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 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
20 Boston, MA 02110-1301, USA. */
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"
50 #include "tree-gimple.h"
52 #include "tm-constrs.h"
55 #ifndef CHECK_STACK_LIMIT
56 #define CHECK_STACK_LIMIT (-1)
59 /* Return index of given mode in mult and division cost tables. */
60 #define MODE_INDEX(mode) \
61 ((mode) == QImode ? 0 \
62 : (mode) == HImode ? 1 \
63 : (mode) == SImode ? 2 \
64 : (mode) == DImode ? 3 \
67 /* Processor costs (relative to an add) */
68 /* We assume COSTS_N_INSNS is defined as (N)*4 and an addition is 2 bytes. */
69 #define COSTS_N_BYTES(N) ((N) * 2)
71 #define DUMMY_STRINGOP_ALGS {libcall, {{-1, libcall}}}
74 struct processor_costs size_cost
= { /* costs for tuning for size */
75 COSTS_N_BYTES (2), /* cost of an add instruction */
76 COSTS_N_BYTES (3), /* cost of a lea instruction */
77 COSTS_N_BYTES (2), /* variable shift costs */
78 COSTS_N_BYTES (3), /* constant shift costs */
79 {COSTS_N_BYTES (3), /* cost of starting multiply for QI */
80 COSTS_N_BYTES (3), /* HI */
81 COSTS_N_BYTES (3), /* SI */
82 COSTS_N_BYTES (3), /* DI */
83 COSTS_N_BYTES (5)}, /* other */
84 0, /* cost of multiply per each bit set */
85 {COSTS_N_BYTES (3), /* cost of a divide/mod for QI */
86 COSTS_N_BYTES (3), /* HI */
87 COSTS_N_BYTES (3), /* SI */
88 COSTS_N_BYTES (3), /* DI */
89 COSTS_N_BYTES (5)}, /* other */
90 COSTS_N_BYTES (3), /* cost of movsx */
91 COSTS_N_BYTES (3), /* cost of movzx */
94 2, /* cost for loading QImode using movzbl */
95 {2, 2, 2}, /* cost of loading integer registers
96 in QImode, HImode and SImode.
97 Relative to reg-reg move (2). */
98 {2, 2, 2}, /* cost of storing integer registers */
99 2, /* cost of reg,reg fld/fst */
100 {2, 2, 2}, /* cost of loading fp registers
101 in SFmode, DFmode and XFmode */
102 {2, 2, 2}, /* cost of storing fp registers
103 in SFmode, DFmode and XFmode */
104 3, /* cost of moving MMX register */
105 {3, 3}, /* cost of loading MMX registers
106 in SImode and DImode */
107 {3, 3}, /* cost of storing MMX registers
108 in SImode and DImode */
109 3, /* cost of moving SSE register */
110 {3, 3, 3}, /* cost of loading SSE registers
111 in SImode, DImode and TImode */
112 {3, 3, 3}, /* cost of storing SSE registers
113 in SImode, DImode and TImode */
114 3, /* MMX or SSE register to integer */
115 0, /* size of prefetch block */
116 0, /* number of parallel prefetches */
118 COSTS_N_BYTES (2), /* cost of FADD and FSUB insns. */
119 COSTS_N_BYTES (2), /* cost of FMUL instruction. */
120 COSTS_N_BYTES (2), /* cost of FDIV instruction. */
121 COSTS_N_BYTES (2), /* cost of FABS instruction. */
122 COSTS_N_BYTES (2), /* cost of FCHS instruction. */
123 COSTS_N_BYTES (2), /* cost of FSQRT instruction. */
124 {{rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}},
125 {rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}}},
126 {{rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}},
127 {rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}}}
130 /* Processor costs (relative to an add) */
132 struct processor_costs i386_cost
= { /* 386 specific costs */
133 COSTS_N_INSNS (1), /* cost of an add instruction */
134 COSTS_N_INSNS (1), /* cost of a lea instruction */
135 COSTS_N_INSNS (3), /* variable shift costs */
136 COSTS_N_INSNS (2), /* constant shift costs */
137 {COSTS_N_INSNS (6), /* cost of starting multiply for QI */
138 COSTS_N_INSNS (6), /* HI */
139 COSTS_N_INSNS (6), /* SI */
140 COSTS_N_INSNS (6), /* DI */
141 COSTS_N_INSNS (6)}, /* other */
142 COSTS_N_INSNS (1), /* cost of multiply per each bit set */
143 {COSTS_N_INSNS (23), /* cost of a divide/mod for QI */
144 COSTS_N_INSNS (23), /* HI */
145 COSTS_N_INSNS (23), /* SI */
146 COSTS_N_INSNS (23), /* DI */
147 COSTS_N_INSNS (23)}, /* other */
148 COSTS_N_INSNS (3), /* cost of movsx */
149 COSTS_N_INSNS (2), /* cost of movzx */
150 15, /* "large" insn */
152 4, /* cost for loading QImode using movzbl */
153 {2, 4, 2}, /* cost of loading integer registers
154 in QImode, HImode and SImode.
155 Relative to reg-reg move (2). */
156 {2, 4, 2}, /* cost of storing integer registers */
157 2, /* cost of reg,reg fld/fst */
158 {8, 8, 8}, /* cost of loading fp registers
159 in SFmode, DFmode and XFmode */
160 {8, 8, 8}, /* cost of storing fp registers
161 in SFmode, DFmode and XFmode */
162 2, /* cost of moving MMX register */
163 {4, 8}, /* cost of loading MMX registers
164 in SImode and DImode */
165 {4, 8}, /* cost of storing MMX registers
166 in SImode and DImode */
167 2, /* cost of moving SSE register */
168 {4, 8, 16}, /* cost of loading SSE registers
169 in SImode, DImode and TImode */
170 {4, 8, 16}, /* cost of storing SSE registers
171 in SImode, DImode and TImode */
172 3, /* MMX or SSE register to integer */
173 0, /* size of prefetch block */
174 0, /* number of parallel prefetches */
176 COSTS_N_INSNS (23), /* cost of FADD and FSUB insns. */
177 COSTS_N_INSNS (27), /* cost of FMUL instruction. */
178 COSTS_N_INSNS (88), /* cost of FDIV instruction. */
179 COSTS_N_INSNS (22), /* cost of FABS instruction. */
180 COSTS_N_INSNS (24), /* cost of FCHS instruction. */
181 COSTS_N_INSNS (122), /* cost of FSQRT instruction. */
182 {{rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}},
183 DUMMY_STRINGOP_ALGS
},
184 {{rep_prefix_1_byte
, {{-1, rep_prefix_1_byte
}}},
185 DUMMY_STRINGOP_ALGS
},
189 struct processor_costs i486_cost
= { /* 486 specific costs */
190 COSTS_N_INSNS (1), /* cost of an add instruction */
191 COSTS_N_INSNS (1), /* cost of a lea instruction */
192 COSTS_N_INSNS (3), /* variable shift costs */
193 COSTS_N_INSNS (2), /* constant shift costs */
194 {COSTS_N_INSNS (12), /* cost of starting multiply for QI */
195 COSTS_N_INSNS (12), /* HI */
196 COSTS_N_INSNS (12), /* SI */
197 COSTS_N_INSNS (12), /* DI */
198 COSTS_N_INSNS (12)}, /* other */
199 1, /* cost of multiply per each bit set */
200 {COSTS_N_INSNS (40), /* cost of a divide/mod for QI */
201 COSTS_N_INSNS (40), /* HI */
202 COSTS_N_INSNS (40), /* SI */
203 COSTS_N_INSNS (40), /* DI */
204 COSTS_N_INSNS (40)}, /* other */
205 COSTS_N_INSNS (3), /* cost of movsx */
206 COSTS_N_INSNS (2), /* cost of movzx */
207 15, /* "large" insn */
209 4, /* cost for loading QImode using movzbl */
210 {2, 4, 2}, /* cost of loading integer registers
211 in QImode, HImode and SImode.
212 Relative to reg-reg move (2). */
213 {2, 4, 2}, /* cost of storing integer registers */
214 2, /* cost of reg,reg fld/fst */
215 {8, 8, 8}, /* cost of loading fp registers
216 in SFmode, DFmode and XFmode */
217 {8, 8, 8}, /* cost of storing fp registers
218 in SFmode, DFmode and XFmode */
219 2, /* cost of moving MMX register */
220 {4, 8}, /* cost of loading MMX registers
221 in SImode and DImode */
222 {4, 8}, /* cost of storing MMX registers
223 in SImode and DImode */
224 2, /* cost of moving SSE register */
225 {4, 8, 16}, /* cost of loading SSE registers
226 in SImode, DImode and TImode */
227 {4, 8, 16}, /* cost of storing SSE registers
228 in SImode, DImode and TImode */
229 3, /* MMX or SSE register to integer */
230 0, /* size of prefetch block */
231 0, /* number of parallel prefetches */
233 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
234 COSTS_N_INSNS (16), /* cost of FMUL instruction. */
235 COSTS_N_INSNS (73), /* cost of FDIV instruction. */
236 COSTS_N_INSNS (3), /* cost of FABS instruction. */
237 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
238 COSTS_N_INSNS (83), /* cost of FSQRT instruction. */
239 {{rep_prefix_4_byte
, {{-1, rep_prefix_4_byte
}}},
240 DUMMY_STRINGOP_ALGS
},
241 {{rep_prefix_4_byte
, {{-1, rep_prefix_4_byte
}}},
246 struct processor_costs pentium_cost
= {
247 COSTS_N_INSNS (1), /* cost of an add instruction */
248 COSTS_N_INSNS (1), /* cost of a lea instruction */
249 COSTS_N_INSNS (4), /* variable shift costs */
250 COSTS_N_INSNS (1), /* constant shift costs */
251 {COSTS_N_INSNS (11), /* cost of starting multiply for QI */
252 COSTS_N_INSNS (11), /* HI */
253 COSTS_N_INSNS (11), /* SI */
254 COSTS_N_INSNS (11), /* DI */
255 COSTS_N_INSNS (11)}, /* other */
256 0, /* cost of multiply per each bit set */
257 {COSTS_N_INSNS (25), /* cost of a divide/mod for QI */
258 COSTS_N_INSNS (25), /* HI */
259 COSTS_N_INSNS (25), /* SI */
260 COSTS_N_INSNS (25), /* DI */
261 COSTS_N_INSNS (25)}, /* other */
262 COSTS_N_INSNS (3), /* cost of movsx */
263 COSTS_N_INSNS (2), /* cost of movzx */
264 8, /* "large" insn */
266 6, /* cost for loading QImode using movzbl */
267 {2, 4, 2}, /* cost of loading integer registers
268 in QImode, HImode and SImode.
269 Relative to reg-reg move (2). */
270 {2, 4, 2}, /* cost of storing integer registers */
271 2, /* cost of reg,reg fld/fst */
272 {2, 2, 6}, /* cost of loading fp registers
273 in SFmode, DFmode and XFmode */
274 {4, 4, 6}, /* cost of storing fp registers
275 in SFmode, DFmode and XFmode */
276 8, /* cost of moving MMX register */
277 {8, 8}, /* cost of loading MMX registers
278 in SImode and DImode */
279 {8, 8}, /* cost of storing MMX registers
280 in SImode and DImode */
281 2, /* cost of moving SSE register */
282 {4, 8, 16}, /* cost of loading SSE registers
283 in SImode, DImode and TImode */
284 {4, 8, 16}, /* cost of storing SSE registers
285 in SImode, DImode and TImode */
286 3, /* MMX or SSE register to integer */
287 0, /* size of prefetch block */
288 0, /* number of parallel prefetches */
290 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
291 COSTS_N_INSNS (3), /* cost of FMUL instruction. */
292 COSTS_N_INSNS (39), /* cost of FDIV instruction. */
293 COSTS_N_INSNS (1), /* cost of FABS instruction. */
294 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
295 COSTS_N_INSNS (70), /* cost of FSQRT instruction. */
296 {{libcall
, {{256, rep_prefix_4_byte
}, {-1, libcall
}}},
297 DUMMY_STRINGOP_ALGS
},
298 {{libcall
, {{-1, rep_prefix_4_byte
}}},
303 struct processor_costs pentiumpro_cost
= {
304 COSTS_N_INSNS (1), /* cost of an add instruction */
305 COSTS_N_INSNS (1), /* cost of a lea instruction */
306 COSTS_N_INSNS (1), /* variable shift costs */
307 COSTS_N_INSNS (1), /* constant shift costs */
308 {COSTS_N_INSNS (4), /* cost of starting multiply for QI */
309 COSTS_N_INSNS (4), /* HI */
310 COSTS_N_INSNS (4), /* SI */
311 COSTS_N_INSNS (4), /* DI */
312 COSTS_N_INSNS (4)}, /* other */
313 0, /* cost of multiply per each bit set */
314 {COSTS_N_INSNS (17), /* cost of a divide/mod for QI */
315 COSTS_N_INSNS (17), /* HI */
316 COSTS_N_INSNS (17), /* SI */
317 COSTS_N_INSNS (17), /* DI */
318 COSTS_N_INSNS (17)}, /* other */
319 COSTS_N_INSNS (1), /* cost of movsx */
320 COSTS_N_INSNS (1), /* cost of movzx */
321 8, /* "large" insn */
323 2, /* cost for loading QImode using movzbl */
324 {4, 4, 4}, /* cost of loading integer registers
325 in QImode, HImode and SImode.
326 Relative to reg-reg move (2). */
327 {2, 2, 2}, /* cost of storing integer registers */
328 2, /* cost of reg,reg fld/fst */
329 {2, 2, 6}, /* cost of loading fp registers
330 in SFmode, DFmode and XFmode */
331 {4, 4, 6}, /* cost of storing fp registers
332 in SFmode, DFmode and XFmode */
333 2, /* cost of moving MMX register */
334 {2, 2}, /* cost of loading MMX registers
335 in SImode and DImode */
336 {2, 2}, /* cost of storing MMX registers
337 in SImode and DImode */
338 2, /* cost of moving SSE register */
339 {2, 2, 8}, /* cost of loading SSE registers
340 in SImode, DImode and TImode */
341 {2, 2, 8}, /* cost of storing SSE registers
342 in SImode, DImode and TImode */
343 3, /* MMX or SSE register to integer */
344 32, /* size of prefetch block */
345 6, /* number of parallel prefetches */
347 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
348 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
349 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
350 COSTS_N_INSNS (2), /* cost of FABS instruction. */
351 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
352 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
353 /* PentiumPro has optimized rep instructions for blocks aligned by 8 bytes (we ensure
354 the alignment). For small blocks inline loop is still a noticeable win, for bigger
355 blocks either rep movsl or rep movsb is way to go. Rep movsb has apparently
356 more expensive startup time in CPU, but after 4K the difference is down in the noise.
358 {{rep_prefix_4_byte
, {{128, loop
}, {1024, unrolled_loop
},
359 {8192, rep_prefix_4_byte
}, {-1, rep_prefix_1_byte
}}},
360 DUMMY_STRINGOP_ALGS
},
361 {{rep_prefix_4_byte
, {{1024, unrolled_loop
},
362 {8192, rep_prefix_4_byte
}, {-1, libcall
}}},
367 struct processor_costs geode_cost
= {
368 COSTS_N_INSNS (1), /* cost of an add instruction */
369 COSTS_N_INSNS (1), /* cost of a lea instruction */
370 COSTS_N_INSNS (2), /* variable shift costs */
371 COSTS_N_INSNS (1), /* constant shift costs */
372 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
373 COSTS_N_INSNS (4), /* HI */
374 COSTS_N_INSNS (7), /* SI */
375 COSTS_N_INSNS (7), /* DI */
376 COSTS_N_INSNS (7)}, /* other */
377 0, /* cost of multiply per each bit set */
378 {COSTS_N_INSNS (15), /* cost of a divide/mod for QI */
379 COSTS_N_INSNS (23), /* HI */
380 COSTS_N_INSNS (39), /* SI */
381 COSTS_N_INSNS (39), /* DI */
382 COSTS_N_INSNS (39)}, /* other */
383 COSTS_N_INSNS (1), /* cost of movsx */
384 COSTS_N_INSNS (1), /* cost of movzx */
385 8, /* "large" insn */
387 1, /* cost for loading QImode using movzbl */
388 {1, 1, 1}, /* cost of loading integer registers
389 in QImode, HImode and SImode.
390 Relative to reg-reg move (2). */
391 {1, 1, 1}, /* cost of storing integer registers */
392 1, /* cost of reg,reg fld/fst */
393 {1, 1, 1}, /* cost of loading fp registers
394 in SFmode, DFmode and XFmode */
395 {4, 6, 6}, /* cost of storing fp registers
396 in SFmode, DFmode and XFmode */
398 1, /* cost of moving MMX register */
399 {1, 1}, /* cost of loading MMX registers
400 in SImode and DImode */
401 {1, 1}, /* cost of storing MMX registers
402 in SImode and DImode */
403 1, /* cost of moving SSE register */
404 {1, 1, 1}, /* cost of loading SSE registers
405 in SImode, DImode and TImode */
406 {1, 1, 1}, /* cost of storing SSE registers
407 in SImode, DImode and TImode */
408 1, /* MMX or SSE register to integer */
409 32, /* size of prefetch block */
410 1, /* number of parallel prefetches */
412 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
413 COSTS_N_INSNS (11), /* cost of FMUL instruction. */
414 COSTS_N_INSNS (47), /* cost of FDIV instruction. */
415 COSTS_N_INSNS (1), /* cost of FABS instruction. */
416 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
417 COSTS_N_INSNS (54), /* cost of FSQRT instruction. */
418 {{libcall
, {{256, rep_prefix_4_byte
}, {-1, libcall
}}},
419 DUMMY_STRINGOP_ALGS
},
420 {{libcall
, {{256, rep_prefix_4_byte
}, {-1, libcall
}}},
425 struct processor_costs k6_cost
= {
426 COSTS_N_INSNS (1), /* cost of an add instruction */
427 COSTS_N_INSNS (2), /* cost of a lea instruction */
428 COSTS_N_INSNS (1), /* variable shift costs */
429 COSTS_N_INSNS (1), /* constant shift costs */
430 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
431 COSTS_N_INSNS (3), /* HI */
432 COSTS_N_INSNS (3), /* SI */
433 COSTS_N_INSNS (3), /* DI */
434 COSTS_N_INSNS (3)}, /* other */
435 0, /* cost of multiply per each bit set */
436 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
437 COSTS_N_INSNS (18), /* HI */
438 COSTS_N_INSNS (18), /* SI */
439 COSTS_N_INSNS (18), /* DI */
440 COSTS_N_INSNS (18)}, /* other */
441 COSTS_N_INSNS (2), /* cost of movsx */
442 COSTS_N_INSNS (2), /* cost of movzx */
443 8, /* "large" insn */
445 3, /* cost for loading QImode using movzbl */
446 {4, 5, 4}, /* cost of loading integer registers
447 in QImode, HImode and SImode.
448 Relative to reg-reg move (2). */
449 {2, 3, 2}, /* cost of storing integer registers */
450 4, /* cost of reg,reg fld/fst */
451 {6, 6, 6}, /* cost of loading fp registers
452 in SFmode, DFmode and XFmode */
453 {4, 4, 4}, /* cost of storing fp registers
454 in SFmode, DFmode and XFmode */
455 2, /* cost of moving MMX register */
456 {2, 2}, /* cost of loading MMX registers
457 in SImode and DImode */
458 {2, 2}, /* cost of storing MMX registers
459 in SImode and DImode */
460 2, /* cost of moving SSE register */
461 {2, 2, 8}, /* cost of loading SSE registers
462 in SImode, DImode and TImode */
463 {2, 2, 8}, /* cost of storing SSE registers
464 in SImode, DImode and TImode */
465 6, /* MMX or SSE register to integer */
466 32, /* size of prefetch block */
467 1, /* number of parallel prefetches */
469 COSTS_N_INSNS (2), /* cost of FADD and FSUB insns. */
470 COSTS_N_INSNS (2), /* cost of FMUL instruction. */
471 COSTS_N_INSNS (56), /* cost of FDIV instruction. */
472 COSTS_N_INSNS (2), /* cost of FABS instruction. */
473 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
474 COSTS_N_INSNS (56), /* cost of FSQRT instruction. */
475 {{libcall
, {{256, rep_prefix_4_byte
}, {-1, libcall
}}},
476 DUMMY_STRINGOP_ALGS
},
477 {{libcall
, {{256, rep_prefix_4_byte
}, {-1, libcall
}}},
482 struct processor_costs athlon_cost
= {
483 COSTS_N_INSNS (1), /* cost of an add instruction */
484 COSTS_N_INSNS (2), /* cost of a lea instruction */
485 COSTS_N_INSNS (1), /* variable shift costs */
486 COSTS_N_INSNS (1), /* constant shift costs */
487 {COSTS_N_INSNS (5), /* cost of starting multiply for QI */
488 COSTS_N_INSNS (5), /* HI */
489 COSTS_N_INSNS (5), /* SI */
490 COSTS_N_INSNS (5), /* DI */
491 COSTS_N_INSNS (5)}, /* other */
492 0, /* cost of multiply per each bit set */
493 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
494 COSTS_N_INSNS (26), /* HI */
495 COSTS_N_INSNS (42), /* SI */
496 COSTS_N_INSNS (74), /* DI */
497 COSTS_N_INSNS (74)}, /* other */
498 COSTS_N_INSNS (1), /* cost of movsx */
499 COSTS_N_INSNS (1), /* cost of movzx */
500 8, /* "large" insn */
502 4, /* cost for loading QImode using movzbl */
503 {3, 4, 3}, /* cost of loading integer registers
504 in QImode, HImode and SImode.
505 Relative to reg-reg move (2). */
506 {3, 4, 3}, /* cost of storing integer registers */
507 4, /* cost of reg,reg fld/fst */
508 {4, 4, 12}, /* cost of loading fp registers
509 in SFmode, DFmode and XFmode */
510 {6, 6, 8}, /* cost of storing fp registers
511 in SFmode, DFmode and XFmode */
512 2, /* cost of moving MMX register */
513 {4, 4}, /* cost of loading MMX registers
514 in SImode and DImode */
515 {4, 4}, /* cost of storing MMX registers
516 in SImode and DImode */
517 2, /* cost of moving SSE register */
518 {4, 4, 6}, /* cost of loading SSE registers
519 in SImode, DImode and TImode */
520 {4, 4, 5}, /* cost of storing SSE registers
521 in SImode, DImode and TImode */
522 5, /* MMX or SSE register to integer */
523 64, /* size of prefetch block */
524 6, /* number of parallel prefetches */
526 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
527 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
528 COSTS_N_INSNS (24), /* cost of FDIV instruction. */
529 COSTS_N_INSNS (2), /* cost of FABS instruction. */
530 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
531 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
532 /* For some reason, Athlon deals better with REP prefix (relative to loops)
533 compared to K8. Alignment becomes important after 8 bytes for memcpy and
534 128 bytes for memset. */
535 {{libcall
, {{2048, rep_prefix_4_byte
}, {-1, libcall
}}},
536 DUMMY_STRINGOP_ALGS
},
537 {{libcall
, {{2048, rep_prefix_4_byte
}, {-1, libcall
}}},
542 struct processor_costs k8_cost
= {
543 COSTS_N_INSNS (1), /* cost of an add instruction */
544 COSTS_N_INSNS (2), /* cost of a lea instruction */
545 COSTS_N_INSNS (1), /* variable shift costs */
546 COSTS_N_INSNS (1), /* constant shift costs */
547 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
548 COSTS_N_INSNS (4), /* HI */
549 COSTS_N_INSNS (3), /* SI */
550 COSTS_N_INSNS (4), /* DI */
551 COSTS_N_INSNS (5)}, /* other */
552 0, /* cost of multiply per each bit set */
553 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
554 COSTS_N_INSNS (26), /* HI */
555 COSTS_N_INSNS (42), /* SI */
556 COSTS_N_INSNS (74), /* DI */
557 COSTS_N_INSNS (74)}, /* other */
558 COSTS_N_INSNS (1), /* cost of movsx */
559 COSTS_N_INSNS (1), /* cost of movzx */
560 8, /* "large" insn */
562 4, /* cost for loading QImode using movzbl */
563 {3, 4, 3}, /* cost of loading integer registers
564 in QImode, HImode and SImode.
565 Relative to reg-reg move (2). */
566 {3, 4, 3}, /* cost of storing integer registers */
567 4, /* cost of reg,reg fld/fst */
568 {4, 4, 12}, /* cost of loading fp registers
569 in SFmode, DFmode and XFmode */
570 {6, 6, 8}, /* cost of storing fp registers
571 in SFmode, DFmode and XFmode */
572 2, /* cost of moving MMX register */
573 {3, 3}, /* cost of loading MMX registers
574 in SImode and DImode */
575 {4, 4}, /* cost of storing MMX registers
576 in SImode and DImode */
577 2, /* cost of moving SSE register */
578 {4, 3, 6}, /* cost of loading SSE registers
579 in SImode, DImode and TImode */
580 {4, 4, 5}, /* cost of storing SSE registers
581 in SImode, DImode and TImode */
582 5, /* MMX or SSE register to integer */
583 64, /* size of prefetch block */
584 /* New AMD processors never drop prefetches; if they cannot be performed
585 immediately, they are queued. We set number of simultaneous prefetches
586 to a large constant to reflect this (it probably is not a good idea not
587 to limit number of prefetches at all, as their execution also takes some
589 100, /* number of parallel prefetches */
591 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
592 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
593 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
594 COSTS_N_INSNS (2), /* cost of FABS instruction. */
595 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
596 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
597 /* K8 has optimized REP instruction for medium sized blocks, but for very small
598 blocks it is better to use loop. For large blocks, libcall can do
599 nontemporary accesses and beat inline considerably. */
600 {{libcall
, {{6, loop
}, {14, unrolled_loop
}, {-1, rep_prefix_4_byte
}}},
601 {libcall
, {{16, loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}},
602 {{libcall
, {{8, loop
}, {24, unrolled_loop
},
603 {2048, rep_prefix_4_byte
}, {-1, libcall
}}},
604 {libcall
, {{48, unrolled_loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}}
607 struct processor_costs amdfam10_cost
= {
608 COSTS_N_INSNS (1), /* cost of an add instruction */
609 COSTS_N_INSNS (2), /* cost of a lea instruction */
610 COSTS_N_INSNS (1), /* variable shift costs */
611 COSTS_N_INSNS (1), /* constant shift costs */
612 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
613 COSTS_N_INSNS (4), /* HI */
614 COSTS_N_INSNS (3), /* SI */
615 COSTS_N_INSNS (4), /* DI */
616 COSTS_N_INSNS (5)}, /* other */
617 0, /* cost of multiply per each bit set */
618 {COSTS_N_INSNS (19), /* cost of a divide/mod for QI */
619 COSTS_N_INSNS (35), /* HI */
620 COSTS_N_INSNS (51), /* SI */
621 COSTS_N_INSNS (83), /* DI */
622 COSTS_N_INSNS (83)}, /* other */
623 COSTS_N_INSNS (1), /* cost of movsx */
624 COSTS_N_INSNS (1), /* cost of movzx */
625 8, /* "large" insn */
627 4, /* cost for loading QImode using movzbl */
628 {3, 4, 3}, /* cost of loading integer registers
629 in QImode, HImode and SImode.
630 Relative to reg-reg move (2). */
631 {3, 4, 3}, /* cost of storing integer registers */
632 4, /* cost of reg,reg fld/fst */
633 {4, 4, 12}, /* cost of loading fp registers
634 in SFmode, DFmode and XFmode */
635 {6, 6, 8}, /* cost of storing fp registers
636 in SFmode, DFmode and XFmode */
637 2, /* cost of moving MMX register */
638 {3, 3}, /* cost of loading MMX registers
639 in SImode and DImode */
640 {4, 4}, /* cost of storing MMX registers
641 in SImode and DImode */
642 2, /* cost of moving SSE register */
643 {4, 4, 3}, /* cost of loading SSE registers
644 in SImode, DImode and TImode */
645 {4, 4, 5}, /* cost of storing SSE registers
646 in SImode, DImode and TImode */
647 3, /* MMX or SSE register to integer */
649 MOVD reg64, xmmreg Double FSTORE 4
650 MOVD reg32, xmmreg Double FSTORE 4
652 MOVD reg64, xmmreg Double FADD 3
654 MOVD reg32, xmmreg Double FADD 3
656 64, /* size of prefetch block */
657 /* New AMD processors never drop prefetches; if they cannot be performed
658 immediately, they are queued. We set number of simultaneous prefetches
659 to a large constant to reflect this (it probably is not a good idea not
660 to limit number of prefetches at all, as their execution also takes some
662 100, /* number of parallel prefetches */
664 COSTS_N_INSNS (4), /* cost of FADD and FSUB insns. */
665 COSTS_N_INSNS (4), /* cost of FMUL instruction. */
666 COSTS_N_INSNS (19), /* cost of FDIV instruction. */
667 COSTS_N_INSNS (2), /* cost of FABS instruction. */
668 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
669 COSTS_N_INSNS (35), /* cost of FSQRT instruction. */
671 /* AMDFAM10 has optimized REP instruction for medium sized blocks, but for
672 very small blocks it is better to use loop. For large blocks, libcall can
673 do nontemporary accesses and beat inline considerably. */
674 {{libcall
, {{6, loop
}, {14, unrolled_loop
}, {-1, rep_prefix_4_byte
}}},
675 {libcall
, {{16, loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}},
676 {{libcall
, {{8, loop
}, {24, unrolled_loop
},
677 {2048, rep_prefix_4_byte
}, {-1, libcall
}}},
678 {libcall
, {{48, unrolled_loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}}
682 struct processor_costs pentium4_cost
= {
683 COSTS_N_INSNS (1), /* cost of an add instruction */
684 COSTS_N_INSNS (3), /* cost of a lea instruction */
685 COSTS_N_INSNS (4), /* variable shift costs */
686 COSTS_N_INSNS (4), /* constant shift costs */
687 {COSTS_N_INSNS (15), /* cost of starting multiply for QI */
688 COSTS_N_INSNS (15), /* HI */
689 COSTS_N_INSNS (15), /* SI */
690 COSTS_N_INSNS (15), /* DI */
691 COSTS_N_INSNS (15)}, /* other */
692 0, /* cost of multiply per each bit set */
693 {COSTS_N_INSNS (56), /* cost of a divide/mod for QI */
694 COSTS_N_INSNS (56), /* HI */
695 COSTS_N_INSNS (56), /* SI */
696 COSTS_N_INSNS (56), /* DI */
697 COSTS_N_INSNS (56)}, /* other */
698 COSTS_N_INSNS (1), /* cost of movsx */
699 COSTS_N_INSNS (1), /* cost of movzx */
700 16, /* "large" insn */
702 2, /* cost for loading QImode using movzbl */
703 {4, 5, 4}, /* cost of loading integer registers
704 in QImode, HImode and SImode.
705 Relative to reg-reg move (2). */
706 {2, 3, 2}, /* cost of storing integer registers */
707 2, /* cost of reg,reg fld/fst */
708 {2, 2, 6}, /* cost of loading fp registers
709 in SFmode, DFmode and XFmode */
710 {4, 4, 6}, /* cost of storing fp registers
711 in SFmode, DFmode and XFmode */
712 2, /* cost of moving MMX register */
713 {2, 2}, /* cost of loading MMX registers
714 in SImode and DImode */
715 {2, 2}, /* cost of storing MMX registers
716 in SImode and DImode */
717 12, /* cost of moving SSE register */
718 {12, 12, 12}, /* cost of loading SSE registers
719 in SImode, DImode and TImode */
720 {2, 2, 8}, /* cost of storing SSE registers
721 in SImode, DImode and TImode */
722 10, /* MMX or SSE register to integer */
723 64, /* size of prefetch block */
724 6, /* number of parallel prefetches */
726 COSTS_N_INSNS (5), /* cost of FADD and FSUB insns. */
727 COSTS_N_INSNS (7), /* cost of FMUL instruction. */
728 COSTS_N_INSNS (43), /* cost of FDIV instruction. */
729 COSTS_N_INSNS (2), /* cost of FABS instruction. */
730 COSTS_N_INSNS (2), /* cost of FCHS instruction. */
731 COSTS_N_INSNS (43), /* cost of FSQRT instruction. */
732 {{libcall
, {{12, loop_1_byte
}, {-1, rep_prefix_4_byte
}}},
733 DUMMY_STRINGOP_ALGS
},
734 {{libcall
, {{6, loop_1_byte
}, {48, loop
}, {20480, rep_prefix_4_byte
},
736 DUMMY_STRINGOP_ALGS
},
740 struct processor_costs nocona_cost
= {
741 COSTS_N_INSNS (1), /* cost of an add instruction */
742 COSTS_N_INSNS (1), /* cost of a lea instruction */
743 COSTS_N_INSNS (1), /* variable shift costs */
744 COSTS_N_INSNS (1), /* constant shift costs */
745 {COSTS_N_INSNS (10), /* cost of starting multiply for QI */
746 COSTS_N_INSNS (10), /* HI */
747 COSTS_N_INSNS (10), /* SI */
748 COSTS_N_INSNS (10), /* DI */
749 COSTS_N_INSNS (10)}, /* other */
750 0, /* cost of multiply per each bit set */
751 {COSTS_N_INSNS (66), /* cost of a divide/mod for QI */
752 COSTS_N_INSNS (66), /* HI */
753 COSTS_N_INSNS (66), /* SI */
754 COSTS_N_INSNS (66), /* DI */
755 COSTS_N_INSNS (66)}, /* other */
756 COSTS_N_INSNS (1), /* cost of movsx */
757 COSTS_N_INSNS (1), /* cost of movzx */
758 16, /* "large" insn */
760 4, /* cost for loading QImode using movzbl */
761 {4, 4, 4}, /* cost of loading integer registers
762 in QImode, HImode and SImode.
763 Relative to reg-reg move (2). */
764 {4, 4, 4}, /* cost of storing integer registers */
765 3, /* cost of reg,reg fld/fst */
766 {12, 12, 12}, /* cost of loading fp registers
767 in SFmode, DFmode and XFmode */
768 {4, 4, 4}, /* cost of storing fp registers
769 in SFmode, DFmode and XFmode */
770 6, /* cost of moving MMX register */
771 {12, 12}, /* cost of loading MMX registers
772 in SImode and DImode */
773 {12, 12}, /* cost of storing MMX registers
774 in SImode and DImode */
775 6, /* cost of moving SSE register */
776 {12, 12, 12}, /* cost of loading SSE registers
777 in SImode, DImode and TImode */
778 {12, 12, 12}, /* cost of storing SSE registers
779 in SImode, DImode and TImode */
780 8, /* MMX or SSE register to integer */
781 128, /* size of prefetch block */
782 8, /* number of parallel prefetches */
784 COSTS_N_INSNS (6), /* cost of FADD and FSUB insns. */
785 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
786 COSTS_N_INSNS (40), /* cost of FDIV instruction. */
787 COSTS_N_INSNS (3), /* cost of FABS instruction. */
788 COSTS_N_INSNS (3), /* cost of FCHS instruction. */
789 COSTS_N_INSNS (44), /* cost of FSQRT instruction. */
790 {{libcall
, {{12, loop_1_byte
}, {-1, rep_prefix_4_byte
}}},
791 {libcall
, {{32, loop
}, {20000, rep_prefix_8_byte
},
792 {100000, unrolled_loop
}, {-1, libcall
}}}},
793 {{libcall
, {{6, loop_1_byte
}, {48, loop
}, {20480, rep_prefix_4_byte
},
795 {libcall
, {{24, loop
}, {64, unrolled_loop
},
796 {8192, rep_prefix_8_byte
}, {-1, libcall
}}}}
800 struct processor_costs core2_cost
= {
801 COSTS_N_INSNS (1), /* cost of an add instruction */
802 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
803 COSTS_N_INSNS (1), /* variable shift costs */
804 COSTS_N_INSNS (1), /* constant shift costs */
805 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
806 COSTS_N_INSNS (3), /* HI */
807 COSTS_N_INSNS (3), /* SI */
808 COSTS_N_INSNS (3), /* DI */
809 COSTS_N_INSNS (3)}, /* other */
810 0, /* cost of multiply per each bit set */
811 {COSTS_N_INSNS (22), /* cost of a divide/mod for QI */
812 COSTS_N_INSNS (22), /* HI */
813 COSTS_N_INSNS (22), /* SI */
814 COSTS_N_INSNS (22), /* DI */
815 COSTS_N_INSNS (22)}, /* other */
816 COSTS_N_INSNS (1), /* cost of movsx */
817 COSTS_N_INSNS (1), /* cost of movzx */
818 8, /* "large" insn */
820 2, /* cost for loading QImode using movzbl */
821 {6, 6, 6}, /* cost of loading integer registers
822 in QImode, HImode and SImode.
823 Relative to reg-reg move (2). */
824 {4, 4, 4}, /* cost of storing integer registers */
825 2, /* cost of reg,reg fld/fst */
826 {6, 6, 6}, /* cost of loading fp registers
827 in SFmode, DFmode and XFmode */
828 {4, 4, 4}, /* cost of loading integer registers */
829 2, /* cost of moving MMX register */
830 {6, 6}, /* cost of loading MMX registers
831 in SImode and DImode */
832 {4, 4}, /* cost of storing MMX registers
833 in SImode and DImode */
834 2, /* cost of moving SSE register */
835 {6, 6, 6}, /* cost of loading SSE registers
836 in SImode, DImode and TImode */
837 {4, 4, 4}, /* cost of storing SSE registers
838 in SImode, DImode and TImode */
839 2, /* MMX or SSE register to integer */
840 128, /* size of prefetch block */
841 8, /* number of parallel prefetches */
843 COSTS_N_INSNS (3), /* cost of FADD and FSUB insns. */
844 COSTS_N_INSNS (5), /* cost of FMUL instruction. */
845 COSTS_N_INSNS (32), /* cost of FDIV instruction. */
846 COSTS_N_INSNS (1), /* cost of FABS instruction. */
847 COSTS_N_INSNS (1), /* cost of FCHS instruction. */
848 COSTS_N_INSNS (58), /* cost of FSQRT instruction. */
849 {{libcall
, {{11, loop
}, {-1, rep_prefix_4_byte
}}},
850 {libcall
, {{32, loop
}, {64, rep_prefix_4_byte
},
851 {8192, rep_prefix_8_byte
}, {-1, libcall
}}}},
852 {{libcall
, {{8, loop
}, {15, unrolled_loop
},
853 {2048, rep_prefix_4_byte
}, {-1, libcall
}}},
854 {libcall
, {{24, loop
}, {32, unrolled_loop
},
855 {8192, rep_prefix_8_byte
}, {-1, libcall
}}}}
858 /* Generic64 should produce code tuned for Nocona and K8. */
860 struct processor_costs generic64_cost
= {
861 COSTS_N_INSNS (1), /* cost of an add instruction */
862 /* On all chips taken into consideration lea is 2 cycles and more. With
863 this cost however our current implementation of synth_mult results in
864 use of unnecessary temporary registers causing regression on several
865 SPECfp benchmarks. */
866 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
867 COSTS_N_INSNS (1), /* variable shift costs */
868 COSTS_N_INSNS (1), /* constant shift costs */
869 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
870 COSTS_N_INSNS (4), /* HI */
871 COSTS_N_INSNS (3), /* SI */
872 COSTS_N_INSNS (4), /* DI */
873 COSTS_N_INSNS (2)}, /* other */
874 0, /* cost of multiply per each bit set */
875 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
876 COSTS_N_INSNS (26), /* HI */
877 COSTS_N_INSNS (42), /* SI */
878 COSTS_N_INSNS (74), /* DI */
879 COSTS_N_INSNS (74)}, /* other */
880 COSTS_N_INSNS (1), /* cost of movsx */
881 COSTS_N_INSNS (1), /* cost of movzx */
882 8, /* "large" insn */
884 4, /* cost for loading QImode using movzbl */
885 {4, 4, 4}, /* cost of loading integer registers
886 in QImode, HImode and SImode.
887 Relative to reg-reg move (2). */
888 {4, 4, 4}, /* cost of storing integer registers */
889 4, /* cost of reg,reg fld/fst */
890 {12, 12, 12}, /* cost of loading fp registers
891 in SFmode, DFmode and XFmode */
892 {6, 6, 8}, /* cost of storing fp registers
893 in SFmode, DFmode and XFmode */
894 2, /* cost of moving MMX register */
895 {8, 8}, /* cost of loading MMX registers
896 in SImode and DImode */
897 {8, 8}, /* cost of storing MMX registers
898 in SImode and DImode */
899 2, /* cost of moving SSE register */
900 {8, 8, 8}, /* cost of loading SSE registers
901 in SImode, DImode and TImode */
902 {8, 8, 8}, /* cost of storing SSE registers
903 in SImode, DImode and TImode */
904 5, /* MMX or SSE register to integer */
905 64, /* size of prefetch block */
906 6, /* number of parallel prefetches */
907 /* Benchmarks shows large regressions on K8 sixtrack benchmark when this value
908 is increased to perhaps more appropriate value of 5. */
910 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
911 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
912 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
913 COSTS_N_INSNS (8), /* cost of FABS instruction. */
914 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
915 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
916 {DUMMY_STRINGOP_ALGS
,
917 {libcall
, {{32, loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}},
918 {DUMMY_STRINGOP_ALGS
,
919 {libcall
, {{32, loop
}, {8192, rep_prefix_8_byte
}, {-1, libcall
}}}}
922 /* Generic32 should produce code tuned for Athlon, PPro, Pentium4, Nocona and K8. */
924 struct processor_costs generic32_cost
= {
925 COSTS_N_INSNS (1), /* cost of an add instruction */
926 COSTS_N_INSNS (1) + 1, /* cost of a lea instruction */
927 COSTS_N_INSNS (1), /* variable shift costs */
928 COSTS_N_INSNS (1), /* constant shift costs */
929 {COSTS_N_INSNS (3), /* cost of starting multiply for QI */
930 COSTS_N_INSNS (4), /* HI */
931 COSTS_N_INSNS (3), /* SI */
932 COSTS_N_INSNS (4), /* DI */
933 COSTS_N_INSNS (2)}, /* other */
934 0, /* cost of multiply per each bit set */
935 {COSTS_N_INSNS (18), /* cost of a divide/mod for QI */
936 COSTS_N_INSNS (26), /* HI */
937 COSTS_N_INSNS (42), /* SI */
938 COSTS_N_INSNS (74), /* DI */
939 COSTS_N_INSNS (74)}, /* other */
940 COSTS_N_INSNS (1), /* cost of movsx */
941 COSTS_N_INSNS (1), /* cost of movzx */
942 8, /* "large" insn */
944 4, /* cost for loading QImode using movzbl */
945 {4, 4, 4}, /* cost of loading integer registers
946 in QImode, HImode and SImode.
