 #define OPTION_MASK_ISA_POPCNT_SET OPTION_MASK_ISA_POPCNT
1970 #define OPTION_MASK_ISA_CX16_SET OPTION_MASK_ISA_CX16
1971 #define OPTION_MASK_ISA_SAHF_SET OPTION_MASK_ISA_SAHF
1972 #define OPTION_MASK_ISA_MOVBE_SET OPTION_MASK_ISA_MOVBE
1973 #define OPTION_MASK_ISA_CRC32_SET OPTION_MASK_ISA_CRC32
1975 /* Define a set of ISAs which aren't available when a given ISA is
1976 disabled. MMX and SSE ISAs are handled separately. */
1978 #define OPTION_MASK_ISA_MMX_UNSET \
1979 (OPTION_MASK_ISA_MMX | OPTION_MASK_ISA_3DNOW_UNSET)
1980 #define OPTION_MASK_ISA_3DNOW_UNSET \
1981 (OPTION_MASK_ISA_3DNOW | OPTION_MASK_ISA_3DNOW_A_UNSET)
1982 #define OPTION_MASK_ISA_3DNOW_A_UNSET OPTION_MASK_ISA_3DNOW_A
1984 #define OPTION_MASK_ISA_SSE_UNSET \
1985 (OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_SSE2_UNSET)
1986 #define OPTION_MASK_ISA_SSE2_UNSET \
1987 (OPTION_MASK_ISA_SSE2 | OPTION_MASK_ISA_SSE3_UNSET)
1988 #define OPTION_MASK_ISA_SSE3_UNSET \
1989 (OPTION_MASK_ISA_SSE3 \
1990 | OPTION_MASK_ISA_SSSE3_UNSET \
1991 | OPTION_MASK_ISA_SSE4A_UNSET )
1992 #define OPTION_MASK_ISA_SSSE3_UNSET \
1993 (OPTION_MASK_ISA_SSSE3 | OPTION_MASK_ISA_SSE4_1_UNSET)
1994 #define OPTION_MASK_ISA_SSE4_1_UNSET \
1995 (OPTION_MASK_ISA_SSE4_1 | OPTION_MASK_ISA_SSE4_2_UNSET)
1996 #define OPTION_MASK_ISA_SSE4_2_UNSET \
1997 (OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_AVX_UNSET )
1998 #define OPTION_MASK_ISA_AVX_UNSET \
1999 (OPTION_MASK_ISA_AVX | OPTION_MASK_ISA_FMA_UNSET)
2000 #define OPTION_MASK_ISA_FMA_UNSET OPTION_MASK_ISA_FMA
2002 /* SSE4 includes both SSE4.1 and SSE4.2. -mno-sse4 should the same
2004 #define OPTION_MASK_ISA_SSE4_UNSET OPTION_MASK_ISA_SSE4_1_UNSET
2006 #define OPTION_MASK_ISA_SSE4A_UNSET \
2007 (OPTION_MASK_ISA_SSE4A | OPTION_MASK_ISA_SSE5_UNSET)
2008 #define OPTION_MASK_ISA_SSE5_UNSET OPTION_MASK_ISA_SSE5
2009 #define OPTION_MASK_ISA_AES_UNSET OPTION_MASK_ISA_AES
2010 #define OPTION_MASK_ISA_PCLMUL_UNSET OPTION_MASK_ISA_PCLMUL
2011 #define OPTION_MASK_ISA_ABM_UNSET OPTION_MASK_ISA_ABM
2012 #define OPTION_MASK_ISA_POPCNT_UNSET OPTION_MASK_ISA_POPCNT
2013 #define OPTION_MASK_ISA_CX16_UNSET OPTION_MASK_ISA_CX16
2014 #define OPTION_MASK_ISA_SAHF_UNSET OPTION_MASK_ISA_SAHF
2015 #define OPTION_MASK_ISA_MOVBE_UNSET OPTION_MASK_ISA_MOVBE
2016 #define OPTION_MASK_ISA_CRC32_UNSET OPTION_MASK_ISA_CRC32
2018 /* Vectorization library interface and handlers. */
2019 tree (*ix86_veclib_handler)(enum built_in_function, tree, tree) = NULL;
2020 static tree ix86_veclibabi_svml (enum built_in_function, tree, tree);
2021 static tree ix86_veclibabi_acml (enum built_in_function, tree, tree);
2023 /* Processor target table, indexed by processor number */
2026 const struct processor_costs *cost; /* Processor costs */
2027 const int align_loop; /* Default alignments. */
2028 const int align_loop_max_skip;
2029 const int align_jump;
2030 const int align_jump_max_skip;
2031 const int align_func;
2034 static const struct ptt processor_target_table[PROCESSOR_max] =
2036 {&i386_cost, 4, 3, 4, 3, 4},
2037 {&i486_cost, 16, 15, 16, 15, 16},
2038 {&pentium_cost, 16, 7, 16, 7, 16},
2039 {&pentiumpro_cost, 16, 15, 16, 10, 16},