947 Relative to reg-reg move (2). */
948 {4, 4, 4}, /* cost of storing integer registers */
949 4, /* cost of reg,reg fld/fst */
950 {12, 12, 12}, /* cost of loading fp registers
951 in SFmode, DFmode and XFmode */
952 {6, 6, 8}, /* cost of storing fp registers
953 in SFmode, DFmode and XFmode */
954 2, /* cost of moving MMX register */
955 {8, 8}, /* cost of loading MMX registers
956 in SImode and DImode */
957 {8, 8}, /* cost of storing MMX registers
958 in SImode and DImode */
959 2, /* cost of moving SSE register */
960 {8, 8, 8}, /* cost of loading SSE registers
961 in SImode, DImode and TImode */
962 {8, 8, 8}, /* cost of storing SSE registers
963 in SImode, DImode and TImode */
964 5, /* MMX or SSE register to integer */
965 64, /* size of prefetch block */
966 6, /* number of parallel prefetches */
968 COSTS_N_INSNS (8), /* cost of FADD and FSUB insns. */
969 COSTS_N_INSNS (8), /* cost of FMUL instruction. */
970 COSTS_N_INSNS (20), /* cost of FDIV instruction. */
971 COSTS_N_INSNS (8), /* cost of FABS instruction. */
972 COSTS_N_INSNS (8), /* cost of FCHS instruction. */
973 COSTS_N_INSNS (40), /* cost of FSQRT instruction. */
974 {{libcall
, {{32, loop
}, {8192, rep_prefix_4_byte
}, {-1, libcall
}}},
975 DUMMY_STRINGOP_ALGS
},
976 {{libcall
, {{32, loop
}, {8192, rep_prefix_4_byte
}, {-1, libcall
}}},
977 DUMMY_STRINGOP_ALGS
},
980 const struct processor_costs
*ix86_cost
= &pentium_cost
;
982 /* Processor feature/optimization bitmasks. */
983 #define m_386 (1<<PROCESSOR_I386)
984 #define m_486 (1<<PROCESSOR_I486)
985 #define m_PENT (1<<PROCESSOR_PENTIUM)
986 #define m_PPRO (1<<PROCESSOR_PENTIUMPRO)
987 #define m_PENT4 (1<<PROCESSOR_PENTIUM4)
988 #define m_NOCONA (1<<PROCESSOR_NOCONA)
989 #define m_CORE2 (1<<PROCESSOR_CORE2)
991 #define m_GEODE (1<<PROCESSOR_GEODE)
992 #define m_K6 (1<<PROCESSOR_K6)
993 #define m_K6_GEODE (m_K6 | m_GEODE)
994 #define m_K8 (1<<PROCESSOR_K8)
995 #define m_ATHLON (1<<PROCESSOR_ATHLON)
996 #define m_ATHLON_K8 (m_K8 | m_ATHLON)
997 #define m_AMDFAM10 (1<<PROCESSOR_AMDFAM10)
998 #define m_ATHLON_K8_AMDFAM10 (m_K8 | m_ATHLON | m_AMDFAM10)
1000 #define m_GENERIC32 (1<<PROCESSOR_GENERIC32)
1001 #define m_GENERIC64 (1<<PROCESSOR_GENERIC64)
1003 /* Generic instruction choice should be common subset of supported CPUs
1004 (PPro/PENT4/NOCONA/CORE2/Athlon/K8). */
1005 #define m_GENERIC (m_GENERIC32 | m_GENERIC64)
1007 /* Feature tests against the various tunings. */
1008 unsigned int ix86_tune_features
[X86_TUNE_LAST
] = {
1009 /* X86_TUNE_USE_LEAVE: Leave does not affect Nocona SPEC2000 results
1010 negatively, so enabling for Generic64 seems like good code size
1011 tradeoff. We can't enable it for 32bit generic because it does not
1012 work well with PPro base chips. */
1013 m_386
| m_K6_GEODE
| m_ATHLON_K8_AMDFAM10
| m_CORE2
| m_GENERIC64
,
1015 /* X86_TUNE_PUSH_MEMORY */
1016 m_386
| m_K6_GEODE
| m_ATHLON_K8_AMDFAM10
| m_PENT4
1017 | m_NOCONA
| m_CORE2
| m_GENERIC
,
1019 /* X86_TUNE_ZERO_EXTEND_WITH_AND */
1022 /* X86_TUNE_USE_BIT_TEST */
1025 /* X86_TUNE_UNROLL_STRLEN */
1026 m_486
| m_PENT
| m_PPRO
| m_ATHLON_K8_AMDFAM10
| m_K6
| m_CORE2
| m_GENERIC
,
1028 /* X86_TUNE_DEEP_BRANCH_PREDICTION */
1029 m_PPRO
| m_K6_GEODE
| m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_GENERIC
,
1031 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
1032 on simulation result. But after P4 was made, no performance benefit
1033 was observed with branch hints. It also increases the code size.
1034 As a result, icc never generates branch hints. */
1037 /* X86_TUNE_DOUBLE_WITH_ADD */
1040 /* X86_TUNE_USE_SAHF */
1041 m_PPRO
| m_K6_GEODE
| m_K8
| m_AMDFAM10
| m_PENT4
1042 | m_NOCONA
| m_CORE2
| m_GENERIC
,
1044 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
1045 partial dependencies. */
1046 m_ATHLON_K8_AMDFAM10
| m_PPRO
| m_PENT4
| m_NOCONA
1047 | m_CORE2
| m_GENERIC
| m_GEODE
/* m_386 | m_K6 */,
1049 /* X86_TUNE_PARTIAL_REG_STALL: We probably ought to watch for partial
1050 register stalls on Generic32 compilation setting as well. However
1051 in current implementation the partial register stalls are not eliminated
1052 very well - they can be introduced via subregs synthesized by combine
1053 and can happen in caller/callee saving sequences. Because this option
1054 pays back little on PPro based chips and is in conflict with partial reg
1055 dependencies used by Athlon/P4 based chips, it is better to leave it off
1056 for generic32 for now. */
1059 /* X86_TUNE_PARTIAL_FLAG_REG_STALL */
1060 m_CORE2
| m_GENERIC
,
1062 /* X86_TUNE_USE_HIMODE_FIOP */
1063 m_386
| m_486
| m_K6_GEODE
,
1065 /* X86_TUNE_USE_SIMODE_FIOP */
1066 ~(m_PPRO
| m_ATHLON_K8_AMDFAM10
| m_PENT
| m_CORE2
| m_GENERIC
),
1068 /* X86_TUNE_USE_MOV0 */
1071 /* X86_TUNE_USE_CLTD */
1072 ~(m_PENT
| m_K6
| m_CORE2
| m_GENERIC
),
1074 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
1077 /* X86_TUNE_SPLIT_LONG_MOVES */
1080 /* X86_TUNE_READ_MODIFY_WRITE */
1083 /* X86_TUNE_READ_MODIFY */
1086 /* X86_TUNE_PROMOTE_QIMODE */
1087 m_K6_GEODE
| m_PENT
| m_386
| m_486
| m_ATHLON_K8_AMDFAM10
| m_CORE2
1088 | m_GENERIC
/* | m_PENT4 ? */,
1090 /* X86_TUNE_FAST_PREFIX */
1091 ~(m_PENT
| m_486
| m_386
),
1093 /* X86_TUNE_SINGLE_STRINGOP */
1094 m_386
| m_PENT4
| m_NOCONA
,
1096 /* X86_TUNE_QIMODE_MATH */
1099 /* X86_TUNE_HIMODE_MATH: On PPro this flag is meant to avoid partial
1100 register stalls. Just like X86_TUNE_PARTIAL_REG_STALL this option
1101 might be considered for Generic32 if our scheme for avoiding partial
1102 stalls was more effective. */
1105 /* X86_TUNE_PROMOTE_QI_REGS */
1108 /* X86_TUNE_PROMOTE_HI_REGS */
1111 /* X86_TUNE_ADD_ESP_4: Enable if add/sub is preferred over 1/2 push/pop. */
1112 m_ATHLON_K8_AMDFAM10
| m_K6_GEODE
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1114 /* X86_TUNE_ADD_ESP_8 */
1115 m_ATHLON_K8_AMDFAM10
| m_PPRO
| m_K6_GEODE
| m_386
1116 | m_486
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1118 /* X86_TUNE_SUB_ESP_4 */
1119 m_ATHLON_K8_AMDFAM10
| m_PPRO
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1121 /* X86_TUNE_SUB_ESP_8 */
1122 m_ATHLON_K8_AMDFAM10
| m_PPRO
| m_386
| m_486
1123 | m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1125 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
1126 for DFmode copies */
1127 ~(m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_NOCONA
| m_PPRO
| m_CORE2
1128 | m_GENERIC
| m_GEODE
),
1130 /* X86_TUNE_PARTIAL_REG_DEPENDENCY */
1131 m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1133 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: In the Generic model we have a
1134 conflict here in between PPro/Pentium4 based chips that thread 128bit
1135 SSE registers as single units versus K8 based chips that divide SSE
1136 registers to two 64bit halves. This knob promotes all store destinations
1137 to be 128bit to allow register renaming on 128bit SSE units, but usually
1138 results in one extra microop on 64bit SSE units. Experimental results
1139 shows that disabling this option on P4 brings over 20% SPECfp regression,
1140 while enabling it on K8 brings roughly 2.4% regression that can be partly
1141 masked by careful scheduling of moves. */
1142 m_PENT4
| m_NOCONA
| m_PPRO
| m_CORE2
| m_GENERIC
| m_AMDFAM10
,
1144 /* X86_TUNE_SSE_UNALIGNED_MOVE_OPTIMAL */
1147 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
1148 are resolved on SSE register parts instead of whole registers, so we may
1149 maintain just lower part of scalar values in proper format leaving the
1150 upper part undefined. */
1153 /* X86_TUNE_SSE_TYPELESS_STORES */
1154 m_ATHLON_K8_AMDFAM10
,
1156 /* X86_TUNE_SSE_LOAD0_BY_PXOR */
1157 m_PPRO
| m_PENT4
| m_NOCONA
,
1159 /* X86_TUNE_MEMORY_MISMATCH_STALL */
1160 m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1162 /* X86_TUNE_PROLOGUE_USING_MOVE */
1163 m_ATHLON_K8
| m_PPRO
| m_CORE2
| m_GENERIC
,
1165 /* X86_TUNE_EPILOGUE_USING_MOVE */
1166 m_ATHLON_K8
| m_PPRO
| m_CORE2
| m_GENERIC
,
1168 /* X86_TUNE_SHIFT1 */
1171 /* X86_TUNE_USE_FFREEP */
1172 m_ATHLON_K8_AMDFAM10
,
1174 /* X86_TUNE_INTER_UNIT_MOVES */
1175 ~(m_ATHLON_K8_AMDFAM10
| m_GENERIC
),
1177 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
1178 than 4 branch instructions in the 16 byte window. */
1179 m_PPRO
| m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_NOCONA
| m_CORE2
| m_GENERIC
,
1181 /* X86_TUNE_SCHEDULE */
1182 m_PPRO
| m_ATHLON_K8_AMDFAM10
| m_K6_GEODE
| m_PENT
| m_CORE2
| m_GENERIC
,
1184 /* X86_TUNE_USE_BT */
1185 m_ATHLON_K8_AMDFAM10
,
1187 /* X86_TUNE_USE_INCDEC */
1188 ~(m_PENT4
| m_NOCONA
| m_GENERIC
),
1190 /* X86_TUNE_PAD_RETURNS */
1191 m_ATHLON_K8_AMDFAM10
| m_CORE2
| m_GENERIC
,
1193 /* X86_TUNE_EXT_80387_CONSTANTS */
1194 m_K6_GEODE
| m_ATHLON_K8
| m_PENT4
| m_NOCONA
| m_PPRO
| m_CORE2
| m_GENERIC
,
1196 /* X86_TUNE_SHORTEN_X87_SSE */
1199 /* X86_TUNE_AVOID_VECTOR_DECODE */
1202 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
1203 and SImode multiply, but 386 and 486 do HImode multiply faster. */
1206 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
1207 vector path on AMD machines. */
1208 m_K8
| m_GENERIC64
| m_AMDFAM10
,
1210 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
1212 m_K8
| m_GENERIC64
| m_AMDFAM10
,
1214 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
1218 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
1219 but one byte longer. */
1222 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
1223 operand that cannot be represented using a modRM byte. The XOR
1224 replacement is long decoded, so this split helps here as well. */
1228 /* Feature tests against the various architecture variations. */
1229 unsigned int ix86_arch_features
[X86_ARCH_LAST
] = {
1230 /* X86_ARCH_CMOVE: Conditional move was added for pentiumpro. */
1231 ~(m_386
| m_486
| m_PENT
| m_K6
),
1233 /* X86_ARCH_CMPXCHG: Compare and exchange was added for 80486. */
1236 /* X86_ARCH_CMPXCHG8B: Compare and exchange 8 bytes was added for pentium. */
1239 /* X86_ARCH_XADD: Exchange and add was added for 80486. */
1242 /* X86_ARCH_BSWAP: Byteswap was added for 80486. */
1246 static const unsigned int x86_accumulate_outgoing_args
1247 = m_ATHLON_K8_AMDFAM10
| m_PENT4
| m_NOCONA
| m_PPRO
| m_CORE2
| m_GENERIC
;
1249 static const unsigned int x86_arch_always_fancy_math_387
1250 = m_PENT
| m_PPRO
| m_ATHLON_K8_AMDFAM10
| m_PENT4
1251 | m_NOCONA
| m_CORE2
| m_GENERIC
;
1253 static enum stringop_alg stringop_alg
= no_stringop
;
1255 /* In case the average insn count for single function invocation is
1256 lower than this constant, emit fast (but longer) prologue and
1258 #define FAST_PROLOGUE_INSN_COUNT 20
1260 /* Names for 8 (low), 8 (high), and 16-bit registers, respectively. */
1261 static const char *const qi_reg_name
[] = QI_REGISTER_NAMES
;
1262 static const char *const qi_high_reg_name
[] = QI_HIGH_REGISTER_NAMES
;
1263 static const char *const hi_reg_name
[] = HI_REGISTER_NAMES
;
1265 /* Array of the smallest class containing reg number REGNO, indexed by
1266 REGNO. Used by REGNO_REG_CLASS in i386.h. */
1268 enum reg_class
const regclass_map
[FIRST_PSEUDO_REGISTER
] =
1270 /* ax, dx, cx, bx */
1271 AREG
, DREG
, CREG
, BREG
,
1272 /* si, di, bp, sp */
1273 SIREG
, DIREG
, NON_Q_REGS
, NON_Q_REGS
,
1275 FP_TOP_REG
, FP_SECOND_REG
, FLOAT_REGS
, FLOAT_REGS
,
1276 FLOAT_REGS
, FLOAT_REGS
, FLOAT_REGS
, FLOAT_REGS
,
1279 /* flags, fpsr, fpcr, frame */
1280 NO_REGS
, NO_REGS
, NO_REGS
, NON_Q_REGS
,
1281 SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
,
1283 MMX_REGS
, MMX_REGS
, MMX_REGS
, MMX_REGS
, MMX_REGS
, MMX_REGS
,
1285 NON_Q_REGS
, NON_Q_REGS
, NON_Q_REGS
, NON_Q_REGS
,
1286 NON_Q_REGS
, NON_Q_REGS
, NON_Q_REGS
, NON_Q_REGS
,
1287 SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
, SSE_REGS
,
1291 /* The "default" register map used in 32bit mode. */
1293 int const dbx_register_map
[FIRST_PSEUDO_REGISTER
] =
1295 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */
1296 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */
1297 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1298 21, 22, 23, 24, 25, 26, 27, 28, /* SSE */
1299 29, 30, 31, 32, 33, 34, 35, 36, /* MMX */
1300 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1301 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1304 static int const x86_64_int_parameter_registers
[6] =
1306 5 /*RDI*/, 4 /*RSI*/, 1 /*RDX*/, 2 /*RCX*/,
1307 FIRST_REX_INT_REG
/*R8 */, FIRST_REX_INT_REG
+ 1 /*R9 */
1310 static int const x86_64_ms_abi_int_parameter_registers
[4] =
1312 2 /*RCX*/, 1 /*RDX*/,
1313 FIRST_REX_INT_REG
/*R8 */, FIRST_REX_INT_REG
+ 1 /*R9 */
1316 static int const x86_64_int_return_registers
[4] =
1318 0 /*RAX*/, 1 /*RDX*/, 5 /*RDI*/, 4 /*RSI*/
1321 /* The "default" register map used in 64bit mode. */
1322 int const dbx64_register_map
[FIRST_PSEUDO_REGISTER
] =
1324 0, 1, 2, 3, 4, 5, 6, 7, /* general regs */
1325 33, 34, 35, 36, 37, 38, 39, 40, /* fp regs */
1326 -1, -1, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1327 17, 18, 19, 20, 21, 22, 23, 24, /* SSE */
1328 41, 42, 43, 44, 45, 46, 47, 48, /* MMX */
1329 8,9,10,11,12,13,14,15, /* extended integer registers */
1330 25, 26, 27, 28, 29, 30, 31, 32, /* extended SSE registers */
1333 /* Define the register numbers to be used in Dwarf debugging information.
1334 The SVR4 reference port C compiler uses the following register numbers
1335 in its Dwarf output code:
1336 0 for %eax (gcc regno = 0)
1337 1 for %ecx (gcc regno = 2)
1338 2 for %edx (gcc regno = 1)
1339 3 for %ebx (gcc regno = 3)
1340 4 for %esp (gcc regno = 7)
1341 5 for %ebp (gcc regno = 6)
1342 6 for %esi (gcc regno = 4)
1343 7 for %edi (gcc regno = 5)
1344 The following three DWARF register numbers are never generated by
1345 the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4
1346 believes these numbers have these meanings.
1347 8 for %eip (no gcc equivalent)
1348 9 for %eflags (gcc regno = 17)
1349 10 for %trapno (no gcc equivalent)
1350 It is not at all clear how we should number the FP stack registers
1351 for the x86 architecture. If the version of SDB on x86/svr4 were
1352 a bit less brain dead with respect to floating-point then we would
1353 have a precedent to follow with respect to DWARF register numbers
1354 for x86 FP registers, but the SDB on x86/svr4 is so completely
1355 broken with respect to FP registers that it is hardly worth thinking
1356 of it as something to strive for compatibility with.
1357 The version of x86/svr4 SDB I have at the moment does (partially)
1358 seem to believe that DWARF register number 11 is associated with
1359 the x86 register %st(0), but that's about all. Higher DWARF
1360 register numbers don't seem to be associated with anything in
1361 particular, and even for DWARF regno 11, SDB only seems to under-
1362 stand that it should say that a variable lives in %st(0) (when
1363 asked via an `=' command) if we said it was in DWARF regno 11,
1364 but SDB still prints garbage when asked for the value of the
1365 variable in question (via a `/' command).
1366 (Also note that the labels SDB prints for various FP stack regs
1367 when doing an `x' command are all wrong.)
1368 Note that these problems generally don't affect the native SVR4
1369 C compiler because it doesn't allow the use of -O with -g and
1370 because when it is *not* optimizing, it allocates a memory
1371 location for each floating-point variable, and the memory
1372 location is what gets described in the DWARF AT_location
1373 attribute for the variable in question.
1374 Regardless of the severe mental illness of the x86/svr4 SDB, we
1375 do something sensible here and we use the following DWARF
1376 register numbers. Note that these are all stack-top-relative
1378 11 for %st(0) (gcc regno = 8)
1379 12 for %st(1) (gcc regno = 9)
1380 13 for %st(2) (gcc regno = 10)
1381 14 for %st(3) (gcc regno = 11)
1382 15 for %st(4) (gcc regno = 12)
1383 16 for %st(5) (gcc regno = 13)
1384 17 for %st(6) (gcc regno = 14)
1385 18 for %st(7) (gcc regno = 15)
1387 int const svr4_dbx_register_map
[FIRST_PSEUDO_REGISTER
] =
1389 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */
1390 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */
1391 -1, 9, -1, -1, -1, /* arg, flags, fpsr, fpcr, frame */
1392 21, 22, 23, 24, 25, 26, 27, 28, /* SSE registers */
1393 29, 30, 31, 32, 33, 34, 35, 36, /* MMX registers */
1394 -1, -1, -1, -1, -1, -1, -1, -1, /* extended integer registers */
1395 -1, -1, -1, -1, -1, -1, -1, -1, /* extended SSE registers */
1398 /* Test and compare insns in i386.md store the information needed to
1399 generate branch and scc insns here. */
1401 rtx ix86_compare_op0
= NULL_RTX
;
1402 rtx ix86_compare_op1
= NULL_RTX
;
1403 rtx ix86_compare_emitted
= NULL_RTX
;
1405 /* Size of the register save area. */
1406 #define X86_64_VARARGS_SIZE (REGPARM_MAX * UNITS_PER_WORD + SSE_REGPARM_MAX * 16)
1408 /* Define the structure for the machine field in struct function. */
1410 struct stack_local_entry
GTY(())
1412 unsigned short mode
;
1415 struct stack_local_entry
*next
;
1418 /* Structure describing stack frame layout.
1419 Stack grows downward:
1425 saved frame pointer if frame_pointer_needed
1426 <- HARD_FRAME_POINTER
1431 [va_arg registers] (
1432 > to_allocate <- FRAME_POINTER
1442 HOST_WIDE_INT frame
;
1444 int outgoing_arguments_size
;
1447 HOST_WIDE_INT to_allocate
;
1448 /* The offsets relative to ARG_POINTER. */
1449 HOST_WIDE_INT frame_pointer_offset
;
1450 HOST_WIDE_INT hard_frame_pointer_offset
;
1451 HOST_WIDE_INT stack_pointer_offset
;
1453 /* When save_regs_using_mov is set, emit prologue using
1454 move instead of push instructions. */
1455 bool save_regs_using_mov
;
1458 /* Code model option. */
1459 enum cmodel ix86_cmodel
;
1461 enum asm_dialect ix86_asm_dialect
= ASM_ATT
;
1463 enum tls_dialect ix86_tls_dialect
= TLS_DIALECT_GNU
;
1465 /* Which unit we are generating floating point math for. */
1466 enum fpmath_unit ix86_fpmath
;
1468 /* Which cpu are we scheduling for. */
1469 enum processor_type ix86_tune
;
1471 /* Which instruction set architecture to use. */
1472 enum processor_type ix86_arch
;
1474 /* true if sse prefetch instruction is not NOOP. */
1475 int x86_prefetch_sse
;
1477 /* ix86_regparm_string as a number */
1478 static int ix86_regparm
;
1480 /* -mstackrealign option */
1481 extern int ix86_force_align_arg_pointer
;
1482 static const char ix86_force_align_arg_pointer_string
[] = "force_align_arg_pointer";
1484 /* Preferred alignment for stack boundary in bits. */
1485 unsigned int ix86_preferred_stack_boundary
;
1487 /* Values 1-5: see jump.c */
1488 int ix86_branch_cost
;
1490 /* Variables which are this size or smaller are put in the data/bss
1491 or ldata/lbss sections. */
1493 int ix86_section_threshold
= 65536;
1495 /* Prefix built by ASM_GENERATE_INTERNAL_LABEL. */
1496 char internal_label_prefix
[16];
1497 int internal_label_prefix_len
;
1499 /* Register class used for passing given 64bit part of the argument.
1500 These represent classes as documented by the PS ABI, with the exception
1501 of SSESF, SSEDF classes, that are basically SSE class, just gcc will
1502 use SF or DFmode move instead of DImode to avoid reformatting penalties.
1504 Similarly we play games with INTEGERSI_CLASS to use cheaper SImode moves
1505 whenever possible (upper half does contain padding). */
1506 enum x86_64_reg_class
1509 X86_64_INTEGER_CLASS
,
1510 X86_64_INTEGERSI_CLASS
,
1517 X86_64_COMPLEX_X87_CLASS
,
1520 static const char * const x86_64_reg_class_name
[] =
1522 "no", "integer", "integerSI", "sse", "sseSF", "sseDF",
1523 "sseup", "x87", "x87up", "cplx87", "no"
1526 #define MAX_CLASSES 4
1528 /* Table of constants used by fldpi, fldln2, etc.... */
1529 static REAL_VALUE_TYPE ext_80387_constants_table
[5];
1530 static bool ext_80387_constants_init
= 0;
1533 static struct machine_function
* ix86_init_machine_status (void);
1534 static rtx
ix86_function_value (tree
, tree
, bool);
1535 static int ix86_function_regparm (tree
, tree
);
1536 static void ix86_compute_frame_layout (struct ix86_frame
*);
1537 static bool ix86_expand_vector_init_one_nonzero (bool, enum machine_mode
,
1541 /* The svr4 ABI for the i386 says that records and unions are returned
1543 #ifndef DEFAULT_PCC_STRUCT_RETURN
1544 #define DEFAULT_PCC_STRUCT_RETURN 1
1547 /* Bit flags that specify the ISA we are compiling for. */
1548 int ix86_isa_flags
= TARGET_64BIT_DEFAULT
| TARGET_SUBTARGET_ISA_DEFAULT
;
1550 /* A mask of ix86_isa_flags that includes bit X if X
1551 was set or cleared on the command line. */
1552 static int ix86_isa_flags_explicit
;
1554 /* Implement TARGET_HANDLE_OPTION. */
1557 ix86_handle_option (size_t code
, const char *arg ATTRIBUTE_UNUSED
, int value
)
1562 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_MMX
;
1566 &= ~(OPTION_MASK_ISA_3DNOW
| OPTION_MASK_ISA_3DNOW_A
);
1567 ix86_isa_flags_explicit
1568 |= OPTION_MASK_ISA_3DNOW
| OPTION_MASK_ISA_3DNOW_A
;
1573 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_3DNOW
;
1576 ix86_isa_flags
&= ~OPTION_MASK_ISA_3DNOW_A
;
1577 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_3DNOW_A
;
1585 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE
;
1589 &= ~(OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_SSE3
1590 | OPTION_MASK_ISA_SSSE3
| OPTION_MASK_ISA_SSE4_1
1591 | OPTION_MASK_ISA_SSE4A
);
1592 ix86_isa_flags_explicit
1593 |= (OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_SSE3
1594 | OPTION_MASK_ISA_SSSE3
| OPTION_MASK_ISA_SSE4_1
1595 | OPTION_MASK_ISA_SSE4A
);
1600 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE2
;
1604 &= ~(OPTION_MASK_ISA_SSE3
| OPTION_MASK_ISA_SSSE3
1605 | OPTION_MASK_ISA_SSE4_1
| OPTION_MASK_ISA_SSE4A
);
1606 ix86_isa_flags_explicit
1607 |= (OPTION_MASK_ISA_SSE3
| OPTION_MASK_ISA_SSSE3
1608 | OPTION_MASK_ISA_SSE4_1
| OPTION_MASK_ISA_SSE4A
);
1613 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE3
;
1617 &= ~(OPTION_MASK_ISA_SSSE3
| OPTION_MASK_ISA_SSE4_1
1618 | OPTION_MASK_ISA_SSE4A
);
1619 ix86_isa_flags_explicit
1620 |= (OPTION_MASK_ISA_SSSE3
| OPTION_MASK_ISA_SSE4_1
1621 | OPTION_MASK_ISA_SSE4A
);
1626 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSSE3
;
1630 &= ~(OPTION_MASK_ISA_SSE4_1
| OPTION_MASK_ISA_SSE4A
);
1631 ix86_isa_flags_explicit
1632 |= OPTION_MASK_ISA_SSE4_1
| OPTION_MASK_ISA_SSE4A
;
1637 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE4_1
;
1640 ix86_isa_flags
&= ~OPTION_MASK_ISA_SSE4A
;
1641 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE4A
;
1646 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE4A
;
1649 ix86_isa_flags
&= ~OPTION_MASK_ISA_SSE4_1
;
1650 ix86_isa_flags_explicit
|= OPTION_MASK_ISA_SSE4_1
;
1659 /* Sometimes certain combinations of command options do not make
1660 sense on a particular target machine. You can define a macro
1661 `OVERRIDE_OPTIONS' to take account of this. This macro, if
1662 defined, is executed once just after all the command options have
1665 Don't use this macro to turn on various extra optimizations for
1666 `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */
1669 override_options (void)
1672 int ix86_tune_defaulted
= 0;
1673 unsigned int ix86_arch_mask
, ix86_tune_mask
;
1675 /* Comes from final.c -- no real reason to change it. */
1676 #define MAX_CODE_ALIGN 16
1680 const struct processor_costs
*cost
; /* Processor costs */
1681 const int align_loop
; /* Default alignments. */
1682 const int align_loop_max_skip
;
1683 const int align_jump
;
1684 const int align_jump_max_skip
;
1685 const int align_func
;
1687 const processor_target_table
[PROCESSOR_max
] =
1689 {&i386_cost
, 4, 3, 4, 3, 4},
1690 {&i486_cost
, 16, 15, 16, 15, 16},
1691 {&pentium_cost
, 16, 7, 16, 7, 16},
1692 {&pentiumpro_cost
, 16, 15, 16, 7, 16},
1693 {&geode_cost
, 0, 0, 0, 0, 0},
1694 {&k6_cost
, 32, 7, 32, 7, 32},
1695 {&athlon_cost
, 16, 7, 16, 7, 16},
1696 {&pentium4_cost
, 0, 0, 0, 0, 0},
1697 {&k8_cost
, 16, 7, 16, 7, 16},
1698 {&nocona_cost
, 0, 0, 0, 0, 0},
1699 {&core2_cost
, 16, 7, 16, 7, 16},
1700 {&generic32_cost
, 16, 7, 16, 7, 16},
1701 {&generic64_cost
, 16, 7, 16, 7, 16},
1702 {&amdfam10_cost
, 32, 24, 32, 7, 32}
1705 static const char * const cpu_names
[] = TARGET_CPU_DEFAULT_NAMES
;
1708 const char *const name
; /* processor name or nickname. */
1709 const enum processor_type processor
;
1710 const enum pta_flags
1716 PTA_PREFETCH_SSE
= 1 << 4,
1718 PTA_3DNOW_A
= 1 << 6,
1722 PTA_POPCNT
= 1 << 10,
1724 PTA_SSE4A
= 1 << 12,
1725 PTA_NO_SAHF
= 1 << 13,
1726 PTA_SSE4_1
= 1 << 14
1729 const processor_alias_table
[] =
1731 {"i386", PROCESSOR_I386
, 0},
1732 {"i486", PROCESSOR_I486
, 0},
1733 {"i586", PROCESSOR_PENTIUM
, 0},
1734 {"pentium", PROCESSOR_PENTIUM
, 0},
1735 {"pentium-mmx", PROCESSOR_PENTIUM
, PTA_MMX
},
1736 {"winchip-c6", PROCESSOR_I486
, PTA_MMX
},
1737 {"winchip2", PROCESSOR_I486
, PTA_MMX
| PTA_3DNOW
},
1738 {"c3", PROCESSOR_I486
, PTA_MMX
| PTA_3DNOW
},
1739 {"c3-2", PROCESSOR_PENTIUMPRO
, PTA_MMX
| PTA_SSE
},
1740 {"i686", PROCESSOR_PENTIUMPRO
, 0},
1741 {"pentiumpro", PROCESSOR_PENTIUMPRO
, 0},
1742 {"pentium2", PROCESSOR_PENTIUMPRO
, PTA_MMX
},
1743 {"pentium3", PROCESSOR_PENTIUMPRO
, PTA_MMX
| PTA_SSE
},
1744 {"pentium3m", PROCESSOR_PENTIUMPRO
, PTA_MMX
| PTA_SSE
},
1745 {"pentium-m", PROCESSOR_PENTIUMPRO
, PTA_MMX
| PTA_SSE
| PTA_SSE2
},
1746 {"pentium4", PROCESSOR_PENTIUM4
, PTA_MMX
|PTA_SSE
| PTA_SSE2
},
1747 {"pentium4m", PROCESSOR_PENTIUM4
, PTA_MMX
| PTA_SSE
| PTA_SSE2
},
1748 {"prescott", PROCESSOR_NOCONA
, PTA_MMX
| PTA_SSE
| PTA_SSE2
| PTA_SSE3
},
1749 {"nocona", PROCESSOR_NOCONA
, (PTA_64BIT
1750 | PTA_MMX
| PTA_SSE
| PTA_SSE2
| PTA_SSE3
1751 | PTA_CX16
| PTA_NO_SAHF
)},
1752 {"core2", PROCESSOR_CORE2
, (PTA_64BIT
1753 | PTA_MMX
| PTA_SSE
| PTA_SSE2
| PTA_SSE3
1756 {"geode", PROCESSOR_GEODE
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1757 |PTA_PREFETCH_SSE
)},
1758 {"k6", PROCESSOR_K6
, PTA_MMX
},
1759 {"k6-2", PROCESSOR_K6
, PTA_MMX
| PTA_3DNOW
},
1760 {"k6-3", PROCESSOR_K6
, PTA_MMX
| PTA_3DNOW
},
1761 {"athlon", PROCESSOR_ATHLON
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1762 | PTA_PREFETCH_SSE
)},
1763 {"athlon-tbird", PROCESSOR_ATHLON
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1764 | PTA_PREFETCH_SSE
)},
1765 {"athlon-4", PROCESSOR_ATHLON
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1767 {"athlon-xp", PROCESSOR_ATHLON
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1769 {"athlon-mp", PROCESSOR_ATHLON
, (PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1771 {"x86-64", PROCESSOR_K8
, (PTA_64BIT
1772 | PTA_MMX
| PTA_SSE
| PTA_SSE2
1774 {"k8", PROCESSOR_K8
, (PTA_64BIT
1775 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1776 | PTA_SSE
| PTA_SSE2
1778 {"k8-sse3", PROCESSOR_K8
, (PTA_64BIT
1779 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1780 | PTA_SSE
| PTA_SSE2
| PTA_SSE3
1782 {"opteron", PROCESSOR_K8
, (PTA_64BIT
1783 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1784 | PTA_SSE
| PTA_SSE2
1786 {"opteron-sse3", PROCESSOR_K8
, (PTA_64BIT
1787 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1788 | PTA_SSE
| PTA_SSE2
| PTA_SSE3
1790 {"athlon64", PROCESSOR_K8
, (PTA_64BIT
1791 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1792 | PTA_SSE
| PTA_SSE2
1794 {"athlon64-sse3", PROCESSOR_K8
, (PTA_64BIT
1795 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1796 | PTA_SSE
| PTA_SSE2
| PTA_SSE3
1798 {"athlon-fx", PROCESSOR_K8
, (PTA_64BIT
1799 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1800 | PTA_SSE
| PTA_SSE2
1802 {"amdfam10", PROCESSOR_AMDFAM10
, (PTA_64BIT
1803 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1804 | PTA_SSE
| PTA_SSE2
| PTA_SSE3
1806 | PTA_CX16
| PTA_ABM
)},
1807 {"barcelona", PROCESSOR_AMDFAM10
, (PTA_64BIT
1808 | PTA_MMX
| PTA_3DNOW
| PTA_3DNOW_A
1809 | PTA_SSE
| PTA_SSE2
| PTA_SSE3
1811 | PTA_CX16
| PTA_ABM
)},
1812 {"generic32", PROCESSOR_GENERIC32
, 0 /* flags are only used for -march switch. */ },
1813 {"generic64", PROCESSOR_GENERIC64
, PTA_64BIT
/* flags are only used for -march switch. */ },
1816 int const pta_size
= ARRAY_SIZE (processor_alias_table
);
1818 #ifdef SUBTARGET_OVERRIDE_OPTIONS
1819 SUBTARGET_OVERRIDE_OPTIONS
;
1822 #ifdef SUBSUBTARGET_OVERRIDE_OPTIONS
1823 SUBSUBTARGET_OVERRIDE_OPTIONS
;
1826 /* -fPIC is the default for x86_64. */
1827 if (TARGET_MACHO
&& TARGET_64BIT
)
1830 /* Set the default values for switches whose default depends on TARGET_64BIT
1831 in case they weren't overwritten by command line options. */
1834 /* Mach-O doesn't support omitting the frame pointer for now. */
1835 if (flag_omit_frame_pointer
== 2)
1836 flag_omit_frame_pointer
= (TARGET_MACHO
? 0 : 1);
1837 if (flag_asynchronous_unwind_tables
== 2)
1838 flag_asynchronous_unwind_tables
= 1;
1839 if (flag_pcc_struct_return
== 2)
1840 flag_pcc_struct_return
= 0;
1844 if (flag_omit_frame_pointer
== 2)
1845 flag_omit_frame_pointer
= 0;
1846 if (flag_asynchronous_unwind_tables
== 2)
1847 flag_asynchronous_unwind_tables
= 0;
1848 if (flag_pcc_struct_return
== 2)
1849 flag_pcc_struct_return
= DEFAULT_PCC_STRUCT_RETURN
;
1852 /* Need to check -mtune=generic first. */
1853 if (ix86_tune_string
)
1855 if (!strcmp (ix86_tune_string
, "generic")
1856 || !strcmp (ix86_tune_string
, "i686")
1857 /* As special support for cross compilers we read -mtune=native
1858 as -mtune=generic. With native compilers we won't see the
1859 -mtune=native, as it was changed by the driver. */
1860 || !strcmp (ix86_tune_string
, "native"))
1863 ix86_tune_string
= "generic64";
1865 ix86_tune_string
= "generic32";
1867 else if (!strncmp (ix86_tune_string
, "generic", 7))
1868 error ("bad value (%s) for -mtune= switch", ix86_tune_string
);
1872 if (ix86_arch_string
)
1873 ix86_tune_string
= ix86_arch_string
;
1874 if (!ix86_tune_string
)
1876 ix86_tune_string
= cpu_names
[TARGET_CPU_DEFAULT
];
1877 ix86_tune_defaulted
= 1;
1880 /* ix86_tune_string is set to ix86_arch_string or defaulted. We
1881 need to use a sensible tune option. */
1882 if (!strcmp (ix86_tune_string
, "generic")
1883 || !strcmp (ix86_tune_string
, "x86-64")
1884 || !strcmp (ix86_tune_string
, "i686"))
1887 ix86_tune_string
= "generic64";
1889 ix86_tune_string
= "generic32";
1892 if (ix86_stringop_string
)
1894 if (!strcmp (ix86_stringop_string
, "rep_byte"))
1895 stringop_alg
= rep_prefix_1_byte
;
1896 else if (!strcmp (ix86_stringop_string
, "libcall"))
1897 stringop_alg
= libcall
;
1898 else if (!strcmp (ix86_stringop_string
, "rep_4byte"))
1899 stringop_alg
= rep_prefix_4_byte
;
1900 else if (!strcmp (ix86_stringop_string
, "rep_8byte"))
1901 stringop_alg
= rep_prefix_8_byte
;
1902 else if (!strcmp (ix86_stringop_string
, "byte_loop"))
1903 stringop_alg
= loop_1_byte
;
1904 else if (!strcmp (ix86_stringop_string
, "loop"))
1905 stringop_alg
= loop
;
1906 else if (!strcmp (ix86_stringop_string
, "unrolled_loop"))
1907 stringop_alg
= unrolled_loop
;
1909 error ("bad value (%s) for -mstringop-strategy= switch", ix86_stringop_string
);
1911 if (!strcmp (ix86_tune_string
, "x86-64"))
1912 warning (OPT_Wdeprecated
, "-mtune=x86-64 is deprecated. Use -mtune=k8 or "
1913 "-mtune=generic instead as appropriate.");
1915 if (!ix86_arch_string
)
1916 ix86_arch_string
= TARGET_64BIT
? "x86-64" : "i386";
1917 if (!strcmp (ix86_arch_string
, "generic"))
1918 error ("generic CPU can be used only for -mtune= switch");
1919 if (!strncmp (ix86_arch_string
, "generic", 7))
1920 error ("bad value (%s) for -march= switch", ix86_arch_string
);
1922 if (ix86_cmodel_string
!= 0)
1924 if (!strcmp (ix86_cmodel_string
, "small"))
1925 ix86_cmodel
= flag_pic
? CM_SMALL_PIC
: CM_SMALL
;
1926 else if (!strcmp (ix86_cmodel_string
, "medium"))
1927 ix86_cmodel
= flag_pic
? CM_MEDIUM_PIC
: CM_MEDIUM
;
1928 else if (!strcmp (ix86_cmodel_string
, "large"))
1929 ix86_cmodel
= flag_pic
? CM_LARGE_PIC
: CM_LARGE
;
1931 error ("code model %s does not support PIC mode", ix86_cmodel_string
);
1932 else if (!strcmp (ix86_cmodel_string
, "32"))
1933 ix86_cmodel
= CM_32
;
1934 else if (!strcmp (ix86_cmodel_string
, "kernel") && !flag_pic
)
1935 ix86_cmodel
= CM_KERNEL
;
1937 error ("bad value (%s) for -mcmodel= switch", ix86_cmodel_string
);
1941 /* For TARGET_64BIT_MS_ABI, force pic on, in order to enable the
1942 use of rip-relative addressing. This eliminates fixups that
1943 would otherwise be needed if this object is to be placed in a
1944 DLL, and is essentially just as efficient as direct addressing. */
1945 if (TARGET_64BIT_MS_ABI
)
1946 ix86_cmodel
= CM_SMALL_PIC
, flag_pic
= 1;
1947 else if (TARGET_64BIT
)
1948 ix86_cmodel
= flag_pic
? CM_SMALL_PIC
: CM_SMALL
;
1950 ix86_cmodel
= CM_32
;
1952 if (ix86_asm_string
!= 0)
1955 && !strcmp (ix86_asm_string
, "intel"))
1956 ix86_asm_dialect
= ASM_INTEL
;
1957 else if (!strcmp (ix86_asm_string
, "att"))
1958 ix86_asm_dialect
= ASM_ATT
;
1960 error ("bad value (%s) for -masm= switch", ix86_asm_string
);
1962 if ((TARGET_64BIT
== 0) != (ix86_cmodel
== CM_32
))
1963 error ("code model %qs not supported in the %s bit mode",
1964 ix86_cmodel_string
, TARGET_64BIT
? "64" : "32");
1965 if ((TARGET_64BIT
!= 0) != ((ix86_isa_flags
& OPTION_MASK_ISA_64BIT
) != 0))
1966 sorry ("%i-bit mode not compiled in",
1967 (ix86_isa_flags
& OPTION_MASK_ISA_64BIT
) ? 64 : 32);
1969 for (i
= 0; i
< pta_size
; i
++)
1970 if (! strcmp (ix86_arch_string
, processor_alias_table
[i
].name
))
1972 ix86_arch
= processor_alias_table
[i
].processor
;
1973 /* Default cpu tuning to the architecture. */
1974 ix86_tune
= ix86_arch
;
1976 if (TARGET_64BIT
&& !(processor_alias_table
[i
].flags
& PTA_64BIT
))
1977 error ("CPU you selected does not support x86-64 "
1980 if (processor_alias_table
[i
].flags
& PTA_MMX
1981 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_MMX
))
1982 ix86_isa_flags
|= OPTION_MASK_ISA_MMX
;
1983 if (processor_alias_table
[i
].flags
& PTA_3DNOW
1984 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_3DNOW
))
1985 ix86_isa_flags
|= OPTION_MASK_ISA_3DNOW
;
1986 if (processor_alias_table
[i
].flags
& PTA_3DNOW_A
1987 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_3DNOW_A
))
1988 ix86_isa_flags
|= OPTION_MASK_ISA_3DNOW_A
;
1989 if (processor_alias_table
[i
].flags
& PTA_SSE
1990 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSE
))
1991 ix86_isa_flags
|= OPTION_MASK_ISA_SSE
;
1992 if (processor_alias_table
[i
].flags
& PTA_SSE2
1993 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSE2
))
1994 ix86_isa_flags
|= OPTION_MASK_ISA_SSE2
;
1995 if (processor_alias_table
[i
].flags
& PTA_SSE3
1996 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSE3
))
1997 ix86_isa_flags
|= OPTION_MASK_ISA_SSE3
;
1998 if (processor_alias_table
[i
].flags
& PTA_SSSE3
1999 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSSE3
))
2000 ix86_isa_flags
|= OPTION_MASK_ISA_SSSE3
;
2001 if (processor_alias_table
[i
].flags
& PTA_SSE4_1
2002 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSE4_1
))
2003 ix86_isa_flags
|= OPTION_MASK_ISA_SSE4_1
;
2004 if (processor_alias_table
[i
].flags
& PTA_SSE4A
2005 && !(ix86_isa_flags_explicit
& OPTION_MASK_ISA_SSE4A
))
2006 ix86_isa_flags
|= OPTION_MASK_ISA_SSE4A
;
2008 if (processor_alias_table
[i
].flags
& PTA_ABM
)
2010 if (processor_alias_table
[i
].flags
& PTA_CX16
)
2011 x86_cmpxchg16b
= true;
2012 if (processor_alias_table
[i
].flags
& (PTA_POPCNT
| PTA_ABM
))
2014 if (processor_alias_table
[i
].flags
& (PTA_PREFETCH_SSE
| PTA_SSE
))
2015 x86_prefetch_sse
= true;
2016 if ((processor_alias_table
[i
].flags
& PTA_NO_SAHF
) && !TARGET_64BIT
)
2023 error ("bad value (%s) for -march= switch", ix86_arch_string
);
2025 ix86_arch_mask
= 1u << ix86_arch
;
2026 for (i
= 0; i
< X86_ARCH_LAST
; ++i
)
2027 ix86_arch_features
[i
] &= ix86_arch_mask
;
2029 for (i
= 0; i
< pta_size
; i
++)
2030 if (! strcmp (ix86_tune_string
, processor_alias_table
[i
].name
))
2032 ix86_tune
= processor_alias_table
[i
].processor
;
2033 if (TARGET_64BIT
&& !(processor_alias_table
[i
].flags
& PTA_64BIT
))
2035 if (ix86_tune_defaulted
)
2037 ix86_tune_string
= "x86-64";
2038 for (i
= 0; i
< pta_size
; i
++)
2039 if (! strcmp (ix86_tune_string
,
2040 processor_alias_table
[i
].name
))
2042 ix86_tune
= processor_alias_table
[i
].processor
;
2045 error ("CPU you selected does not support x86-64 "
2048 /* Intel CPUs have always interpreted SSE prefetch instructions as
2049 NOPs; so, we can enable SSE prefetch instructions even when
2050 -mtune (rather than -march) points us to a processor that has them.
2051 However, the VIA C3 gives a SIGILL, so we only do that for i686 and
2052 higher processors. */
2054 && (processor_alias_table
[i
].flags
& (PTA_PREFETCH_SSE
| PTA_SSE
)))
2055 x86_prefetch_sse
= true;
2059 error ("bad value (%s) for -mtune= switch", ix86_tune_string
);
2061 ix86_tune_mask
= 1u << ix86_tune
;
2062 for (i
= 0; i
< X86_TUNE_LAST
; ++i
)
2063 ix86_tune_features
[i
] &= ix86_tune_mask
;
2066 ix86_cost
= &size_cost
;
2068 ix86_cost
= processor_target_table
[ix86_tune
].cost
;
2070 /* Arrange to set up i386_stack_locals for all functions. */
2071 init_machine_status
= ix86_init_machine_status
;
2073 /* Validate -mregparm= value. */
2074 if (ix86_regparm_string
)
2077 warning (0, "-mregparm is ignored in 64-bit mode");
2078 i
= atoi (ix86_regparm_string
);
2079 if (i
< 0 || i
> REGPARM_MAX
)
2080 error ("-mregparm=%d is not between 0 and %d", i
, REGPARM_MAX
);
2085 ix86_regparm
= REGPARM_MAX
;
2087 /* If the user has provided any of the -malign-* options,
2088 warn and use that value only if -falign-* is not set.
2089 Remove this code in GCC 3.2 or later. */
2090 if (ix86_align_loops_string
)
2092 warning (0, "-malign-loops is obsolete, use -falign-loops");
2093 if (align_loops
== 0)
2095 i
= atoi (ix86_align_loops_string
);
2096 if (i
< 0 || i
> MAX_CODE_ALIGN
)
2097 error ("-malign-loops=%d is not between 0 and %d", i
, MAX_CODE_ALIGN
);
2099 align_loops
= 1 << i
;
2103 if (ix86_align_jumps_string
)
2105 warning (0, "-malign-jumps is obsolete, use -falign-jumps");
2106 if (align_jumps
== 0)
2108 i
= atoi (ix86_align_jumps_string
);
2109 if (i
< 0 || i
> MAX_CODE_ALIGN
)
2110 error ("-malign-loops=%d is not between 0 and %d", i
, MAX_CODE_ALIGN
);
2112 align_jumps
= 1 << i
;
2116 if (ix86_align_funcs_string
)
2118 warning (0, "-malign-functions is obsolete, use -falign-functions");
2119 if (align_functions
== 0)
2121 i
= atoi (ix86_align_funcs_string
);
2122 if (i
< 0 || i
> MAX_CODE_ALIGN
)
2123 error ("-malign-loops=%d is not between 0 and %d", i
, MAX_CODE_ALIGN
);
2125 align_functions
= 1 << i
;
2129 /* Default align_* from the processor table. */
2130 if (align_loops
== 0)
2132 align_loops
= processor_target_table
[ix86_tune
].align_loop
;
2133 align_loops_max_skip
= processor_target_table
[ix86_tune
].align_loop_max_skip
;
2135 if (align_jumps
== 0)
2137 align_jumps
= processor_target_table
[ix86_tune
].align_jump
;
2138 align_jumps_max_skip
= processor_target_table
[ix86_tune
].align_jump_max_skip
;
2140 if (align_functions
== 0)
2142 align_functions
= processor_target_table
[ix86_tune
].align_func
;
2145 /* Validate -mbranch-cost= value, or provide default. */
2146 ix86_branch_cost
= ix86_cost
->branch_cost
;
2147 if (ix86_branch_cost_string
)
2149 i
= atoi (ix86_branch_cost_string
);
2151 error ("-mbranch-cost=%d is not between 0 and 5", i
);
2153 ix86_branch_cost
= i
;
2155 if (ix86_section_threshold_string
)
2157 i
= atoi (ix86_section_threshold_string
);
2159 error ("-mlarge-data-threshold=%d is negative", i
);
2161 ix86_section_threshold
= i
;
2164 if (ix86_tls_dialect_string
)
2166 if (strcmp (ix86_tls_dialect_string
, "gnu") == 0)
2167 ix86_tls_dialect
= TLS_DIALECT_GNU
;
2168 else if (strcmp (ix86_tls_dialect_string
, "gnu2") == 0)
2169 ix86_tls_dialect
= TLS_DIALECT_GNU2
;
2170 else if (strcmp (ix86_tls_dialect_string
, "sun") == 0)
2171 ix86_tls_dialect
= TLS_DIALECT_SUN
;
2173 error ("bad value (%s) for -mtls-dialect= switch",
2174 ix86_tls_dialect_string
);
2177 if (ix87_precision_string
)
2179 i
= atoi (ix87_precision_string
);
2180 if (i
!= 32 && i
!= 64 && i
!= 80)
2181 error ("pc%d is not valid precision setting (32, 64 or 80)", i
);
2184 /* Keep nonleaf frame pointers. */
2185 if (flag_omit_frame_pointer
)
2186 target_flags
&= ~MASK_OMIT_LEAF_FRAME_POINTER
;
2187 else if (TARGET_OMIT_LEAF_FRAME_POINTER
)
2188 flag_omit_frame_pointer
= 1;
2190 /* If we're doing fast math, we don't care about comparison order
2191 wrt NaNs. This lets us use a shorter comparison sequence. */
2192 if (flag_finite_math_only
)
2193 target_flags
&= ~MASK_IEEE_FP
;
2195 /* If the architecture always has an FPU, turn off NO_FANCY_MATH_387,
2196 since the insns won't need emulation. */
2197 if (x86_arch_always_fancy_math_387
& ix86_arch_mask
)
2198 target_flags
&= ~MASK_NO_FANCY_MATH_387
;
2200 /* Likewise, if the target doesn't have a 387, or we've specified
2201 software floating point, don't use 387 inline intrinsics. */
2203 target_flags
|= MASK_NO_FANCY_MATH_387
;
2205 /* Turn on SSSE3 builtins for -msse4.1. */
2207 ix86_isa_flags
|= OPTION_MASK_ISA_SSSE3
;
2209 /* Turn on SSE3 builtins for -mssse3. */
2211 ix86_isa_flags
|= OPTION_MASK_ISA_SSE3
;
2213 /* Turn on SSE3 builtins for -msse4a. */
2215 ix86_isa_flags
|= OPTION_MASK_ISA_SSE3
;
2217 /* Turn on SSE2 builtins for -msse3. */
2219 ix86_isa_flags
|= OPTION_MASK_ISA_SSE2
;
2221 /* Turn on SSE builtins for -msse2. */
2223 ix86_isa_flags
|= OPTION_MASK_ISA_SSE
;
2225 /* Turn on MMX builtins for -msse. */
2228 ix86_isa_flags
|= OPTION_MASK_ISA_MMX
& ~ix86_isa_flags_explicit
;
2229 x86_prefetch_sse
= true;
2232 /* Turn on MMX builtins for 3Dnow. */
2234 ix86_isa_flags
|= OPTION_MASK_ISA_MMX
;
2236 /* Turn on POPCNT builtins for -mabm. */
2242 target_flags
|= TARGET_SUBTARGET64_DEFAULT
& ~target_flags_explicit
;
2244 /* Enable by default the SSE and MMX builtins. Do allow the user to
2245 explicitly disable any of these. In particular, disabling SSE and
2246 MMX for kernel code is extremely useful. */
2248 |= ((OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_MMX
2249 | TARGET_SUBTARGET64_ISA_DEFAULT
) & ~ix86_isa_flags_explicit
);
2252 warning (0, "-mrtd is ignored in 64bit mode");
2256 target_flags
|= TARGET_SUBTARGET32_DEFAULT
& ~target_flags_explicit
;
2259 |= TARGET_SUBTARGET32_DEFAULT
& ~ix86_isa_flags_explicit
;
2261 /* i386 ABI does not specify red zone. It still makes sense to use it
2262 when programmer takes care to stack from being destroyed. */
2263 if (!(target_flags_explicit
& MASK_NO_RED_ZONE
))
2264 target_flags
|= MASK_NO_RED_ZONE
;
2267 /* Validate -mpreferred-stack-boundary= value, or provide default.
2268 The default of 128 bits is for Pentium III's SSE __m128. We can't
2269 change it because of optimize_size. Otherwise, we can't mix object
2270 files compiled with -Os and -On. */
2271 ix86_preferred_stack_boundary
= 128;
2272 if (ix86_preferred_stack_boundary_string
)
2274 i
= atoi (ix86_preferred_stack_boundary_string
);
2275 if (i
< (TARGET_64BIT
? 4 : 2) || i
> 12)
2276 error ("-mpreferred-stack-boundary=%d is not between %d and 12", i
,
2277 TARGET_64BIT
? 4 : 2);
2279 ix86_preferred_stack_boundary
= (1 << i
) * BITS_PER_UNIT
;
2282 /* Accept -msseregparm only if at least SSE support is enabled. */
2283 if (TARGET_SSEREGPARM
2285 error ("-msseregparm used without SSE enabled");
2287 ix86_fpmath
= TARGET_FPMATH_DEFAULT
;
2288 if (ix86_fpmath_string
!= 0)
2290 if (! strcmp (ix86_fpmath_string
, "387"))
2291 ix86_fpmath
= FPMATH_387
;
2292 else if (! strcmp (ix86_fpmath_string
, "sse"))
2296 warning (0, "SSE instruction set disabled, using 387 arithmetics");
2297 ix86_fpmath
= FPMATH_387
;
2300 ix86_fpmath
= FPMATH_SSE
;
2302 else if (! strcmp (ix86_fpmath_string
, "387,sse")
2303 || ! strcmp (ix86_fpmath_string
, "sse,387"))
2307 warning (0, "SSE instruction set disabled, using 387 arithmetics");
2308 ix86_fpmath
= FPMATH_387
;
2310 else if (!TARGET_80387
)
2312 warning (0, "387 instruction set disabled, using SSE arithmetics");
2313 ix86_fpmath
= FPMATH_SSE
;
2316 ix86_fpmath
= FPMATH_SSE
| FPMATH_387
;
2319 error ("bad value (%s) for -mfpmath= switch", ix86_fpmath_string
);
2322 /* If the i387 is disabled, then do not return values in it. */
2324 target_flags
&= ~MASK_FLOAT_RETURNS
;
2326 if ((x86_accumulate_outgoing_args
& ix86_tune_mask
)
2327 && !(target_flags_explicit
& MASK_ACCUMULATE_OUTGOING_ARGS
)
2329 target_flags
|= MASK_ACCUMULATE_OUTGOING_ARGS
;
2331 /* ??? Unwind info is not correct around the CFG unless either a frame
2332 pointer is present or M_A_O_A is set. Fixing this requires rewriting
2333 unwind info generation to be aware of the CFG and propagating states
2335 if ((flag_unwind_tables
|| flag_asynchronous_unwind_tables
2336 || flag_exceptions
|| flag_non_call_exceptions
)
2337 && flag_omit_frame_pointer
2338 && !(target_flags
& MASK_ACCUMULATE_OUTGOING_ARGS
))
2340 if (target_flags_explicit
& MASK_ACCUMULATE_OUTGOING_ARGS
)
2341 warning (0, "unwind tables currently require either a frame pointer "
2342 "or -maccumulate-outgoing-args for correctness");
2343 target_flags
|= MASK_ACCUMULATE_OUTGOING_ARGS
;
2346 /* For sane SSE instruction set generation we need fcomi instruction.
2347 It is safe to enable all CMOVE instructions. */
2351 /* Figure out what ASM_GENERATE_INTERNAL_LABEL builds as a prefix. */
2354 ASM_GENERATE_INTERNAL_LABEL (internal_label_prefix
, "LX", 0);
2355 p
= strchr (internal_label_prefix
, 'X');
2356 internal_label_prefix_len
= p
- internal_label_prefix
;
2360 /* When scheduling description is not available, disable scheduler pass
2361 so it won't slow down the compilation and make x87 code slower. */
2362 if (!TARGET_SCHEDULE
)
2363 flag_schedule_insns_after_reload
= flag_schedule_insns
= 0;
2365 if (!PARAM_SET_P (PARAM_SIMULTANEOUS_PREFETCHES
))
2366 set_param_value ("simultaneous-prefetches",
2367 ix86_cost
->simultaneous_prefetches
);
2368 if (!PARAM_SET_P (PARAM_L1_CACHE_LINE_SIZE
))
2369 set_param_value ("l1-cache-line-size", ix86_cost
->prefetch_block
);
2372 /* Return true if this goes in large data/bss. */
2375 ix86_in_large_data_p (tree exp
)
2377 if (ix86_cmodel
!= CM_MEDIUM
&& ix86_cmodel
!= CM_MEDIUM_PIC
)
2380 /* Functions are never large data. */
2381 if (TREE_CODE (exp
) == FUNCTION_DECL
)
2384 if (TREE_CODE (exp
) == VAR_DECL
&& DECL_SECTION_NAME (exp
))
2386 const char *section
= TREE_STRING_POINTER (DECL_SECTION_NAME (exp
));
2387 if (strcmp (section
, ".ldata") == 0
2388 || strcmp (section
, ".lbss") == 0)
2394 HOST_WIDE_INT size
= int_size_in_bytes (TREE_TYPE (exp
));
2396 /* If this is an incomplete type with size 0, then we can't put it
2397 in data because it might be too big when completed. */
2398 if (!size
|| size
> ix86_section_threshold
)
2405 /* Switch to the appropriate section for output of DECL.
2406 DECL is either a `VAR_DECL' node or a constant of some sort.
2407 RELOC indicates whether forming the initial value of DECL requires
2408 link-time relocations. */
2410 static section
* x86_64_elf_select_section (tree
, int, unsigned HOST_WIDE_INT
)
2414 x86_64_elf_select_section (tree decl
, int reloc
,
2415 unsigned HOST_WIDE_INT align
)
2417 if ((ix86_cmodel
== CM_MEDIUM
|| ix86_cmodel
== CM_MEDIUM_PIC
)
2418 && ix86_in_large_data_p (decl
))
2420 const char *sname
= NULL
;
2421 unsigned int flags
= SECTION_WRITE
;
2422 switch (categorize_decl_for_section (decl
, reloc
))
2427 case SECCAT_DATA_REL
:
2428 sname
= ".ldata.rel";
2430 case SECCAT_DATA_REL_LOCAL
:
2431 sname
= ".ldata.rel.local";
2433 case SECCAT_DATA_REL_RO
:
2434 sname
= ".ldata.rel.ro";
2436 case SECCAT_DATA_REL_RO_LOCAL
:
2437 sname
= ".ldata.rel.ro.local";
2441 flags
|= SECTION_BSS
;
2444 case SECCAT_RODATA_MERGE_STR
:
2445 case SECCAT_RODATA_MERGE_STR_INIT
:
2446 case SECCAT_RODATA_MERGE_CONST
:
2450 case SECCAT_SRODATA
:
2457 /* We don't split these for medium model. Place them into
2458 default sections and hope for best. */
2463 /* We might get called with string constants, but get_named_section
2464 doesn't like them as they are not DECLs. Also, we need to set
2465 flags in that case. */
2467 return get_section (sname
, flags
, NULL
);
2468 return get_named_section (decl
, sname
, reloc
);
2471 return default_elf_select_section (decl
, reloc
, align
);
2474 /* Build up a unique section name, expressed as a
2475 STRING_CST node, and assign it to DECL_SECTION_NAME (decl).
2476 RELOC indicates whether the initial value of EXP requires
2477 link-time relocations. */
2479 static void ATTRIBUTE_UNUSED
2480 x86_64_elf_unique_section (tree decl
, int reloc
)
2482 if ((ix86_cmodel
== CM_MEDIUM
|| ix86_cmodel
== CM_MEDIUM_PIC
)
2483 && ix86_in_large_data_p (decl
))
2485 const char *prefix
= NULL
;
2486 /* We only need to use .gnu.linkonce if we don't have COMDAT groups. */
2487 bool one_only
= DECL_ONE_ONLY (decl
) && !HAVE_COMDAT_GROUP
;
2489 switch (categorize_decl_for_section (decl
, reloc
))
2492 case SECCAT_DATA_REL
:
2493 case SECCAT_DATA_REL_LOCAL
:
2494 case SECCAT_DATA_REL_RO
:
2495 case SECCAT_DATA_REL_RO_LOCAL
:
2496 prefix
= one_only
? ".gnu.linkonce.ld." : ".ldata.";
2499 prefix
= one_only
? ".gnu.linkonce.lb." : ".lbss.";
2502 case SECCAT_RODATA_MERGE_STR
:
2503 case SECCAT_RODATA_MERGE_STR_INIT
:
2504 case SECCAT_RODATA_MERGE_CONST
:
2505 prefix
= one_only
? ".gnu.linkonce.lr." : ".lrodata.";
2507 case SECCAT_SRODATA
:
2514 /* We don't split these for medium model. Place them into
2515 default sections and hope for best. */
2523 plen
= strlen (prefix
);
2525 name
= IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl
));
2526 name
= targetm
.strip_name_encoding (name
);
2527 nlen
= strlen (name
);
2529 string
= alloca (nlen
+ plen
+ 1);
2530 memcpy (string
, prefix
, plen
);
2531 memcpy (string
+ plen
, name
, nlen
+ 1);
2533 DECL_SECTION_NAME (decl
) = build_string (nlen
+ plen
, string
);
2537 default_unique_section (decl
, reloc
);
2540 #ifdef COMMON_ASM_OP
2541 /* This says how to output assembler code to declare an
2542 uninitialized external linkage data object.
2544 For medium model x86-64 we need to use .largecomm opcode for
2547 x86_elf_aligned_common (FILE *file
,
2548 const char *name
, unsigned HOST_WIDE_INT size
,
2551 if ((ix86_cmodel
== CM_MEDIUM
|| ix86_cmodel
== CM_MEDIUM_PIC
)
2552 && size
> (unsigned int)ix86_section_threshold
)
2553 fprintf (file
, ".largecomm\t");
2555 fprintf (file
, "%s", COMMON_ASM_OP
);
2556 assemble_name (file
, name
);
2557 fprintf (file
, ","HOST_WIDE_INT_PRINT_UNSIGNED
",%u\n",
2558 size
, align
/ BITS_PER_UNIT
);
2562 /* Utility function for targets to use in implementing
2563 ASM_OUTPUT_ALIGNED_BSS. */
2566 x86_output_aligned_bss (FILE *file
, tree decl ATTRIBUTE_UNUSED
,
2567 const char *name
, unsigned HOST_WIDE_INT size
,
2570 if ((ix86_cmodel
== CM_MEDIUM
|| ix86_cmodel
== CM_MEDIUM_PIC
)
2571 && size
> (unsigned int)ix86_section_threshold
)
2572 switch_to_section (get_named_section (decl
, ".lbss", 0));
2574 switch_to_section (bss_section
);
2575 ASM_OUTPUT_ALIGN (file
, floor_log2 (align
/ BITS_PER_UNIT
));
2576 #ifdef ASM_DECLARE_OBJECT_NAME
2577 last_assemble_variable_decl
= decl
;
2578 ASM_DECLARE_OBJECT_NAME (file
, name
, decl
);
2580 /* Standard thing is just output label for the object. */
2581 ASM_OUTPUT_LABEL (file
, name
);
2582 #endif /* ASM_DECLARE_OBJECT_NAME */
2583 ASM_OUTPUT_SKIP (file
, size
? size
: 1);
2587 optimization_options (int level
, int size ATTRIBUTE_UNUSED
)
2589 /* For -O2 and beyond, turn off -fschedule-insns by default. It tends to
2590 make the problem with not enough registers even worse. */
2591 #ifdef INSN_SCHEDULING
2593 flag_schedule_insns
= 0;
2597 /* The Darwin libraries never set errno, so we might as well
2598 avoid calling them when that's the only reason we would. */
2599 flag_errno_math
= 0;
2601 /* The default values of these switches depend on the TARGET_64BIT
2602 that is not known at this moment. Mark these values with 2 and
2603 let user the to override these. In case there is no command line option
2604 specifying them, we will set the defaults in override_options. */
2606 flag_omit_frame_pointer
= 2;
2607 flag_pcc_struct_return
= 2;
2608 flag_asynchronous_unwind_tables
= 2;
2609 #ifdef SUBTARGET_OPTIMIZATION_OPTIONS
2610 SUBTARGET_OPTIMIZATION_OPTIONS
;
2614 /* Decide whether we can make a sibling call to a function. DECL is the
2615 declaration of the function being targeted by the call and EXP is the
2616 CALL_EXPR representing the call. */
2619 ix86_function_ok_for_sibcall (tree decl
, tree exp
)
2624 /* If we are generating position-independent code, we cannot sibcall
2625 optimize any indirect call, or a direct call to a global function,
2626 as the PLT requires %ebx be live. */
2627 if (!TARGET_64BIT
&& flag_pic
&& (!decl
|| !targetm
.binds_local_p (decl
)))
2634 func
= TREE_TYPE (CALL_EXPR_FN (exp
));
2635 if (POINTER_TYPE_P (func
))
2636 func
= TREE_TYPE (func
);
2639 /* Check that the return value locations are the same. Like
2640 if we are returning floats on the 80387 register stack, we cannot
2641 make a sibcall from a function that doesn't return a float to a
2642 function that does or, conversely, from a function that does return
2643 a float to a function that doesn't; the necessary stack adjustment
2644 would not be executed. This is also the place we notice
2645 differences in the return value ABI. Note that it is ok for one
2646 of the functions to have void return type as long as the return
2647 value of the other is passed in a register. */
2648 a
= ix86_function_value (TREE_TYPE (exp
), func
, false);
2649 b
= ix86_function_value (TREE_TYPE (DECL_RESULT (cfun
->decl
)),
2651 if (STACK_REG_P (a
) || STACK_REG_P (b
))
2653 if (!rtx_equal_p (a
, b
))
2656 else if (VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun
->decl
))))
2658 else if (!rtx_equal_p (a
, b
))
2661 /* If this call is indirect, we'll need to be able to use a call-clobbered
2662 register for the address of the target function. Make sure that all
2663 such registers are not used for passing parameters. */
2664 if (!decl
&& !TARGET_64BIT
)
2668 /* We're looking at the CALL_EXPR, we need the type of the function. */
2669 type
= CALL_EXPR_FN (exp
); /* pointer expression */
2670 type
= TREE_TYPE (type
); /* pointer type */
2671 type
= TREE_TYPE (type
); /* function type */
2673 if (ix86_function_regparm (type
, NULL
) >= 3)
2675 /* ??? Need to count the actual number of registers to be used,
2676 not the possible number of registers. Fix later. */
2681 /* Dllimport'd functions are also called indirectly. */
2682 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
2683 && decl
&& DECL_DLLIMPORT_P (decl
)
2684 && ix86_function_regparm (TREE_TYPE (decl
), NULL
) >= 3)
2687 /* If we forced aligned the stack, then sibcalling would unalign the
2688 stack, which may break the called function. */
2689 if (cfun
->machine
->force_align_arg_pointer
)
2692 /* Otherwise okay. That also includes certain types of indirect calls. */
2696 /* Handle "cdecl", "stdcall", "fastcall", "regparm" and "sseregparm"
2697 calling convention attributes;
2698 arguments as in struct attribute_spec.handler. */
2701 ix86_handle_cconv_attribute (tree
*node
, tree name
,
2703 int flags ATTRIBUTE_UNUSED
,
2706 if (TREE_CODE (*node
) != FUNCTION_TYPE
2707 && TREE_CODE (*node
) != METHOD_TYPE
2708 && TREE_CODE (*node
) != FIELD_DECL
2709 && TREE_CODE (*node
) != TYPE_DECL
)
2711 warning (OPT_Wattributes
, "%qs attribute only applies to functions",
2712 IDENTIFIER_POINTER (name
));
2713 *no_add_attrs
= true;
2717 /* Can combine regparm with all attributes but fastcall. */
2718 if (is_attribute_p ("regparm", name
))
2722 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node
)))
2724 error ("fastcall and regparm attributes are not compatible");
2727 cst
= TREE_VALUE (args
);
2728 if (TREE_CODE (cst
) != INTEGER_CST
)
2730 warning (OPT_Wattributes
,
2731 "%qs attribute requires an integer constant argument",
2732 IDENTIFIER_POINTER (name
));
2733 *no_add_attrs
= true;
2735 else if (compare_tree_int (cst
, REGPARM_MAX
) > 0)
2737 warning (OPT_Wattributes
, "argument to %qs attribute larger than %d",
2738 IDENTIFIER_POINTER (name
), REGPARM_MAX
);
2739 *no_add_attrs
= true;
2743 && lookup_attribute (ix86_force_align_arg_pointer_string
,
2744 TYPE_ATTRIBUTES (*node
))
2745 && compare_tree_int (cst
, REGPARM_MAX
-1))
2747 error ("%s functions limited to %d register parameters",
2748 ix86_force_align_arg_pointer_string
, REGPARM_MAX
-1);
2756 /* Do not warn when emulating the MS ABI. */
2757 if (!TARGET_64BIT_MS_ABI
)
2758 warning (OPT_Wattributes
, "%qs attribute ignored",
2759 IDENTIFIER_POINTER (name
));
2760 *no_add_attrs
= true;
2764 /* Can combine fastcall with stdcall (redundant) and sseregparm. */
2765 if (is_attribute_p ("fastcall", name
))
2767 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node
)))
2769 error ("fastcall and cdecl attributes are not compatible");
2771 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node
)))
2773 error ("fastcall and stdcall attributes are not compatible");
2775 if (lookup_attribute ("regparm", TYPE_ATTRIBUTES (*node
)))
2777 error ("fastcall and regparm attributes are not compatible");
2781 /* Can combine stdcall with fastcall (redundant), regparm and
2783 else if (is_attribute_p ("stdcall", name
))
2785 if (lookup_attribute ("cdecl", TYPE_ATTRIBUTES (*node
)))
2787 error ("stdcall and cdecl attributes are not compatible");
2789 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node
)))
2791 error ("stdcall and fastcall attributes are not compatible");
2795 /* Can combine cdecl with regparm and sseregparm. */
2796 else if (is_attribute_p ("cdecl", name
))
2798 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (*node
)))
2800 error ("stdcall and cdecl attributes are not compatible");
2802 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (*node
)))
2804 error ("fastcall and cdecl attributes are not compatible");
2808 /* Can combine sseregparm with all attributes. */
2813 /* Return 0 if the attributes for two types are incompatible, 1 if they
2814 are compatible, and 2 if they are nearly compatible (which causes a
2815 warning to be generated). */
2818 ix86_comp_type_attributes (tree type1
, tree type2
)
2820 /* Check for mismatch of non-default calling convention. */
2821 const char *const rtdstr
= TARGET_RTD
? "cdecl" : "stdcall";
2823 if (TREE_CODE (type1
) != FUNCTION_TYPE
)
2826 /* Check for mismatched fastcall/regparm types. */
2827 if ((!lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type1
))
2828 != !lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type2
)))
2829 || (ix86_function_regparm (type1
, NULL
)
2830 != ix86_function_regparm (type2
, NULL
)))
2833 /* Check for mismatched sseregparm types. */
2834 if (!lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type1
))
2835 != !lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type2
)))
2838 /* Check for mismatched return types (cdecl vs stdcall). */
2839 if (!lookup_attribute (rtdstr
, TYPE_ATTRIBUTES (type1
))
2840 != !lookup_attribute (rtdstr
, TYPE_ATTRIBUTES (type2
)))
2846 /* Return the regparm value for a function with the indicated TYPE and DECL.
2847 DECL may be NULL when calling function indirectly
2848 or considering a libcall. */
2851 ix86_function_regparm (tree type
, tree decl
)
2854 int regparm
= ix86_regparm
;
2859 attr
= lookup_attribute ("regparm", TYPE_ATTRIBUTES (type
));
2861 return TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attr
)));
2863 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type
)))
2866 /* Use register calling convention for local functions when possible. */
2867 if (decl
&& TREE_CODE (decl
) == FUNCTION_DECL
2868 && flag_unit_at_a_time
&& !profile_flag
)
2870 struct cgraph_local_info
*i
= cgraph_local_info (decl
);
2873 int local_regparm
, globals
= 0, regno
;
2876 /* Make sure no regparm register is taken by a
2877 global register variable. */
2878 for (local_regparm
= 0; local_regparm
< 3; local_regparm
++)
2879 if (global_regs
[local_regparm
])
2882 /* We can't use regparm(3) for nested functions as these use
2883 static chain pointer in third argument. */
2884 if (local_regparm
== 3
2885 && (decl_function_context (decl
)
2886 || ix86_force_align_arg_pointer
)
2887 && !DECL_NO_STATIC_CHAIN (decl
))
2890 /* If the function realigns its stackpointer, the prologue will
2891 clobber %ecx. If we've already generated code for the callee,
2892 the callee DECL_STRUCT_FUNCTION is gone, so we fall back to
2893 scanning the attributes for the self-realigning property. */
2894 f
= DECL_STRUCT_FUNCTION (decl
);
2895 if (local_regparm
== 3
2896 && (f
? !!f
->machine
->force_align_arg_pointer
2897 : !!lookup_attribute (ix86_force_align_arg_pointer_string
,
2898 TYPE_ATTRIBUTES (TREE_TYPE (decl
)))))
2901 /* Each global register variable increases register preassure,
2902 so the more global reg vars there are, the smaller regparm
2903 optimization use, unless requested by the user explicitly. */
2904 for (regno
= 0; regno
< 6; regno
++)
2905 if (global_regs
[regno
])
2908 = globals
< local_regparm
? local_regparm
- globals
: 0;
2910 if (local_regparm
> regparm
)
2911 regparm
= local_regparm
;
2918 /* Return 1 or 2, if we can pass up to SSE_REGPARM_MAX SFmode (1) and
2919 DFmode (2) arguments in SSE registers for a function with the
2920 indicated TYPE and DECL. DECL may be NULL when calling function
2921 indirectly or considering a libcall. Otherwise return 0. */
2924 ix86_function_sseregparm (tree type
, tree decl
)
2926 gcc_assert (!TARGET_64BIT
);
2928 /* Use SSE registers to pass SFmode and DFmode arguments if requested
2929 by the sseregparm attribute. */
2930 if (TARGET_SSEREGPARM
2931 || (type
&& lookup_attribute ("sseregparm", TYPE_ATTRIBUTES (type
))))
2936 error ("Calling %qD with attribute sseregparm without "
2937 "SSE/SSE2 enabled", decl
);
2939 error ("Calling %qT with attribute sseregparm without "
2940 "SSE/SSE2 enabled", type
);
2947 /* For local functions, pass up to SSE_REGPARM_MAX SFmode
2948 (and DFmode for SSE2) arguments in SSE registers. */
2949 if (decl
&& TARGET_SSE_MATH
&& flag_unit_at_a_time
&& !profile_flag
)
2951 struct cgraph_local_info
*i
= cgraph_local_info (decl
);
2953 return TARGET_SSE2
? 2 : 1;
2959 /* Return true if EAX is live at the start of the function. Used by
2960 ix86_expand_prologue to determine if we need special help before
2961 calling allocate_stack_worker. */
2964 ix86_eax_live_at_start_p (void)
2966 /* Cheat. Don't bother working forward from ix86_function_regparm
2967 to the function type to whether an actual argument is located in
2968 eax. Instead just look at cfg info, which is still close enough
2969 to correct at this point. This gives false positives for broken
2970 functions that might use uninitialized data that happens to be
2971 allocated in eax, but who cares? */
2972 return REGNO_REG_SET_P (ENTRY_BLOCK_PTR
->il
.rtl
->global_live_at_end
, 0);
2975 /* Return true if TYPE has a variable argument list. */
2978 type_has_variadic_args_p (tree type
)
2980 tree n
, t
= TYPE_ARG_TYPES (type
);
2985 while ((n
= TREE_CHAIN (t
)) != NULL
)
2988 return TREE_VALUE (t
) != void_type_node
;
2991 /* Value is the number of bytes of arguments automatically
2992 popped when returning from a subroutine call.
2993 FUNDECL is the declaration node of the function (as a tree),
2994 FUNTYPE is the data type of the function (as a tree),
2995 or for a library call it is an identifier node for the subroutine name.
2996 SIZE is the number of bytes of arguments passed on the stack.
2998 On the 80386, the RTD insn may be used to pop them if the number
2999 of args is fixed, but if the number is variable then the caller
3000 must pop them all. RTD can't be used for library calls now
3001 because the library is compiled with the Unix compiler.
3002 Use of RTD is a selectable option, since it is incompatible with
3003 standard Unix calling sequences. If the option is not selected,
3004 the caller must always pop the args.
3006 The attribute stdcall is equivalent to RTD on a per module basis. */
3009 ix86_return_pops_args (tree fundecl
, tree funtype
, int size
)
3013 /* None of the 64-bit ABIs pop arguments. */
3017 rtd
= TARGET_RTD
&& (!fundecl
|| TREE_CODE (fundecl
) != IDENTIFIER_NODE
);
3019 /* Cdecl functions override -mrtd, and never pop the stack. */
3020 if (! lookup_attribute ("cdecl", TYPE_ATTRIBUTES (funtype
)))
3022 /* Stdcall and fastcall functions will pop the stack if not
3024 if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (funtype
))
3025 || lookup_attribute ("fastcall", TYPE_ATTRIBUTES (funtype
)))
3028 if (rtd
&& ! type_has_variadic_args_p (funtype
))
3032 /* Lose any fake structure return argument if it is passed on the stack. */
3033 if (aggregate_value_p (TREE_TYPE (funtype
), fundecl
)
3034 && !KEEP_AGGREGATE_RETURN_POINTER
)
3036 int nregs
= ix86_function_regparm (funtype
, fundecl
);
3038 return GET_MODE_SIZE (Pmode
);
3044 /* Argument support functions. */
3046 /* Return true when register may be used to pass function parameters. */
3048 ix86_function_arg_regno_p (int regno
)
3051 const int *parm_regs
;
3056 return (regno
< REGPARM_MAX
3057 || (TARGET_SSE
&& SSE_REGNO_P (regno
) && !fixed_regs
[regno
]));
3059 return (regno
< REGPARM_MAX
3060 || (TARGET_MMX
&& MMX_REGNO_P (regno
)
3061 && (regno
< FIRST_MMX_REG
+ MMX_REGPARM_MAX
))
3062 || (TARGET_SSE
&& SSE_REGNO_P (regno
)
3063 && (regno
< FIRST_SSE_REG
+ SSE_REGPARM_MAX
)));
3068 if (SSE_REGNO_P (regno
) && TARGET_SSE
)
3073 if (TARGET_SSE
&& SSE_REGNO_P (regno
)
3074 && (regno
< FIRST_SSE_REG
+ SSE_REGPARM_MAX
))
3078 /* RAX is used as hidden argument to va_arg functions. */
3079 if (!TARGET_64BIT_MS_ABI
&& regno
== 0)
3082 if (TARGET_64BIT_MS_ABI
)
3083 parm_regs
= x86_64_ms_abi_int_parameter_registers
;
3085 parm_regs
= x86_64_int_parameter_registers
;
3086 for (i
= 0; i
< REGPARM_MAX
; i
++)
3087 if (regno
== parm_regs
[i
])
3092 /* Return if we do not know how to pass TYPE solely in registers. */
3095 ix86_must_pass_in_stack (enum machine_mode mode
, tree type
)
3097 if (must_pass_in_stack_var_size_or_pad (mode
, type
))
3100 /* For 32-bit, we want TImode aggregates to go on the stack. But watch out!
3101 The layout_type routine is crafty and tries to trick us into passing
3102 currently unsupported vector types on the stack by using TImode. */
3103 return (!TARGET_64BIT
&& mode
== TImode
3104 && type
&& TREE_CODE (type
) != VECTOR_TYPE
);
3107 /* Initialize a variable CUM of type CUMULATIVE_ARGS
3108 for a call to a function whose data type is FNTYPE.
3109 For a library call, FNTYPE is 0. */
3112 init_cumulative_args (CUMULATIVE_ARGS
*cum
, /* Argument info to initialize */
3113 tree fntype
, /* tree ptr for function decl */
3114 rtx libname
, /* SYMBOL_REF of library name or 0 */
3117 memset (cum
, 0, sizeof (*cum
));
3119 /* Set up the number of registers to use for passing arguments. */
3120 cum
->nregs
= ix86_regparm
;
3122 cum
->sse_nregs
= SSE_REGPARM_MAX
;
3124 cum
->mmx_nregs
= MMX_REGPARM_MAX
;
3125 cum
->warn_sse
= true;
3126 cum
->warn_mmx
= true;
3127 cum
->maybe_vaarg
= (fntype
3128 ? (!TYPE_ARG_TYPES (fntype
)
3129 || type_has_variadic_args_p (fntype
))
3134 /* If there are variable arguments, then we won't pass anything
3135 in registers in 32-bit mode. */
3136 if (cum
->maybe_vaarg
)
3146 /* Use ecx and edx registers if function has fastcall attribute,
3147 else look for regparm information. */
3150 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (fntype
)))
3156 cum
->nregs
= ix86_function_regparm (fntype
, fndecl
);
3159 /* Set up the number of SSE registers used for passing SFmode
3160 and DFmode arguments. Warn for mismatching ABI. */
3161 cum
->float_in_sse
= ix86_function_sseregparm (fntype
, fndecl
);
3165 /* Return the "natural" mode for TYPE. In most cases, this is just TYPE_MODE.
3166 But in the case of vector types, it is some vector mode.
3168 When we have only some of our vector isa extensions enabled, then there
3169 are some modes for which vector_mode_supported_p is false. For these
3170 modes, the generic vector support in gcc will choose some non-vector mode
3171 in order to implement the type. By computing the natural mode, we'll
3172 select the proper ABI location for the operand and not depend on whatever
3173 the middle-end decides to do with these vector types. */
3175 static enum machine_mode
3176 type_natural_mode (tree type
)
3178 enum machine_mode mode
= TYPE_MODE (type
);
3180 if (TREE_CODE (type
) == VECTOR_TYPE
&& !VECTOR_MODE_P (mode
))
3182 HOST_WIDE_INT size
= int_size_in_bytes (type
);
3183 if ((size
== 8 || size
== 16)
3184 /* ??? Generic code allows us to create width 1 vectors. Ignore. */
3185 && TYPE_VECTOR_SUBPARTS (type
) > 1)
3187 enum machine_mode innermode
= TYPE_MODE (TREE_TYPE (type
));
3189 if (TREE_CODE (TREE_TYPE (type
)) == REAL_TYPE
)
3190 mode
= MIN_MODE_VECTOR_FLOAT
;
3192 mode
= MIN_MODE_VECTOR_INT
;
3194 /* Get the mode which has this inner mode and number of units. */
3195 for (; mode
!= VOIDmode
; mode
= GET_MODE_WIDER_MODE (mode
))
3196 if (GET_MODE_NUNITS (mode
) == TYPE_VECTOR_SUBPARTS (type
)
3197 && GET_MODE_INNER (mode
) == innermode
)
3207 /* We want to pass a value in REGNO whose "natural" mode is MODE. However,
3208 this may not agree with the mode that the type system has chosen for the
3209 register, which is ORIG_MODE. If ORIG_MODE is not BLKmode, then we can
3210 go ahead and use it. Otherwise we have to build a PARALLEL instead. */
3213 gen_reg_or_parallel (enum machine_mode mode
, enum machine_mode orig_mode
,
3218 if (orig_mode
!= BLKmode
)
3219 tmp
= gen_rtx_REG (orig_mode
, regno
);
3222 tmp
= gen_rtx_REG (mode
, regno
);
3223 tmp
= gen_rtx_EXPR_LIST (VOIDmode
, tmp
, const0_rtx
);
3224 tmp
= gen_rtx_PARALLEL (orig_mode
, gen_rtvec (1, tmp
));
3230 /* x86-64 register passing implementation. See x86-64 ABI for details. Goal
3231 of this code is to classify each 8bytes of incoming argument by the register
3232 class and assign registers accordingly. */
3234 /* Return the union class of CLASS1 and CLASS2.
3235 See the x86-64 PS ABI for details. */
3237 static enum x86_64_reg_class
3238 merge_classes (enum x86_64_reg_class class1
, enum x86_64_reg_class class2
)
3240 /* Rule #1: If both classes are equal, this is the resulting class. */
3241 if (class1
== class2
)
3244 /* Rule #2: If one of the classes is NO_CLASS, the resulting class is
3246 if (class1
== X86_64_NO_CLASS
)
3248 if (class2
== X86_64_NO_CLASS
)
3251 /* Rule #3: If one of the classes is MEMORY, the result is MEMORY. */
3252 if (class1
== X86_64_MEMORY_CLASS
|| class2
== X86_64_MEMORY_CLASS
)
3253 return X86_64_MEMORY_CLASS
;
3255 /* Rule #4: If one of the classes is INTEGER, the result is INTEGER. */
3256 if ((class1
== X86_64_INTEGERSI_CLASS
&& class2
== X86_64_SSESF_CLASS
)
3257 || (class2
== X86_64_INTEGERSI_CLASS
&& class1
== X86_64_SSESF_CLASS
))
3258 return X86_64_INTEGERSI_CLASS
;
3259 if (class1
== X86_64_INTEGER_CLASS
|| class1
== X86_64_INTEGERSI_CLASS
3260 || class2
== X86_64_INTEGER_CLASS
|| class2
== X86_64_INTEGERSI_CLASS
)
3261 return X86_64_INTEGER_CLASS
;
3263 /* Rule #5: If one of the classes is X87, X87UP, or COMPLEX_X87 class,
3265 if (class1
== X86_64_X87_CLASS
3266 || class1
== X86_64_X87UP_CLASS
3267 || class1
== X86_64_COMPLEX_X87_CLASS
3268 || class2
== X86_64_X87_CLASS
3269 || class2
== X86_64_X87UP_CLASS
3270 || class2
== X86_64_COMPLEX_X87_CLASS
)
3271 return X86_64_MEMORY_CLASS
;
3273 /* Rule #6: Otherwise class SSE is used. */
3274 return X86_64_SSE_CLASS
;
3277 /* Classify the argument of type TYPE and mode MODE.
3278 CLASSES will be filled by the register class used to pass each word
3279 of the operand. The number of words is returned. In case the parameter
3280 should be passed in memory, 0 is returned. As a special case for zero
3281 sized containers, classes[0] will be NO_CLASS and 1 is returned.
3283 BIT_OFFSET is used internally for handling records and specifies offset
3284 of the offset in bits modulo 256 to avoid overflow cases.