2040 {&geode_cost, 0, 0, 0, 0, 0},
2041 {&k6_cost, 32, 7, 32, 7, 32},
2042 {&athlon_cost, 16, 7, 16, 7, 16},
2043 {&pentium4_cost, 0, 0, 0, 0, 0},
2044 {&k8_cost, 16, 7, 16, 7, 16},
2045 {&nocona_cost, 0, 0, 0, 0, 0},
2046 {&core2_cost, 16, 10, 16, 10, 16},
2047 {&generic32_cost, 16, 7, 16, 7, 16},
2048 {&generic64_cost, 16, 10, 16, 10, 16},
2049 {&amdfam10_cost, 32, 24, 32, 7, 32},
2050 {&atom_cost, 16, 7, 16, 7, 16}
2053 static const char *const cpu_names[TARGET_CPU_DEFAULT_max] =
2079 /* Implement TARGET_HANDLE_OPTION. */
2082 ix86_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, int value)
2089 ix86_isa_flags |= OPTION_MASK_ISA_MMX_SET;
2090 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_SET;
2094 ix86_isa_flags &= ~OPTION_MASK_ISA_MMX_UNSET;
2095 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MMX_UNSET;
2102 ix86_isa_flags |= OPTION_MASK_ISA_3DNOW_SET;
2103 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_SET;
2107 ix86_isa_flags &= ~OPTION_MASK_ISA_3DNOW_UNSET;
2108 ix86_isa_flags_explicit |= OPTION_MASK_ISA_3DNOW_UNSET;
2118 ix86_isa_flags |= OPTION_MASK_ISA_SSE_SET;
2119 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_SET;
2123 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE_UNSET;
2124 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE_UNSET;
2131 ix86_isa_flags |= OPTION_MASK_ISA_SSE2_SET;
2132 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_SET;
2136 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE2_UNSET;
2137 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE2_UNSET;
2144 ix86_isa_flags |= OPTION_MASK_ISA_SSE3_SET;
2145 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_SET;
2149 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE3_UNSET;
2150 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE3_UNSET;
2157 ix86_isa_flags |= OPTION_MASK_ISA_SSSE3_SET;
2158 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_SET;
2162 ix86_isa_flags &= ~OPTION_MASK_ISA_SSSE3_UNSET;
2163 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSSE3_UNSET;
2170 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_1_SET;
2171 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_SET;
2175 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_1_UNSET;
2176 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_1_UNSET;
2183 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_2_SET;
2184 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_SET;
2188 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_2_UNSET;
2189 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_2_UNSET;
2196 ix86_isa_flags |= OPTION_MASK_ISA_AVX_SET;
2197 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_SET;
2201 ix86_isa_flags &= ~OPTION_MASK_ISA_AVX_UNSET;
2202 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AVX_UNSET;
2209 ix86_isa_flags |= OPTION_MASK_ISA_FMA_SET;
2210 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_SET;
2214 ix86_isa_flags &= ~OPTION_MASK_ISA_FMA_UNSET;
2215 ix86_isa_flags_explicit |= OPTION_MASK_ISA_FMA_UNSET;
2220 ix86_isa_flags |= OPTION_MASK_ISA_SSE4_SET;
2221 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_SET;