3286 See the x86-64 PS ABI for details.
3290 classify_argument (enum machine_mode mode
, tree type
,
3291 enum x86_64_reg_class classes
[MAX_CLASSES
], int bit_offset
)
3293 HOST_WIDE_INT bytes
=
3294 (mode
== BLKmode
) ? int_size_in_bytes (type
) : (int) GET_MODE_SIZE (mode
);
3295 int words
= (bytes
+ (bit_offset
% 64) / 8 + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
3297 /* Variable sized entities are always passed/returned in memory. */
3301 if (mode
!= VOIDmode
3302 && targetm
.calls
.must_pass_in_stack (mode
, type
))
3305 if (type
&& AGGREGATE_TYPE_P (type
))
3309 enum x86_64_reg_class subclasses
[MAX_CLASSES
];
3311 /* On x86-64 we pass structures larger than 16 bytes on the stack. */
3315 for (i
= 0; i
< words
; i
++)
3316 classes
[i
] = X86_64_NO_CLASS
;
3318 /* Zero sized arrays or structures are NO_CLASS. We return 0 to
3319 signalize memory class, so handle it as special case. */
3322 classes
[0] = X86_64_NO_CLASS
;
3326 /* Classify each field of record and merge classes. */
3327 switch (TREE_CODE (type
))
3330 /* And now merge the fields of structure. */
3331 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
3333 if (TREE_CODE (field
) == FIELD_DECL
)
3337 if (TREE_TYPE (field
) == error_mark_node
)
3340 /* Bitfields are always classified as integer. Handle them
3341 early, since later code would consider them to be
3342 misaligned integers. */
3343 if (DECL_BIT_FIELD (field
))
3345 for (i
= (int_bit_position (field
) + (bit_offset
% 64)) / 8 / 8;
3346 i
< ((int_bit_position (field
) + (bit_offset
% 64))
3347 + tree_low_cst (DECL_SIZE (field
), 0)
3350 merge_classes (X86_64_INTEGER_CLASS
,
3355 num
= classify_argument (TYPE_MODE (TREE_TYPE (field
)),
3356 TREE_TYPE (field
), subclasses
,
3357 (int_bit_position (field
)
3358 + bit_offset
) % 256);
3361 for (i
= 0; i
< num
; i
++)
3364 (int_bit_position (field
) + (bit_offset
% 64)) / 8 / 8;
3366 merge_classes (subclasses
[i
], classes
[i
+ pos
]);
3374 /* Arrays are handled as small records. */
3377 num
= classify_argument (TYPE_MODE (TREE_TYPE (type
)),
3378 TREE_TYPE (type
), subclasses
, bit_offset
);
3382 /* The partial classes are now full classes. */
3383 if (subclasses
[0] == X86_64_SSESF_CLASS
&& bytes
!= 4)
3384 subclasses
[0] = X86_64_SSE_CLASS
;
3385 if (subclasses
[0] == X86_64_INTEGERSI_CLASS
&& bytes
!= 4)
3386 subclasses
[0] = X86_64_INTEGER_CLASS
;
3388 for (i
= 0; i
< words
; i
++)
3389 classes
[i
] = subclasses
[i
% num
];
3394 case QUAL_UNION_TYPE
:
3395 /* Unions are similar to RECORD_TYPE but offset is always 0.
3397 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
3399 if (TREE_CODE (field
) == FIELD_DECL
)
3403 if (TREE_TYPE (field
) == error_mark_node
)
3406 num
= classify_argument (TYPE_MODE (TREE_TYPE (field
)),
3407 TREE_TYPE (field
), subclasses
,
3411 for (i
= 0; i
< num
; i
++)
3412 classes
[i
] = merge_classes (subclasses
[i
], classes
[i
]);
3421 /* Final merger cleanup. */
3422 for (i
= 0; i
< words
; i
++)
3424 /* If one class is MEMORY, everything should be passed in
3426 if (classes
[i
] == X86_64_MEMORY_CLASS
)
3429 /* The X86_64_SSEUP_CLASS should be always preceded by
3430 X86_64_SSE_CLASS. */
3431 if (classes
[i
] == X86_64_SSEUP_CLASS
3432 && (i
== 0 || classes
[i
- 1] != X86_64_SSE_CLASS
))
3433 classes
[i
] = X86_64_SSE_CLASS
;
3435 /* X86_64_X87UP_CLASS should be preceded by X86_64_X87_CLASS. */
3436 if (classes
[i
] == X86_64_X87UP_CLASS
3437 && (i
== 0 || classes
[i
- 1] != X86_64_X87_CLASS
))
3438 classes
[i
] = X86_64_SSE_CLASS
;
3443 /* Compute alignment needed. We align all types to natural boundaries with
3444 exception of XFmode that is aligned to 64bits. */
3445 if (mode
!= VOIDmode
&& mode
!= BLKmode
)
3447 int mode_alignment
= GET_MODE_BITSIZE (mode
);
3450 mode_alignment
= 128;
3451 else if (mode
== XCmode
)
3452 mode_alignment
= 256;
3453 if (COMPLEX_MODE_P (mode
))
3454 mode_alignment
/= 2;
3455 /* Misaligned fields are always returned in memory. */
3456 if (bit_offset
% mode_alignment
)
3460 /* for V1xx modes, just use the base mode */
3461 if (VECTOR_MODE_P (mode
)
3462 && GET_MODE_SIZE (GET_MODE_INNER (mode
)) == bytes
)
3463 mode
= GET_MODE_INNER (mode
);
3465 /* Classification of atomic types. */
3470 classes
[0] = X86_64_SSE_CLASS
;
3473 classes
[0] = X86_64_SSE_CLASS
;
3474 classes
[1] = X86_64_SSEUP_CLASS
;
3483 if (bit_offset
+ GET_MODE_BITSIZE (mode
) <= 32)
3484 classes
[0] = X86_64_INTEGERSI_CLASS
;
3486 classes
[0] = X86_64_INTEGER_CLASS
;
3490 classes
[0] = classes
[1] = X86_64_INTEGER_CLASS
;
3495 if (!(bit_offset
% 64))
3496 classes
[0] = X86_64_SSESF_CLASS
;
3498 classes
[0] = X86_64_SSE_CLASS
;
3501 classes
[0] = X86_64_SSEDF_CLASS
;
3504 classes
[0] = X86_64_X87_CLASS
;
3505 classes
[1] = X86_64_X87UP_CLASS
;
3508 classes
[0] = X86_64_SSE_CLASS
;
3509 classes
[1] = X86_64_SSEUP_CLASS
;
3512 classes
[0] = X86_64_SSE_CLASS
;
3515 classes
[0] = X86_64_SSEDF_CLASS
;
3516 classes
[1] = X86_64_SSEDF_CLASS
;
3519 classes
[0] = X86_64_COMPLEX_X87_CLASS
;
3522 /* This modes is larger than 16 bytes. */
3530 classes
[0] = X86_64_SSE_CLASS
;
3531 classes
[1] = X86_64_SSEUP_CLASS
;
3537 classes
[0] = X86_64_SSE_CLASS
;
3543 gcc_assert (VECTOR_MODE_P (mode
));
3548 gcc_assert (GET_MODE_CLASS (GET_MODE_INNER (mode
)) == MODE_INT
);
3550 if (bit_offset
+ GET_MODE_BITSIZE (mode
) <= 32)
3551 classes
[0] = X86_64_INTEGERSI_CLASS
;
3553 classes
[0] = X86_64_INTEGER_CLASS
;
3554 classes
[1] = X86_64_INTEGER_CLASS
;
3555 return 1 + (bytes
> 8);
3559 /* Examine the argument and return set number of register required in each
3560 class. Return 0 iff parameter should be passed in memory. */
3562 examine_argument (enum machine_mode mode
, tree type
, int in_return
,
3563 int *int_nregs
, int *sse_nregs
)
3565 enum x86_64_reg_class
class[MAX_CLASSES
];
3566 int n
= classify_argument (mode
, type
, class, 0);
3572 for (n
--; n
>= 0; n
--)
3575 case X86_64_INTEGER_CLASS
:
3576 case X86_64_INTEGERSI_CLASS
:
3579 case X86_64_SSE_CLASS
:
3580 case X86_64_SSESF_CLASS
:
3581 case X86_64_SSEDF_CLASS
:
3584 case X86_64_NO_CLASS
:
3585 case X86_64_SSEUP_CLASS
:
3587 case X86_64_X87_CLASS
:
3588 case X86_64_X87UP_CLASS
:
3592 case X86_64_COMPLEX_X87_CLASS
:
3593 return in_return
? 2 : 0;
3594 case X86_64_MEMORY_CLASS
:
3600 /* Construct container for the argument used by GCC interface. See
3601 FUNCTION_ARG for the detailed description. */
3604 construct_container (enum machine_mode mode
, enum machine_mode orig_mode
,
3605 tree type
, int in_return
, int nintregs
, int nsseregs
,
3606 const int *intreg
, int sse_regno
)
3608 /* The following variables hold the static issued_error state. */
3609 static bool issued_sse_arg_error
;
3610 static bool issued_sse_ret_error
;
3611 static bool issued_x87_ret_error
;
3613 enum machine_mode tmpmode
;
3615 (mode
== BLKmode
) ? int_size_in_bytes (type
) : (int) GET_MODE_SIZE (mode
);
3616 enum x86_64_reg_class
class[MAX_CLASSES
];
3620 int needed_sseregs
, needed_intregs
;
3621 rtx exp
[MAX_CLASSES
];
3624 n
= classify_argument (mode
, type
, class, 0);
3627 if (!examine_argument (mode
, type
, in_return
, &needed_intregs
,
3630 if (needed_intregs
> nintregs
|| needed_sseregs
> nsseregs
)
3633 /* We allowed the user to turn off SSE for kernel mode. Don't crash if
3634 some less clueful developer tries to use floating-point anyway. */
3635 if (needed_sseregs
&& !TARGET_SSE
)
3639 if (!issued_sse_ret_error
)
3641 error ("SSE register return with SSE disabled");
3642 issued_sse_ret_error
= true;
3645 else if (!issued_sse_arg_error
)
3647 error ("SSE register argument with SSE disabled");
3648 issued_sse_arg_error
= true;
3653 /* Likewise, error if the ABI requires us to return values in the
3654 x87 registers and the user specified -mno-80387. */
3655 if (!TARGET_80387
&& in_return
)
3656 for (i
= 0; i
< n
; i
++)
3657 if (class[i
] == X86_64_X87_CLASS
3658 || class[i
] == X86_64_X87UP_CLASS
3659 || class[i
] == X86_64_COMPLEX_X87_CLASS
)
3661 if (!issued_x87_ret_error
)
3663 error ("x87 register return with x87 disabled");
3664 issued_x87_ret_error
= true;
3669 /* First construct simple cases. Avoid SCmode, since we want to use
3670 single register to pass this type. */
3671 if (n
== 1 && mode
!= SCmode
)
3674 case X86_64_INTEGER_CLASS
:
3675 case X86_64_INTEGERSI_CLASS
:
3676 return gen_rtx_REG (mode
, intreg
[0]);
3677 case X86_64_SSE_CLASS
:
3678 case X86_64_SSESF_CLASS
:
3679 case X86_64_SSEDF_CLASS
:
3680 return gen_reg_or_parallel (mode
, orig_mode
, SSE_REGNO (sse_regno
));
3681 case X86_64_X87_CLASS
:
3682 case X86_64_COMPLEX_X87_CLASS
:
3683 return gen_rtx_REG (mode
, FIRST_STACK_REG
);
3684 case X86_64_NO_CLASS
:
3685 /* Zero sized array, struct or class. */
3690 if (n
== 2 && class[0] == X86_64_SSE_CLASS
&& class[1] == X86_64_SSEUP_CLASS
3692 return gen_rtx_REG (mode
, SSE_REGNO (sse_regno
));
3695 && class[0] == X86_64_X87_CLASS
&& class[1] == X86_64_X87UP_CLASS
)
3696 return gen_rtx_REG (XFmode
, FIRST_STACK_REG
);
3697 if (n
== 2 && class[0] == X86_64_INTEGER_CLASS
3698 && class[1] == X86_64_INTEGER_CLASS
3699 && (mode
== CDImode
|| mode
== TImode
|| mode
== TFmode
)
3700 && intreg
[0] + 1 == intreg
[1])
3701 return gen_rtx_REG (mode
, intreg
[0]);
3703 /* Otherwise figure out the entries of the PARALLEL. */
3704 for (i
= 0; i
< n
; i
++)
3708 case X86_64_NO_CLASS
:
3710 case X86_64_INTEGER_CLASS
:
3711 case X86_64_INTEGERSI_CLASS
:
3712 /* Merge TImodes on aligned occasions here too. */
3713 if (i
* 8 + 8 > bytes
)
3714 tmpmode
= mode_for_size ((bytes
- i
* 8) * BITS_PER_UNIT
, MODE_INT
, 0);
3715 else if (class[i
] == X86_64_INTEGERSI_CLASS
)
3719 /* We've requested 24 bytes we don't have mode for. Use DImode. */
3720 if (tmpmode
== BLKmode
)
3722 exp
[nexps
++] = gen_rtx_EXPR_LIST (VOIDmode
,
3723 gen_rtx_REG (tmpmode
, *intreg
),
3727 case X86_64_SSESF_CLASS
:
3728 exp
[nexps
++] = gen_rtx_EXPR_LIST (VOIDmode
,
3729 gen_rtx_REG (SFmode
,
3730 SSE_REGNO (sse_regno
)),
3734 case X86_64_SSEDF_CLASS
:
3735 exp
[nexps
++] = gen_rtx_EXPR_LIST (VOIDmode
,
3736 gen_rtx_REG (DFmode
,
3737 SSE_REGNO (sse_regno
)),
3741 case X86_64_SSE_CLASS
:
3742 if (i
< n
- 1 && class[i
+ 1] == X86_64_SSEUP_CLASS
)
3746 exp
[nexps
++] = gen_rtx_EXPR_LIST (VOIDmode
,
3747 gen_rtx_REG (tmpmode
,
3748 SSE_REGNO (sse_regno
)),
3750 if (tmpmode
== TImode
)
3759 /* Empty aligned struct, union or class. */
3763 ret
= gen_rtx_PARALLEL (mode
, rtvec_alloc (nexps
));
3764 for (i
= 0; i
< nexps
; i
++)
3765 XVECEXP (ret
, 0, i
) = exp
[i
];
3769 /* Update the data in CUM to advance over an argument of mode MODE
3770 and data type TYPE. (TYPE is null for libcalls where that information
3771 may not be available.) */
3774 function_arg_advance_32 (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
3775 tree type
, HOST_WIDE_INT bytes
, HOST_WIDE_INT words
)
3791 cum
->words
+= words
;
3792 cum
->nregs
-= words
;
3793 cum
->regno
+= words
;
3795 if (cum
->nregs
<= 0)
3803 if (cum
->float_in_sse
< 2)
3806 if (cum
->float_in_sse
< 1)
3817 if (!type
|| !AGGREGATE_TYPE_P (type
))
3819 cum
->sse_words
+= words
;
3820 cum
->sse_nregs
-= 1;
3821 cum
->sse_regno
+= 1;
3822 if (cum
->sse_nregs
<= 0)
3834 if (!type
|| !AGGREGATE_TYPE_P (type
))
3836 cum
->mmx_words
+= words
;
3837 cum
->mmx_nregs
-= 1;
3838 cum
->mmx_regno
+= 1;
3839 if (cum
->mmx_nregs
<= 0)
3850 function_arg_advance_64 (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
3851 tree type
, HOST_WIDE_INT words
)
3853 int int_nregs
, sse_nregs
;
3855 if (!examine_argument (mode
, type
, 0, &int_nregs
, &sse_nregs
))
3856 cum
->words
+= words
;
3857 else if (sse_nregs
<= cum
->sse_nregs
&& int_nregs
<= cum
->nregs
)
3859 cum
->nregs
-= int_nregs
;
3860 cum
->sse_nregs
-= sse_nregs
;
3861 cum
->regno
+= int_nregs
;
3862 cum
->sse_regno
+= sse_nregs
;
3865 cum
->words
+= words
;
3869 function_arg_advance_ms_64 (CUMULATIVE_ARGS
*cum
, HOST_WIDE_INT bytes
,
3870 HOST_WIDE_INT words
)
3872 /* Otherwise, this should be passed indirect. */
3873 gcc_assert (bytes
== 1 || bytes
== 2 || bytes
== 4 || bytes
== 8);
3875 cum
->words
+= words
;
3884 function_arg_advance (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
3885 tree type
, int named ATTRIBUTE_UNUSED
)
3887 HOST_WIDE_INT bytes
, words
;
3889 if (mode
== BLKmode
)
3890 bytes
= int_size_in_bytes (type
);
3892 bytes
= GET_MODE_SIZE (mode
);
3893 words
= (bytes
+ UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
3896 mode
= type_natural_mode (type
);
3898 if (TARGET_64BIT_MS_ABI
)
3899 function_arg_advance_ms_64 (cum
, bytes
, words
);
3900 else if (TARGET_64BIT
)
3901 function_arg_advance_64 (cum
, mode
, type
, words
);
3903 function_arg_advance_32 (cum
, mode
, type
, bytes
, words
);
3906 /* Define where to put the arguments to a function.
3907 Value is zero to push the argument on the stack,
3908 or a hard register in which to store the argument.
3910 MODE is the argument's machine mode.
3911 TYPE is the data type of the argument (as a tree).
3912 This is null for libcalls where that information may
3914 CUM is a variable of type CUMULATIVE_ARGS which gives info about
3915 the preceding args and about the function being called.
3916 NAMED is nonzero if this argument is a named parameter
3917 (otherwise it is an extra parameter matching an ellipsis). */
3920 function_arg_32 (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
3921 enum machine_mode orig_mode
, tree type
,
3922 HOST_WIDE_INT bytes
, HOST_WIDE_INT words
)
3924 static bool warnedsse
, warnedmmx
;
3926 /* Avoid the AL settings for the Unix64 ABI. */
3927 if (mode
== VOIDmode
)
3943 if (words
<= cum
->nregs
)
3945 int regno
= cum
->regno
;
3947 /* Fastcall allocates the first two DWORD (SImode) or
3948 smaller arguments to ECX and EDX. */
3951 if (mode
== BLKmode
|| mode
== DImode
)
3954 /* ECX not EAX is the first allocated register. */
3958 return gen_rtx_REG (mode
, regno
);
3963 if (cum
->float_in_sse
< 2)
3966 if (cum
->float_in_sse
< 1)
3976 if (!type
|| !AGGREGATE_TYPE_P (type
))
3978 if (!TARGET_SSE
&& !warnedsse
&& cum
->warn_sse
)
3981 warning (0, "SSE vector argument without SSE enabled "
3985 return gen_reg_or_parallel (mode
, orig_mode
,
3986 cum
->sse_regno
+ FIRST_SSE_REG
);
3994 if (!type
|| !AGGREGATE_TYPE_P (type
))
3996 if (!TARGET_MMX
&& !warnedmmx
&& cum
->warn_mmx
)
3999 warning (0, "MMX vector argument without MMX enabled "
4003 return gen_reg_or_parallel (mode
, orig_mode
,
4004 cum
->mmx_regno
+ FIRST_MMX_REG
);
4013 function_arg_64 (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
4014 enum machine_mode orig_mode
, tree type
)
4016 /* Handle a hidden AL argument containing number of registers
4017 for varargs x86-64 functions. */
4018 if (mode
== VOIDmode
)
4019 return GEN_INT (cum
->maybe_vaarg
4020 ? (cum
->sse_nregs
< 0
4025 return construct_container (mode
, orig_mode
, type
, 0, cum
->nregs
,
4027 &x86_64_int_parameter_registers
[cum
->regno
],
4032 function_arg_ms_64 (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
4033 enum machine_mode orig_mode
, int named
)
4037 /* Avoid the AL settings for the Unix64 ABI. */
4038 if (mode
== VOIDmode
)
4041 /* If we've run out of registers, it goes on the stack. */
4042 if (cum
->nregs
== 0)
4045 regno
= x86_64_ms_abi_int_parameter_registers
[cum
->regno
];
4047 /* Only floating point modes are passed in anything but integer regs. */
4048 if (TARGET_SSE
&& (mode
== SFmode
|| mode
== DFmode
))
4051 regno
= cum
->regno
+ FIRST_SSE_REG
;
4056 /* Unnamed floating parameters are passed in both the
4057 SSE and integer registers. */
4058 t1
= gen_rtx_REG (mode
, cum
->regno
+ FIRST_SSE_REG
);
4059 t2
= gen_rtx_REG (mode
, regno
);
4060 t1
= gen_rtx_EXPR_LIST (VOIDmode
, t1
, const0_rtx
);
4061 t2
= gen_rtx_EXPR_LIST (VOIDmode
, t2
, const0_rtx
);
4062 return gen_rtx_PARALLEL (mode
, gen_rtvec (2, t1
, t2
));
4066 return gen_reg_or_parallel (mode
, orig_mode
, regno
);
4070 function_arg (CUMULATIVE_ARGS
*cum
, enum machine_mode omode
,
4071 tree type
, int named
)
4073 enum machine_mode mode
= omode
;
4074 HOST_WIDE_INT bytes
, words
;
4076 if (mode
== BLKmode
)
4077 bytes
= int_size_in_bytes (type
);
4079 bytes
= GET_MODE_SIZE (mode
);
4080 words
= (bytes
+ UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
4082 /* To simplify the code below, represent vector types with a vector mode
4083 even if MMX/SSE are not active. */
4084 if (type
&& TREE_CODE (type
) == VECTOR_TYPE
)
4085 mode
= type_natural_mode (type
);
4087 if (TARGET_64BIT_MS_ABI
)
4088 return function_arg_ms_64 (cum
, mode
, omode
, named
);
4089 else if (TARGET_64BIT
)
4090 return function_arg_64 (cum
, mode
, omode
, type
);
4092 return function_arg_32 (cum
, mode
, omode
, type
, bytes
, words
);
4095 /* A C expression that indicates when an argument must be passed by
4096 reference. If nonzero for an argument, a copy of that argument is
4097 made in memory and a pointer to the argument is passed instead of
4098 the argument itself. The pointer is passed in whatever way is
4099 appropriate for passing a pointer to that type. */
4102 ix86_pass_by_reference (CUMULATIVE_ARGS
*cum ATTRIBUTE_UNUSED
,
4103 enum machine_mode mode ATTRIBUTE_UNUSED
,
4104 tree type
, bool named ATTRIBUTE_UNUSED
)
4106 if (TARGET_64BIT_MS_ABI
)
4110 /* Arrays are passed by reference. */
4111 if (TREE_CODE (type
) == ARRAY_TYPE
)
4114 if (AGGREGATE_TYPE_P (type
))
4116 /* Structs/unions of sizes other than 8, 16, 32, or 64 bits
4117 are passed by reference. */
4118 int el2
= exact_log2 (int_size_in_bytes (type
));
4119 return !(el2
>= 0 && el2
<= 3);
4123 /* __m128 is passed by reference. */
4124 /* ??? How to handle complex? For now treat them as structs,
4125 and pass them by reference if they're too large. */
4126 if (GET_MODE_SIZE (mode
) > 8)
4129 else if (TARGET_64BIT
&& type
&& int_size_in_bytes (type
) == -1)
4135 /* Return true when TYPE should be 128bit aligned for 32bit argument passing
4136 ABI. Only called if TARGET_SSE. */
4138 contains_128bit_aligned_vector_p (tree type
)
4140 enum machine_mode mode
= TYPE_MODE (type
);
4141 if (SSE_REG_MODE_P (mode
)
4142 && (!TYPE_USER_ALIGN (type
) || TYPE_ALIGN (type
) > 128))
4144 if (TYPE_ALIGN (type
) < 128)
4147 if (AGGREGATE_TYPE_P (type
))
4149 /* Walk the aggregates recursively. */
4150 switch (TREE_CODE (type
))
4154 case QUAL_UNION_TYPE
:
4158 /* Walk all the structure fields. */
4159 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
4161 if (TREE_CODE (field
) == FIELD_DECL
4162 && contains_128bit_aligned_vector_p (TREE_TYPE (field
)))
4169 /* Just for use if some languages passes arrays by value. */
4170 if (contains_128bit_aligned_vector_p (TREE_TYPE (type
)))
4181 /* Gives the alignment boundary, in bits, of an argument with the
4182 specified mode and type. */
4185 ix86_function_arg_boundary (enum machine_mode mode
, tree type
)
4189 align
= TYPE_ALIGN (type
);
4191 align
= GET_MODE_ALIGNMENT (mode
);
4192 if (align
< PARM_BOUNDARY
)
4193 align
= PARM_BOUNDARY
;
4196 /* i386 ABI defines all arguments to be 4 byte aligned. We have to
4197 make an exception for SSE modes since these require 128bit
4200 The handling here differs from field_alignment. ICC aligns MMX
4201 arguments to 4 byte boundaries, while structure fields are aligned
4202 to 8 byte boundaries. */
4204 align
= PARM_BOUNDARY
;
4207 if (!SSE_REG_MODE_P (mode
))
4208 align
= PARM_BOUNDARY
;
4212 if (!contains_128bit_aligned_vector_p (type
))
4213 align
= PARM_BOUNDARY
;
4221 /* Return true if N is a possible register number of function value. */
4224 ix86_function_value_regno_p (int regno
)
4231 case FIRST_FLOAT_REG
:
4232 if (TARGET_64BIT_MS_ABI
)
4234 return TARGET_FLOAT_RETURNS_IN_80387
;
4240 if (TARGET_MACHO
|| TARGET_64BIT
)
4248 /* Define how to find the value returned by a function.
4249 VALTYPE is the data type of the value (as a tree).
4250 If the precise function being called is known, FUNC is its FUNCTION_DECL;
4251 otherwise, FUNC is 0. */
4254 function_value_32 (enum machine_mode orig_mode
, enum machine_mode mode
,
4255 tree fntype
, tree fn
)
4259 /* 8-byte vector modes in %mm0. See ix86_return_in_memory for where
4260 we normally prevent this case when mmx is not available. However
4261 some ABIs may require the result to be returned like DImode. */
4262 if (VECTOR_MODE_P (mode
) && GET_MODE_SIZE (mode
) == 8)
4263 regno
= TARGET_MMX
? FIRST_MMX_REG
: 0;
4265 /* 16-byte vector modes in %xmm0. See ix86_return_in_memory for where
4266 we prevent this case when sse is not available. However some ABIs
4267 may require the result to be returned like integer TImode. */
4268 else if (mode
== TImode
4269 || (VECTOR_MODE_P (mode
) && GET_MODE_SIZE (mode
) == 16))
4270 regno
= TARGET_SSE
? FIRST_SSE_REG
: 0;
4272 /* Floating point return values in %st(0) (unless -mno-fp-ret-in-387). */
4273 else if (X87_FLOAT_MODE_P (mode
) && TARGET_FLOAT_RETURNS_IN_80387
)
4274 regno
= FIRST_FLOAT_REG
;
4276 /* Most things go in %eax. */
4279 /* Override FP return register with %xmm0 for local functions when
4280 SSE math is enabled or for functions with sseregparm attribute. */
4281 if ((fn
|| fntype
) && (mode
== SFmode
|| mode
== DFmode
))
4283 int sse_level
= ix86_function_sseregparm (fntype
, fn
);
4284 if ((sse_level
>= 1 && mode
== SFmode
)
4285 || (sse_level
== 2 && mode
== DFmode
))
4286 regno
= FIRST_SSE_REG
;
4289 return gen_rtx_REG (orig_mode
, regno
);
4293 function_value_64 (enum machine_mode orig_mode
, enum machine_mode mode
,
4298 /* Handle libcalls, which don't provide a type node. */
4299 if (valtype
== NULL
)
4311 return gen_rtx_REG (mode
, FIRST_SSE_REG
);
4314 return gen_rtx_REG (mode
, FIRST_FLOAT_REG
);
4318 return gen_rtx_REG (mode
, 0);
4322 ret
= construct_container (mode
, orig_mode
, valtype
, 1,
4323 REGPARM_MAX
, SSE_REGPARM_MAX
,
4324 x86_64_int_return_registers
, 0);
4326 /* For zero sized structures, construct_container returns NULL, but we
4327 need to keep rest of compiler happy by returning meaningful value. */
4329 ret
= gen_rtx_REG (orig_mode
, 0);
4335 function_value_ms_64 (enum machine_mode orig_mode
, enum machine_mode mode
)
4337 unsigned int regno
= 0;
4341 if (mode
== SFmode
|| mode
== DFmode
)
4342 regno
= FIRST_SSE_REG
;
4343 else if (VECTOR_MODE_P (mode
) || GET_MODE_SIZE (mode
) == 16)
4344 regno
= FIRST_SSE_REG
;
4347 return gen_rtx_REG (orig_mode
, regno
);
4351 ix86_function_value_1 (tree valtype
, tree fntype_or_decl
,
4352 enum machine_mode orig_mode
, enum machine_mode mode
)
4357 if (fntype_or_decl
&& DECL_P (fntype_or_decl
))
4358 fn
= fntype_or_decl
;
4359 fntype
= fn
? TREE_TYPE (fn
) : fntype_or_decl
;
4361 if (TARGET_64BIT_MS_ABI
)
4362 return function_value_ms_64 (orig_mode
, mode
);
4363 else if (TARGET_64BIT
)
4364 return function_value_64 (orig_mode
, mode
, valtype
);
4366 return function_value_32 (orig_mode
, mode
, fntype
, fn
);
4370 ix86_function_value (tree valtype
, tree fntype_or_decl
,
4371 bool outgoing ATTRIBUTE_UNUSED
)
4373 enum machine_mode mode
, orig_mode
;
4375 orig_mode
= TYPE_MODE (valtype
);
4376 mode
= type_natural_mode (valtype
);
4377 return ix86_function_value_1 (valtype
, fntype_or_decl
, orig_mode
, mode
);
4381 ix86_libcall_value (enum machine_mode mode
)
4383 return ix86_function_value_1 (NULL
, NULL
, mode
, mode
);
4386 /* Return true iff type is returned in memory. */
4389 return_in_memory_32 (tree type
, enum machine_mode mode
)
4393 if (mode
== BLKmode
)
4396 size
= int_size_in_bytes (type
);
4398 if (MS_AGGREGATE_RETURN
&& AGGREGATE_TYPE_P (type
) && size
<= 8)
4401 if (VECTOR_MODE_P (mode
) || mode
== TImode
)
4403 /* User-created vectors small enough to fit in EAX. */
4407 /* MMX/3dNow values are returned in MM0,
4408 except when it doesn't exits. */
4410 return (TARGET_MMX
? 0 : 1);
4412 /* SSE values are returned in XMM0, except when it doesn't exist. */
4414 return (TARGET_SSE
? 0 : 1);
4429 return_in_memory_64 (tree type
, enum machine_mode mode
)
4431 int needed_intregs
, needed_sseregs
;
4432 return !examine_argument (mode
, type
, 1, &needed_intregs
, &needed_sseregs
);
4436 return_in_memory_ms_64 (tree type
, enum machine_mode mode
)
4438 HOST_WIDE_INT size
= int_size_in_bytes (type
);
4440 /* __m128 and friends are returned in xmm0. */
4441 if (size
== 16 && VECTOR_MODE_P (mode
))
4444 /* Otherwise, the size must be exactly in [1248]. */
4445 return (size
!= 1 && size
!= 2 && size
!= 4 && size
!= 8);
4449 ix86_return_in_memory (tree type
)
4451 enum machine_mode mode
= type_natural_mode (type
);
4453 if (TARGET_64BIT_MS_ABI
)
4454 return return_in_memory_ms_64 (type
, mode
);
4455 else if (TARGET_64BIT
)
4456 return return_in_memory_64 (type
, mode
);
4458 return return_in_memory_32 (type
, mode
);
4461 /* Return false iff TYPE is returned in memory. This version is used
4462 on Solaris 10. It is similar to the generic ix86_return_in_memory,
4463 but differs notably in that when MMX is available, 8-byte vectors
4464 are returned in memory, rather than in MMX registers. */
4467 ix86_sol10_return_in_memory (tree type
)
4470 enum machine_mode mode
= type_natural_mode (type
);
4473 return return_in_memory_64 (type
, mode
);
4475 if (mode
== BLKmode
)
4478 size
= int_size_in_bytes (type
);
4480 if (VECTOR_MODE_P (mode
))
4482 /* Return in memory only if MMX registers *are* available. This
4483 seems backwards, but it is consistent with the existing
4490 else if (mode
== TImode
)
4492 else if (mode
== XFmode
)
4498 /* When returning SSE vector types, we have a choice of either
4499 (1) being abi incompatible with a -march switch, or
4500 (2) generating an error.
4501 Given no good solution, I think the safest thing is one warning.
4502 The user won't be able to use -Werror, but....
4504 Choose the STRUCT_VALUE_RTX hook because that's (at present) only
4505 called in response to actually generating a caller or callee that
4506 uses such a type. As opposed to RETURN_IN_MEMORY, which is called
4507 via aggregate_value_p for general type probing from tree-ssa. */
4510 ix86_struct_value_rtx (tree type
, int incoming ATTRIBUTE_UNUSED
)
4512 static bool warnedsse
, warnedmmx
;
4514 if (!TARGET_64BIT
&& type
)
4516 /* Look at the return type of the function, not the function type. */
4517 enum machine_mode mode
= TYPE_MODE (TREE_TYPE (type
));
4519 if (!TARGET_SSE
&& !warnedsse
)
4522 || (VECTOR_MODE_P (mode
) && GET_MODE_SIZE (mode
) == 16))
4525 warning (0, "SSE vector return without SSE enabled "
4530 if (!TARGET_MMX
&& !warnedmmx
)
4532 if (VECTOR_MODE_P (mode
) && GET_MODE_SIZE (mode
) == 8)
4535 warning (0, "MMX vector return without MMX enabled "
4545 /* Create the va_list data type. */
4548 ix86_build_builtin_va_list (void)
4550 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
4552 /* For i386 we use plain pointer to argument area. */
4553 if (!TARGET_64BIT
|| TARGET_64BIT_MS_ABI
)
4554 return build_pointer_type (char_type_node
);
4556 record
= (*lang_hooks
.types
.make_type
) (RECORD_TYPE
);
4557 type_decl
= build_decl (TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
4559 f_gpr
= build_decl (FIELD_DECL
, get_identifier ("gp_offset"),
4560 unsigned_type_node
);
4561 f_fpr
= build_decl (FIELD_DECL
, get_identifier ("fp_offset"),
4562 unsigned_type_node
);
4563 f_ovf
= build_decl (FIELD_DECL
, get_identifier ("overflow_arg_area"),
4565 f_sav
= build_decl (FIELD_DECL
, get_identifier ("reg_save_area"),
4568 va_list_gpr_counter_field
= f_gpr
;
4569 va_list_fpr_counter_field
= f_fpr
;
4571 DECL_FIELD_CONTEXT (f_gpr
) = record
;
4572 DECL_FIELD_CONTEXT (f_fpr
) = record
;
4573 DECL_FIELD_CONTEXT (f_ovf
) = record
;
4574 DECL_FIELD_CONTEXT (f_sav
) = record
;
4576 TREE_CHAIN (record
) = type_decl
;
4577 TYPE_NAME (record
) = type_decl
;
4578 TYPE_FIELDS (record
) = f_gpr
;
4579 TREE_CHAIN (f_gpr
) = f_fpr
;
4580 TREE_CHAIN (f_fpr
) = f_ovf
;
4581 TREE_CHAIN (f_ovf
) = f_sav
;
4583 layout_type (record
);
4585 /* The correct type is an array type of one element. */
4586 return build_array_type (record
, build_index_type (size_zero_node
));
4589 /* Worker function for TARGET_SETUP_INCOMING_VARARGS. */
4592 setup_incoming_varargs_64 (CUMULATIVE_ARGS
*cum
)
4602 if (! cfun
->va_list_gpr_size
&& ! cfun
->va_list_fpr_size
)
4605 /* Indicate to allocate space on the stack for varargs save area. */
4606 ix86_save_varrargs_registers
= 1;
4607 cfun
->stack_alignment_needed
= 128;
4609 save_area
= frame_pointer_rtx
;
4610 set
= get_varargs_alias_set ();
4612 for (i
= cum
->regno
;
4614 && i
< cum
->regno
+ cfun
->va_list_gpr_size
/ UNITS_PER_WORD
;
4617 mem
= gen_rtx_MEM (Pmode
,
4618 plus_constant (save_area
, i
* UNITS_PER_WORD
));
4619 MEM_NOTRAP_P (mem
) = 1;
4620 set_mem_alias_set (mem
, set
);
4621 emit_move_insn (mem
, gen_rtx_REG (Pmode
,
4622 x86_64_int_parameter_registers
[i
]));
4625 if (cum
->sse_nregs
&& cfun
->va_list_fpr_size
)
4627 /* Now emit code to save SSE registers. The AX parameter contains number
4628 of SSE parameter registers used to call this function. We use
4629 sse_prologue_save insn template that produces computed jump across
4630 SSE saves. We need some preparation work to get this working. */
4632 label
= gen_label_rtx ();
4633 label_ref
= gen_rtx_LABEL_REF (Pmode
, label
);
4635 /* Compute address to jump to :
4636 label - 5*eax + nnamed_sse_arguments*5 */
4637 tmp_reg
= gen_reg_rtx (Pmode
);
4638 nsse_reg
= gen_reg_rtx (Pmode
);
4639 emit_insn (gen_zero_extendqidi2 (nsse_reg
, gen_rtx_REG (QImode
, 0)));
4640 emit_insn (gen_rtx_SET (VOIDmode
, tmp_reg
,
4641 gen_rtx_MULT (Pmode
, nsse_reg
,
4646 gen_rtx_CONST (DImode
,
4647 gen_rtx_PLUS (DImode
,
4649 GEN_INT (cum
->sse_regno
* 4))));
4651 emit_move_insn (nsse_reg
, label_ref
);
4652 emit_insn (gen_subdi3 (nsse_reg
, nsse_reg
, tmp_reg
));
4654 /* Compute address of memory block we save into. We always use pointer
4655 pointing 127 bytes after first byte to store - this is needed to keep
4656 instruction size limited by 4 bytes. */
4657 tmp_reg
= gen_reg_rtx (Pmode
);
4658 emit_insn (gen_rtx_SET (VOIDmode
, tmp_reg
,
4659 plus_constant (save_area
,
4660 8 * REGPARM_MAX
+ 127)));
4661 mem
= gen_rtx_MEM (BLKmode
, plus_constant (tmp_reg
, -127));
4662 MEM_NOTRAP_P (mem
) = 1;
4663 set_mem_alias_set (mem
, set
);
4664 set_mem_align (mem
, BITS_PER_WORD
);
4666 /* And finally do the dirty job! */
4667 emit_insn (gen_sse_prologue_save (mem
, nsse_reg
,
4668 GEN_INT (cum
->sse_regno
), label
));
4673 setup_incoming_varargs_ms_64 (CUMULATIVE_ARGS
*cum
)
4675 int set
= get_varargs_alias_set ();
4678 for (i
= cum
->regno
; i
< REGPARM_MAX
; i
++)
4682 mem
= gen_rtx_MEM (Pmode
,
4683 plus_constant (virtual_incoming_args_rtx
,
4684 i
* UNITS_PER_WORD
));
4685 MEM_NOTRAP_P (mem
) = 1;
4686 set_mem_alias_set (mem
, set
);
4688 reg
= gen_rtx_REG (Pmode
, x86_64_ms_abi_int_parameter_registers
[i
]);
4689 emit_move_insn (mem
, reg
);
4694 ix86_setup_incoming_varargs (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
4695 tree type
, int *pretend_size ATTRIBUTE_UNUSED
,
4698 CUMULATIVE_ARGS next_cum
;
4702 /* This argument doesn't appear to be used anymore. Which is good,
4703 because the old code here didn't suppress rtl generation. */
4704 gcc_assert (!no_rtl
);
4709 fntype
= TREE_TYPE (current_function_decl
);
4710 stdarg_p
= (TYPE_ARG_TYPES (fntype
) != 0
4711 && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (fntype
)))
4712 != void_type_node
));
4714 /* For varargs, we do not want to skip the dummy va_dcl argument.
4715 For stdargs, we do want to skip the last named argument. */
4718 function_arg_advance (&next_cum
, mode
, type
, 1);
4720 if (TARGET_64BIT_MS_ABI
)
4721 setup_incoming_varargs_ms_64 (&next_cum
);
4723 setup_incoming_varargs_64 (&next_cum
);
4726 /* Implement va_start. */
4729 ix86_va_start (tree valist
, rtx nextarg
)
4731 HOST_WIDE_INT words
, n_gpr
, n_fpr
;
4732 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
4733 tree gpr
, fpr
, ovf
, sav
, t
;
4736 /* Only 64bit target needs something special. */
4737 if (!TARGET_64BIT
|| TARGET_64BIT_MS_ABI
)
4739 std_expand_builtin_va_start (valist
, nextarg
);
4743 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
4744 f_fpr
= TREE_CHAIN (f_gpr
);
4745 f_ovf
= TREE_CHAIN (f_fpr
);
4746 f_sav
= TREE_CHAIN (f_ovf
);
4748 valist
= build1 (INDIRECT_REF
, TREE_TYPE (TREE_TYPE (valist
)), valist
);
4749 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
4750 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
4751 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
4752 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
4754 /* Count number of gp and fp argument registers used. */
4755 words
= current_function_args_info
.words
;
4756 n_gpr
= current_function_args_info
.regno
;
4757 n_fpr
= current_function_args_info
.sse_regno
;
4759 if (cfun
->va_list_gpr_size
)
4761 type
= TREE_TYPE (gpr
);
4762 t
= build2 (GIMPLE_MODIFY_STMT
, type
, gpr
,
4763 build_int_cst (type
, n_gpr
* 8));
4764 TREE_SIDE_EFFECTS (t
) = 1;
4765 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
4768 if (cfun
->va_list_fpr_size
)
4770 type
= TREE_TYPE (fpr
);
4771 t
= build2 (GIMPLE_MODIFY_STMT
, type
, fpr
,
4772 build_int_cst (type
, n_fpr
* 16 + 8*REGPARM_MAX
));
4773 TREE_SIDE_EFFECTS (t
) = 1;
4774 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
4777 /* Find the overflow area. */
4778 type
= TREE_TYPE (ovf
);
4779 t
= make_tree (type
, virtual_incoming_args_rtx
);
4781 t
= build2 (PLUS_EXPR
, type
, t
,
4782 build_int_cst (type
, words
* UNITS_PER_WORD
));
4783 t
= build2 (GIMPLE_MODIFY_STMT
, type
, ovf
, t
);
4784 TREE_SIDE_EFFECTS (t
) = 1;
4785 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
4787 if (cfun
->va_list_gpr_size
|| cfun
->va_list_fpr_size
)
4789 /* Find the register save area.
4790 Prologue of the function save it right above stack frame. */
4791 type
= TREE_TYPE (sav
);
4792 t
= make_tree (type
, frame_pointer_rtx
);
4793 t
= build2 (GIMPLE_MODIFY_STMT
, type
, sav
, t
);
4794 TREE_SIDE_EFFECTS (t
) = 1;
4795 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
4799 /* Implement va_arg. */
4802 ix86_gimplify_va_arg (tree valist
, tree type
, tree
*pre_p
, tree
*post_p
)
4804 static const int intreg
[6] = { 0, 1, 2, 3, 4, 5 };
4805 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
4806 tree gpr
, fpr
, ovf
, sav
, t
;
4808 tree lab_false
, lab_over
= NULL_TREE
;
4813 enum machine_mode nat_mode
;
4815 /* Only 64bit target needs something special. */
4816 if (!TARGET_64BIT
|| TARGET_64BIT_MS_ABI
)
4817 return std_gimplify_va_arg_expr (valist
, type
, pre_p
, post_p
);
4819 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
4820 f_fpr
= TREE_CHAIN (f_gpr
);
4821 f_ovf
= TREE_CHAIN (f_fpr
);
4822 f_sav
= TREE_CHAIN (f_ovf
);
4824 valist
= build_va_arg_indirect_ref (valist
);
4825 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
4826 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
4827 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
4828 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
4830 indirect_p
= pass_by_reference (NULL
, TYPE_MODE (type
), type
, false);
4832 type
= build_pointer_type (type
);
4833 size
= int_size_in_bytes (type
);
4834 rsize
= (size
+ UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
4836 nat_mode
= type_natural_mode (type
);
4837 container
= construct_container (nat_mode
, TYPE_MODE (type
), type
, 0,
4838 REGPARM_MAX
, SSE_REGPARM_MAX
, intreg
, 0);
4840 /* Pull the value out of the saved registers. */
4842 addr
= create_tmp_var (ptr_type_node
, "addr");
4843 DECL_POINTER_ALIAS_SET (addr
) = get_varargs_alias_set ();
4847 int needed_intregs
, needed_sseregs
;
4849 tree int_addr
, sse_addr
;
4851 lab_false
= create_artificial_label ();
4852 lab_over
= create_artificial_label ();
4854 examine_argument (nat_mode
, type
, 0, &needed_intregs
, &needed_sseregs
);
4856 need_temp
= (!REG_P (container
)
4857 && ((needed_intregs
&& TYPE_ALIGN (type
) > 64)
4858 || TYPE_ALIGN (type
) > 128));
4860 /* In case we are passing structure, verify that it is consecutive block
4861 on the register save area. If not we need to do moves. */
4862 if (!need_temp
&& !REG_P (container
))
4864 /* Verify that all registers are strictly consecutive */
4865 if (SSE_REGNO_P (REGNO (XEXP (XVECEXP (container
, 0, 0), 0))))
4869 for (i
= 0; i
< XVECLEN (container
, 0) && !need_temp
; i
++)
4871 rtx slot
= XVECEXP (container
, 0, i
);
4872 if (REGNO (XEXP (slot
, 0)) != FIRST_SSE_REG
+ (unsigned int) i
4873 || INTVAL (XEXP (slot
, 1)) != i
* 16)
4881 for (i
= 0; i
< XVECLEN (container
, 0) && !need_temp
; i
++)
4883 rtx slot
= XVECEXP (container
, 0, i
);
4884 if (REGNO (XEXP (slot
, 0)) != (unsigned int) i
4885 || INTVAL (XEXP (slot
, 1)) != i
* 8)
4897 int_addr
= create_tmp_var (ptr_type_node
, "int_addr");
4898 DECL_POINTER_ALIAS_SET (int_addr
) = get_varargs_alias_set ();
4899 sse_addr
= create_tmp_var (ptr_type_node
, "sse_addr");
4900 DECL_POINTER_ALIAS_SET (sse_addr
) = get_varargs_alias_set ();
4903 /* First ensure that we fit completely in registers. */
4906 t
= build_int_cst (TREE_TYPE (gpr
),
4907 (REGPARM_MAX
- needed_intregs
+ 1) * 8);
4908 t
= build2 (GE_EXPR
, boolean_type_node
, gpr
, t
);
4909 t2
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
4910 t
= build3 (COND_EXPR
, void_type_node
, t
, t2
, NULL_TREE
);
4911 gimplify_and_add (t
, pre_p
);
4915 t
= build_int_cst (TREE_TYPE (fpr
),
4916 (SSE_REGPARM_MAX
- needed_sseregs
+ 1) * 16
4918 t
= build2 (GE_EXPR
, boolean_type_node
, fpr
, t
);
4919 t2
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
4920 t
= build3 (COND_EXPR
, void_type_node
, t
, t2
, NULL_TREE
);
4921 gimplify_and_add (t
, pre_p
);
4924 /* Compute index to start of area used for integer regs. */
4927 /* int_addr = gpr + sav; */
4928 t
= fold_convert (ptr_type_node
, fold_convert (size_type_node
, gpr
));
4929 t
= build2 (PLUS_EXPR
, ptr_type_node
, sav
, t
);
4930 t
= build2 (GIMPLE_MODIFY_STMT
, void_type_node
, int_addr
, t
);
4931 gimplify_and_add (t
, pre_p
);
4935 /* sse_addr = fpr + sav; */
4936 t
= fold_convert (ptr_type_node
, fold_convert (size_type_node
, fpr
));
4937 t
= build2 (PLUS_EXPR
, ptr_type_node
, sav
, t
);
4938 t
= build2 (GIMPLE_MODIFY_STMT
, void_type_node
, sse_addr
, t
);
4939 gimplify_and_add (t
, pre_p
);
4944 tree temp
= create_tmp_var (type
, "va_arg_tmp");
4947 t
= build1 (ADDR_EXPR
, build_pointer_type (type
), temp
);
4948 t
= build2 (GIMPLE_MODIFY_STMT
, void_type_node
, addr
, t
);
4949 gimplify_and_add (t
, pre_p
);
4951 for (i
= 0; i
< XVECLEN (container
, 0); i
++)
4953 rtx slot
= XVECEXP (container
, 0, i
);
4954 rtx reg
= XEXP (slot
, 0);
4955 enum machine_mode mode
= GET_MODE (reg
);
4956 tree piece_type
= lang_hooks
.types
.type_for_mode (mode
, 1);
4957 tree addr_type
= build_pointer_type (piece_type
);
4960 tree dest_addr
, dest
;
4962 if (SSE_REGNO_P (REGNO (reg
)))
4964 src_addr
= sse_addr
;
4965 src_offset
= (REGNO (reg
) - FIRST_SSE_REG
) * 16;
4969 src_addr
= int_addr
;
4970 src_offset
= REGNO (reg
) * 8;
4972 src_addr
= fold_convert (addr_type
, src_addr
);
4973 src_addr
= fold_build2 (PLUS_EXPR
, addr_type
, src_addr
,
4974 build_int_cst (addr_type
, src_offset
));
4975 src
= build_va_arg_indirect_ref (src_addr
);
4977 dest_addr
= fold_convert (addr_type
, addr
);
4978 dest_addr
= fold_build2 (PLUS_EXPR
, addr_type
, dest_addr
,
4979 build_int_cst (addr_type
, INTVAL (XEXP (slot
, 1))));
4980 dest
= build_va_arg_indirect_ref (dest_addr
);
4982 t
= build2 (GIMPLE_MODIFY_STMT
, void_type_node
, dest
, src
);
4983 gimplify_and_add (t
, pre_p
);
4989 t
= build2 (PLUS_EXPR
, TREE_TYPE (gpr
), gpr
,
4990 build_int_cst (TREE_TYPE (gpr
), needed_intregs
* 8));
4991 t
= build2 (GIMPLE_MODIFY_STMT
, TREE_TYPE (gpr
), gpr
, t
);
4992 gimplify_and_add (t
, pre_p
);
4996 t
= build2 (PLUS_EXPR
, TREE_TYPE (fpr
), fpr
,
4997 build_int_cst (TREE_TYPE (fpr
), needed_sseregs
* 16));
4998 t
= build2 (GIMPLE_MODIFY_STMT
, TREE_TYPE (fpr
), fpr
, t
);
4999 gimplify_and_add (t
, pre_p
);
5002 t
= build1 (GOTO_EXPR
, void_type_node
, lab_over
);
5003 gimplify_and_add (t
, pre_p
);
5005 t
= build1 (LABEL_EXPR
, void_type_node
, lab_false
);
5006 append_to_statement_list (t
, pre_p
);
5009 /* ... otherwise out of the overflow area. */
5011 /* Care for on-stack alignment if needed. */
5012 if (FUNCTION_ARG_BOUNDARY (VOIDmode
, type
) <= 64
5013 || integer_zerop (TYPE_SIZE (type
)))
5017 HOST_WIDE_INT align
= FUNCTION_ARG_BOUNDARY (VOIDmode
, type
) / 8;
5018 t
= build2 (PLUS_EXPR
, TREE_TYPE (ovf
), ovf
,
5019 build_int_cst (TREE_TYPE (ovf
), align
- 1));
5020 t
= build2 (BIT_AND_EXPR
, TREE_TYPE (t
), t
,
5021 build_int_cst (TREE_TYPE (t
), -align
));
5023 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
5025 t2
= build2 (GIMPLE_MODIFY_STMT
, void_type_node
, addr
, t
);
5026 gimplify_and_add (t2
, pre_p
);
5028 t
= build2 (PLUS_EXPR
, TREE_TYPE (t
), t
,
5029 build_int_cst (TREE_TYPE (t
), rsize
* UNITS_PER_WORD
));
5030 t
= build2 (GIMPLE_MODIFY_STMT
, TREE_TYPE (ovf
), ovf
, t
);
5031 gimplify_and_add (t
, pre_p
);
5035 t
= build1 (LABEL_EXPR
, void_type_node
, lab_over
);
5036 append_to_statement_list (t
, pre_p
);
5039 ptrtype
= build_pointer_type (type
);
5040 addr
= fold_convert (ptrtype
, addr
);
5043 addr
= build_va_arg_indirect_ref (addr
);
5044 return build_va_arg_indirect_ref (addr
);
5047 /* Return nonzero if OPNUM's MEM should be matched
5048 in movabs* patterns. */
5051 ix86_check_movabs (rtx insn
, int opnum
)
5055 set
= PATTERN (insn
);
5056 if (GET_CODE (set
) == PARALLEL
)
5057 set
= XVECEXP (set
, 0, 0);
5058 gcc_assert (GET_CODE (set
) == SET
);
5059 mem
= XEXP (set
, opnum
);
5060 while (GET_CODE (mem
) == SUBREG
)
5061 mem
= SUBREG_REG (mem
);
5062 gcc_assert (MEM_P (mem
));
5063 return (volatile_ok
|| !MEM_VOLATILE_P (mem
));
5066 /* Initialize the table of extra 80387 mathematical constants. */
5069 init_ext_80387_constants (void)
5071 static const char * cst
[5] =
5073 "0.3010299956639811952256464283594894482", /* 0: fldlg2 */
5074 "0.6931471805599453094286904741849753009", /* 1: fldln2 */
5075 "1.4426950408889634073876517827983434472", /* 2: fldl2e */
5076 "3.3219280948873623478083405569094566090", /* 3: fldl2t */
5077 "3.1415926535897932385128089594061862044", /* 4: fldpi */
5081 for (i
= 0; i
< 5; i
++)
5083 real_from_string (&ext_80387_constants_table
[i
], cst
[i
]);
5084 /* Ensure each constant is rounded to XFmode precision. */
5085 real_convert (&ext_80387_constants_table
[i
],
5086 XFmode
, &ext_80387_constants_table
[i
]);
5089 ext_80387_constants_init
= 1;
5092 /* Return true if the constant is something that can be loaded with
5093 a special instruction. */
5096 standard_80387_constant_p (rtx x
)
5098 enum machine_mode mode
= GET_MODE (x
);
5102 if (!(X87_FLOAT_MODE_P (mode
) && (GET_CODE (x
) == CONST_DOUBLE
)))
5105 if (x
== CONST0_RTX (mode
))
5107 if (x
== CONST1_RTX (mode
))
5110 REAL_VALUE_FROM_CONST_DOUBLE (r
, x
);
5112 /* For XFmode constants, try to find a special 80387 instruction when
5113 optimizing for size or on those CPUs that benefit from them. */
5115 && (optimize_size
|| TARGET_EXT_80387_CONSTANTS
))
5119 if (! ext_80387_constants_init
)
5120 init_ext_80387_constants ();
5122 for (i
= 0; i
< 5; i
++)
5123 if (real_identical (&r
, &ext_80387_constants_table
[i
]))
5127 /* Load of the constant -0.0 or -1.0 will be split as
5128 fldz;fchs or fld1;fchs sequence. */
5129 if (real_isnegzero (&r
))
5131 if (real_identical (&r
, &dconstm1
))
5137 /* Return the opcode of the special instruction to be used to load
5141 standard_80387_constant_opcode (rtx x
)
5143 switch (standard_80387_constant_p (x
))
5167 /* Return the CONST_DOUBLE representing the 80387 constant that is
5168 loaded by the specified special instruction. The argument IDX
5169 matches the return value from standard_80387_constant_p. */
5172 standard_80387_constant_rtx (int idx
)
5176 if (! ext_80387_constants_init
)
5177 init_ext_80387_constants ();
5193 return CONST_DOUBLE_FROM_REAL_VALUE (ext_80387_constants_table
[i
],
5197 /* Return 1 if mode is a valid mode for sse. */
5199 standard_sse_mode_p (enum machine_mode mode
)
5216 /* Return 1 if X is FP constant we can load to SSE register w/o using memory.
5219 standard_sse_constant_p (rtx x
)
5221 enum machine_mode mode
= GET_MODE (x
);
5223 if (x
== const0_rtx
|| x
== CONST0_RTX (GET_MODE (x
)))
5225 if (vector_all_ones_operand (x
, mode
)
5226 && standard_sse_mode_p (mode
))
5227 return TARGET_SSE2
? 2 : -1;
5232 /* Return the opcode of the special instruction to be used to load
5236 standard_sse_constant_opcode (rtx insn
, rtx x
)
5238 switch (standard_sse_constant_p (x
))
5241 if (get_attr_mode (insn
) == MODE_V4SF
)
5242 return "xorps\t%0, %0";
5243 else if (get_attr_mode (insn
) == MODE_V2DF
)
5244 return "xorpd\t%0, %0";
5246 return "pxor\t%0, %0";
5248 return "pcmpeqd\t%0, %0";
5253 /* Returns 1 if OP contains a symbol reference */
5256 symbolic_reference_mentioned_p (rtx op
)
5261 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
5264 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
5265 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
5271 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
5272 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
5276 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
5283 /* Return 1 if it is appropriate to emit `ret' instructions in the
5284 body of a function. Do this only if the epilogue is simple, needing a
5285 couple of insns. Prior to reloading, we can't tell how many registers
5286 must be saved, so return 0 then. Return 0 if there is no frame
5287 marker to de-allocate. */
5290 ix86_can_use_return_insn_p (void)
5292 struct ix86_frame frame
;
5294 if (! reload_completed
|| frame_pointer_needed
)
5297 /* Don't allow more than 32 pop, since that's all we can do
5298 with one instruction. */
5299 if (current_function_pops_args
5300 && current_function_args_size
>= 32768)
5303 ix86_compute_frame_layout (&frame
);
5304 return frame
.to_allocate
== 0 && frame
.nregs
== 0;
5307 /* Value should be nonzero if functions must have frame pointers.
5308 Zero means the frame pointer need not be set up (and parms may
5309 be accessed via the stack pointer) in functions that seem suitable. */
5312 ix86_frame_pointer_required (void)
5314 /* If we accessed previous frames, then the generated code expects
5315 to be able to access the saved ebp value in our frame. */
5316 if (cfun
->machine
->accesses_prev_frame
)
5319 /* Several x86 os'es need a frame pointer for other reasons,
5320 usually pertaining to setjmp. */
5321 if (SUBTARGET_FRAME_POINTER_REQUIRED
)
5324 /* In override_options, TARGET_OMIT_LEAF_FRAME_POINTER turns off
5325 the frame pointer by default. Turn it back on now if we've not
5326 got a leaf function. */
5327 if (TARGET_OMIT_LEAF_FRAME_POINTER
5328 && (!current_function_is_leaf
5329 || ix86_current_function_calls_tls_descriptor
))
5332 if (current_function_profile
)
5338 /* Record that the current function accesses previous call frames. */
5341 ix86_setup_frame_addresses (void)
5343 cfun
->machine
->accesses_prev_frame
= 1;
5346 #if (defined(HAVE_GAS_HIDDEN) && (SUPPORTS_ONE_ONLY - 0)) || TARGET_MACHO
5347 # define USE_HIDDEN_LINKONCE 1
5349 # define USE_HIDDEN_LINKONCE 0
5352 static int pic_labels_used
;
5354 /* Fills in the label name that should be used for a pc thunk for
5355 the given register. */
5358 get_pc_thunk_name (char name
[32], unsigned int regno
)
5360 gcc_assert (!TARGET_64BIT
);
5362 if (USE_HIDDEN_LINKONCE
)
5363 sprintf (name
, "__i686.get_pc_thunk.%s", reg_names
[regno
]);
5365 ASM_GENERATE_INTERNAL_LABEL (name
, "LPR", regno
);
5369 /* This function generates code for -fpic that loads %ebx with
5370 the return address of the caller and then returns. */
5373 ix86_file_end (void)
5378 for (regno
= 0; regno
< 8; ++regno
)
5382 if (! ((pic_labels_used
>> regno
) & 1))
5385 get_pc_thunk_name (name
, regno
);
5390 switch_to_section (darwin_sections
[text_coal_section
]);
5391 fputs ("\t.weak_definition\t", asm_out_file
);
5392 assemble_name (asm_out_file
, name
);
5393 fputs ("\n\t.private_extern\t", asm_out_file
);
5394 assemble_name (asm_out_file
, name
);
5395 fputs ("\n", asm_out_file
);
5396 ASM_OUTPUT_LABEL (asm_out_file
, name
);
5400 if (USE_HIDDEN_LINKONCE
)
5404 decl
= build_decl (FUNCTION_DECL
, get_identifier (name
),
5406 TREE_PUBLIC (decl
) = 1;
5407 TREE_STATIC (decl
) = 1;
5408 DECL_ONE_ONLY (decl
) = 1;
5410 (*targetm
.asm_out
.unique_section
) (decl
, 0);
5411 switch_to_section (get_named_section (decl
, NULL
, 0));
5413 (*targetm
.asm_out
.globalize_label
) (asm_out_file
, name
);
5414 fputs ("\t.hidden\t", asm_out_file
);
5415 assemble_name (asm_out_file
, name
);
5416 fputc ('\n', asm_out_file
);
5417 ASM_DECLARE_FUNCTION_NAME (asm_out_file
, name
, decl
);
5421 switch_to_section (text_section
);
5422 ASM_OUTPUT_LABEL (asm_out_file
, name
);
5425 xops
[0] = gen_rtx_REG (SImode
, regno
);
5426 xops
[1] = gen_rtx_MEM (SImode
, stack_pointer_rtx
);
5427 output_asm_insn ("mov{l}\t{%1, %0|%0, %1}", xops
);
5428 output_asm_insn ("ret", xops
);
5431 if (NEED_INDICATE_EXEC_STACK
)
5432 file_end_indicate_exec_stack ();
5435 /* Emit code for the SET_GOT patterns. */
5438 output_set_got (rtx dest
, rtx label ATTRIBUTE_UNUSED
)
5444 if (TARGET_VXWORKS_RTP
&& flag_pic
)
5446 /* Load (*VXWORKS_GOTT_BASE) into the PIC register. */
5447 xops
[2] = gen_rtx_MEM (Pmode
,
5448 gen_rtx_SYMBOL_REF (Pmode
, VXWORKS_GOTT_BASE
));
5449 output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops
);
5451 /* Load (*VXWORKS_GOTT_BASE)[VXWORKS_GOTT_INDEX] into the PIC register.
5452 Use %P and a local symbol in order to print VXWORKS_GOTT_INDEX as
5453 an unadorned address. */
5454 xops
[2] = gen_rtx_SYMBOL_REF (Pmode
, VXWORKS_GOTT_INDEX
);
5455 SYMBOL_REF_FLAGS (xops
[2]) |= SYMBOL_FLAG_LOCAL
;
5456 output_asm_insn ("mov{l}\t{%P2(%0), %0|%0, DWORD PTR %P2[%0]}", xops
);
5460 xops
[1] = gen_rtx_SYMBOL_REF (Pmode
, GOT_SYMBOL_NAME
);
5462 if (! TARGET_DEEP_BRANCH_PREDICTION
|| !flag_pic
)
5464 xops
[2] = gen_rtx_LABEL_REF (Pmode
, label
? label
: gen_label_rtx ());
5467 output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops
);
5469 output_asm_insn ("call\t%a2", xops
);
5472 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
5473 is what will be referenced by the Mach-O PIC subsystem. */
5475 ASM_OUTPUT_LABEL (asm_out_file
, machopic_function_base_name ());
5478 (*targetm
.asm_out
.internal_label
) (asm_out_file
, "L",
5479 CODE_LABEL_NUMBER (XEXP (xops
[2], 0)));
5482 output_asm_insn ("pop{l}\t%0", xops
);
5487 get_pc_thunk_name (name
, REGNO (dest
));
5488 pic_labels_used
|= 1 << REGNO (dest
);
5490 xops
[2] = gen_rtx_SYMBOL_REF (Pmode
, ggc_strdup (name
));
5491 xops
[2] = gen_rtx_MEM (QImode
, xops
[2]);
5492 output_asm_insn ("call\t%X2", xops
);
5493 /* Output the Mach-O "canonical" label name ("Lxx$pb") here too. This
5494 is what will be referenced by the Mach-O PIC subsystem. */
5497 ASM_OUTPUT_LABEL (asm_out_file
, machopic_function_base_name ());
5499 targetm
.asm_out
.internal_label (asm_out_file
, "L",
5500 CODE_LABEL_NUMBER (label
));
5507 if (!flag_pic
|| TARGET_DEEP_BRANCH_PREDICTION
)
5508 output_asm_insn ("add{l}\t{%1, %0|%0, %1}", xops
);
5510 output_asm_insn ("add{l}\t{%1+[.-%a2], %0|%0, %1+(.-%a2)}", xops
);
5515 /* Generate an "push" pattern for input ARG. */
5520 return gen_rtx_SET (VOIDmode
,
5522 gen_rtx_PRE_DEC (Pmode
,
5523 stack_pointer_rtx
)),
5527 /* Return >= 0 if there is an unused call-clobbered register available
5528 for the entire function. */
5531 ix86_select_alt_pic_regnum (void)
5533 if (current_function_is_leaf
&& !current_function_profile
5534 && !ix86_current_function_calls_tls_descriptor
)
5537 for (i
= 2; i
>= 0; --i
)
5538 if (!regs_ever_live
[i
])
5542 return INVALID_REGNUM
;
5545 /* Return 1 if we need to save REGNO. */
5547 ix86_save_reg (unsigned int regno
, int maybe_eh_return
)
5549 if (pic_offset_table_rtx
5550 && regno
== REAL_PIC_OFFSET_TABLE_REGNUM
5551 && (regs_ever_live
[REAL_PIC_OFFSET_TABLE_REGNUM
]
5552 || current_function_profile
5553 || current_function_calls_eh_return
5554 || current_function_uses_const_pool
))
5556 if (ix86_select_alt_pic_regnum () != INVALID_REGNUM
)
5561 if (current_function_calls_eh_return
&& maybe_eh_return
)
5566 unsigned test
= EH_RETURN_DATA_REGNO (i
);
5567 if (test
== INVALID_REGNUM
)
5574 if (cfun
->machine
->force_align_arg_pointer
5575 && regno
== REGNO (cfun
->machine
->force_align_arg_pointer
))
5578 return (regs_ever_live
[regno
]
5579 && !call_used_regs
[regno
]
5580 && !fixed_regs
[regno
]
5581 && (regno
!= HARD_FRAME_POINTER_REGNUM
|| !frame_pointer_needed
));
5584 /* Return number of registers to be saved on the stack. */
5587 ix86_nsaved_regs (void)
5592 for (regno
= FIRST_PSEUDO_REGISTER
- 1; regno
>= 0; regno
--)
5593 if (ix86_save_reg (regno
, true))
5598 /* Return the offset between two registers, one to be eliminated, and the other
5599 its replacement, at the start of a routine. */
5602 ix86_initial_elimination_offset (int from
, int to
)
5604 struct ix86_frame frame
;
5605 ix86_compute_frame_layout (&frame
);
5607 if (from
== ARG_POINTER_REGNUM
&& to
== HARD_FRAME_POINTER_REGNUM
)
5608 return frame
.hard_frame_pointer_offset
;
5609 else if (from
== FRAME_POINTER_REGNUM
5610 && to
== HARD_FRAME_POINTER_REGNUM
)
5611 return frame
.hard_frame_pointer_offset
- frame
.frame_pointer_offset
;
5614 gcc_assert (to
== STACK_POINTER_REGNUM
);
5616 if (from
== ARG_POINTER_REGNUM
)
5617 return frame
.stack_pointer_offset
;
5619 gcc_assert (from
== FRAME_POINTER_REGNUM
);
5620 return frame
.stack_pointer_offset
- frame
.frame_pointer_offset
;
5624 /* Fill structure ix86_frame about frame of currently computed function. */
5627 ix86_compute_frame_layout (struct ix86_frame
*frame
)
5629 HOST_WIDE_INT total_size
;
5630 unsigned int stack_alignment_needed
;
5631 HOST_WIDE_INT offset
;
5632 unsigned int preferred_alignment
;
5633 HOST_WIDE_INT size
= get_frame_size ();
5635 frame
->nregs
= ix86_nsaved_regs ();
5638 stack_alignment_needed
= cfun
->stack_alignment_needed
/ BITS_PER_UNIT
;
5639 preferred_alignment
= cfun
->preferred_stack_boundary
/ BITS_PER_UNIT
;
5641 /* During reload iteration the amount of registers saved can change.
5642 Recompute the value as needed. Do not recompute when amount of registers
5643 didn't change as reload does multiple calls to the function and does not
5644 expect the decision to change within single iteration. */
5646 && cfun
->machine
->use_fast_prologue_epilogue_nregs
!= frame
->nregs
)
5648 int count
= frame
->nregs
;
5650 cfun
->machine
->use_fast_prologue_epilogue_nregs
= count
;
5651 /* The fast prologue uses move instead of push to save registers. This
5652 is significantly longer, but also executes faster as modern hardware
5653 can execute the moves in parallel, but can't do that for push/pop.
5655 Be careful about choosing what prologue to emit: When function takes
5656 many instructions to execute we may use slow version as well as in
5657 case function is known to be outside hot spot (this is known with
5658 feedback only). Weight the size of function by number of registers
5659 to save as it is cheap to use one or two push instructions but very
5660 slow to use many of them. */
5662 count
= (count
- 1) * FAST_PROLOGUE_INSN_COUNT
;
5663 if (cfun
->function_frequency
< FUNCTION_FREQUENCY_NORMAL
5664 || (flag_branch_probabilities
5665 && cfun
->function_frequency
< FUNCTION_FREQUENCY_HOT
))
5666 cfun
->machine
->use_fast_prologue_epilogue
= false;
5668 cfun
->machine
->use_fast_prologue_epilogue
5669 = !expensive_function_p (count
);
5671 if (TARGET_PROLOGUE_USING_MOVE
5672 && cfun
->machine
->use_fast_prologue_epilogue
)
5673 frame
->save_regs_using_mov
= true;
5675 frame
->save_regs_using_mov
= false;
5678 /* Skip return address and saved base pointer. */
5679 offset
= frame_pointer_needed
? UNITS_PER_WORD
* 2 : UNITS_PER_WORD
;
5681 frame
->hard_frame_pointer_offset
= offset
;
5683 /* Do some sanity checking of stack_alignment_needed and
5684 preferred_alignment, since i386 port is the only using those features
5685 that may break easily. */
5687 gcc_assert (!size
|| stack_alignment_needed
);
5688 gcc_assert (preferred_alignment
>= STACK_BOUNDARY
/ BITS_PER_UNIT
);
5689 gcc_assert (preferred_alignment
<= PREFERRED_STACK_BOUNDARY
/ BITS_PER_UNIT
);
5690 gcc_assert (stack_alignment_needed
5691 <= PREFERRED_STACK_BOUNDARY
/ BITS_PER_UNIT
);
5693 if (stack_alignment_needed
< STACK_BOUNDARY
/ BITS_PER_UNIT
)
5694 stack_alignment_needed
= STACK_BOUNDARY
/ BITS_PER_UNIT
;
5696 /* Register save area */
5697 offset
+= frame
->nregs
* UNITS_PER_WORD
;
5700 if (ix86_save_varrargs_registers
)
5702 offset
+= X86_64_VARARGS_SIZE
;
5703 frame
->va_arg_size
= X86_64_VARARGS_SIZE
;
5706 frame
->va_arg_size
= 0;
5708 /* Align start of frame for local function. */
5709 frame
->padding1
= ((offset
+ stack_alignment_needed
- 1)
5710 & -stack_alignment_needed
) - offset
;
5712 offset
+= frame
->padding1
;
5714 /* Frame pointer points here. */
5715 frame
->frame_pointer_offset
= offset
;
5719 /* Add outgoing arguments area. Can be skipped if we eliminated
5720 all the function calls as dead code.
5721 Skipping is however impossible when function calls alloca. Alloca
5722 expander assumes that last current_function_outgoing_args_size
5723 of stack frame are unused. */
5724 if (ACCUMULATE_OUTGOING_ARGS
5725 && (!current_function_is_leaf
|| current_function_calls_alloca
5726 || ix86_current_function_calls_tls_descriptor
))
5728 offset
+= current_function_outgoing_args_size
;
5729 frame
->outgoing_arguments_size
= current_function_outgoing_args_size
;
5732 frame
->outgoing_arguments_size
= 0;
5734 /* Align stack boundary. Only needed if we're calling another function
5736 if (!current_function_is_leaf
|| current_function_calls_alloca
5737 || ix86_current_function_calls_tls_descriptor
)
5738 frame
->padding2
= ((offset
+ preferred_alignment
- 1)
5739 & -preferred_alignment
) - offset
;
5741 frame
->padding2
= 0;
5743 offset
+= frame
->padding2
;
5745 /* We've reached end of stack frame. */
5746 frame
->stack_pointer_offset
= offset
;
5748 /* Size prologue needs to allocate. */
5749 frame
->to_allocate
=
5750 (size
+ frame
->padding1
+ frame
->padding2
5751 + frame
->outgoing_arguments_size
+ frame
->va_arg_size
);
5753 if ((!frame
->to_allocate
&& frame
->nregs
<= 1)
5754 || (TARGET_64BIT
&& frame
->to_allocate
>= (HOST_WIDE_INT
) 0x80000000))
5755 frame
->save_regs_using_mov
= false;
5757 if (TARGET_RED_ZONE
&& current_function_sp_is_unchanging
5758 && current_function_is_leaf
5759 && !ix86_current_function_calls_tls_descriptor
)
5761 frame
->red_zone_size
= frame
->to_allocate
;
5762 if (frame
->save_regs_using_mov
)
5763 frame
->red_zone_size
+= frame
->nregs
* UNITS_PER_WORD
;
5764 if (frame
->red_zone_size
> RED_ZONE_SIZE
- RED_ZONE_RESERVE
)
5765 frame
->red_zone_size
= RED_ZONE_SIZE
- RED_ZONE_RESERVE
;
5768 frame
->red_zone_size
= 0;
5769 frame
->to_allocate
-= frame
->red_zone_size
;
5770 frame
->stack_pointer_offset
-= frame
->red_zone_size
;
5772 fprintf (stderr
, "\n");
5773 fprintf (stderr
, "nregs: %ld\n", (long)frame
->nregs
);
5774 fprintf (stderr
, "size: %ld\n", (long)size
);
5775 fprintf (stderr
, "alignment1: %ld\n", (long)stack_alignment_needed
);
5776 fprintf (stderr
, "padding1: %ld\n", (long)frame
->padding1
);
5777 fprintf (stderr
, "va_arg: %ld\n", (long)frame
->va_arg_size
);
5778 fprintf (stderr
, "padding2: %ld\n", (long)frame
->padding2
);
5779 fprintf (stderr
, "to_allocate: %ld\n", (long)frame
->to_allocate
);
5780 fprintf (stderr
, "red_zone_size: %ld\n", (long)frame
->red_zone_size
);
5781 fprintf (stderr
, "frame_pointer_offset: %ld\n", (long)frame
->frame_pointer_offset
);
5782 fprintf (stderr
, "hard_frame_pointer_offset: %ld\n",
5783 (long)frame
->hard_frame_pointer_offset
);
5784 fprintf (stderr
, "stack_pointer_offset: %ld\n", (long)frame
->stack_pointer_offset
);
5785 fprintf (stderr
, "current_function_is_leaf: %ld\n", (long)current_function_is_leaf
);
5786 fprintf (stderr
, "current_function_calls_alloca: %ld\n", (long)current_function_calls_alloca
);
5787 fprintf (stderr
, "x86_current_function_calls_tls_descriptor: %ld\n", (long)ix86_current_function_calls_tls_descriptor
);
5791 /* Emit code to save registers in the prologue. */
5794 ix86_emit_save_regs (void)
5799 for (regno
= FIRST_PSEUDO_REGISTER
; regno
-- > 0; )
5800 if (ix86_save_reg (regno
, true))
5802 insn
= emit_insn (gen_push (gen_rtx_REG (Pmode
, regno
)));
5803 RTX_FRAME_RELATED_P (insn
) = 1;
5807 /* Emit code to save registers using MOV insns. First register
5808 is restored from POINTER + OFFSET. */
5810 ix86_emit_save_regs_using_mov (rtx pointer
, HOST_WIDE_INT offset
)
5815 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5816 if (ix86_save_reg (regno
, true))
5818 insn
= emit_move_insn (adjust_address (gen_rtx_MEM (Pmode
, pointer
),
5820 gen_rtx_REG (Pmode
, regno
));
5821 RTX_FRAME_RELATED_P (insn
) = 1;
5822 offset
+= UNITS_PER_WORD
;
5826 /* Expand prologue or epilogue stack adjustment.
5827 The pattern exist to put a dependency on all ebp-based memory accesses.
5828 STYLE should be negative if instructions should be marked as frame related,
5829 zero if %r11 register is live and cannot be freely used and positive
5833 pro_epilogue_adjust_stack (rtx dest
, rtx src
, rtx offset
, int style
)
5838 insn
= emit_insn (gen_pro_epilogue_adjust_stack_1 (dest
, src
, offset
));
5839 else if (x86_64_immediate_operand (offset
, DImode
))
5840 insn
= emit_insn (gen_pro_epilogue_adjust_stack_rex64 (dest
, src
, offset
));
5844 /* r11 is used by indirect sibcall return as well, set before the
5845 epilogue and used after the epilogue. ATM indirect sibcall
5846 shouldn't be used together with huge frame sizes in one
5847 function because of the frame_size check in sibcall.c. */
5849 r11
= gen_rtx_REG (DImode
, R11_REG
);
5850 insn
= emit_insn (gen_rtx_SET (DImode
, r11
, offset
));
5852 RTX_FRAME_RELATED_P (insn
) = 1;
5853 insn
= emit_insn (gen_pro_epilogue_adjust_stack_rex64_2 (dest
, src
, r11
,
5857 RTX_FRAME_RELATED_P (insn
) = 1;
5860 /* Handle the TARGET_INTERNAL_ARG_POINTER hook. */
5863 ix86_internal_arg_pointer (void)
5865 bool has_force_align_arg_pointer
=
5866 (0 != lookup_attribute (ix86_force_align_arg_pointer_string
,
5867 TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl
))));
5868 if ((FORCE_PREFERRED_STACK_BOUNDARY_IN_MAIN
5869 && DECL_NAME (current_function_decl
)
5870 && MAIN_NAME_P (DECL_NAME (current_function_decl
))
5871 && DECL_FILE_SCOPE_P (current_function_decl
))
5872 || ix86_force_align_arg_pointer
5873 || has_force_align_arg_pointer
)
5875 /* Nested functions can't realign the stack due to a register
5877 if (DECL_CONTEXT (current_function_decl
)
5878 && TREE_CODE (DECL_CONTEXT (current_function_decl
)) == FUNCTION_DECL
)
5880 if (ix86_force_align_arg_pointer
)
5881 warning (0, "-mstackrealign ignored for nested functions");
5882 if (has_force_align_arg_pointer
)
5883 error ("%s not supported for nested functions",
5884 ix86_force_align_arg_pointer_string
);
5885 return virtual_incoming_args_rtx
;
5887 cfun
->machine
->force_align_arg_pointer
= gen_rtx_REG (Pmode
, 2);
5888 return copy_to_reg (cfun
->machine
->force_align_arg_pointer
);
5891 return virtual_incoming_args_rtx
;
5894 /* Handle the TARGET_DWARF_HANDLE_FRAME_UNSPEC hook.
5895 This is called from dwarf2out.c to emit call frame instructions
5896 for frame-related insns containing UNSPECs and UNSPEC_VOLATILEs. */
5898 ix86_dwarf_handle_frame_unspec (const char *label
, rtx pattern
, int index
)
5900 rtx unspec
= SET_SRC (pattern
);
5901 gcc_assert (GET_CODE (unspec
) == UNSPEC
);
5905 case UNSPEC_REG_SAVE
:
5906 dwarf2out_reg_save_reg (label
, XVECEXP (unspec
, 0, 0),
5907 SET_DEST (pattern
));
5909 case UNSPEC_DEF_CFA
:
5910 dwarf2out_def_cfa (label
, REGNO (SET_DEST (pattern
)),
5911 INTVAL (XVECEXP (unspec
, 0, 0)));
5918 /* Expand the prologue into a bunch of separate insns. */
5921 ix86_expand_prologue (void)
5925 struct ix86_frame frame
;
5926 HOST_WIDE_INT allocate
;
5928 ix86_compute_frame_layout (&frame
);
5930 if (cfun
->machine
->force_align_arg_pointer
)
5934 /* Grab the argument pointer. */
5935 x
= plus_constant (stack_pointer_rtx
, 4);
5936 y
= cfun
->machine
->force_align_arg_pointer
;
5937 insn
= emit_insn (gen_rtx_SET (VOIDmode
, y
, x
));
5938 RTX_FRAME_RELATED_P (insn
) = 1;
5940 /* The unwind info consists of two parts: install the fafp as the cfa,
5941 and record the fafp as the "save register" of the stack pointer.
5942 The later is there in order that the unwinder can see where it
5943 should restore the stack pointer across the and insn. */
5944 x
= gen_rtx_UNSPEC (VOIDmode
, gen_rtvec (1, const0_rtx
), UNSPEC_DEF_CFA
);
5945 x
= gen_rtx_SET (VOIDmode
, y
, x
);
5946 RTX_FRAME_RELATED_P (x
) = 1;
5947 y
= gen_rtx_UNSPEC (VOIDmode
, gen_rtvec (1, stack_pointer_rtx
),
5949 y
= gen_rtx_SET (VOIDmode
, cfun
->machine
->force_align_arg_pointer
, y
);
5950 RTX_FRAME_RELATED_P (y
) = 1;
5951 x
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, x
, y
));
5952 x
= gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR
, x
, NULL
);
5953 REG_NOTES (insn
) = x
;
5955 /* Align the stack. */
5956 emit_insn (gen_andsi3 (stack_pointer_rtx
, stack_pointer_rtx
,
5959 /* And here we cheat like madmen with the unwind info. We force the
5960 cfa register back to sp+4, which is exactly what it was at the
5961 start of the function. Re-pushing the return address results in
5962 the return at the same spot relative to the cfa, and thus is
5963 correct wrt the unwind info. */
5964 x
= cfun
->machine
->force_align_arg_pointer
;
5965 x
= gen_frame_mem (Pmode
, plus_constant (x
, -4));
5966 insn
= emit_insn (gen_push (x
));
5967 RTX_FRAME_RELATED_P (insn
) = 1;
5970 x
= gen_rtx_UNSPEC (VOIDmode
, gen_rtvec (1, x
), UNSPEC_DEF_CFA
);
5971 x
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
, x
);
5972 x
= gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR
, x
, NULL
);
5973 REG_NOTES (insn
) = x
;
5976 /* Note: AT&T enter does NOT have reversed args. Enter is probably
5977 slower on all targets. Also sdb doesn't like it. */
5979 if (frame_pointer_needed
)
5981 insn
= emit_insn (gen_push (hard_frame_pointer_rtx
));
5982 RTX_FRAME_RELATED_P (insn
) = 1;
5984 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
5985 RTX_FRAME_RELATED_P (insn
) = 1;
5988 allocate
= frame
.to_allocate
;
5990 if (!frame
.save_regs_using_mov
)
5991 ix86_emit_save_regs ();
5993 allocate
+= frame
.nregs
* UNITS_PER_WORD
;
5995 /* When using red zone we may start register saving before allocating
5996 the stack frame saving one cycle of the prologue. */
5997 if (TARGET_RED_ZONE
&& frame
.save_regs_using_mov
)
5998 ix86_emit_save_regs_using_mov (frame_pointer_needed
? hard_frame_pointer_rtx
5999 : stack_pointer_rtx
,
6000 -frame
.nregs
* UNITS_PER_WORD
);
6004 else if (! TARGET_STACK_PROBE
|| allocate
< CHECK_STACK_LIMIT
)
6005 pro_epilogue_adjust_stack (stack_pointer_rtx
, stack_pointer_rtx
,
6006 GEN_INT (-allocate
), -1);
6009 /* Only valid for Win32. */
6010 rtx eax
= gen_rtx_REG (Pmode
, 0);
6014 gcc_assert (!TARGET_64BIT
|| TARGET_64BIT_MS_ABI
);
6016 if (TARGET_64BIT_MS_ABI
)
6019 eax_live
= ix86_eax_live_at_start_p ();
6023 emit_insn (gen_push (eax
));
6024 allocate
-= UNITS_PER_WORD
;
6027 emit_move_insn (eax
, GEN_INT (allocate
));
6030 insn
= gen_allocate_stack_worker_64 (eax
);
6032 insn
= gen_allocate_stack_worker_32 (eax
);
6033 insn
= emit_insn (insn
);
6034 RTX_FRAME_RELATED_P (insn
) = 1;
6035 t
= gen_rtx_PLUS (Pmode
, stack_pointer_rtx
, GEN_INT (-allocate
));
6036 t
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
, t
);
6037 REG_NOTES (insn
) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR
,
6038 t
, REG_NOTES (insn
));
6042 if (frame_pointer_needed
)
6043 t
= plus_constant (hard_frame_pointer_rtx
,
6046 - frame
.nregs
* UNITS_PER_WORD
);
6048 t
= plus_constant (stack_pointer_rtx
, allocate
);
6049 emit_move_insn (eax
, gen_rtx_MEM (Pmode
, t
));
6053 if (frame
.save_regs_using_mov
&& !TARGET_RED_ZONE
)
6055 if (!frame_pointer_needed
|| !frame
.to_allocate
)
6056 ix86_emit_save_regs_using_mov (stack_pointer_rtx
, frame
.to_allocate
);
6058 ix86_emit_save_regs_using_mov (hard_frame_pointer_rtx
,
6059 -frame
.nregs
* UNITS_PER_WORD
);
6062 pic_reg_used
= false;
6063 if (pic_offset_table_rtx
6064 && (regs_ever_live
[REAL_PIC_OFFSET_TABLE_REGNUM
]
6065 || current_function_profile
))
6067 unsigned int alt_pic_reg_used
= ix86_select_alt_pic_regnum ();
6069 if (alt_pic_reg_used
!= INVALID_REGNUM
)
6070 REGNO (pic_offset_table_rtx
) = alt_pic_reg_used
;
6072 pic_reg_used
= true;
6079 if (ix86_cmodel
== CM_LARGE_PIC
)
6081 rtx tmp_reg
= gen_rtx_REG (DImode
,
6082 FIRST_REX_INT_REG
+ 3 /* R11 */);
6083 rtx label
= gen_label_rtx ();
6085 LABEL_PRESERVE_P (label
) = 1;
6086 gcc_assert (REGNO (pic_offset_table_rtx
) != REGNO (tmp_reg
));
6087 insn
= emit_insn (gen_set_rip_rex64 (pic_offset_table_rtx
, label
));
6088 REG_NOTES (insn
) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD
, const0_rtx
, NULL
);
6089 insn
= emit_insn (gen_set_got_offset_rex64 (tmp_reg
, label
));
6090 REG_NOTES (insn
) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD
, const0_rtx
, NULL
);
6091 insn
= emit_insn (gen_adddi3 (pic_offset_table_rtx
,
6092 pic_offset_table_rtx
, tmp_reg
));
6095 insn
= emit_insn (gen_set_got_rex64 (pic_offset_table_rtx
));
6098 insn
= emit_insn (gen_set_got (pic_offset_table_rtx
));
6100 /* Even with accurate pre-reload life analysis, we can wind up
6101 deleting all references to the pic register after reload.
6102 Consider if cross-jumping unifies two sides of a branch
6103 controlled by a comparison vs the only read from a global.
6104 In which case, allow the set_got to be deleted, though we're
6105 too late to do anything about the ebx save in the prologue. */
6106 REG_NOTES (insn
) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD
, const0_rtx
, NULL
);
6109 /* Prevent function calls from be scheduled before the call to mcount.