2225 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4_UNSET;
2226 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4_UNSET;
2232 ix86_isa_flags |= OPTION_MASK_ISA_SSE4A_SET;
2233 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_SET;
2237 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE4A_UNSET;
2238 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE4A_UNSET;
2245 ix86_isa_flags |= OPTION_MASK_ISA_SSE5_SET;
2246 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_SET;
2250 ix86_isa_flags &= ~OPTION_MASK_ISA_SSE5_UNSET;
2251 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SSE5_UNSET;
2258 ix86_isa_flags |= OPTION_MASK_ISA_ABM_SET;
2259 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_SET;
2263 ix86_isa_flags &= ~OPTION_MASK_ISA_ABM_UNSET;
2264 ix86_isa_flags_explicit |= OPTION_MASK_ISA_ABM_UNSET;
2271 ix86_isa_flags |= OPTION_MASK_ISA_POPCNT_SET;
2272 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_SET;
2276 ix86_isa_flags &= ~OPTION_MASK_ISA_POPCNT_UNSET;
2277 ix86_isa_flags_explicit |= OPTION_MASK_ISA_POPCNT_UNSET;
2284 ix86_isa_flags |= OPTION_MASK_ISA_SAHF_SET;
2285 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_SET;
2289 ix86_isa_flags &= ~OPTION_MASK_ISA_SAHF_UNSET;
2290 ix86_isa_flags_explicit |= OPTION_MASK_ISA_SAHF_UNSET;
2297 ix86_isa_flags |= OPTION_MASK_ISA_CX16_SET;
2298 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_SET;
2302 ix86_isa_flags &= ~OPTION_MASK_ISA_CX16_UNSET;
2303 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CX16_UNSET;
2310 ix86_isa_flags |= OPTION_MASK_ISA_MOVBE_SET;
2311 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MOVBE_SET;
2315 ix86_isa_flags &= ~OPTION_MASK_ISA_MOVBE_UNSET;
2316 ix86_isa_flags_explicit |= OPTION_MASK_ISA_MOVBE_UNSET;
2323 ix86_isa_flags |= OPTION_MASK_ISA_CRC32_SET;
2324 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CRC32_SET;
2328 ix86_isa_flags &= ~OPTION_MASK_ISA_CRC32_UNSET;
2329 ix86_isa_flags_explicit |= OPTION_MASK_ISA_CRC32_UNSET;
2336 ix86_isa_flags |= OPTION_MASK_ISA_AES_SET;
2337 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_SET;
2341 ix86_isa_flags &= ~OPTION_MASK_ISA_AES_UNSET;
2342 ix86_isa_flags_explicit |= OPTION_MASK_ISA_AES_UNSET;
2349 ix86_isa_flags |= OPTION_MASK_ISA_PCLMUL_SET;
2350 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_SET;
2354 ix86_isa_flags &= ~OPTION_MASK_ISA_PCLMUL_UNSET;
2355 ix86_isa_flags_explicit |= OPTION_MASK_ISA_PCLMUL_UNSET;
2364 /* Return a string the documents the current -m options. The caller is
2365 responsible for freeing the string. */
2368 ix86_target_string (int isa, int flags, const char *arch, const char *tune,
2369 const char *fpmath, bool add_nl_p)
2371 struct ix86_target_opts
2373 const char *option; /* option string */
2374 int mask; /* isa mask options */
2377 /* This table is ordered so that options like -msse5 or -msse4.2 that imply
2378 preceding options while match those first. */
2379 static struct ix86_target_opts isa_opts[] =
2381 { "-m64", OPTION_MASK_ISA_64BIT },
2382 { "-msse5", OPTION_MASK_ISA_SSE5 },
2383 { "-msse4a", OPTION_MASK_ISA_SSE4A },