6110 In the pic_reg_used case, make sure that the got load isn't deleted. */
6111 if (current_function_profile
)
6112 emit_insn (gen_blockage (pic_reg_used
? pic_offset_table_rtx
: const0_rtx
));
6115 /* Emit code to restore saved registers using MOV insns. First register
6116 is restored from POINTER + OFFSET. */
6118 ix86_emit_restore_regs_using_mov (rtx pointer
, HOST_WIDE_INT offset
,
6119 int maybe_eh_return
)
6122 rtx base_address
= gen_rtx_MEM (Pmode
, pointer
);
6124 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
6125 if (ix86_save_reg (regno
, maybe_eh_return
))
6127 /* Ensure that adjust_address won't be forced to produce pointer
6128 out of range allowed by x86-64 instruction set. */
6129 if (TARGET_64BIT
&& offset
!= trunc_int_for_mode (offset
, SImode
))
6133 r11
= gen_rtx_REG (DImode
, R11_REG
);
6134 emit_move_insn (r11
, GEN_INT (offset
));
6135 emit_insn (gen_adddi3 (r11
, r11
, pointer
));
6136 base_address
= gen_rtx_MEM (Pmode
, r11
);
6139 emit_move_insn (gen_rtx_REG (Pmode
, regno
),
6140 adjust_address (base_address
, Pmode
, offset
));
6141 offset
+= UNITS_PER_WORD
;
6145 /* Restore function stack, frame, and registers. */
6148 ix86_expand_epilogue (int style
)
6151 int sp_valid
= !frame_pointer_needed
|| current_function_sp_is_unchanging
;
6152 struct ix86_frame frame
;
6153 HOST_WIDE_INT offset
;
6155 ix86_compute_frame_layout (&frame
);
6157 /* Calculate start of saved registers relative to ebp. Special care
6158 must be taken for the normal return case of a function using
6159 eh_return: the eax and edx registers are marked as saved, but not
6160 restored along this path. */
6161 offset
= frame
.nregs
;
6162 if (current_function_calls_eh_return
&& style
!= 2)
6164 offset
*= -UNITS_PER_WORD
;
6166 /* If we're only restoring one register and sp is not valid then
6167 using a move instruction to restore the register since it's
6168 less work than reloading sp and popping the register.
6170 The default code result in stack adjustment using add/lea instruction,
6171 while this code results in LEAVE instruction (or discrete equivalent),
6172 so it is profitable in some other cases as well. Especially when there
6173 are no registers to restore. We also use this code when TARGET_USE_LEAVE
6174 and there is exactly one register to pop. This heuristic may need some
6175 tuning in future. */
6176 if ((!sp_valid
&& frame
.nregs
<= 1)
6177 || (TARGET_EPILOGUE_USING_MOVE
6178 && cfun
->machine
->use_fast_prologue_epilogue
6179 && (frame
.nregs
> 1 || frame
.to_allocate
))
6180 || (frame_pointer_needed
&& !frame
.nregs
&& frame
.to_allocate
)
6181 || (frame_pointer_needed
&& TARGET_USE_LEAVE
6182 && cfun
->machine
->use_fast_prologue_epilogue
6183 && frame
.nregs
== 1)
6184 || current_function_calls_eh_return
)
6186 /* Restore registers. We can use ebp or esp to address the memory
6187 locations. If both are available, default to ebp, since offsets
6188 are known to be small. Only exception is esp pointing directly to the
6189 end of block of saved registers, where we may simplify addressing
6192 if (!frame_pointer_needed
|| (sp_valid
&& !frame
.to_allocate
))
6193 ix86_emit_restore_regs_using_mov (stack_pointer_rtx
,
6194 frame
.to_allocate
, style
== 2);
6196 ix86_emit_restore_regs_using_mov (hard_frame_pointer_rtx
,
6197 offset
, style
== 2);
6199 /* eh_return epilogues need %ecx added to the stack pointer. */
6202 rtx tmp
, sa
= EH_RETURN_STACKADJ_RTX
;
6204 if (frame_pointer_needed
)
6206 tmp
= gen_rtx_PLUS (Pmode
, hard_frame_pointer_rtx
, sa
);
6207 tmp
= plus_constant (tmp
, UNITS_PER_WORD
);
6208 emit_insn (gen_rtx_SET (VOIDmode
, sa
, tmp
));
6210 tmp
= gen_rtx_MEM (Pmode
, hard_frame_pointer_rtx
);
6211 emit_move_insn (hard_frame_pointer_rtx
, tmp
);
6213 pro_epilogue_adjust_stack (stack_pointer_rtx
, sa
,
6218 tmp
= gen_rtx_PLUS (Pmode
, stack_pointer_rtx
, sa
);
6219 tmp
= plus_constant (tmp
, (frame
.to_allocate
6220 + frame
.nregs
* UNITS_PER_WORD
));
6221 emit_insn (gen_rtx_SET (VOIDmode
, stack_pointer_rtx
, tmp
));
6224 else if (!frame_pointer_needed
)
6225 pro_epilogue_adjust_stack (stack_pointer_rtx
, stack_pointer_rtx
,
6226 GEN_INT (frame
.to_allocate
6227 + frame
.nregs
* UNITS_PER_WORD
),
6229 /* If not an i386, mov & pop is faster than "leave". */
6230 else if (TARGET_USE_LEAVE
|| optimize_size
6231 || !cfun
->machine
->use_fast_prologue_epilogue
)
6232 emit_insn (TARGET_64BIT
? gen_leave_rex64 () : gen_leave ());
6235 pro_epilogue_adjust_stack (stack_pointer_rtx
,
6236 hard_frame_pointer_rtx
,
6239 emit_insn (gen_popdi1 (hard_frame_pointer_rtx
));
6241 emit_insn (gen_popsi1 (hard_frame_pointer_rtx
));
6246 /* First step is to deallocate the stack frame so that we can
6247 pop the registers. */
6250 gcc_assert (frame_pointer_needed
);
6251 pro_epilogue_adjust_stack (stack_pointer_rtx
,
6252 hard_frame_pointer_rtx
,
6253 GEN_INT (offset
), style
);
6255 else if (frame
.to_allocate
)
6256 pro_epilogue_adjust_stack (stack_pointer_rtx
, stack_pointer_rtx
,
6257 GEN_INT (frame
.to_allocate
), style
);
6259 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
6260 if (ix86_save_reg (regno
, false))
6263 emit_insn (gen_popdi1 (gen_rtx_REG (Pmode
, regno
)));
6265 emit_insn (gen_popsi1 (gen_rtx_REG (Pmode
, regno
)));
6267 if (frame_pointer_needed
)
6269 /* Leave results in shorter dependency chains on CPUs that are
6270 able to grok it fast. */
6271 if (TARGET_USE_LEAVE
)
6272 emit_insn (TARGET_64BIT
? gen_leave_rex64 () : gen_leave ());
6273 else if (TARGET_64BIT
)
6274 emit_insn (gen_popdi1 (hard_frame_pointer_rtx
));
6276 emit_insn (gen_popsi1 (hard_frame_pointer_rtx
));
6280 if (cfun
->machine
->force_align_arg_pointer
)
6282 emit_insn (gen_addsi3 (stack_pointer_rtx
,
6283 cfun
->machine
->force_align_arg_pointer
,
6287 /* Sibcall epilogues don't want a return instruction. */
6291 if (current_function_pops_args
&& current_function_args_size
)
6293 rtx popc
= GEN_INT (current_function_pops_args
);
6295 /* i386 can only pop 64K bytes. If asked to pop more, pop
6296 return address, do explicit add, and jump indirectly to the
6299 if (current_function_pops_args
>= 65536)
6301 rtx ecx
= gen_rtx_REG (SImode
, 2);
6303 /* There is no "pascal" calling convention in any 64bit ABI. */
6304 gcc_assert (!TARGET_64BIT
);
6306 emit_insn (gen_popsi1 (ecx
));
6307 emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
, popc
));
6308 emit_jump_insn (gen_return_indirect_internal (ecx
));
6311 emit_jump_insn (gen_return_pop_internal (popc
));
6314 emit_jump_insn (gen_return_internal ());
6317 /* Reset from the function's potential modifications. */
6320 ix86_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED
,
6321 HOST_WIDE_INT size ATTRIBUTE_UNUSED
)
6323 if (pic_offset_table_rtx
)
6324 REGNO (pic_offset_table_rtx
) = REAL_PIC_OFFSET_TABLE_REGNUM
;
6326 /* Mach-O doesn't support labels at the end of objects, so if
6327 it looks like we might want one, insert a NOP. */
6329 rtx insn
= get_last_insn ();
6332 && NOTE_KIND (insn
) != NOTE_INSN_DELETED_LABEL
)
6333 insn
= PREV_INSN (insn
);
6337 && NOTE_KIND (insn
) == NOTE_INSN_DELETED_LABEL
)))
6338 fputs ("\tnop\n", file
);
6344 /* Extract the parts of an RTL expression that is a valid memory address
6345 for an instruction. Return 0 if the structure of the address is
6346 grossly off. Return -1 if the address contains ASHIFT, so it is not
6347 strictly valid, but still used for computing length of lea instruction. */
6350 ix86_decompose_address (rtx addr
, struct ix86_address
*out
)
6352 rtx base
= NULL_RTX
, index
= NULL_RTX
, disp
= NULL_RTX
;
6353 rtx base_reg
, index_reg
;
6354 HOST_WIDE_INT scale
= 1;
6355 rtx scale_rtx
= NULL_RTX
;
6357 enum ix86_address_seg seg
= SEG_DEFAULT
;
6359 if (REG_P (addr
) || GET_CODE (addr
) == SUBREG
)
6361 else if (GET_CODE (addr
) == PLUS
)
6371 addends
[n
++] = XEXP (op
, 1);
6374 while (GET_CODE (op
) == PLUS
);
6379 for (i
= n
; i
>= 0; --i
)
6382 switch (GET_CODE (op
))
6387 index
= XEXP (op
, 0);
6388 scale_rtx
= XEXP (op
, 1);
6392 if (XINT (op
, 1) == UNSPEC_TP
6393 && TARGET_TLS_DIRECT_SEG_REFS
6394 && seg
== SEG_DEFAULT
)
6395 seg
= TARGET_64BIT
? SEG_FS
: SEG_GS
;
6424 else if (GET_CODE (addr
) == MULT
)
6426 index
= XEXP (addr
, 0); /* index*scale */
6427 scale_rtx
= XEXP (addr
, 1);
6429 else if (GET_CODE (addr
) == ASHIFT
)
6433 /* We're called for lea too, which implements ashift on occasion. */
6434 index
= XEXP (addr
, 0);
6435 tmp
= XEXP (addr
, 1);
6436 if (!CONST_INT_P (tmp
))
6438 scale
= INTVAL (tmp
);
6439 if ((unsigned HOST_WIDE_INT
) scale
> 3)
6445 disp
= addr
; /* displacement */
6447 /* Extract the integral value of scale. */
6450 if (!CONST_INT_P (scale_rtx
))
6452 scale
= INTVAL (scale_rtx
);
6455 base_reg
= base
&& GET_CODE (base
) == SUBREG
? SUBREG_REG (base
) : base
;
6456 index_reg
= index
&& GET_CODE (index
) == SUBREG
? SUBREG_REG (index
) : index
;
6458 /* Allow arg pointer and stack pointer as index if there is not scaling. */
6459 if (base_reg
&& index_reg
&& scale
== 1
6460 && (index_reg
== arg_pointer_rtx
6461 || index_reg
== frame_pointer_rtx
6462 || (REG_P (index_reg
) && REGNO (index_reg
) == STACK_POINTER_REGNUM
)))
6465 tmp
= base
, base
= index
, index
= tmp
;
6466 tmp
= base_reg
, base_reg
= index_reg
, index_reg
= tmp
;
6469 /* Special case: %ebp cannot be encoded as a base without a displacement. */
6470 if ((base_reg
== hard_frame_pointer_rtx
6471 || base_reg
== frame_pointer_rtx
6472 || base_reg
== arg_pointer_rtx
) && !disp
)
6475 /* Special case: on K6, [%esi] makes the instruction vector decoded.
6476 Avoid this by transforming to [%esi+0]. */
6477 if (ix86_tune
== PROCESSOR_K6
&& !optimize_size
6478 && base_reg
&& !index_reg
&& !disp
6480 && REGNO_REG_CLASS (REGNO (base_reg
)) == SIREG
)
6483 /* Special case: encode reg+reg instead of reg*2. */
6484 if (!base
&& index
&& scale
&& scale
== 2)
6485 base
= index
, base_reg
= index_reg
, scale
= 1;
6487 /* Special case: scaling cannot be encoded without base or displacement. */
6488 if (!base
&& !disp
&& index
&& scale
!= 1)
6500 /* Return cost of the memory address x.
6501 For i386, it is better to use a complex address than let gcc copy
6502 the address into a reg and make a new pseudo. But not if the address
6503 requires to two regs - that would mean more pseudos with longer
6506 ix86_address_cost (rtx x
)
6508 struct ix86_address parts
;
6510 int ok
= ix86_decompose_address (x
, &parts
);
6514 if (parts
.base
&& GET_CODE (parts
.base
) == SUBREG
)
6515 parts
.base
= SUBREG_REG (parts
.base
);
6516 if (parts
.index
&& GET_CODE (parts
.index
) == SUBREG
)
6517 parts
.index
= SUBREG_REG (parts
.index
);
6519 /* More complex memory references are better. */
6520 if (parts
.disp
&& parts
.disp
!= const0_rtx
)
6522 if (parts
.seg
!= SEG_DEFAULT
)
6525 /* Attempt to minimize number of registers in the address. */
6527 && (!REG_P (parts
.base
) || REGNO (parts
.base
) >= FIRST_PSEUDO_REGISTER
))
6529 && (!REG_P (parts
.index
)
6530 || REGNO (parts
.index
) >= FIRST_PSEUDO_REGISTER
)))
6534 && (!REG_P (parts
.base
) || REGNO (parts
.base
) >= FIRST_PSEUDO_REGISTER
)
6536 && (!REG_P (parts
.index
) || REGNO (parts
.index
) >= FIRST_PSEUDO_REGISTER
)
6537 && parts
.base
!= parts
.index
)
6540 /* AMD-K6 don't like addresses with ModR/M set to 00_xxx_100b,
6541 since it's predecode logic can't detect the length of instructions
6542 and it degenerates to vector decoded. Increase cost of such
6543 addresses here. The penalty is minimally 2 cycles. It may be worthwhile
6544 to split such addresses or even refuse such addresses at all.
6546 Following addressing modes are affected:
6551 The first and last case may be avoidable by explicitly coding the zero in
6552 memory address, but I don't have AMD-K6 machine handy to check this
6556 && ((!parts
.disp
&& parts
.base
&& parts
.index
&& parts
.scale
!= 1)
6557 || (parts
.disp
&& !parts
.base
&& parts
.index
&& parts
.scale
!= 1)
6558 || (!parts
.disp
&& parts
.base
&& parts
.index
&& parts
.scale
== 1)))
6564 /* Allow {LABEL | SYMBOL}_REF - SYMBOL_REF-FOR-PICBASE for Mach-O as
6565 this is used for to form addresses to local data when -fPIC is in
6569 darwin_local_data_pic (rtx disp
)
6571 if (GET_CODE (disp
) == MINUS
)
6573 if (GET_CODE (XEXP (disp
, 0)) == LABEL_REF
6574 || GET_CODE (XEXP (disp
, 0)) == SYMBOL_REF
)
6575 if (GET_CODE (XEXP (disp
, 1)) == SYMBOL_REF
)
6577 const char *sym_name
= XSTR (XEXP (disp
, 1), 0);
6578 if (! strcmp (sym_name
, "<pic base>"))
6586 /* Determine if a given RTX is a valid constant. We already know this
6587 satisfies CONSTANT_P. */
6590 legitimate_constant_p (rtx x
)
6592 switch (GET_CODE (x
))
6597 if (GET_CODE (x
) == PLUS
)
6599 if (!CONST_INT_P (XEXP (x
, 1)))
6604 if (TARGET_MACHO
&& darwin_local_data_pic (x
))
6607 /* Only some unspecs are valid as "constants". */
6608 if (GET_CODE (x
) == UNSPEC
)
6609 switch (XINT (x
, 1))
6614 return TARGET_64BIT
;
6617 x
= XVECEXP (x
, 0, 0);
6618 return (GET_CODE (x
) == SYMBOL_REF
6619 && SYMBOL_REF_TLS_MODEL (x
) == TLS_MODEL_LOCAL_EXEC
);
6621 x
= XVECEXP (x
, 0, 0);
6622 return (GET_CODE (x
) == SYMBOL_REF
6623 && SYMBOL_REF_TLS_MODEL (x
) == TLS_MODEL_LOCAL_DYNAMIC
);
6628 /* We must have drilled down to a symbol. */
6629 if (GET_CODE (x
) == LABEL_REF
)
6631 if (GET_CODE (x
) != SYMBOL_REF
)
6636 /* TLS symbols are never valid. */
6637 if (SYMBOL_REF_TLS_MODEL (x
))
6640 /* DLLIMPORT symbols are never valid. */
6641 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
6642 && SYMBOL_REF_DLLIMPORT_P (x
))
6647 if (GET_MODE (x
) == TImode
6648 && x
!= CONST0_RTX (TImode
)
6654 if (x
== CONST0_RTX (GET_MODE (x
)))
6662 /* Otherwise we handle everything else in the move patterns. */
6666 /* Determine if it's legal to put X into the constant pool. This
6667 is not possible for the address of thread-local symbols, which
6668 is checked above. */
6671 ix86_cannot_force_const_mem (rtx x
)
6673 /* We can always put integral constants and vectors in memory. */
6674 switch (GET_CODE (x
))
6684 return !legitimate_constant_p (x
);
6687 /* Determine if a given RTX is a valid constant address. */
6690 constant_address_p (rtx x
)
6692 return CONSTANT_P (x
) && legitimate_address_p (Pmode
, x
, 1);
6695 /* Nonzero if the constant value X is a legitimate general operand
6696 when generating PIC code. It is given that flag_pic is on and
6697 that X satisfies CONSTANT_P or is a CONST_DOUBLE. */
6700 legitimate_pic_operand_p (rtx x
)
6704 switch (GET_CODE (x
))
6707 inner
= XEXP (x
, 0);
6708 if (GET_CODE (inner
) == PLUS
6709 && CONST_INT_P (XEXP (inner
, 1)))
6710 inner
= XEXP (inner
, 0);
6712 /* Only some unspecs are valid as "constants". */
6713 if (GET_CODE (inner
) == UNSPEC
)
6714 switch (XINT (inner
, 1))
6719 return TARGET_64BIT
;
6721 x
= XVECEXP (inner
, 0, 0);
6722 return (GET_CODE (x
) == SYMBOL_REF
6723 && SYMBOL_REF_TLS_MODEL (x
) == TLS_MODEL_LOCAL_EXEC
);
6731 return legitimate_pic_address_disp_p (x
);
6738 /* Determine if a given CONST RTX is a valid memory displacement
6742 legitimate_pic_address_disp_p (rtx disp
)
6746 /* In 64bit mode we can allow direct addresses of symbols and labels
6747 when they are not dynamic symbols. */
6750 rtx op0
= disp
, op1
;
6752 switch (GET_CODE (disp
))
6758 if (GET_CODE (XEXP (disp
, 0)) != PLUS
)
6760 op0
= XEXP (XEXP (disp
, 0), 0);
6761 op1
= XEXP (XEXP (disp
, 0), 1);
6762 if (!CONST_INT_P (op1
)
6763 || INTVAL (op1
) >= 16*1024*1024
6764 || INTVAL (op1
) < -16*1024*1024)
6766 if (GET_CODE (op0
) == LABEL_REF
)
6768 if (GET_CODE (op0
) != SYMBOL_REF
)
6773 /* TLS references should always be enclosed in UNSPEC. */
6774 if (SYMBOL_REF_TLS_MODEL (op0
))
6776 if (!SYMBOL_REF_FAR_ADDR_P (op0
) && SYMBOL_REF_LOCAL_P (op0
)
6777 && ix86_cmodel
!= CM_LARGE_PIC
)
6785 if (GET_CODE (disp
) != CONST
)
6787 disp
= XEXP (disp
, 0);
6791 /* We are unsafe to allow PLUS expressions. This limit allowed distance
6792 of GOT tables. We should not need these anyway. */
6793 if (GET_CODE (disp
) != UNSPEC
6794 || (XINT (disp
, 1) != UNSPEC_GOTPCREL
6795 && XINT (disp
, 1) != UNSPEC_GOTOFF
6796 && XINT (disp
, 1) != UNSPEC_PLTOFF
))
6799 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
6800 && GET_CODE (XVECEXP (disp
, 0, 0)) != LABEL_REF
)
6806 if (GET_CODE (disp
) == PLUS
)
6808 if (!CONST_INT_P (XEXP (disp
, 1)))
6810 disp
= XEXP (disp
, 0);
6814 if (TARGET_MACHO
&& darwin_local_data_pic (disp
))
6817 if (GET_CODE (disp
) != UNSPEC
)
6820 switch (XINT (disp
, 1))
6825 /* We need to check for both symbols and labels because VxWorks loads
6826 text labels with @GOT rather than @GOTOFF. See gotoff_operand for
6828 return (GET_CODE (XVECEXP (disp
, 0, 0)) == SYMBOL_REF
6829 || GET_CODE (XVECEXP (disp
, 0, 0)) == LABEL_REF
);
6831 /* Refuse GOTOFF in 64bit mode since it is always 64bit when used.
6832 While ABI specify also 32bit relocation but we don't produce it in
6833 small PIC model at all. */
6834 if ((GET_CODE (XVECEXP (disp
, 0, 0)) == SYMBOL_REF
6835 || GET_CODE (XVECEXP (disp
, 0, 0)) == LABEL_REF
)
6837 return gotoff_operand (XVECEXP (disp
, 0, 0), Pmode
);
6839 case UNSPEC_GOTTPOFF
:
6840 case UNSPEC_GOTNTPOFF
:
6841 case UNSPEC_INDNTPOFF
:
6844 disp
= XVECEXP (disp
, 0, 0);
6845 return (GET_CODE (disp
) == SYMBOL_REF
6846 && SYMBOL_REF_TLS_MODEL (disp
) == TLS_MODEL_INITIAL_EXEC
);
6848 disp
= XVECEXP (disp
, 0, 0);
6849 return (GET_CODE (disp
) == SYMBOL_REF
6850 && SYMBOL_REF_TLS_MODEL (disp
) == TLS_MODEL_LOCAL_EXEC
);
6852 disp
= XVECEXP (disp
, 0, 0);
6853 return (GET_CODE (disp
) == SYMBOL_REF
6854 && SYMBOL_REF_TLS_MODEL (disp
) == TLS_MODEL_LOCAL_DYNAMIC
);
6860 /* GO_IF_LEGITIMATE_ADDRESS recognizes an RTL expression that is a valid
6861 memory address for an instruction. The MODE argument is the machine mode
6862 for the MEM expression that wants to use this address.
6864 It only recognizes address in canonical form. LEGITIMIZE_ADDRESS should
6865 convert common non-canonical forms to canonical form so that they will
6869 legitimate_address_p (enum machine_mode mode ATTRIBUTE_UNUSED
,
6870 rtx addr
, int strict
)
6872 struct ix86_address parts
;
6873 rtx base
, index
, disp
;
6874 HOST_WIDE_INT scale
;
6875 const char *reason
= NULL
;
6876 rtx reason_rtx
= NULL_RTX
;
6878 if (ix86_decompose_address (addr
, &parts
) <= 0)
6880 reason
= "decomposition failed";
6885 index
= parts
.index
;
6887 scale
= parts
.scale
;
6889 /* Validate base register.
6891 Don't allow SUBREG's that span more than a word here. It can lead to spill
6892 failures when the base is one word out of a two word structure, which is
6893 represented internally as a DImode int. */
6902 else if (GET_CODE (base
) == SUBREG
6903 && REG_P (SUBREG_REG (base
))
6904 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (base
)))
6906 reg
= SUBREG_REG (base
);
6909 reason
= "base is not a register";
6913 if (GET_MODE (base
) != Pmode
)
6915 reason
= "base is not in Pmode";
6919 if ((strict
&& ! REG_OK_FOR_BASE_STRICT_P (reg
))
6920 || (! strict
&& ! REG_OK_FOR_BASE_NONSTRICT_P (reg
)))
6922 reason
= "base is not valid";
6927 /* Validate index register.
6929 Don't allow SUBREG's that span more than a word here -- same as above. */
6938 else if (GET_CODE (index
) == SUBREG
6939 && REG_P (SUBREG_REG (index
))
6940 && GET_MODE_SIZE (GET_MODE (SUBREG_REG (index
)))
6942 reg
= SUBREG_REG (index
);
6945 reason
= "index is not a register";
6949 if (GET_MODE (index
) != Pmode
)
6951 reason
= "index is not in Pmode";
6955 if ((strict
&& ! REG_OK_FOR_INDEX_STRICT_P (reg
))
6956 || (! strict
&& ! REG_OK_FOR_INDEX_NONSTRICT_P (reg
)))
6958 reason
= "index is not valid";
6963 /* Validate scale factor. */
6966 reason_rtx
= GEN_INT (scale
);
6969 reason
= "scale without index";
6973 if (scale
!= 2 && scale
!= 4 && scale
!= 8)
6975 reason
= "scale is not a valid multiplier";
6980 /* Validate displacement. */
6985 if (GET_CODE (disp
) == CONST
6986 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
)
6987 switch (XINT (XEXP (disp
, 0), 1))
6989 /* Refuse GOTOFF and GOT in 64bit mode since it is always 64bit when
6990 used. While ABI specify also 32bit relocations, we don't produce
6991 them at all and use IP relative instead. */
6994 gcc_assert (flag_pic
);
6996 goto is_legitimate_pic
;
6997 reason
= "64bit address unspec";
7000 case UNSPEC_GOTPCREL
:
7001 gcc_assert (flag_pic
);
7002 goto is_legitimate_pic
;
7004 case UNSPEC_GOTTPOFF
:
7005 case UNSPEC_GOTNTPOFF
:
7006 case UNSPEC_INDNTPOFF
:
7012 reason
= "invalid address unspec";
7016 else if (SYMBOLIC_CONST (disp
)
7020 && MACHOPIC_INDIRECT
7021 && !machopic_operand_p (disp
)
7027 if (TARGET_64BIT
&& (index
|| base
))
7029 /* foo@dtpoff(%rX) is ok. */
7030 if (GET_CODE (disp
) != CONST
7031 || GET_CODE (XEXP (disp
, 0)) != PLUS
7032 || GET_CODE (XEXP (XEXP (disp
, 0), 0)) != UNSPEC
7033 || !CONST_INT_P (XEXP (XEXP (disp
, 0), 1))
7034 || (XINT (XEXP (XEXP (disp
, 0), 0), 1) != UNSPEC_DTPOFF
7035 && XINT (XEXP (XEXP (disp
, 0), 0), 1) != UNSPEC_NTPOFF
))
7037 reason
= "non-constant pic memory reference";
7041 else if (! legitimate_pic_address_disp_p (disp
))
7043 reason
= "displacement is an invalid pic construct";
7047 /* This code used to verify that a symbolic pic displacement
7048 includes the pic_offset_table_rtx register.
7050 While this is good idea, unfortunately these constructs may
7051 be created by "adds using lea" optimization for incorrect
7060 This code is nonsensical, but results in addressing
7061 GOT table with pic_offset_table_rtx base. We can't
7062 just refuse it easily, since it gets matched by
7063 "addsi3" pattern, that later gets split to lea in the
7064 case output register differs from input. While this
7065 can be handled by separate addsi pattern for this case
7066 that never results in lea, this seems to be easier and
7067 correct fix for crash to disable this test. */
7069 else if (GET_CODE (disp
) != LABEL_REF
7070 && !CONST_INT_P (disp
)
7071 && (GET_CODE (disp
) != CONST
7072 || !legitimate_constant_p (disp
))
7073 && (GET_CODE (disp
) != SYMBOL_REF
7074 || !legitimate_constant_p (disp
)))
7076 reason
= "displacement is not constant";
7079 else if (TARGET_64BIT
7080 && !x86_64_immediate_operand (disp
, VOIDmode
))
7082 reason
= "displacement is out of range";
7087 /* Everything looks valid. */
7094 /* Return a unique alias set for the GOT. */
7096 static HOST_WIDE_INT
7097 ix86_GOT_alias_set (void)
7099 static HOST_WIDE_INT set
= -1;
7101 set
= new_alias_set ();
7105 /* Return a legitimate reference for ORIG (an address) using the
7106 register REG. If REG is 0, a new pseudo is generated.
7108 There are two types of references that must be handled:
7110 1. Global data references must load the address from the GOT, via
7111 the PIC reg. An insn is emitted to do this load, and the reg is
7114 2. Static data references, constant pool addresses, and code labels
7115 compute the address as an offset from the GOT, whose base is in
7116 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
7117 differentiate them from global data objects. The returned
7118 address is the PIC reg + an unspec constant.
7120 GO_IF_LEGITIMATE_ADDRESS rejects symbolic references unless the PIC
7121 reg also appears in the address. */
7124 legitimize_pic_address (rtx orig
, rtx reg
)
7131 if (TARGET_MACHO
&& !TARGET_64BIT
)
7134 reg
= gen_reg_rtx (Pmode
);
7135 /* Use the generic Mach-O PIC machinery. */
7136 return machopic_legitimize_pic_address (orig
, GET_MODE (orig
), reg
);
7140 if (TARGET_64BIT
&& legitimate_pic_address_disp_p (addr
))
7142 else if (TARGET_64BIT
7143 && ix86_cmodel
!= CM_SMALL_PIC
7144 && gotoff_operand (addr
, Pmode
))
7147 /* This symbol may be referenced via a displacement from the PIC
7148 base address (@GOTOFF). */
7150 if (reload_in_progress
)
7151 regs_ever_live
[PIC_OFFSET_TABLE_REGNUM
] = 1;
7152 if (GET_CODE (addr
) == CONST
)
7153 addr
= XEXP (addr
, 0);
7154 if (GET_CODE (addr
) == PLUS
)
7156 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (addr
, 0)),
7158 new = gen_rtx_PLUS (Pmode
, new, XEXP (addr
, 1));
7161 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
7162 new = gen_rtx_CONST (Pmode
, new);
7164 tmpreg
= gen_reg_rtx (Pmode
);
7167 emit_move_insn (tmpreg
, new);
7171 new = expand_simple_binop (Pmode
, PLUS
, reg
, pic_offset_table_rtx
,
7172 tmpreg
, 1, OPTAB_DIRECT
);
7175 else new = gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, tmpreg
);
7177 else if (!TARGET_64BIT
&& gotoff_operand (addr
, Pmode
))
7179 /* This symbol may be referenced via a displacement from the PIC
7180 base address (@GOTOFF). */
7182 if (reload_in_progress
)
7183 regs_ever_live
[PIC_OFFSET_TABLE_REGNUM
] = 1;
7184 if (GET_CODE (addr
) == CONST
)
7185 addr
= XEXP (addr
, 0);
7186 if (GET_CODE (addr
) == PLUS
)
7188 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (addr
, 0)),
7190 new = gen_rtx_PLUS (Pmode
, new, XEXP (addr
, 1));
7193 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
7194 new = gen_rtx_CONST (Pmode
, new);
7195 new = gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new);
7199 emit_move_insn (reg
, new);
7203 else if ((GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (addr
) == 0)
7204 /* We can't use @GOTOFF for text labels on VxWorks;
7205 see gotoff_operand. */
7206 || (TARGET_VXWORKS_RTP
&& GET_CODE (addr
) == LABEL_REF
))
7208 /* Given that we've already handled dllimport variables separately
7209 in legitimize_address, and all other variables should satisfy
7210 legitimate_pic_address_disp_p, we should never arrive here. */
7211 gcc_assert (!TARGET_64BIT_MS_ABI
);
7213 if (TARGET_64BIT
&& ix86_cmodel
!= CM_LARGE_PIC
)
7215 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTPCREL
);
7216 new = gen_rtx_CONST (Pmode
, new);
7217 new = gen_const_mem (Pmode
, new);
7218 set_mem_alias_set (new, ix86_GOT_alias_set ());
7221 reg
= gen_reg_rtx (Pmode
);
7222 /* Use directly gen_movsi, otherwise the address is loaded
7223 into register for CSE. We don't want to CSE this addresses,
7224 instead we CSE addresses from the GOT table, so skip this. */
7225 emit_insn (gen_movsi (reg
, new));
7230 /* This symbol must be referenced via a load from the
7231 Global Offset Table (@GOT). */
7233 if (reload_in_progress
)
7234 regs_ever_live
[PIC_OFFSET_TABLE_REGNUM
] = 1;
7235 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
7236 new = gen_rtx_CONST (Pmode
, new);
7238 new = force_reg (Pmode
, new);
7239 new = gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new);
7240 new = gen_const_mem (Pmode
, new);
7241 set_mem_alias_set (new, ix86_GOT_alias_set ());
7244 reg
= gen_reg_rtx (Pmode
);
7245 emit_move_insn (reg
, new);
7251 if (CONST_INT_P (addr
)
7252 && !x86_64_immediate_operand (addr
, VOIDmode
))
7256 emit_move_insn (reg
, addr
);
7260 new = force_reg (Pmode
, addr
);
7262 else if (GET_CODE (addr
) == CONST
)
7264 addr
= XEXP (addr
, 0);
7266 /* We must match stuff we generate before. Assume the only
7267 unspecs that can get here are ours. Not that we could do
7268 anything with them anyway.... */
7269 if (GET_CODE (addr
) == UNSPEC
7270 || (GET_CODE (addr
) == PLUS
7271 && GET_CODE (XEXP (addr
, 0)) == UNSPEC
))
7273 gcc_assert (GET_CODE (addr
) == PLUS
);
7275 if (GET_CODE (addr
) == PLUS
)
7277 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
7279 /* Check first to see if this is a constant offset from a @GOTOFF
7280 symbol reference. */
7281 if (gotoff_operand (op0
, Pmode
)
7282 && CONST_INT_P (op1
))
7286 if (reload_in_progress
)
7287 regs_ever_live
[PIC_OFFSET_TABLE_REGNUM
] = 1;
7288 new = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
7290 new = gen_rtx_PLUS (Pmode
, new, op1
);
7291 new = gen_rtx_CONST (Pmode
, new);
7292 new = gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new);
7296 emit_move_insn (reg
, new);
7302 if (INTVAL (op1
) < -16*1024*1024
7303 || INTVAL (op1
) >= 16*1024*1024)
7305 if (!x86_64_immediate_operand (op1
, Pmode
))
7306 op1
= force_reg (Pmode
, op1
);
7307 new = gen_rtx_PLUS (Pmode
, force_reg (Pmode
, op0
), op1
);
7313 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
7314 new = legitimize_pic_address (XEXP (addr
, 1),
7315 base
== reg
? NULL_RTX
: reg
);
7317 if (CONST_INT_P (new))
7318 new = plus_constant (base
, INTVAL (new));
7321 if (GET_CODE (new) == PLUS
&& CONSTANT_P (XEXP (new, 1)))
7323 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new, 0));
7324 new = XEXP (new, 1);
7326 new = gen_rtx_PLUS (Pmode
, base
, new);
7334 /* Load the thread pointer. If TO_REG is true, force it into a register. */
7337 get_thread_pointer (int to_reg
)
7341 tp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TP
);
7345 reg
= gen_reg_rtx (Pmode
);
7346 insn
= gen_rtx_SET (VOIDmode
, reg
, tp
);
7347 insn
= emit_insn (insn
);
7352 /* A subroutine of legitimize_address and ix86_expand_move. FOR_MOV is
7353 false if we expect this to be used for a memory address and true if
7354 we expect to load the address into a register. */
7357 legitimize_tls_address (rtx x
, enum tls_model model
, int for_mov
)
7359 rtx dest
, base
, off
, pic
, tp
;
7364 case TLS_MODEL_GLOBAL_DYNAMIC
:
7365 dest
= gen_reg_rtx (Pmode
);
7366 tp
= TARGET_GNU2_TLS
? get_thread_pointer (1) : 0;
7368 if (TARGET_64BIT
&& ! TARGET_GNU2_TLS
)
7370 rtx rax
= gen_rtx_REG (Pmode
, 0), insns
;
7373 emit_call_insn (gen_tls_global_dynamic_64 (rax
, x
));
7374 insns
= get_insns ();
7377 CONST_OR_PURE_CALL_P (insns
) = 1;
7378 emit_libcall_block (insns
, dest
, rax
, x
);
7380 else if (TARGET_64BIT
&& TARGET_GNU2_TLS
)
7381 emit_insn (gen_tls_global_dynamic_64 (dest
, x
));
7383 emit_insn (gen_tls_global_dynamic_32 (dest
, x
));
7385 if (TARGET_GNU2_TLS
)
7387 dest
= force_reg (Pmode
, gen_rtx_PLUS (Pmode
, tp
, dest
));
7389 set_unique_reg_note (get_last_insn (), REG_EQUIV
, x
);
7393 case TLS_MODEL_LOCAL_DYNAMIC
:
7394 base
= gen_reg_rtx (Pmode
);
7395 tp
= TARGET_GNU2_TLS
? get_thread_pointer (1) : 0;
7397 if (TARGET_64BIT
&& ! TARGET_GNU2_TLS
)
7399 rtx rax
= gen_rtx_REG (Pmode
, 0), insns
, note
;
7402 emit_call_insn (gen_tls_local_dynamic_base_64 (rax
));
7403 insns
= get_insns ();
7406 note
= gen_rtx_EXPR_LIST (VOIDmode
, const0_rtx
, NULL
);
7407 note
= gen_rtx_EXPR_LIST (VOIDmode
, ix86_tls_get_addr (), note
);
7408 CONST_OR_PURE_CALL_P (insns
) = 1;
7409 emit_libcall_block (insns
, base
, rax
, note
);
7411 else if (TARGET_64BIT
&& TARGET_GNU2_TLS
)
7412 emit_insn (gen_tls_local_dynamic_base_64 (base
));
7414 emit_insn (gen_tls_local_dynamic_base_32 (base
));
7416 if (TARGET_GNU2_TLS
)
7418 rtx x
= ix86_tls_module_base ();
7420 set_unique_reg_note (get_last_insn (), REG_EQUIV
,
7421 gen_rtx_MINUS (Pmode
, x
, tp
));
7424 off
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
), UNSPEC_DTPOFF
);
7425 off
= gen_rtx_CONST (Pmode
, off
);
7427 dest
= force_reg (Pmode
, gen_rtx_PLUS (Pmode
, base
, off
));
7429 if (TARGET_GNU2_TLS
)
7431 dest
= force_reg (Pmode
, gen_rtx_PLUS (Pmode
, dest
, tp
));
7433 set_unique_reg_note (get_last_insn (), REG_EQUIV
, x
);
7438 case TLS_MODEL_INITIAL_EXEC
:
7442 type
= UNSPEC_GOTNTPOFF
;
7446 if (reload_in_progress
)
7447 regs_ever_live
[PIC_OFFSET_TABLE_REGNUM
] = 1;
7448 pic
= pic_offset_table_rtx
;
7449 type
= TARGET_ANY_GNU_TLS
? UNSPEC_GOTNTPOFF
: UNSPEC_GOTTPOFF
;
7451 else if (!TARGET_ANY_GNU_TLS
)
7453 pic
= gen_reg_rtx (Pmode
);
7454 emit_insn (gen_set_got (pic
));
7455 type
= UNSPEC_GOTTPOFF
;
7460 type
= UNSPEC_INDNTPOFF
;
7463 off
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
), type
);
7464 off
= gen_rtx_CONST (Pmode
, off
);
7466 off
= gen_rtx_PLUS (Pmode
, pic
, off
);
7467 off
= gen_const_mem (Pmode
, off
);
7468 set_mem_alias_set (off
, ix86_GOT_alias_set ());
7470 if (TARGET_64BIT
|| TARGET_ANY_GNU_TLS
)
7472 base
= get_thread_pointer (for_mov
|| !TARGET_TLS_DIRECT_SEG_REFS
);
7473 off
= force_reg (Pmode
, off
);
7474 return gen_rtx_PLUS (Pmode
, base
, off
);
7478 base
= get_thread_pointer (true);
7479 dest
= gen_reg_rtx (Pmode
);
7480 emit_insn (gen_subsi3 (dest
, base
, off
));
7484 case TLS_MODEL_LOCAL_EXEC
:
7485 off
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, x
),
7486 (TARGET_64BIT
|| TARGET_ANY_GNU_TLS
)
7487 ? UNSPEC_NTPOFF
: UNSPEC_TPOFF
);
7488 off
= gen_rtx_CONST (Pmode
, off
);
7490 if (TARGET_64BIT
|| TARGET_ANY_GNU_TLS
)
7492 base
= get_thread_pointer (for_mov
|| !TARGET_TLS_DIRECT_SEG_REFS
);
7493 return gen_rtx_PLUS (Pmode
, base
, off
);
7497 base
= get_thread_pointer (true);
7498 dest
= gen_reg_rtx (Pmode
);
7499 emit_insn (gen_subsi3 (dest
, base
, off
));
7510 /* Create or return the unique __imp_DECL dllimport symbol corresponding
7513 static GTY((if_marked ("tree_map_marked_p"), param_is (struct tree_map
)))
7514 htab_t dllimport_map
;
7517 get_dllimport_decl (tree decl
)
7519 struct tree_map
*h
, in
;
7523 size_t namelen
, prefixlen
;
7529 dllimport_map
= htab_create_ggc (512, tree_map_hash
, tree_map_eq
, 0);
7531 in
.hash
= htab_hash_pointer (decl
);
7532 in
.base
.from
= decl
;
7533 loc
= htab_find_slot_with_hash (dllimport_map
, &in
, in
.hash
, INSERT
);
7538 *loc
= h
= ggc_alloc (sizeof (struct tree_map
));
7540 h
->base
.from
= decl
;
7541 h
->to
= to
= build_decl (VAR_DECL
, NULL
, ptr_type_node
);
7542 DECL_ARTIFICIAL (to
) = 1;
7543 DECL_IGNORED_P (to
) = 1;
7544 DECL_EXTERNAL (to
) = 1;
7545 TREE_READONLY (to
) = 1;
7547 name
= IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl
));
7548 name
= targetm
.strip_name_encoding (name
);
7549 if (name
[0] == FASTCALL_PREFIX
)
7555 prefix
= "*__imp__";
7557 namelen
= strlen (name
);
7558 prefixlen
= strlen (prefix
);
7559 imp_name
= alloca (namelen
+ prefixlen
+ 1);
7560 memcpy (imp_name
, prefix
, prefixlen
);
7561 memcpy (imp_name
+ prefixlen
, name
, namelen
+ 1);
7563 name
= ggc_alloc_string (imp_name
, namelen
+ prefixlen
);
7564 rtl
= gen_rtx_SYMBOL_REF (Pmode
, name
);
7565 SET_SYMBOL_REF_DECL (rtl
, to
);
7566 SYMBOL_REF_FLAGS (rtl
) = SYMBOL_FLAG_LOCAL
;
7568 rtl
= gen_const_mem (Pmode
, rtl
);
7569 set_mem_alias_set (rtl
, ix86_GOT_alias_set ());
7571 SET_DECL_RTL (to
, rtl
);
7576 /* Expand SYMBOL into its corresponding dllimport symbol. WANT_REG is
7577 true if we require the result be a register. */
7580 legitimize_dllimport_symbol (rtx symbol
, bool want_reg
)
7585 gcc_assert (SYMBOL_REF_DECL (symbol
));
7586 imp_decl
= get_dllimport_decl (SYMBOL_REF_DECL (symbol
));
7588 x
= DECL_RTL (imp_decl
);
7590 x
= force_reg (Pmode
, x
);
7594 /* Try machine-dependent ways of modifying an illegitimate address
7595 to be legitimate. If we find one, return the new, valid address.
7596 This macro is used in only one place: `memory_address' in explow.c.
7598 OLDX is the address as it was before break_out_memory_refs was called.
7599 In some cases it is useful to look at this to decide what needs to be done.
7601 MODE and WIN are passed so that this macro can use
7602 GO_IF_LEGITIMATE_ADDRESS.
7604 It is always safe for this macro to do nothing. It exists to recognize
7605 opportunities to optimize the output.