2384 { "-msse4.2", OPTION_MASK_ISA_SSE4_2 },
2385 { "-msse4.1", OPTION_MASK_ISA_SSE4_1 },
2386 { "-mssse3", OPTION_MASK_ISA_SSSE3 },
2387 { "-msse3", OPTION_MASK_ISA_SSE3 },
2388 { "-msse2", OPTION_MASK_ISA_SSE2 },
2389 { "-msse", OPTION_MASK_ISA_SSE },
2390 { "-m3dnow", OPTION_MASK_ISA_3DNOW },
2391 { "-m3dnowa", OPTION_MASK_ISA_3DNOW_A },
2392 { "-mmmx", OPTION_MASK_ISA_MMX },
2393 { "-mabm", OPTION_MASK_ISA_ABM },
2394 { "-mpopcnt", OPTION_MASK_ISA_POPCNT },
2395 { "-mmovbe", OPTION_MASK_ISA_MOVBE },
2396 { "-mcrc32", OPTION_MASK_ISA_CRC32 },
2397 { "-maes", OPTION_MASK_ISA_AES },
2398 { "-mpclmul", OPTION_MASK_ISA_PCLMUL },
2402 static struct ix86_target_opts flag_opts[] =
2404 { "-m128bit-long-double", MASK_128BIT_LONG_DOUBLE },
2405 { "-m80387", MASK_80387 },
2406 { "-maccumulate-outgoing-args", MASK_ACCUMULATE_OUTGOING_ARGS },
2407 { "-malign-double", MASK_ALIGN_DOUBLE },
2408 { "-mcld", MASK_CLD },
2409 { "-mfp-ret-in-387", MASK_FLOAT_RETURNS },
2410 { "-mieee-fp", MASK_IEEE_FP },
2411 { "-minline-all-stringops", MASK_INLINE_ALL_STRINGOPS },
2412 { "-minline-stringops-dynamically", MASK_INLINE_STRINGOPS_DYNAMICALLY },
2413 { "-mms-bitfields", MASK_MS_BITFIELD_LAYOUT },
2414 { "-mno-align-stringops", MASK_NO_ALIGN_STRINGOPS },
2415 { "-mno-fancy-math-387", MASK_NO_FANCY_MATH_387 },
2416 { "-mno-fused-madd", MASK_NO_FUSED_MADD },
2417 { "-mno-push-args", MASK_NO_PUSH_ARGS },
2418 { "-mno-red-zone", MASK_NO_RED_ZONE },
2419 { "-momit-leaf-frame-pointer", MASK_OMIT_LEAF_FRAME_POINTER },
2420 { "-mrecip", MASK_RECIP },
2421 { "-mrtd", MASK_RTD },
2422 { "-msseregparm", MASK_SSEREGPARM },
2423 { "-mstack-arg-probe", MASK_STACK_PROBE },
2424 { "-mtls-direct-seg-refs", MASK_TLS_DIRECT_SEG_REFS },
2427 const char *opts[ARRAY_SIZE (isa_opts) + ARRAY_SIZE (flag_opts) + 6][2];
2430 char target_other[40];
2439 memset (opts, '\0', sizeof (opts));
2441 /* Add -march= option. */
2444 opts[num][0] = "-march=";
2445 opts[num++][1] = arch;
2448 /* Add -mtune= option. */
2451 opts[num][0] = "-mtune=";
2452 opts[num++][1] = tune;
2455 /* Pick out the options in isa options. */
2456 for (i = 0; i < ARRAY_SIZE (isa_opts); i++)
2458 if ((isa & isa_opts[i].mask) != 0)
2460 opts[num++][0] = isa_opts[i].option;
2461 isa &= ~ isa_opts[i].mask;
2465 if (isa && add_nl_p)
2467 opts[num++][0] = isa_other;
2468 sprintf (isa_other, "(other isa: 0x%x)", isa);
2471 /* Add flag options. */
2472 for (i = 0; i < ARRAY_SIZE (flag_opts); i++)
2474 if ((flags & flag_opts[i].mask) != 0)
2476 opts[num++][0] = flag_opts[i].option;
2477 flags &= ~ flag_opts[i].mask;
2481 if (flags && add_nl_p)
2483 opts[num++][0] = target_other;
2484 sprintf (target_other, "(other flags: 0x%x)", isa);
2487 /* Add -fpmath= option. */
2490 opts[num][0] = "-mfpmath=";
2491 opts[num++][1] = fpmath;
2498 gcc_assert (num < ARRAY_SIZE (opts));
2500 /* Size the string. */
2502 sep_len = (add_nl_p) ? 3 : 1;
2503 for (i = 0; i < num; i++)
2506 for (j = 0; j < 2; j++)
2508 len += strlen (opts[i][j]);
2511 /* Build the string. */
2512 ret = ptr = (char *) xmalloc (len);
2515 for (i = 0; i < num; i++)
2519 for (j = 0; j < 2; j++)
2520 len2[j] = (opts[i][j]) ? strlen (opts[i][j]) : 0;
2527 if (add_nl_p && line_len + len2[0] + len2[1] > 70)
2535 for (j = 0; j < 2; j++)
2538 memcpy (ptr, opts[i][j], len2[j]);
2540 line_len += len2[j];
2545 gcc_assert (ret + len >= ptr);
2550 /* Function that is callable from the debugger to print the current
2553 ix86_debug_options (void)
2555 char *opts = ix86_target_string (ix86_isa_flags, target_flags,
2556 ix86_arch_string, ix86_tune_string,
2557 ix86_fpmath_string, true);
2561 fprintf (stderr, "%s\n\n", opts);
2565 fputs ("<no options>\n\n", stderr);
2570 /* Sometimes certain combinations of command options do not make
2571 sense on a particular target machine. You can define a macro
2572 `OVERRIDE_OPTIONS' to take account of this. This macro, if
2573 defined, is executed once just after all the command options have
2576 Don't use this macro to turn on various extra optimizations for
2577 `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */
2580 override_options (bool main_args_p)
2583 unsigned int ix86_arch_mask, ix86_tune_mask;
2588 /* Comes from final.c -- no real reason to change it. */
2589 #define MAX_CODE_ALIGN 16
2597 PTA_PREFETCH_SSE = 1 << 4,
2599 PTA_3DNOW_A = 1 << 6,
2603 PTA_POPCNT = 1 << 10,
2605 PTA_SSE4A = 1 << 12,
2606 PTA_NO_SAHF = 1 << 13,
2607 PTA_SSE4_1 = 1 << 14,
2608 PTA_SSE4_2 = 1 << 15,
2611 PTA_PCLMUL = 1 << 18,
2619 const char *const name; /* processor name or nickname. */
2620 const enum processor_type processor;
2621 const enum attr_cpu schedule;
2622 const unsigned /*enum pta_flags*/ flags;
2624 const processor_alias_table[] =
2626 {"i386", PROCESSOR_I386, CPU_NONE, 0},
2627 {"i486", PROCESSOR_I486, CPU_NONE, 0},
2628 {"i586", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2629 {"pentium", PROCESSOR_PENTIUM, CPU_PENTIUM, 0},
2630 {"pentium-mmx", PROCESSOR_PENTIUM, CPU_PENTIUM, PTA_MMX},
2631 {"winchip-c6", PROCESSOR_I486, CPU_NONE, PTA_MMX},
2632 {"winchip2", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2633 {"c3", PROCESSOR_I486, CPU_NONE, PTA_MMX | PTA_3DNOW},
2634 {"c3-2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX | PTA_SSE},
2635 {"i686", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2636 {"pentiumpro", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, 0},
2637 {"pentium2", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO, PTA_MMX},
2638 {"pentium3", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2640 {"pentium3m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2642 {"pentium-m", PROCESSOR_PENTIUMPRO, CPU_PENTIUMPRO,
2643 PTA_MMX | PTA_SSE | PTA_SSE2},
2644 {"pentium4", PROCESSOR_PENTIUM4, CPU_NONE,
2645 PTA_MMX |PTA_SSE | PTA_SSE2},
2646 {"pentium4m", PROCESSOR_PENTIUM4, CPU_NONE,
2647 PTA_MMX | PTA_SSE | PTA_SSE2},
2648 {"prescott", PROCESSOR_NOCONA, CPU_NONE,
2649 PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3},
2650 {"nocona", PROCESSOR_NOCONA, CPU_NONE,
2651 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2652 | PTA_CX16 | PTA_NO_SAHF},
2653 {"core2", PROCESSOR_CORE2, CPU_CORE2,
2654 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2655 | PTA_SSSE3 | PTA_CX16},
2656 {"atom", PROCESSOR_ATOM, CPU_ATOM,