7607 For the 80386, we handle X+REG by loading X into a register R and
7608 using R+REG. R will go in a general reg and indexing will be used.
7609 However, if REG is a broken-out memory address or multiplication,
7610 nothing needs to be done because REG can certainly go in a general reg.
7612 When -fpic is used, special handling is needed for symbolic references.
7613 See comments by legitimize_pic_address in i386.c for details. */
7616 legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
, enum machine_mode mode
)
7621 log
= GET_CODE (x
) == SYMBOL_REF
? SYMBOL_REF_TLS_MODEL (x
) : 0;
7623 return legitimize_tls_address (x
, log
, false);
7624 if (GET_CODE (x
) == CONST
7625 && GET_CODE (XEXP (x
, 0)) == PLUS
7626 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
7627 && (log
= SYMBOL_REF_TLS_MODEL (XEXP (XEXP (x
, 0), 0))))
7629 rtx t
= legitimize_tls_address (XEXP (XEXP (x
, 0), 0), log
, false);
7630 return gen_rtx_PLUS (Pmode
, t
, XEXP (XEXP (x
, 0), 1));
7633 if (flag_pic
&& SYMBOLIC_CONST (x
))
7634 return legitimize_pic_address (x
, 0);
7636 if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
)
7638 if (GET_CODE (x
) == SYMBOL_REF
&& SYMBOL_REF_DLLIMPORT_P (x
))
7639 return legitimize_dllimport_symbol (x
, true);
7640 if (GET_CODE (x
) == CONST
7641 && GET_CODE (XEXP (x
, 0)) == PLUS
7642 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
7643 && SYMBOL_REF_DLLIMPORT_P (XEXP (XEXP (x
, 0), 0)))
7645 rtx t
= legitimize_dllimport_symbol (XEXP (XEXP (x
, 0), 0), true);
7646 return gen_rtx_PLUS (Pmode
, t
, XEXP (XEXP (x
, 0), 1));
7650 /* Canonicalize shifts by 0, 1, 2, 3 into multiply */
7651 if (GET_CODE (x
) == ASHIFT
7652 && CONST_INT_P (XEXP (x
, 1))
7653 && (unsigned HOST_WIDE_INT
) INTVAL (XEXP (x
, 1)) < 4)
7656 log
= INTVAL (XEXP (x
, 1));
7657 x
= gen_rtx_MULT (Pmode
, force_reg (Pmode
, XEXP (x
, 0)),
7658 GEN_INT (1 << log
));
7661 if (GET_CODE (x
) == PLUS
)
7663 /* Canonicalize shifts by 0, 1, 2, 3 into multiply. */
7665 if (GET_CODE (XEXP (x
, 0)) == ASHIFT
7666 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
7667 && (unsigned HOST_WIDE_INT
) INTVAL (XEXP (XEXP (x
, 0), 1)) < 4)
7670 log
= INTVAL (XEXP (XEXP (x
, 0), 1));
7671 XEXP (x
, 0) = gen_rtx_MULT (Pmode
,
7672 force_reg (Pmode
, XEXP (XEXP (x
, 0), 0)),
7673 GEN_INT (1 << log
));
7676 if (GET_CODE (XEXP (x
, 1)) == ASHIFT
7677 && CONST_INT_P (XEXP (XEXP (x
, 1), 1))
7678 && (unsigned HOST_WIDE_INT
) INTVAL (XEXP (XEXP (x
, 1), 1)) < 4)
7681 log
= INTVAL (XEXP (XEXP (x
, 1), 1));
7682 XEXP (x
, 1) = gen_rtx_MULT (Pmode
,
7683 force_reg (Pmode
, XEXP (XEXP (x
, 1), 0)),
7684 GEN_INT (1 << log
));
7687 /* Put multiply first if it isn't already. */
7688 if (GET_CODE (XEXP (x
, 1)) == MULT
)
7690 rtx tmp
= XEXP (x
, 0);
7691 XEXP (x
, 0) = XEXP (x
, 1);
7696 /* Canonicalize (plus (mult (reg) (const)) (plus (reg) (const)))
7697 into (plus (plus (mult (reg) (const)) (reg)) (const)). This can be
7698 created by virtual register instantiation, register elimination, and
7699 similar optimizations. */
7700 if (GET_CODE (XEXP (x
, 0)) == MULT
&& GET_CODE (XEXP (x
, 1)) == PLUS
)
7703 x
= gen_rtx_PLUS (Pmode
,
7704 gen_rtx_PLUS (Pmode
, XEXP (x
, 0),
7705 XEXP (XEXP (x
, 1), 0)),
7706 XEXP (XEXP (x
, 1), 1));
7710 (plus (plus (mult (reg) (const)) (plus (reg) (const))) const)
7711 into (plus (plus (mult (reg) (const)) (reg)) (const)). */
7712 else if (GET_CODE (x
) == PLUS
&& GET_CODE (XEXP (x
, 0)) == PLUS
7713 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
7714 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == PLUS
7715 && CONSTANT_P (XEXP (x
, 1)))
7718 rtx other
= NULL_RTX
;
7720 if (CONST_INT_P (XEXP (x
, 1)))
7722 constant
= XEXP (x
, 1);
7723 other
= XEXP (XEXP (XEXP (x
, 0), 1), 1);
7725 else if (CONST_INT_P (XEXP (XEXP (XEXP (x
, 0), 1), 1)))
7727 constant
= XEXP (XEXP (XEXP (x
, 0), 1), 1);
7728 other
= XEXP (x
, 1);
7736 x
= gen_rtx_PLUS (Pmode
,
7737 gen_rtx_PLUS (Pmode
, XEXP (XEXP (x
, 0), 0),
7738 XEXP (XEXP (XEXP (x
, 0), 1), 0)),
7739 plus_constant (other
, INTVAL (constant
)));
7743 if (changed
&& legitimate_address_p (mode
, x
, FALSE
))
7746 if (GET_CODE (XEXP (x
, 0)) == MULT
)
7749 XEXP (x
, 0) = force_operand (XEXP (x
, 0), 0);
7752 if (GET_CODE (XEXP (x
, 1)) == MULT
)
7755 XEXP (x
, 1) = force_operand (XEXP (x
, 1), 0);
7759 && REG_P (XEXP (x
, 1))
7760 && REG_P (XEXP (x
, 0)))
7763 if (flag_pic
&& SYMBOLIC_CONST (XEXP (x
, 1)))
7766 x
= legitimize_pic_address (x
, 0);
7769 if (changed
&& legitimate_address_p (mode
, x
, FALSE
))
7772 if (REG_P (XEXP (x
, 0)))
7774 rtx temp
= gen_reg_rtx (Pmode
);
7775 rtx val
= force_operand (XEXP (x
, 1), temp
);
7777 emit_move_insn (temp
, val
);
7783 else if (REG_P (XEXP (x
, 1)))
7785 rtx temp
= gen_reg_rtx (Pmode
);
7786 rtx val
= force_operand (XEXP (x
, 0), temp
);
7788 emit_move_insn (temp
, val
);
7798 /* Print an integer constant expression in assembler syntax. Addition
7799 and subtraction are the only arithmetic that may appear in these
7800 expressions. FILE is the stdio stream to write to, X is the rtx, and
7801 CODE is the operand print code from the output string. */
7804 output_pic_addr_const (FILE *file
, rtx x
, int code
)
7808 switch (GET_CODE (x
))
7811 gcc_assert (flag_pic
);
7816 if (! TARGET_MACHO
|| TARGET_64BIT
)
7817 output_addr_const (file
, x
);
7820 const char *name
= XSTR (x
, 0);
7822 /* Mark the decl as referenced so that cgraph will
7823 output the function. */
7824 if (SYMBOL_REF_DECL (x
))
7825 mark_decl_referenced (SYMBOL_REF_DECL (x
));
7828 if (MACHOPIC_INDIRECT
7829 && machopic_classify_symbol (x
) == MACHOPIC_UNDEFINED_FUNCTION
)
7830 name
= machopic_indirection_name (x
, /*stub_p=*/true);
7832 assemble_name (file
, name
);
7834 if (!TARGET_MACHO
&& !TARGET_64BIT_MS_ABI
7835 && code
== 'P' && ! SYMBOL_REF_LOCAL_P (x
))
7836 fputs ("@PLT", file
);
7843 ASM_GENERATE_INTERNAL_LABEL (buf
, "L", CODE_LABEL_NUMBER (x
));
7844 assemble_name (asm_out_file
, buf
);
7848 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
7852 /* This used to output parentheses around the expression,
7853 but that does not work on the 386 (either ATT or BSD assembler). */
7854 output_pic_addr_const (file
, XEXP (x
, 0), code
);
7858 if (GET_MODE (x
) == VOIDmode
)
7860 /* We can use %d if the number is <32 bits and positive. */
7861 if (CONST_DOUBLE_HIGH (x
) || CONST_DOUBLE_LOW (x
) < 0)
7862 fprintf (file
, "0x%lx%08lx",
7863 (unsigned long) CONST_DOUBLE_HIGH (x
),
7864 (unsigned long) CONST_DOUBLE_LOW (x
));
7866 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
));
7869 /* We can't handle floating point constants;
7870 PRINT_OPERAND must handle them. */
7871 output_operand_lossage ("floating constant misused");
7875 /* Some assemblers need integer constants to appear first. */
7876 if (CONST_INT_P (XEXP (x
, 0)))
7878 output_pic_addr_const (file
, XEXP (x
, 0), code
);
7880 output_pic_addr_const (file
, XEXP (x
, 1), code
);
7884 gcc_assert (CONST_INT_P (XEXP (x
, 1)));
7885 output_pic_addr_const (file
, XEXP (x
, 1), code
);
7887 output_pic_addr_const (file
, XEXP (x
, 0), code
);
7893 putc (ASSEMBLER_DIALECT
== ASM_INTEL
? '(' : '[', file
);
7894 output_pic_addr_const (file
, XEXP (x
, 0), code
);
7896 output_pic_addr_const (file
, XEXP (x
, 1), code
);
7898 putc (ASSEMBLER_DIALECT
== ASM_INTEL
? ')' : ']', file
);
7902 gcc_assert (XVECLEN (x
, 0) == 1);
7903 output_pic_addr_const (file
, XVECEXP (x
, 0, 0), code
);
7904 switch (XINT (x
, 1))
7907 fputs ("@GOT", file
);
7910 fputs ("@GOTOFF", file
);
7913 fputs ("@PLTOFF", file
);
7915 case UNSPEC_GOTPCREL
:
7916 fputs ("@GOTPCREL(%rip)", file
);
7918 case UNSPEC_GOTTPOFF
:
7919 /* FIXME: This might be @TPOFF in Sun ld too. */
7920 fputs ("@GOTTPOFF", file
);
7923 fputs ("@TPOFF", file
);
7927 fputs ("@TPOFF", file
);
7929 fputs ("@NTPOFF", file
);
7932 fputs ("@DTPOFF", file
);
7934 case UNSPEC_GOTNTPOFF
:
7936 fputs ("@GOTTPOFF(%rip)", file
);
7938 fputs ("@GOTNTPOFF", file
);
7940 case UNSPEC_INDNTPOFF
:
7941 fputs ("@INDNTPOFF", file
);
7944 output_operand_lossage ("invalid UNSPEC as operand");
7950 output_operand_lossage ("invalid expression as operand");
7954 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
7955 We need to emit DTP-relative relocations. */
7957 static void ATTRIBUTE_UNUSED
7958 i386_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
7960 fputs (ASM_LONG
, file
);
7961 output_addr_const (file
, x
);
7962 fputs ("@DTPOFF", file
);
7968 fputs (", 0", file
);
7975 /* In the name of slightly smaller debug output, and to cater to
7976 general assembler lossage, recognize PIC+GOTOFF and turn it back
7977 into a direct symbol reference.
7979 On Darwin, this is necessary to avoid a crash, because Darwin
7980 has a different PIC label for each routine but the DWARF debugging
7981 information is not associated with any particular routine, so it's
7982 necessary to remove references to the PIC label from RTL stored by
7983 the DWARF output code. */
7986 ix86_delegitimize_address (rtx orig_x
)
7989 /* reg_addend is NULL or a multiple of some register. */
7990 rtx reg_addend
= NULL_RTX
;
7991 /* const_addend is NULL or a const_int. */
7992 rtx const_addend
= NULL_RTX
;
7993 /* This is the result, or NULL. */
7994 rtx result
= NULL_RTX
;
8001 if (GET_CODE (x
) != CONST
8002 || GET_CODE (XEXP (x
, 0)) != UNSPEC
8003 || XINT (XEXP (x
, 0), 1) != UNSPEC_GOTPCREL
8006 return XVECEXP (XEXP (x
, 0), 0, 0);
8009 if (GET_CODE (x
) != PLUS
8010 || GET_CODE (XEXP (x
, 1)) != CONST
)
8013 if (REG_P (XEXP (x
, 0))
8014 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
8015 /* %ebx + GOT/GOTOFF */
8017 else if (GET_CODE (XEXP (x
, 0)) == PLUS
)
8019 /* %ebx + %reg * scale + GOT/GOTOFF */
8020 reg_addend
= XEXP (x
, 0);
8021 if (REG_P (XEXP (reg_addend
, 0))
8022 && REGNO (XEXP (reg_addend
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
8023 reg_addend
= XEXP (reg_addend
, 1);
8024 else if (REG_P (XEXP (reg_addend
, 1))
8025 && REGNO (XEXP (reg_addend
, 1)) == PIC_OFFSET_TABLE_REGNUM
)
8026 reg_addend
= XEXP (reg_addend
, 0);
8029 if (!REG_P (reg_addend
)
8030 && GET_CODE (reg_addend
) != MULT
8031 && GET_CODE (reg_addend
) != ASHIFT
)
8037 x
= XEXP (XEXP (x
, 1), 0);
8038 if (GET_CODE (x
) == PLUS
8039 && CONST_INT_P (XEXP (x
, 1)))
8041 const_addend
= XEXP (x
, 1);
8045 if (GET_CODE (x
) == UNSPEC
8046 && ((XINT (x
, 1) == UNSPEC_GOT
&& MEM_P (orig_x
))
8047 || (XINT (x
, 1) == UNSPEC_GOTOFF
&& !MEM_P (orig_x
))))
8048 result
= XVECEXP (x
, 0, 0);
8050 if (TARGET_MACHO
&& darwin_local_data_pic (x
)
8052 result
= XEXP (x
, 0);
8058 result
= gen_rtx_PLUS (Pmode
, result
, const_addend
);
8060 result
= gen_rtx_PLUS (Pmode
, reg_addend
, result
);
8064 /* If X is a machine specific address (i.e. a symbol or label being
8065 referenced as a displacement from the GOT implemented using an
8066 UNSPEC), then return the base term. Otherwise return X. */
8069 ix86_find_base_term (rtx x
)
8075 if (GET_CODE (x
) != CONST
)
8078 if (GET_CODE (term
) == PLUS
8079 && (CONST_INT_P (XEXP (term
, 1))
8080 || GET_CODE (XEXP (term
, 1)) == CONST_DOUBLE
))
8081 term
= XEXP (term
, 0);
8082 if (GET_CODE (term
) != UNSPEC
8083 || XINT (term
, 1) != UNSPEC_GOTPCREL
)
8086 term
= XVECEXP (term
, 0, 0);
8088 if (GET_CODE (term
) != SYMBOL_REF
8089 && GET_CODE (term
) != LABEL_REF
)
8095 term
= ix86_delegitimize_address (x
);
8097 if (GET_CODE (term
) != SYMBOL_REF
8098 && GET_CODE (term
) != LABEL_REF
)
8105 put_condition_code (enum rtx_code code
, enum machine_mode mode
, int reverse
,
8110 if (mode
== CCFPmode
|| mode
== CCFPUmode
)
8112 enum rtx_code second_code
, bypass_code
;
8113 ix86_fp_comparison_codes (code
, &bypass_code
, &code
, &second_code
);
8114 gcc_assert (bypass_code
== UNKNOWN
&& second_code
== UNKNOWN
);
8115 code
= ix86_fp_compare_code_to_integer (code
);
8119 code
= reverse_condition (code
);
8130 gcc_assert (mode
== CCmode
|| mode
== CCNOmode
|| mode
== CCGCmode
);
8134 /* ??? Use "nbe" instead of "a" for fcmov lossage on some assemblers.
8135 Those same assemblers have the same but opposite lossage on cmov. */
8136 gcc_assert (mode
== CCmode
);
8137 suffix
= fp
? "nbe" : "a";
8157 gcc_assert (mode
== CCmode
);
8179 gcc_assert (mode
== CCmode
);
8180 suffix
= fp
? "nb" : "ae";
8183 gcc_assert (mode
== CCmode
|| mode
== CCGCmode
|| mode
== CCNOmode
);
8187 gcc_assert (mode
== CCmode
);
8191 suffix
= fp
? "u" : "p";
8194 suffix
= fp
? "nu" : "np";
8199 fputs (suffix
, file
);
8202 /* Print the name of register X to FILE based on its machine mode and number.
8203 If CODE is 'w', pretend the mode is HImode.
8204 If CODE is 'b', pretend the mode is QImode.
8205 If CODE is 'k', pretend the mode is SImode.
8206 If CODE is 'q', pretend the mode is DImode.
8207 If CODE is 'h', pretend the reg is the 'high' byte register.
8208 If CODE is 'y', print "st(0)" instead of "st", if the reg is stack op. */
8211 print_reg (rtx x
, int code
, FILE *file
)
8213 gcc_assert (REGNO (x
) != ARG_POINTER_REGNUM
8214 && REGNO (x
) != FRAME_POINTER_REGNUM
8215 && REGNO (x
) != FLAGS_REG
8216 && REGNO (x
) != FPSR_REG
8217 && REGNO (x
) != FPCR_REG
);
8219 if (ASSEMBLER_DIALECT
== ASM_ATT
|| USER_LABEL_PREFIX
[0] == 0)
8222 if (code
== 'w' || MMX_REG_P (x
))
8224 else if (code
== 'b')
8226 else if (code
== 'k')
8228 else if (code
== 'q')
8230 else if (code
== 'y')
8232 else if (code
== 'h')
8235 code
= GET_MODE_SIZE (GET_MODE (x
));
8237 /* Irritatingly, AMD extended registers use different naming convention
8238 from the normal registers. */
8239 if (REX_INT_REG_P (x
))
8241 gcc_assert (TARGET_64BIT
);
8245 error ("extended registers have no high halves");
8248 fprintf (file
, "r%ib", REGNO (x
) - FIRST_REX_INT_REG
+ 8);
8251 fprintf (file
, "r%iw", REGNO (x
) - FIRST_REX_INT_REG
+ 8);
8254 fprintf (file
, "r%id", REGNO (x
) - FIRST_REX_INT_REG
+ 8);
8257 fprintf (file
, "r%i", REGNO (x
) - FIRST_REX_INT_REG
+ 8);
8260 error ("unsupported operand size for extended register");
8268 if (STACK_TOP_P (x
))
8270 fputs ("st(0)", file
);
8277 if (! ANY_FP_REG_P (x
))
8278 putc (code
== 8 && TARGET_64BIT
? 'r' : 'e', file
);
8283 fputs (hi_reg_name
[REGNO (x
)], file
);
8286 if (REGNO (x
) >= ARRAY_SIZE (qi_reg_name
))
8288 fputs (qi_reg_name
[REGNO (x
)], file
);
8291 if (REGNO (x
) >= ARRAY_SIZE (qi_high_reg_name
))
8293 fputs (qi_high_reg_name
[REGNO (x
)], file
);
8300 /* Locate some local-dynamic symbol still in use by this function
8301 so that we can print its name in some tls_local_dynamic_base
8305 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
8309 if (GET_CODE (x
) == SYMBOL_REF
8310 && SYMBOL_REF_TLS_MODEL (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
8312 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
8320 get_some_local_dynamic_name (void)
8324 if (cfun
->machine
->some_ld_name
)
8325 return cfun
->machine
->some_ld_name
;
8327 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
8329 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
8330 return cfun
->machine
->some_ld_name
;
8336 L,W,B,Q,S,T -- print the opcode suffix for specified size of operand.
8337 C -- print opcode suffix for set/cmov insn.
8338 c -- like C, but print reversed condition
8339 F,f -- likewise, but for floating-point.
8340 O -- if HAVE_AS_IX86_CMOV_SUN_SYNTAX, expand to "w.", "l." or "q.",
8342 R -- print the prefix for register names.
8343 z -- print the opcode suffix for the size of the current operand.
8344 * -- print a star (in certain assembler syntax)
8345 A -- print an absolute memory reference.
8346 w -- print the operand as if it's a "word" (HImode) even if it isn't.
8347 s -- print a shift double count, followed by the assemblers argument
8349 b -- print the QImode name of the register for the indicated operand.
8350 %b0 would print %al if operands[0] is reg 0.
8351 w -- likewise, print the HImode name of the register.
8352 k -- likewise, print the SImode name of the register.
8353 q -- likewise, print the DImode name of the register.
8354 h -- print the QImode name for a "high" register, either ah, bh, ch or dh.
8355 y -- print "st(0)" instead of "st" as a register.
8356 D -- print condition for SSE cmp instruction.
8357 P -- if PIC, print an @PLT suffix.
8358 X -- don't print any sort of PIC '@' suffix for a symbol.
8359 & -- print some in-use local-dynamic symbol name.
8360 H -- print a memory address offset by 8; used for sse high-parts
8364 print_operand (FILE *file
, rtx x
, int code
)
8371 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8376 assemble_name (file
, get_some_local_dynamic_name ());
8380 switch (ASSEMBLER_DIALECT
)
8387 /* Intel syntax. For absolute addresses, registers should not
8388 be surrounded by braces. */
8392 PRINT_OPERAND (file
, x
, 0);
8402 PRINT_OPERAND (file
, x
, 0);
8407 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8412 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8417 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8422 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8427 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8432 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8437 /* 387 opcodes don't get size suffixes if the operands are
8439 if (STACK_REG_P (x
))
8442 /* Likewise if using Intel opcodes. */
8443 if (ASSEMBLER_DIALECT
== ASM_INTEL
)
8446 /* This is the size of op from size of operand. */
8447 switch (GET_MODE_SIZE (GET_MODE (x
)))
8456 #ifdef HAVE_GAS_FILDS_FISTS
8466 if (GET_MODE (x
) == SFmode
)
8481 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_INT
)
8483 #ifdef GAS_MNEMONICS
8509 if (CONST_INT_P (x
) || ! SHIFT_DOUBLE_OMITS_COUNT
)
8511 PRINT_OPERAND (file
, x
, 0);
8517 /* Little bit of braindamage here. The SSE compare instructions
8518 does use completely different names for the comparisons that the
8519 fp conditional moves. */
8520 switch (GET_CODE (x
))
8535 fputs ("unord", file
);
8539 fputs ("neq", file
);
8543 fputs ("nlt", file
);
8547 fputs ("nle", file
);
8550 fputs ("ord", file
);
8557 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
8558 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8560 switch (GET_MODE (x
))
8562 case HImode
: putc ('w', file
); break;
8564 case SFmode
: putc ('l', file
); break;
8566 case DFmode
: putc ('q', file
); break;
8567 default: gcc_unreachable ();
8574 put_condition_code (GET_CODE (x
), GET_MODE (XEXP (x
, 0)), 0, 0, file
);
8577 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
8578 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8581 put_condition_code (GET_CODE (x
), GET_MODE (XEXP (x
, 0)), 0, 1, file
);
8584 /* Like above, but reverse condition */
8586 /* Check to see if argument to %c is really a constant
8587 and not a condition code which needs to be reversed. */
8588 if (!COMPARISON_P (x
))
8590 output_operand_lossage ("operand is neither a constant nor a condition code, invalid operand code 'c'");
8593 put_condition_code (GET_CODE (x
), GET_MODE (XEXP (x
, 0)), 1, 0, file
);
8596 #ifdef HAVE_AS_IX86_CMOV_SUN_SYNTAX
8597 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8600 put_condition_code (GET_CODE (x
), GET_MODE (XEXP (x
, 0)), 1, 1, file
);
8604 /* It doesn't actually matter what mode we use here, as we're
8605 only going to use this for printing. */
8606 x
= adjust_address_nv (x
, DImode
, 8);
8613 if (!optimize
|| optimize_size
|| !TARGET_BRANCH_PREDICTION_HINTS
)
8616 x
= find_reg_note (current_output_insn
, REG_BR_PROB
, 0);
8619 int pred_val
= INTVAL (XEXP (x
, 0));
8621 if (pred_val
< REG_BR_PROB_BASE
* 45 / 100
8622 || pred_val
> REG_BR_PROB_BASE
* 55 / 100)
8624 int taken
= pred_val
> REG_BR_PROB_BASE
/ 2;
8625 int cputaken
= final_forward_branch_p (current_output_insn
) == 0;
8627 /* Emit hints only in the case default branch prediction
8628 heuristics would fail. */
8629 if (taken
!= cputaken
)
8631 /* We use 3e (DS) prefix for taken branches and
8632 2e (CS) prefix for not taken branches. */
8634 fputs ("ds ; ", file
);
8636 fputs ("cs ; ", file
);
8643 output_operand_lossage ("invalid operand code '%c'", code
);
8648 print_reg (x
, code
, file
);
8652 /* No `byte ptr' prefix for call instructions. */
8653 if (ASSEMBLER_DIALECT
== ASM_INTEL
&& code
!= 'X' && code
!= 'P')
8656 switch (GET_MODE_SIZE (GET_MODE (x
)))
8658 case 1: size
= "BYTE"; break;
8659 case 2: size
= "WORD"; break;
8660 case 4: size
= "DWORD"; break;
8661 case 8: size
= "QWORD"; break;
8662 case 12: size
= "XWORD"; break;
8663 case 16: size
= "XMMWORD"; break;
8668 /* Check for explicit size override (codes 'b', 'w' and 'k') */
8671 else if (code
== 'w')
8673 else if (code
== 'k')
8677 fputs (" PTR ", file
);
8681 /* Avoid (%rip) for call operands. */
8682 if (CONSTANT_ADDRESS_P (x
) && code
== 'P'
8683 && !CONST_INT_P (x
))
8684 output_addr_const (file
, x
);
8685 else if (this_is_asm_operands
&& ! address_operand (x
, VOIDmode
))
8686 output_operand_lossage ("invalid constraints for operand");
8691 else if (GET_CODE (x
) == CONST_DOUBLE
&& GET_MODE (x
) == SFmode
)
8696 REAL_VALUE_FROM_CONST_DOUBLE (r
, x
);
8697 REAL_VALUE_TO_TARGET_SINGLE (r
, l
);
8699 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8701 fprintf (file
, "0x%08lx", l
);
8704 /* These float cases don't actually occur as immediate operands. */
8705 else if (GET_CODE (x
) == CONST_DOUBLE
&& GET_MODE (x
) == DFmode
)
8709 real_to_decimal (dstr
, CONST_DOUBLE_REAL_VALUE (x
), sizeof (dstr
), 0, 1);
8710 fprintf (file
, "%s", dstr
);
8713 else if (GET_CODE (x
) == CONST_DOUBLE
8714 && GET_MODE (x
) == XFmode
)
8718 real_to_decimal (dstr
, CONST_DOUBLE_REAL_VALUE (x
), sizeof (dstr
), 0, 1);
8719 fprintf (file
, "%s", dstr
);
8724 /* We have patterns that allow zero sets of memory, for instance.
8725 In 64-bit mode, we should probably support all 8-byte vectors,
8726 since we can in fact encode that into an immediate. */
8727 if (GET_CODE (x
) == CONST_VECTOR
)
8729 gcc_assert (x
== CONST0_RTX (GET_MODE (x
)));
8735 if (CONST_INT_P (x
) || GET_CODE (x
) == CONST_DOUBLE
)
8737 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8740 else if (GET_CODE (x
) == CONST
|| GET_CODE (x
) == SYMBOL_REF
8741 || GET_CODE (x
) == LABEL_REF
)
8743 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8746 fputs ("OFFSET FLAT:", file
);
8749 if (CONST_INT_P (x
))
8750 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
8752 output_pic_addr_const (file
, x
, code
);
8754 output_addr_const (file
, x
);
8758 /* Print a memory operand whose address is ADDR. */
8761 print_operand_address (FILE *file
, rtx addr
)
8763 struct ix86_address parts
;
8764 rtx base
, index
, disp
;
8766 int ok
= ix86_decompose_address (addr
, &parts
);
8771 index
= parts
.index
;
8773 scale
= parts
.scale
;
8781 if (USER_LABEL_PREFIX
[0] == 0)
8783 fputs ((parts
.seg
== SEG_FS
? "fs:" : "gs:"), file
);
8789 if (!base
&& !index
)
8791 /* Displacement only requires special attention. */
8793 if (CONST_INT_P (disp
))
8795 if (ASSEMBLER_DIALECT
== ASM_INTEL
&& parts
.seg
== SEG_DEFAULT
)
8797 if (USER_LABEL_PREFIX
[0] == 0)
8799 fputs ("ds:", file
);
8801 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (disp
));
8804 output_pic_addr_const (file
, disp
, 0);
8806 output_addr_const (file
, disp
);
8808 /* Use one byte shorter RIP relative addressing for 64bit mode. */
8811 if (GET_CODE (disp
) == CONST
8812 && GET_CODE (XEXP (disp
, 0)) == PLUS
8813 && CONST_INT_P (XEXP (XEXP (disp
, 0), 1)))
8814 disp
= XEXP (XEXP (disp
, 0), 0);
8815 if (GET_CODE (disp
) == LABEL_REF
8816 || (GET_CODE (disp
) == SYMBOL_REF
8817 && SYMBOL_REF_TLS_MODEL (disp
) == 0))
8818 fputs ("(%rip)", file
);
8823 if (ASSEMBLER_DIALECT
== ASM_ATT
)
8828 output_pic_addr_const (file
, disp
, 0);
8829 else if (GET_CODE (disp
) == LABEL_REF
)
8830 output_asm_label (disp
);
8832 output_addr_const (file
, disp
);
8837 print_reg (base
, 0, file
);
8841 print_reg (index
, 0, file
);
8843 fprintf (file
, ",%d", scale
);
8849 rtx offset
= NULL_RTX
;
8853 /* Pull out the offset of a symbol; print any symbol itself. */
8854 if (GET_CODE (disp
) == CONST
8855 && GET_CODE (XEXP (disp
, 0)) == PLUS
8856 && CONST_INT_P (XEXP (XEXP (disp
, 0), 1)))
8858 offset
= XEXP (XEXP (disp
, 0), 1);
8859 disp
= gen_rtx_CONST (VOIDmode
,
8860 XEXP (XEXP (disp
, 0), 0));
8864 output_pic_addr_const (file
, disp
, 0);
8865 else if (GET_CODE (disp
) == LABEL_REF
)
8866 output_asm_label (disp
);
8867 else if (CONST_INT_P (disp
))
8870 output_addr_const (file
, disp
);
8876 print_reg (base
, 0, file
);
8879 if (INTVAL (offset
) >= 0)
8881 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (offset
));
8885 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (offset
));
8892 print_reg (index
, 0, file
);
8894 fprintf (file
, "*%d", scale
);
8902 output_addr_const_extra (FILE *file
, rtx x
)
8906 if (GET_CODE (x
) != UNSPEC
)
8909 op
= XVECEXP (x
, 0, 0);
8910 switch (XINT (x
, 1))
8912 case UNSPEC_GOTTPOFF
:
8913 output_addr_const (file
, op
);
8914 /* FIXME: This might be @TPOFF in Sun ld. */
8915 fputs ("@GOTTPOFF", file
);
8918 output_addr_const (file
, op
);
8919 fputs ("@TPOFF", file
);
8922 output_addr_const (file
, op
);
8924 fputs ("@TPOFF", file
);
8926 fputs ("@NTPOFF", file
);
8929 output_addr_const (file
, op
);
8930 fputs ("@DTPOFF", file
);
8932 case UNSPEC_GOTNTPOFF
:
8933 output_addr_const (file
, op
);
8935 fputs ("@GOTTPOFF(%rip)", file
);
8937 fputs ("@GOTNTPOFF", file
);
8939 case UNSPEC_INDNTPOFF
:
8940 output_addr_const (file
, op
);
8941 fputs ("@INDNTPOFF", file
);
8951 /* Split one or more DImode RTL references into pairs of SImode
8952 references. The RTL can be REG, offsettable MEM, integer constant, or
8953 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
8954 split and "num" is its length. lo_half and hi_half are output arrays
8955 that parallel "operands". */
8958 split_di (rtx operands
[], int num
, rtx lo_half
[], rtx hi_half
[])
8962 rtx op
= operands
[num
];
8964 /* simplify_subreg refuse to split volatile memory addresses,
8965 but we still have to handle it. */
8968 lo_half
[num
] = adjust_address (op
, SImode
, 0);
8969 hi_half
[num
] = adjust_address (op
, SImode
, 4);
8973 lo_half
[num
] = simplify_gen_subreg (SImode
, op
,
8974 GET_MODE (op
) == VOIDmode
8975 ? DImode
: GET_MODE (op
), 0);
8976 hi_half
[num
] = simplify_gen_subreg (SImode
, op
,
8977 GET_MODE (op
) == VOIDmode
8978 ? DImode
: GET_MODE (op
), 4);
8982 /* Split one or more TImode RTL references into pairs of DImode
8983 references. The RTL can be REG, offsettable MEM, integer constant, or
8984 CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
8985 split and "num" is its length. lo_half and hi_half are output arrays
8986 that parallel "operands". */
8989 split_ti (rtx operands
[], int num
, rtx lo_half
[], rtx hi_half
[])
8993 rtx op
= operands
[num
];
8995 /* simplify_subreg refuse to split volatile memory addresses, but we
8996 still have to handle it. */
8999 lo_half
[num
] = adjust_address (op
, DImode
, 0);
9000 hi_half
[num
] = adjust_address (op
, DImode
, 8);
9004 lo_half
[num
] = simplify_gen_subreg (DImode
, op
, TImode
, 0);
9005 hi_half
[num
] = simplify_gen_subreg (DImode
, op
, TImode
, 8);
9010 /* Output code to perform a 387 binary operation in INSN, one of PLUS,
9011 MINUS, MULT or DIV. OPERANDS are the insn operands, where operands[3]
9012 is the expression of the binary operation. The output may either be
9013 emitted here, or returned to the caller, like all output_* functions.
9015 There is no guarantee that the operands are the same mode, as they
9016 might be within FLOAT or FLOAT_EXTEND expressions. */
9018 #ifndef SYSV386_COMPAT
9019 /* Set to 1 for compatibility with brain-damaged assemblers. No-one
9020 wants to fix the assemblers because that causes incompatibility
9021 with gcc. No-one wants to fix gcc because that causes
9022 incompatibility with assemblers... You can use the option of
9023 -DSYSV386_COMPAT=0 if you recompile both gcc and gas this way. */
9024 #define SYSV386_COMPAT 1
9028 output_387_binary_op (rtx insn
, rtx
*operands
)
9030 static char buf
[30];
9033 int is_sse
= SSE_REG_P (operands
[0]) || SSE_REG_P (operands
[1]) || SSE_REG_P (operands
[2]);
9035 #ifdef ENABLE_CHECKING
9036 /* Even if we do not want to check the inputs, this documents input
9037 constraints. Which helps in understanding the following code. */
9038 if (STACK_REG_P (operands
[0])
9039 && ((REG_P (operands
[1])
9040 && REGNO (operands
[0]) == REGNO (operands
[1])
9041 && (STACK_REG_P (operands
[2]) || MEM_P (operands
[2])))
9042 || (REG_P (operands
[2])
9043 && REGNO (operands
[0]) == REGNO (operands
[2])
9044 && (STACK_REG_P (operands
[1]) || MEM_P (operands
[1]))))
9045 && (STACK_TOP_P (operands
[1]) || STACK_TOP_P (operands
[2])))
9048 gcc_assert (is_sse
);
9051 switch (GET_CODE (operands
[3]))
9054 if (GET_MODE_CLASS (GET_MODE (operands
[1])) == MODE_INT
9055 || GET_MODE_CLASS (GET_MODE (operands
[2])) == MODE_INT
)
9063 if (GET_MODE_CLASS (GET_MODE (operands
[1])) == MODE_INT
9064 || GET_MODE_CLASS (GET_MODE (operands
[2])) == MODE_INT
)
9072 if (GET_MODE_CLASS (GET_MODE (operands
[1])) == MODE_INT
9073 || GET_MODE_CLASS (GET_MODE (operands
[2])) == MODE_INT
)
9081 if (GET_MODE_CLASS (GET_MODE (operands
[1])) == MODE_INT
9082 || GET_MODE_CLASS (GET_MODE (operands
[2])) == MODE_INT
)
9096 if (GET_MODE (operands
[0]) == SFmode
)
9097 strcat (buf
, "ss\t{%2, %0|%0, %2}");
9099 strcat (buf
, "sd\t{%2, %0|%0, %2}");
9104 switch (GET_CODE (operands
[3]))
9108 if (REG_P (operands
[2]) && REGNO (operands
[0]) == REGNO (operands
[2]))
9110 rtx temp
= operands
[2];
9111 operands
[2] = operands
[1];
9115 /* know operands[0] == operands[1]. */
9117 if (MEM_P (operands
[2]))
9123 if (find_regno_note (insn
, REG_DEAD
, REGNO (operands
[2])))
9125 if (STACK_TOP_P (operands
[0]))
9126 /* How is it that we are storing to a dead operand[2]?
9127 Well, presumably operands[1] is dead too. We can't
9128 store the result to st(0) as st(0) gets popped on this
9129 instruction. Instead store to operands[2] (which I
9130 think has to be st(1)). st(1) will be popped later.