2657 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_SSE3
2658 | PTA_SSSE3 | PTA_CX16 | PTA_MOVBE},
2659 {"geode", PROCESSOR_GEODE, CPU_GEODE,
2660 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A |PTA_PREFETCH_SSE},
2661 {"k6", PROCESSOR_K6, CPU_K6, PTA_MMX},
2662 {"k6-2", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2663 {"k6-3", PROCESSOR_K6, CPU_K6, PTA_MMX | PTA_3DNOW},
2664 {"athlon", PROCESSOR_ATHLON, CPU_ATHLON,
2665 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2666 {"athlon-tbird", PROCESSOR_ATHLON, CPU_ATHLON,
2667 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_PREFETCH_SSE},
2668 {"athlon-4", PROCESSOR_ATHLON, CPU_ATHLON,
2669 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2670 {"athlon-xp", PROCESSOR_ATHLON, CPU_ATHLON,
2671 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2672 {"athlon-mp", PROCESSOR_ATHLON, CPU_ATHLON,
2673 PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE},
2674 {"x86-64", PROCESSOR_K8, CPU_K8,
2675 PTA_64BIT | PTA_MMX | PTA_SSE | PTA_SSE2 | PTA_NO_SAHF},
2676 {"k8", PROCESSOR_K8, CPU_K8,
2677 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2678 | PTA_SSE2 | PTA_NO_SAHF},
2679 {"k8-sse3", PROCESSOR_K8, CPU_K8,
2680 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2681 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2682 {"opteron", PROCESSOR_K8, CPU_K8,
2683 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2684 | PTA_SSE2 | PTA_NO_SAHF},
2685 {"opteron-sse3", PROCESSOR_K8, CPU_K8,
2686 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2687 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2688 {"athlon64", PROCESSOR_K8, CPU_K8,
2689 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2690 | PTA_SSE2 | PTA_NO_SAHF},
2691 {"athlon64-sse3", PROCESSOR_K8, CPU_K8,
2692 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2693 | PTA_SSE2 | PTA_SSE3 | PTA_NO_SAHF},
2694 {"athlon-fx", PROCESSOR_K8, CPU_K8,
2695 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2696 | PTA_SSE2 | PTA_NO_SAHF},
2697 {"amdfam10", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2698 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2699 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2700 {"barcelona", PROCESSOR_AMDFAM10, CPU_AMDFAM10,
2701 PTA_64BIT | PTA_MMX | PTA_3DNOW | PTA_3DNOW_A | PTA_SSE
2702 | PTA_SSE2 | PTA_SSE3 | PTA_SSE4A | PTA_CX16 | PTA_ABM},
2703 {"generic32", PROCESSOR_GENERIC32, CPU_PENTIUMPRO,
2704 0 /* flags are only used for -march switch. */ },
2705 {"generic64", PROCESSOR_GENERIC64, CPU_GENERIC64,
2706 PTA_64BIT /* flags are only used for -march switch. */ },
2709 int const pta_size = ARRAY_SIZE (processor_alias_table);
2711 /* Set up prefix/suffix so the error messages refer to either the command
2712 line argument, or the attribute(target). */
2721 prefix = "option(\"";
2726 #ifdef SUBTARGET_OVERRIDE_OPTIONS
2727 SUBTARGET_OVERRIDE_OPTIONS;
2730 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
2731 SUBSUBTARGET_OVERRIDE_OPTIONS;
2734 /* -fPIC is the default for x86_64. */
2735 if (TARGET_MACHO && TARGET_64BIT)
2738 /* Set the default values for switches whose default depends on TARGET_64BIT
2739 in case they weren't overwritten by command line options. */
2742 /* Mach-O doesn't support omitting the frame pointer for now. */
2743 if (flag_omit_frame_pointer == 2)
2744 flag_omit_frame_pointer = (TARGET_MACHO ? 0 : 1);
2745 if (flag_asynchronous_unwind_tables == 2)
2746 flag_asynchronous_unwind_tables = 1;
2747 if (flag_pcc_struct_return == 2)
2748 flag_pcc_struct_return = 0;
2752 if (flag_omit_frame_pointer == 2)
2753 flag_omit_frame_pointer = 0;
2754 if (flag_asynchronous_unwind_tables == 2)
2755 flag_asynchronous_unwind_tables = 0;
2756 if (flag_pcc_struct_return == 2)
2757 flag_pcc_struct_return = DEFAULT_PCC_STRUCT_RETURN;
2760 /* Need to check -mtune=generic first. */
2761 if (ix86_tune_string)
2763 if (!strcmp (ix86_tune_string, "generic")
2764 || !strcmp (ix86_tune_string, "i686")
2765 /* As special support for cross compilers we read -mtune=native
2766 as -mtune=generic. With native compilers we won't see the
2767 -mtune=native, as it was changed by the driver. */
2768 || !strcmp (ix86_tune_string, "native"))
2771 ix86_tune_string = "generic64";
2773 ix86_tune_string = "generic32";
2775 /* If this call is for setting the option attribute, allow the
2776 generic32/generic64 that was previously set. */
2777 else if (!main_args_p
2778 && (!strcmp (ix86_tune_string, "generic32")
2779 || !strcmp (ix86_tune_string, "generic64")))
2781 else if (!strncmp (ix86_tune_string, "generic", 7))
2782 error ("bad value (%s) for %stune=%s %s",
2783 ix86_tune_string, prefix, suffix, sw);
2787 if (ix86_arch_string)
2788 ix86_tune_string = ix86_arch_string;
2789 if (!ix86_tune_string)
2791 ix86_tune_string = cpu_names[TARGET_CPU_DEFAULT];
2792 ix86_tune_defaulted = 1;
2795 /* ix86_tune_string is set to ix86_arch_string or defaulted. We
2796 need to use a sensible tune option. */
2797 if (!strcmp (ix86_tune_string, "generic")
2798 || !strcmp (ix86_tune_string, "x86-64")
2799 || !strcmp (ix86_tune_string, "i686"))
2802 ix86_tune_string = "generic64";
2804 ix86_tune_string = "generic32";
2807 if (ix86_stringop_string)
2809 if (!strcmp (ix86_stringop_string, "rep_byte"))
2810 stringop_alg = rep_prefix_1_byte;
2811 else if (!strcmp (ix86_stringop_string, "libcall"))
2812 stringop_alg = libcall;
2813 else if (!strcmp (ix86_stringop_string, "rep_4byte"))
2814 stringop_alg = rep_prefix_4_byte;
2815 else if (!strcmp (ix86_stringop_string, "rep_8byte")
2817 /* rep; movq isn't available in 32-bit code. */
2818 stringop_alg = rep_prefix_8_byte;
2819 else if (!strcmp (ix86_stringop_string, "byte_loop"))
2820 stringop_alg = loop_1_byte;
2821 else if (!strcmp (ix86_stringop_string, "loop"))
2822 stringop_alg = loop;
2823 else if (!strcmp (ix86_stringop_string, "unrolled_loop"))
2824 stringop_alg = unrolled_loop;
2826 error ("bad value (%s) for %sstringop-strategy=%s %s",
2827 ix86_stringop_string, prefix, suffix, sw);
2829 if (!strcmp (ix86_tune_string, "x86-64"))
2830 warning (OPT_Wdeprecated, "%stune=x86-64%s is deprecated. Use "
2831 "%stune=k8%s or %stune=generic%s instead as appropriate.",
2832 prefix, suffix, prefix, suffix, prefix, suffix);
OSDN Git Service