9131 gcc <= 2.8.1 didn't have this check and generated
9132 assembly code that the Unixware assembler rejected. */
9133 p
= "p\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
9135 p
= "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
9139 if (STACK_TOP_P (operands
[0]))
9140 p
= "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
9142 p
= "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
9147 if (MEM_P (operands
[1]))
9153 if (MEM_P (operands
[2]))
9159 if (find_regno_note (insn
, REG_DEAD
, REGNO (operands
[2])))
9162 /* The SystemV/386 SVR3.2 assembler, and probably all AT&T
9163 derived assemblers, confusingly reverse the direction of
9164 the operation for fsub{r} and fdiv{r} when the
9165 destination register is not st(0). The Intel assembler
9166 doesn't have this brain damage. Read !SYSV386_COMPAT to
9167 figure out what the hardware really does. */
9168 if (STACK_TOP_P (operands
[0]))
9169 p
= "{p\t%0, %2|rp\t%2, %0}";
9171 p
= "{rp\t%2, %0|p\t%0, %2}";
9173 if (STACK_TOP_P (operands
[0]))
9174 /* As above for fmul/fadd, we can't store to st(0). */
9175 p
= "rp\t{%0, %2|%2, %0}"; /* st(1) = st(0) op st(1); pop */
9177 p
= "p\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0); pop */
9182 if (find_regno_note (insn
, REG_DEAD
, REGNO (operands
[1])))
9185 if (STACK_TOP_P (operands
[0]))
9186 p
= "{rp\t%0, %1|p\t%1, %0}";
9188 p
= "{p\t%1, %0|rp\t%0, %1}";
9190 if (STACK_TOP_P (operands
[0]))
9191 p
= "p\t{%0, %1|%1, %0}"; /* st(1) = st(1) op st(0); pop */
9193 p
= "rp\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2); pop */
9198 if (STACK_TOP_P (operands
[0]))
9200 if (STACK_TOP_P (operands
[1]))
9201 p
= "\t{%y2, %0|%0, %y2}"; /* st(0) = st(0) op st(r2) */
9203 p
= "r\t{%y1, %0|%0, %y1}"; /* st(0) = st(r1) op st(0) */
9206 else if (STACK_TOP_P (operands
[1]))
9209 p
= "{\t%1, %0|r\t%0, %1}";
9211 p
= "r\t{%1, %0|%0, %1}"; /* st(r2) = st(0) op st(r2) */
9217 p
= "{r\t%2, %0|\t%0, %2}";
9219 p
= "\t{%2, %0|%0, %2}"; /* st(r1) = st(r1) op st(0) */
9232 /* Return needed mode for entity in optimize_mode_switching pass. */
9235 ix86_mode_needed (int entity
, rtx insn
)
9237 enum attr_i387_cw mode
;
9239 /* The mode UNINITIALIZED is used to store control word after a
9240 function call or ASM pattern. The mode ANY specify that function
9241 has no requirements on the control word and make no changes in the
9242 bits we are interested in. */
9245 || (NONJUMP_INSN_P (insn
)
9246 && (asm_noperands (PATTERN (insn
)) >= 0
9247 || GET_CODE (PATTERN (insn
)) == ASM_INPUT
)))
9248 return I387_CW_UNINITIALIZED
;
9250 if (recog_memoized (insn
) < 0)
9253 mode
= get_attr_i387_cw (insn
);
9258 if (mode
== I387_CW_TRUNC
)
9263 if (mode
== I387_CW_FLOOR
)
9268 if (mode
== I387_CW_CEIL
)
9273 if (mode
== I387_CW_MASK_PM
)
9284 /* Output code to initialize control word copies used by trunc?f?i and
9285 rounding patterns. CURRENT_MODE is set to current control word,
9286 while NEW_MODE is set to new control word. */
9289 emit_i387_cw_initialization (int mode
)
9291 rtx stored_mode
= assign_386_stack_local (HImode
, SLOT_CW_STORED
);
9296 rtx reg
= gen_reg_rtx (HImode
);
9298 emit_insn (gen_x86_fnstcw_1 (stored_mode
));
9299 emit_move_insn (reg
, copy_rtx (stored_mode
));
9301 if (TARGET_64BIT
|| TARGET_PARTIAL_REG_STALL
|| optimize_size
)
9306 /* round toward zero (truncate) */
9307 emit_insn (gen_iorhi3 (reg
, reg
, GEN_INT (0x0c00)));
9308 slot
= SLOT_CW_TRUNC
;
9312 /* round down toward -oo */
9313 emit_insn (gen_andhi3 (reg
, reg
, GEN_INT (~0x0c00)));
9314 emit_insn (gen_iorhi3 (reg
, reg
, GEN_INT (0x0400)));
9315 slot
= SLOT_CW_FLOOR
;
9319 /* round up toward +oo */
9320 emit_insn (gen_andhi3 (reg
, reg
, GEN_INT (~0x0c00)));
9321 emit_insn (gen_iorhi3 (reg
, reg
, GEN_INT (0x0800)));
9322 slot
= SLOT_CW_CEIL
;
9325 case I387_CW_MASK_PM
:
9326 /* mask precision exception for nearbyint() */
9327 emit_insn (gen_iorhi3 (reg
, reg
, GEN_INT (0x0020)));
9328 slot
= SLOT_CW_MASK_PM
;
9340 /* round toward zero (truncate) */
9341 emit_insn (gen_movsi_insv_1 (reg
, GEN_INT (0xc)));
9342 slot
= SLOT_CW_TRUNC
;
9346 /* round down toward -oo */
9347 emit_insn (gen_movsi_insv_1 (reg
, GEN_INT (0x4)));
9348 slot
= SLOT_CW_FLOOR
;
9352 /* round up toward +oo */
9353 emit_insn (gen_movsi_insv_1 (reg
, GEN_INT (0x8)));
9354 slot
= SLOT_CW_CEIL
;
9357 case I387_CW_MASK_PM
:
9358 /* mask precision exception for nearbyint() */
9359 emit_insn (gen_iorhi3 (reg
, reg
, GEN_INT (0x0020)));
9360 slot
= SLOT_CW_MASK_PM
;
9368 gcc_assert (slot
< MAX_386_STACK_LOCALS
);
9370 new_mode
= assign_386_stack_local (HImode
, slot
);
9371 emit_move_insn (new_mode
, reg
);
9374 /* Output code for INSN to convert a float to a signed int. OPERANDS
9375 are the insn operands. The output may be [HSD]Imode and the input
9376 operand may be [SDX]Fmode. */
9379 output_fix_trunc (rtx insn
, rtx
*operands
, int fisttp
)
9381 int stack_top_dies
= find_regno_note (insn
, REG_DEAD
, FIRST_STACK_REG
) != 0;
9382 int dimode_p
= GET_MODE (operands
[0]) == DImode
;
9383 int round_mode
= get_attr_i387_cw (insn
);
9385 /* Jump through a hoop or two for DImode, since the hardware has no
9386 non-popping instruction. We used to do this a different way, but
9387 that was somewhat fragile and broke with post-reload splitters. */
9388 if ((dimode_p
|| fisttp
) && !stack_top_dies
)
9389 output_asm_insn ("fld\t%y1", operands
);
9391 gcc_assert (STACK_TOP_P (operands
[1]));
9392 gcc_assert (MEM_P (operands
[0]));
9393 gcc_assert (GET_MODE (operands
[1]) != TFmode
);
9396 output_asm_insn ("fisttp%z0\t%0", operands
);
9399 if (round_mode
!= I387_CW_ANY
)
9400 output_asm_insn ("fldcw\t%3", operands
);
9401 if (stack_top_dies
|| dimode_p
)
9402 output_asm_insn ("fistp%z0\t%0", operands
);
9404 output_asm_insn ("fist%z0\t%0", operands
);
9405 if (round_mode
!= I387_CW_ANY
)
9406 output_asm_insn ("fldcw\t%2", operands
);
9412 /* Output code for x87 ffreep insn. The OPNO argument, which may only
9413 have the values zero or one, indicates the ffreep insn's operand
9414 from the OPERANDS array. */
9417 output_387_ffreep (rtx
*operands ATTRIBUTE_UNUSED
, int opno
)
9419 if (TARGET_USE_FFREEP
)
9420 #if HAVE_AS_IX86_FFREEP
9421 return opno
? "ffreep\t%y1" : "ffreep\t%y0";
9424 static char retval
[] = ".word\t0xc_df";
9425 int regno
= REGNO (operands
[opno
]);
9427 gcc_assert (FP_REGNO_P (regno
));
9429 retval
[9] = '0' + (regno
- FIRST_STACK_REG
);
9434 return opno
? "fstp\t%y1" : "fstp\t%y0";
9438 /* Output code for INSN to compare OPERANDS. EFLAGS_P is 1 when fcomi
9439 should be used. UNORDERED_P is true when fucom should be used. */
9442 output_fp_compare (rtx insn
, rtx
*operands
, int eflags_p
, int unordered_p
)
9445 rtx cmp_op0
, cmp_op1
;
9446 int is_sse
= SSE_REG_P (operands
[0]) || SSE_REG_P (operands
[1]);
9450 cmp_op0
= operands
[0];
9451 cmp_op1
= operands
[1];
9455 cmp_op0
= operands
[1];
9456 cmp_op1
= operands
[2];
9461 if (GET_MODE (operands
[0]) == SFmode
)
9463 return "ucomiss\t{%1, %0|%0, %1}";
9465 return "comiss\t{%1, %0|%0, %1}";
9468 return "ucomisd\t{%1, %0|%0, %1}";
9470 return "comisd\t{%1, %0|%0, %1}";
9473 gcc_assert (STACK_TOP_P (cmp_op0
));
9475 stack_top_dies
= find_regno_note (insn
, REG_DEAD
, FIRST_STACK_REG
) != 0;
9477 if (cmp_op1
== CONST0_RTX (GET_MODE (cmp_op1
)))
9481 output_asm_insn ("ftst\n\tfnstsw\t%0", operands
);
9482 return output_387_ffreep (operands
, 1);
9485 return "ftst\n\tfnstsw\t%0";
9488 if (STACK_REG_P (cmp_op1
)
9490 && find_regno_note (insn
, REG_DEAD
, REGNO (cmp_op1
))
9491 && REGNO (cmp_op1
) != FIRST_STACK_REG
)
9493 /* If both the top of the 387 stack dies, and the other operand
9494 is also a stack register that dies, then this must be a
9495 `fcompp' float compare */
9499 /* There is no double popping fcomi variant. Fortunately,
9500 eflags is immune from the fstp's cc clobbering. */
9502 output_asm_insn ("fucomip\t{%y1, %0|%0, %y1}", operands
);
9504 output_asm_insn ("fcomip\t{%y1, %0|%0, %y1}", operands
);
9505 return output_387_ffreep (operands
, 0);
9510 return "fucompp\n\tfnstsw\t%0";
9512 return "fcompp\n\tfnstsw\t%0";
9517 /* Encoded here as eflags_p | intmode | unordered_p | stack_top_dies. */
9519 static const char * const alt
[16] =
9521 "fcom%z2\t%y2\n\tfnstsw\t%0",
9522 "fcomp%z2\t%y2\n\tfnstsw\t%0",
9523 "fucom%z2\t%y2\n\tfnstsw\t%0",
9524 "fucomp%z2\t%y2\n\tfnstsw\t%0",
9526 "ficom%z2\t%y2\n\tfnstsw\t%0",
9527 "ficomp%z2\t%y2\n\tfnstsw\t%0",
9531 "fcomi\t{%y1, %0|%0, %y1}",
9532 "fcomip\t{%y1, %0|%0, %y1}",
9533 "fucomi\t{%y1, %0|%0, %y1}",
9534 "fucomip\t{%y1, %0|%0, %y1}",
9545 mask
= eflags_p
<< 3;
9546 mask
|= (GET_MODE_CLASS (GET_MODE (cmp_op1
)) == MODE_INT
) << 2;
9547 mask
|= unordered_p
<< 1;
9548 mask
|= stack_top_dies
;
9550 gcc_assert (mask
< 16);
9559 ix86_output_addr_vec_elt (FILE *file
, int value
)
9561 const char *directive
= ASM_LONG
;
9565 directive
= ASM_QUAD
;
9567 gcc_assert (!TARGET_64BIT
);
9570 fprintf (file
, "%s%s%d\n", directive
, LPREFIX
, value
);
9574 ix86_output_addr_diff_elt (FILE *file
, int value
, int rel
)
9576 const char *directive
= ASM_LONG
;
9579 if (TARGET_64BIT
&& CASE_VECTOR_MODE
== DImode
)
9580 directive
= ASM_QUAD
;
9582 gcc_assert (!TARGET_64BIT
);
9584 /* We can't use @GOTOFF for text labels on VxWorks; see gotoff_operand. */
9585 if (TARGET_64BIT
|| TARGET_VXWORKS_RTP
)
9586 fprintf (file
, "%s%s%d-%s%d\n",
9587 directive
, LPREFIX
, value
, LPREFIX
, rel
);
9588 else if (HAVE_AS_GOTOFF_IN_DATA
)
9589 fprintf (file
, "%s%s%d@GOTOFF\n", ASM_LONG
, LPREFIX
, value
);
9591 else if (TARGET_MACHO
)
9593 fprintf (file
, "%s%s%d-", ASM_LONG
, LPREFIX
, value
);
9594 machopic_output_function_base_name (file
);
9595 fprintf(file
, "\n");
9599 asm_fprintf (file
, "%s%U%s+[.-%s%d]\n",
9600 ASM_LONG
, GOT_SYMBOL_NAME
, LPREFIX
, value
);
9603 /* Generate either "mov $0, reg" or "xor reg, reg", as appropriate
9607 ix86_expand_clear (rtx dest
)
9611 /* We play register width games, which are only valid after reload. */
9612 gcc_assert (reload_completed
);
9614 /* Avoid HImode and its attendant prefix byte. */
9615 if (GET_MODE_SIZE (GET_MODE (dest
)) < 4)
9616 dest
= gen_rtx_REG (SImode
, REGNO (dest
));
9617 tmp
= gen_rtx_SET (VOIDmode
, dest
, const0_rtx
);
9619 /* This predicate should match that for movsi_xor and movdi_xor_rex64. */
9620 if (reload_completed
&& (!TARGET_USE_MOV0
|| optimize_size
))
9622 rtx clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, 17));
9623 tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, tmp
, clob
));
9629 /* X is an unchanging MEM. If it is a constant pool reference, return
9630 the constant pool rtx, else NULL. */
9633 maybe_get_pool_constant (rtx x
)
9635 x
= ix86_delegitimize_address (XEXP (x
, 0));
9637 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
9638 return get_pool_constant (x
);
9644 ix86_expand_move (enum machine_mode mode
, rtx operands
[])
9646 int strict
= (reload_in_progress
|| reload_completed
);
9648 enum tls_model model
;
9653 if (GET_CODE (op1
) == SYMBOL_REF
)
9655 model
= SYMBOL_REF_TLS_MODEL (op1
);
9658 op1
= legitimize_tls_address (op1
, model
, true);
9659 op1
= force_operand (op1
, op0
);
9663 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
9664 && SYMBOL_REF_DLLIMPORT_P (op1
))
9665 op1
= legitimize_dllimport_symbol (op1
, false);
9667 else if (GET_CODE (op1
) == CONST
9668 && GET_CODE (XEXP (op1
, 0)) == PLUS
9669 && GET_CODE (XEXP (XEXP (op1
, 0), 0)) == SYMBOL_REF
)
9671 rtx addend
= XEXP (XEXP (op1
, 0), 1);
9672 rtx symbol
= XEXP (XEXP (op1
, 0), 0);
9675 model
= SYMBOL_REF_TLS_MODEL (symbol
);
9677 tmp
= legitimize_tls_address (symbol
, model
, true);
9678 else if (TARGET_DLLIMPORT_DECL_ATTRIBUTES
9679 && SYMBOL_REF_DLLIMPORT_P (symbol
))
9680 tmp
= legitimize_dllimport_symbol (symbol
, true);
9684 tmp
= force_operand (tmp
, NULL
);
9685 tmp
= expand_simple_binop (Pmode
, PLUS
, tmp
, addend
,
9686 op0
, 1, OPTAB_DIRECT
);
9692 if (flag_pic
&& mode
== Pmode
&& symbolic_operand (op1
, Pmode
))
9694 if (TARGET_MACHO
&& !TARGET_64BIT
)
9699 rtx temp
= ((reload_in_progress
9700 || ((op0
&& REG_P (op0
))
9702 ? op0
: gen_reg_rtx (Pmode
));
9703 op1
= machopic_indirect_data_reference (op1
, temp
);
9704 op1
= machopic_legitimize_pic_address (op1
, mode
,
9705 temp
== op1
? 0 : temp
);
9707 else if (MACHOPIC_INDIRECT
)
9708 op1
= machopic_indirect_data_reference (op1
, 0);
9716 op1
= force_reg (Pmode
, op1
);
9717 else if (!TARGET_64BIT
|| !x86_64_movabs_operand (op1
, Pmode
))
9719 rtx reg
= no_new_pseudos
? op0
: NULL_RTX
;
9720 op1
= legitimize_pic_address (op1
, reg
);
9729 && (PUSH_ROUNDING (GET_MODE_SIZE (mode
)) != GET_MODE_SIZE (mode
)
9730 || !push_operand (op0
, mode
))
9732 op1
= force_reg (mode
, op1
);
9734 if (push_operand (op0
, mode
)
9735 && ! general_no_elim_operand (op1
, mode
))
9736 op1
= copy_to_mode_reg (mode
, op1
);
9738 /* Force large constants in 64bit compilation into register
9739 to get them CSEed. */
9740 if (TARGET_64BIT
&& mode
== DImode
9741 && immediate_operand (op1
, mode
)
9742 && !x86_64_zext_immediate_operand (op1
, VOIDmode
)
9743 && !register_operand (op0
, mode
)
9744 && optimize
&& !reload_completed
&& !reload_in_progress
)
9745 op1
= copy_to_mode_reg (mode
, op1
);
9747 if (FLOAT_MODE_P (mode
))
9749 /* If we are loading a floating point constant to a register,
9750 force the value to memory now, since we'll get better code
9751 out the back end. */
9755 else if (GET_CODE (op1
) == CONST_DOUBLE
)
9757 op1
= validize_mem (force_const_mem (mode
, op1
));
9758 if (!register_operand (op0
, mode
))
9760 rtx temp
= gen_reg_rtx (mode
);
9761 emit_insn (gen_rtx_SET (VOIDmode
, temp
, op1
));
9762 emit_move_insn (op0
, temp
);
9769 emit_insn (gen_rtx_SET (VOIDmode
, op0
, op1
));
9773 ix86_expand_vector_move (enum machine_mode mode
, rtx operands
[])
9775 rtx op0
= operands
[0], op1
= operands
[1];
9776 unsigned int align
= GET_MODE_ALIGNMENT (mode
);
9778 /* Force constants other than zero into memory. We do not know how
9779 the instructions used to build constants modify the upper 64 bits
9780 of the register, once we have that information we may be able
9781 to handle some of them more efficiently. */
9782 if ((reload_in_progress
| reload_completed
) == 0
9783 && register_operand (op0
, mode
)
9784 && (CONSTANT_P (op1
)
9785 || (GET_CODE (op1
) == SUBREG
9786 && CONSTANT_P (SUBREG_REG (op1
))))
9787 && standard_sse_constant_p (op1
) <= 0)
9788 op1
= validize_mem (force_const_mem (mode
, op1
));
9790 /* TDmode values are passed as TImode on the stack. Timode values
9791 are moved via xmm registers, and moving them to stack can result in
9792 unaligned memory access. Use ix86_expand_vector_move_misalign()
9793 if memory operand is not aligned correctly. */
9795 && (mode
== TImode
) && !TARGET_64BIT
9796 && ((MEM_P (op0
) && (MEM_ALIGN (op0
) < align
))
9797 || (MEM_P (op1
) && (MEM_ALIGN (op1
) < align
))))
9801 /* ix86_expand_vector_move_misalign() does not like constants ... */
9802 if (CONSTANT_P (op1
)
9803 || (GET_CODE (op1
) == SUBREG
9804 && CONSTANT_P (SUBREG_REG (op1
))))
9805 op1
= validize_mem (force_const_mem (mode
, op1
));
9807 /* ... nor both arguments in memory. */
9808 if (!register_operand (op0
, mode
)
9809 && !register_operand (op1
, mode
))
9810 op1
= force_reg (mode
, op1
);
9812 tmp
[0] = op0
; tmp
[1] = op1
;
9813 ix86_expand_vector_move_misalign (mode
, tmp
);
9817 /* Make operand1 a register if it isn't already. */
9819 && !register_operand (op0
, mode
)
9820 && !register_operand (op1
, mode
))
9822 emit_move_insn (op0
, force_reg (GET_MODE (op0
), op1
));
9826 emit_insn (gen_rtx_SET (VOIDmode
, op0
, op1
));
9829 /* Implement the movmisalign patterns for SSE. Non-SSE modes go
9830 straight to ix86_expand_vector_move. */
9831 /* Code generation for scalar reg-reg moves of single and double precision data:
9832 if (x86_sse_partial_reg_dependency == true | x86_sse_split_regs == true)
9836 if (x86_sse_partial_reg_dependency == true)
9841 Code generation for scalar loads of double precision data:
9842 if (x86_sse_split_regs == true)
9843 movlpd mem, reg (gas syntax)
9847 Code generation for unaligned packed loads of single precision data
9848 (x86_sse_unaligned_move_optimal overrides x86_sse_partial_reg_dependency):
9849 if (x86_sse_unaligned_move_optimal)
9852 if (x86_sse_partial_reg_dependency == true)
9864 Code generation for unaligned packed loads of double precision data
9865 (x86_sse_unaligned_move_optimal overrides x86_sse_split_regs):
9866 if (x86_sse_unaligned_move_optimal)
9869 if (x86_sse_split_regs == true)
9882 ix86_expand_vector_move_misalign (enum machine_mode mode
, rtx operands
[])
9891 /* If we're optimizing for size, movups is the smallest. */
9894 op0
= gen_lowpart (V4SFmode
, op0
);
9895 op1
= gen_lowpart (V4SFmode
, op1
);
9896 emit_insn (gen_sse_movups (op0
, op1
));
9900 /* ??? If we have typed data, then it would appear that using
9901 movdqu is the only way to get unaligned data loaded with
9903 if (TARGET_SSE2
&& GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
)
9905 op0
= gen_lowpart (V16QImode
, op0
);
9906 op1
= gen_lowpart (V16QImode
, op1
);
9907 emit_insn (gen_sse2_movdqu (op0
, op1
));
9911 if (TARGET_SSE2
&& mode
== V2DFmode
)
9915 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL
)
9917 op0
= gen_lowpart (V2DFmode
, op0
);
9918 op1
= gen_lowpart (V2DFmode
, op1
);
9919 emit_insn (gen_sse2_movupd (op0
, op1
));
9923 /* When SSE registers are split into halves, we can avoid
9924 writing to the top half twice. */
9925 if (TARGET_SSE_SPLIT_REGS
)
9927 emit_insn (gen_rtx_CLOBBER (VOIDmode
, op0
));
9932 /* ??? Not sure about the best option for the Intel chips.
9933 The following would seem to satisfy; the register is
9934 entirely cleared, breaking the dependency chain. We
9935 then store to the upper half, with a dependency depth
9936 of one. A rumor has it that Intel recommends two movsd
9937 followed by an unpacklpd, but this is unconfirmed. And
9938 given that the dependency depth of the unpacklpd would
9939 still be one, I'm not sure why this would be better. */
9940 zero
= CONST0_RTX (V2DFmode
);
9943 m
= adjust_address (op1
, DFmode
, 0);
9944 emit_insn (gen_sse2_loadlpd (op0
, zero
, m
));
9945 m
= adjust_address (op1
, DFmode
, 8);
9946 emit_insn (gen_sse2_loadhpd (op0
, op0
, m
));
9950 if (TARGET_SSE_UNALIGNED_MOVE_OPTIMAL
)
9952 op0
= gen_lowpart (V4SFmode
, op0
);
9953 op1
= gen_lowpart (V4SFmode
, op1
);
9954 emit_insn (gen_sse_movups (op0
, op1
));
9958 if (TARGET_SSE_PARTIAL_REG_DEPENDENCY
)
9959 emit_move_insn (op0
, CONST0_RTX (mode
));
9961 emit_insn (gen_rtx_CLOBBER (VOIDmode
, op0
));
9963 if (mode
!= V4SFmode
)
9964 op0
= gen_lowpart (V4SFmode
, op0
);
9965 m
= adjust_address (op1
, V2SFmode
, 0);
9966 emit_insn (gen_sse_loadlps (op0
, op0
, m
));
9967 m
= adjust_address (op1
, V2SFmode
, 8);
9968 emit_insn (gen_sse_loadhps (op0
, op0
, m
));
9971 else if (MEM_P (op0
))
9973 /* If we're optimizing for size, movups is the smallest. */
9976 op0
= gen_lowpart (V4SFmode
, op0
);
9977 op1
= gen_lowpart (V4SFmode
, op1
);
9978 emit_insn (gen_sse_movups (op0
, op1
));
9982 /* ??? Similar to above, only less clear because of quote
9983 typeless stores unquote. */
9984 if (TARGET_SSE2
&& !TARGET_SSE_TYPELESS_STORES
9985 && GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
)
9987 op0
= gen_lowpart (V16QImode
, op0
);
9988 op1
= gen_lowpart (V16QImode
, op1
);
9989 emit_insn (gen_sse2_movdqu (op0
, op1
));
9993 if (TARGET_SSE2
&& mode
== V2DFmode
)
9995 m
= adjust_address (op0
, DFmode
, 0);
9996 emit_insn (gen_sse2_storelpd (m
, op1
));
9997 m
= adjust_address (op0
, DFmode
, 8);
9998 emit_insn (gen_sse2_storehpd (m
, op1
));
10002 if (mode
!= V4SFmode
)
10003 op1
= gen_lowpart (V4SFmode
, op1
);
10004 m
= adjust_address (op0
, V2SFmode
, 0);
10005 emit_insn (gen_sse_storelps (m
, op1
));
10006 m
= adjust_address (op0
, V2SFmode
, 8);
10007 emit_insn (gen_sse_storehps (m
, op1
));
10011 gcc_unreachable ();
10014 /* Expand a push in MODE. This is some mode for which we do not support
10015 proper push instructions, at least from the registers that we expect
10016 the value to live in. */
10019 ix86_expand_push (enum machine_mode mode
, rtx x
)
10023 tmp
= expand_simple_binop (Pmode
, PLUS
, stack_pointer_rtx
,
10024 GEN_INT (-GET_MODE_SIZE (mode
)),
10025 stack_pointer_rtx
, 1, OPTAB_DIRECT
);
10026 if (tmp
!= stack_pointer_rtx
)
10027 emit_move_insn (stack_pointer_rtx
, tmp
);
10029 tmp
= gen_rtx_MEM (mode
, stack_pointer_rtx
);
10030 emit_move_insn (tmp
, x
);
10033 /* Helper function of ix86_fixup_binary_operands to canonicalize
10034 operand order. Returns true if the operands should be swapped. */
10037 ix86_swap_binary_operands_p (enum rtx_code code
, enum machine_mode mode
,
10040 rtx dst
= operands
[0];
10041 rtx src1
= operands
[1];
10042 rtx src2
= operands
[2];
10044 /* If the operation is not commutative, we can't do anything. */
10045 if (GET_RTX_CLASS (code
) != RTX_COMM_ARITH
)
10048 /* Highest priority is that src1 should match dst. */
10049 if (rtx_equal_p (dst
, src1
))
10051 if (rtx_equal_p (dst
, src2
))
10054 /* Next highest priority is that immediate constants come second. */
10055 if (immediate_operand (src2
, mode
))
10057 if (immediate_operand (src1
, mode
))
10060 /* Lowest priority is that memory references should come second. */
10070 /* Fix up OPERANDS to satisfy ix86_binary_operator_ok. Return the
10071 destination to use for the operation. If different from the true
10072 destination in operands[0], a copy operation will be required. */
10075 ix86_fixup_binary_operands (enum rtx_code code
, enum machine_mode mode
,
10078 rtx dst
= operands
[0];
10079 rtx src1
= operands
[1];
10080 rtx src2
= operands
[2];
10082 /* Canonicalize operand order. */
10083 if (ix86_swap_binary_operands_p (code
, mode
, operands
))
10090 /* Both source operands cannot be in memory. */
10091 if (MEM_P (src1
) && MEM_P (src2
))
10093 /* Optimization: Only read from memory once. */
10094 if (rtx_equal_p (src1
, src2
))
10096 src2
= force_reg (mode
, src2
);
10100 src2
= force_reg (mode
, src2
);
10103 /* If the destination is memory, and we do not have matching source
10104 operands, do things in registers. */
10105 if (MEM_P (dst
) && !rtx_equal_p (dst
, src1
))
10106 dst
= gen_reg_rtx (mode
);
10108 /* Source 1 cannot be a constant. */
10109 if (CONSTANT_P (src1
))
10110 src1
= force_reg (mode
, src1
);
10112 /* Source 1 cannot be a non-matching memory. */
10113 if (MEM_P (src1
) && !rtx_equal_p (dst
, src1
))
10114 src1
= force_reg (mode
, src1
);
10116 operands
[1] = src1
;
10117 operands
[2] = src2
;
10121 /* Similarly, but assume that the destination has already been
10122 set up properly. */
10125 ix86_fixup_binary_operands_no_copy (enum rtx_code code
,
10126 enum machine_mode mode
, rtx operands
[])
10128 rtx dst
= ix86_fixup_binary_operands (code
, mode
, operands
);
10129 gcc_assert (dst
== operands
[0]);
10132 /* Attempt to expand a binary operator. Make the expansion closer to the
10133 actual machine, then just general_operand, which will allow 3 separate
10134 memory references (one output, two input) in a single insn. */
10137 ix86_expand_binary_operator (enum rtx_code code
, enum machine_mode mode
,
10140 rtx src1
, src2
, dst
, op
, clob
;
10142 dst
= ix86_fixup_binary_operands (code
, mode
, operands
);
10143 src1
= operands
[1];
10144 src2
= operands
[2];
10146 /* Emit the instruction. */
10148 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, mode
, src1
, src2
));
10149 if (reload_in_progress
)
10151 /* Reload doesn't know about the flags register, and doesn't know that
10152 it doesn't want to clobber it. We can only do this with PLUS. */
10153 gcc_assert (code
== PLUS
);
10158 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, FLAGS_REG
));
10159 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
10162 /* Fix up the destination if needed. */
10163 if (dst
!= operands
[0])
10164 emit_move_insn (operands
[0], dst
);
10167 /* Return TRUE or FALSE depending on whether the binary operator meets the
10168 appropriate constraints. */
10171 ix86_binary_operator_ok (enum rtx_code code
, enum machine_mode mode
,
10174 rtx dst
= operands
[0];
10175 rtx src1
= operands
[1];
10176 rtx src2
= operands
[2];
10178 /* Both source operands cannot be in memory. */
10179 if (MEM_P (src1
) && MEM_P (src2
))
10182 /* Canonicalize operand order for commutative operators. */
10183 if (ix86_swap_binary_operands_p (code
, mode
, operands
))
10190 /* If the destination is memory, we must have a matching source operand. */
10191 if (MEM_P (dst
) && !rtx_equal_p (dst
, src1
))
10194 /* Source 1 cannot be a constant. */
10195 if (CONSTANT_P (src1
))
10198 /* Source 1 cannot be a non-matching memory. */
10199 if (MEM_P (src1
) && !rtx_equal_p (dst
, src1
))
10205 /* Attempt to expand a unary operator. Make the expansion closer to the
10206 actual machine, then just general_operand, which will allow 2 separate
10207 memory references (one output, one input) in a single insn. */
10210 ix86_expand_unary_operator (enum rtx_code code
, enum machine_mode mode
,
10213 int matching_memory
;
10214 rtx src
, dst
, op
, clob
;
10219 /* If the destination is memory, and we do not have matching source
10220 operands, do things in registers. */
10221 matching_memory
= 0;
10224 if (rtx_equal_p (dst
, src
))
10225 matching_memory
= 1;
10227 dst
= gen_reg_rtx (mode
);
10230 /* When source operand is memory, destination must match. */
10231 if (MEM_P (src
) && !matching_memory
)
10232 src
= force_reg (mode
, src
);
10234 /* Emit the instruction. */
10236 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_e (code
, mode
, src
));
10237 if (reload_in_progress
|| code
== NOT
)
10239 /* Reload doesn't know about the flags register, and doesn't know that
10240 it doesn't want to clobber it. */
10241 gcc_assert (code
== NOT
);
10246 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, FLAGS_REG
));
10247 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
10250 /* Fix up the destination if needed. */
10251 if (dst
!= operands
[0])
10252 emit_move_insn (operands
[0], dst
);
10255 /* Return TRUE or FALSE depending on whether the unary operator meets the
10256 appropriate constraints. */
10259 ix86_unary_operator_ok (enum rtx_code code ATTRIBUTE_UNUSED
,
10260 enum machine_mode mode ATTRIBUTE_UNUSED
,
10261 rtx operands
[2] ATTRIBUTE_UNUSED
)
10263 /* If one of operands is memory, source and destination must match. */
10264 if ((MEM_P (operands
[0])
10265 || MEM_P (operands
[1]))
10266 && ! rtx_equal_p (operands
[0], operands
[1]))
10271 /* Post-reload splitter for converting an SF or DFmode value in an
10272 SSE register into an unsigned SImode. */
10275 ix86_split_convert_uns_si_sse (rtx operands
[])
10277 enum machine_mode vecmode
;
10278 rtx value
, large
, zero_or_two31
, input
, two31
, x
;
10280 large
= operands
[1];
10281 zero_or_two31
= operands
[2];
10282 input
= operands
[3];
10283 two31
= operands
[4];
10284 vecmode
= GET_MODE (large
);
10285 value
= gen_rtx_REG (vecmode
, REGNO (operands
[0]));
10287 /* Load up the value into the low element. We must ensure that the other
10288 elements are valid floats -- zero is the easiest such value. */
10291 if (vecmode
== V4SFmode
)
10292 emit_insn (gen_vec_setv4sf_0 (value
, CONST0_RTX (V4SFmode
), input
));
10294 emit_insn (gen_sse2_loadlpd (value
, CONST0_RTX (V2DFmode
), input
));
10298 input
= gen_rtx_REG (vecmode
, REGNO (input
));
10299 emit_move_insn (value
, CONST0_RTX (vecmode
));
10300 if (vecmode
== V4SFmode
)
10301 emit_insn (gen_sse_movss (value
, value
, input
));
10303 emit_insn (gen_sse2_movsd (value
, value
, input
));
10306 emit_move_insn (large
, two31
);
10307 emit_move_insn (zero_or_two31
, MEM_P (two31
) ? large
: two31
);
10309 x
= gen_rtx_fmt_ee (LE
, vecmode
, large
, value
);
10310 emit_insn (gen_rtx_SET (VOIDmode
, large
, x
));
10312 x
= gen_rtx_AND (vecmode
, zero_or_two31
, large
);
10313 emit_insn (gen_rtx_SET (VOIDmode
, zero_or_two31
, x
));
10315 x
= gen_rtx_MINUS (vecmode
, value
, zero_or_two31
);
10316 emit_insn (gen_rtx_SET (VOIDmode
, value
, x
));
10318 large
= gen_rtx_REG (V4SImode
, REGNO (large
));
10319 emit_insn (gen_ashlv4si3 (large
, large
, GEN_INT (31)));
10321 x
= gen_rtx_REG (V4SImode
, REGNO (value
));
10322 if (vecmode
== V4SFmode
)
10323 emit_insn (gen_sse2_cvttps2dq (x
, value
));
10325 emit_insn (gen_sse2_cvttpd2dq (x
, value
));
10328 emit_insn (gen_xorv4si3 (value
, value
, large
));
10331 /* Convert an unsigned DImode value into a DFmode, using only SSE.
10332 Expects the 64-bit DImode to be supplied in a pair of integral
10333 registers. Requires SSE2; will use SSE3 if available. For x86_32,
10334 -mfpmath=sse, !optimize_size only. */
10337 ix86_expand_convert_uns_didf_sse (rtx target
, rtx input
)
10339 REAL_VALUE_TYPE bias_lo_rvt
, bias_hi_rvt
;
10340 rtx int_xmm
, fp_xmm
;
10341 rtx biases
, exponents
;
10344 int_xmm
= gen_reg_rtx (V4SImode
);
10345 if (TARGET_INTER_UNIT_MOVES
)
10346 emit_insn (gen_movdi_to_sse (int_xmm
, input
));
10347 else if (TARGET_SSE_SPLIT_REGS
)
10349 emit_insn (gen_rtx_CLOBBER (VOIDmode
, int_xmm
));
10350 emit_move_insn (gen_lowpart (DImode
, int_xmm
), input
);
10354 x
= gen_reg_rtx (V2DImode
);
10355 ix86_expand_vector_init_one_nonzero (false, V2DImode
, x
, input
, 0);
10356 emit_move_insn (int_xmm
, gen_lowpart (V4SImode
, x
));
10359 x
= gen_rtx_CONST_VECTOR (V4SImode
,
10360 gen_rtvec (4, GEN_INT (0x43300000UL
),
10361 GEN_INT (0x45300000UL
),
10362 const0_rtx
, const0_rtx
));
10363 exponents
= validize_mem (force_const_mem (V4SImode
, x
));
10365 /* int_xmm = {0x45300000UL, fp_xmm/hi, 0x43300000, fp_xmm/lo } */
10366 emit_insn (gen_sse2_punpckldq (int_xmm
, int_xmm
, exponents
));
10368 /* Concatenating (juxtaposing) (0x43300000UL ## fp_value_low_xmm)
10369 yields a valid DF value equal to (0x1.0p52 + double(fp_value_lo_xmm)).
10370 Similarly (0x45300000UL ## fp_value_hi_xmm) yields
10371 (0x1.0p84 + double(fp_value_hi_xmm)).
10372 Note these exponents differ by 32. */
10374 fp_xmm
= copy_to_mode_reg (V2DFmode
, gen_lowpart (V2DFmode
, int_xmm
));
10376 /* Subtract off those 0x1.0p52 and 0x1.0p84 biases, to produce values
10377 in [0,2**32-1] and [0]+[2**32,2**64-1] respectively. */
10378 real_ldexp (&bias_lo_rvt
, &dconst1
, 52);
10379 real_ldexp (&bias_hi_rvt
, &dconst1
, 84);
10380 biases
= const_double_from_real_value (bias_lo_rvt
, DFmode
);
10381 x
= const_double_from_real_value (bias_hi_rvt
, DFmode
);
10382 biases
= gen_rtx_CONST_VECTOR (V2DFmode
, gen_rtvec (2, biases
, x
));
10383 biases
= validize_mem (force_const_mem (V2DFmode
, biases
));
10384 emit_insn (gen_subv2df3 (fp_xmm
, fp_xmm
, biases
));
10386 /* Add the upper and lower DFmode values together. */
10388 emit_insn (gen_sse3_haddv2df3 (fp_xmm
, fp_xmm
, fp_xmm
));
10391 x
= copy_to_mode_reg (V2DFmode
, fp_xmm
);
10392 emit_insn (gen_sse2_unpckhpd (fp_xmm
, fp_xmm
, fp_xmm
));
10393 emit_insn (gen_addv2df3 (fp_xmm
, fp_xmm
, x
));
10396 ix86_expand_vector_extract (false, target
, fp_xmm
, 0);
10399 /* Convert an unsigned SImode value into a DFmode. Only currently used
10400 for SSE, but applicable anywhere. */
10403 ix86_expand_convert_uns_sidf_sse (rtx target
, rtx input
)
10405 REAL_VALUE_TYPE TWO31r
;
10408 x
= expand_simple_binop (SImode
, PLUS
, input
, GEN_INT (-2147483647 - 1),
10409 NULL
, 1, OPTAB_DIRECT
);
10411 fp
= gen_reg_rtx (DFmode
);
10412 emit_insn (gen_floatsidf2 (fp
, x
));
10414 real_ldexp (&TWO31r
, &dconst1
, 31);
10415 x
= const_double_from_real_value (TWO31r
, DFmode
);
10417 x
= expand_simple_binop (DFmode
, PLUS
, fp
, x
, target
, 0, OPTAB_DIRECT
);
10419 emit_move_insn (target
, x
);
10422 /* Convert a signed DImode value into a DFmode. Only used for SSE in
10423 32-bit mode; otherwise we have a direct convert instruction. */
10426 ix86_expand_convert_sign_didf_sse (rtx target
, rtx input
)
10428 REAL_VALUE_TYPE TWO32r
;
10429 rtx fp_lo
, fp_hi
, x
;
10431 fp_lo
= gen_reg_rtx (DFmode
);
10432 fp_hi
= gen_reg_rtx (DFmode
);
10434 emit_insn (gen_floatsidf2 (fp_hi
, gen_highpart (SImode
, input
)));
10436 real_ldexp (&TWO32r
, &dconst1
, 32);
10437 x
= const_double_from_real_value (TWO32r
, DFmode
);
10438 fp_hi
= expand_simple_binop (DFmode
, MULT
, fp_hi
, x
, fp_hi
, 0, OPTAB_DIRECT
);
10440 ix86_expand_convert_uns_sidf_sse (fp_lo
, gen_lowpart (SImode
, input
));
10442 x
= expand_simple_binop (DFmode
, PLUS
, fp_hi
, fp_lo
, target
,
10445 emit_move_insn (target
, x
);
10448 /* Convert an unsigned SImode value into a SFmode, using only SSE.
10449 For x86_32, -mfpmath=sse, !optimize_size only. */
10451 ix86_expand_convert_uns_sisf_sse (rtx target
, rtx input
)
10453 REAL_VALUE_TYPE ONE16r
;
10454 rtx fp_hi
, fp_lo
, int_hi
, int_lo
, x
;
10456 real_ldexp (&ONE16r
, &dconst1
, 16);
10457 x
= const_double_from_real_value (ONE16r
, SFmode
);
10458 int_lo
= expand_simple_binop (SImode
, AND
, input
, GEN_INT(0xffff),
10459 NULL
, 0, OPTAB_DIRECT
);
10460 int_hi
= expand_simple_binop (SImode
, LSHIFTRT
, input
, GEN_INT(16),
10461 NULL
, 0, OPTAB_DIRECT
);
10462 fp_hi
= gen_reg_rtx (SFmode
);
10463 fp_lo
= gen_reg_rtx (SFmode
);
10464 emit_insn (gen_floatsisf2 (fp_hi
, int_hi
));
10465 emit_insn (gen_floatsisf2 (fp_lo
, int_lo
));
10466 fp_hi
= expand_simple_binop (SFmode
, MULT
, fp_hi
, x
, fp_hi
,
10468 fp_hi
= expand_simple_binop (SFmode
, PLUS
, fp_hi
, fp_lo
, target
,
10470 if (!rtx_equal_p (target
, fp_hi
))
10471 emit_move_insn (target
, fp_hi
);
10474 /* A subroutine of ix86_build_signbit_mask_vector. If VECT is true,
10475 then replicate the value for all elements of the vector
10479 ix86_build_const_vector (enum machine_mode mode
, bool vect
, rtx value
)
10486 v
= gen_rtvec (4, value
, value
, value
, value
);
10488 v
= gen_rtvec (4, value
, CONST0_RTX (SFmode
),
10489 CONST0_RTX (SFmode
), CONST0_RTX (SFmode
));
10490 return gen_rtx_CONST_VECTOR (V4SFmode
, v
);
10494 v
= gen_rtvec (2, value
, value
);
10496 v
= gen_rtvec (2, value
, CONST0_RTX (DFmode
));
10497 return gen_rtx_CONST_VECTOR (V2DFmode
, v
);
10500 gcc_unreachable ();
10504 /* A subroutine of ix86_expand_fp_absneg_operator and copysign expanders.
10505 Create a mask for the sign bit in MODE for an SSE register. If VECT is
10506 true, then replicate the mask for all elements of the vector register.
10507 If INVERT is true, then create a mask excluding the sign bit. */
10510 ix86_build_signbit_mask (enum machine_mode mode
, bool vect
, bool invert
)
10512 enum machine_mode vec_mode
;
10513 HOST_WIDE_INT hi
, lo
;
10518 /* Find the sign bit, sign extended to 2*HWI. */
10519 if (mode
== SFmode
)
10520 lo
= 0x80000000, hi
= lo
< 0;
10521 else if (HOST_BITS_PER_WIDE_INT
>= 64)
10522 lo
= (HOST_WIDE_INT
)1 << shift
, hi
= -1;
10524 lo
= 0, hi
= (HOST_WIDE_INT
)1 << (shift
- HOST_BITS_PER_WIDE_INT
);
10527 lo
= ~lo
, hi
= ~hi
;
10529 /* Force this value into the low part of a fp vector constant. */
10530 mask
= immed_double_const (lo
, hi
, mode
== SFmode
? SImode
: DImode
);
10531 mask
= gen_lowpart (mode
, mask
);
10533 v
= ix86_build_const_vector (mode
, vect
, mask
);
10534 vec_mode
= (mode
== SFmode
) ? V4SFmode
: V2DFmode
;
10535 return force_reg (vec_mode
, v
);
10538 /* Generate code for floating point ABS or NEG. */
10541 ix86_expand_fp_absneg_operator (enum rtx_code code
, enum machine_mode mode
,
10544 rtx mask
, set
, use
, clob
, dst
, src
;
10545 bool matching_memory
;
10546 bool use_sse
= false;
10547 bool vector_mode
= VECTOR_MODE_P (mode
);
10548 enum machine_mode elt_mode
= mode
;
10552 elt_mode
= GET_MODE_INNER (mode
);
10555 else if (TARGET_SSE_MATH
)
10556 use_sse
= SSE_FLOAT_MODE_P (mode
);
10558 /* NEG and ABS performed with SSE use bitwise mask operations.
10559 Create the appropriate mask now. */
10561 mask
= ix86_build_signbit_mask (elt_mode
, vector_mode
, code
== ABS
);
10568 /* If the destination is memory, and we don't have matching source
10569 operands or we're using the x87, do things in registers. */
10570 matching_memory
= false;
10573 if (use_sse
&& rtx_equal_p (dst
, src
))
10574 matching_memory
= true;
10576 dst
= gen_reg_rtx (mode
);
10578 if (MEM_P (src
) && !matching_memory
)
10579 src
= force_reg (mode
, src
);
10583 set
= gen_rtx_fmt_ee (code
== NEG
? XOR
: AND
, mode
, src
, mask
);
10584 set
= gen_rtx_SET (VOIDmode
, dst
, set
);
10589 set
= gen_rtx_fmt_e (code
, mode
, src
);
10590 set
= gen_rtx_SET (VOIDmode
, dst
, set
);
10593 use
= gen_rtx_USE (VOIDmode
, mask
);
10594 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, FLAGS_REG
));
10595 emit_insn (gen_rtx_PARALLEL (VOIDmode
,
10596 gen_rtvec (3, set
, use
, clob
)));
10602 if (dst
!= operands
[0])
10603 emit_move_insn (operands
[0], dst
);
10606 /* Expand a copysign operation. Special case operand 0 being a constant. */
10609 ix86_expand_copysign (rtx operands
[])
10611 enum machine_mode mode
, vmode
;
10612 rtx dest
, op0
, op1
, mask
, nmask
;
10614 dest
= operands
[0];
10618 mode
= GET_MODE (dest
);
10619 vmode
= mode
== SFmode
? V4SFmode
: V2DFmode
;
10621 if (GET_CODE (op0
) == CONST_DOUBLE
)
10625 if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0
)))
10626 op0
= simplify_unary_operation (ABS
, mode
, op0
, mode
);
10628 if (op0
== CONST0_RTX (mode
))
10629 op0
= CONST0_RTX (vmode
);
10632 if (mode
== SFmode
)
10633 v
= gen_rtvec (4, op0
, CONST0_RTX (SFmode
),
10634 CONST0_RTX (SFmode
), CONST0_RTX (SFmode
));
10636 v
= gen_rtvec (2, op0
, CONST0_RTX (DFmode
));
10637 op0
= force_reg (vmode
, gen_rtx_CONST_VECTOR (vmode
, v
));
10640 mask
= ix86_build_signbit_mask (mode
, 0, 0);
10642 if (mode
== SFmode
)
10643 emit_insn (gen_copysignsf3_const (dest
, op0
, op1
, mask
));
10645 emit_insn (gen_copysigndf3_const (dest
, op0
, op1
, mask
));
10649 nmask
= ix86_build_signbit_mask (mode
, 0, 1);
10650 mask
= ix86_build_signbit_mask (mode
, 0, 0);
10652 if (mode
== SFmode
)
10653 emit_insn (gen_copysignsf3_var (dest
, NULL
, op0
, op1
, nmask
, mask
));
10655 emit_insn (gen_copysigndf3_var (dest
, NULL
, op0
, op1
, nmask
, mask
));
10659 /* Deconstruct a copysign operation into bit masks. Operand 0 is known to
10660 be a constant, and so has already been expanded into a vector constant. */
10663 ix86_split_copysign_const (rtx operands
[])
10665 enum machine_mode mode
, vmode
;
10666 rtx dest
, op0
, op1
, mask
, x
;
10668 dest
= operands
[0];
10671 mask
= operands
[3];
10673 mode
= GET_MODE (dest
);
10674 vmode
= GET_MODE (mask
);
10676 dest
= simplify_gen_subreg (vmode
, dest
, mode
, 0);
10677 x
= gen_rtx_AND (vmode
, dest
, mask
);
10678 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
10680 if (op0
!= CONST0_RTX (vmode
))
10682 x
= gen_rtx_IOR (vmode
, dest
, op0
);
10683 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
10687 /* Deconstruct a copysign operation into bit masks. Operand 0 is variable,
10688 so we have to do two masks. */
10691 ix86_split_copysign_var (rtx operands
[])
10693 enum machine_mode mode
, vmode
;
10694 rtx dest
, scratch
, op0
, op1
, mask
, nmask
, x
;
10696 dest
= operands
[0];
10697 scratch
= operands
[1];
10700 nmask
= operands
[4];
10701 mask
= operands
[5];
10703 mode
= GET_MODE (dest
);
10704 vmode
= GET_MODE (mask
);
10706 if (rtx_equal_p (op0
, op1
))
10708 /* Shouldn't happen often (it's useless, obviously), but when it does
10709 we'd generate incorrect code if we continue below. */
10710 emit_move_insn (dest
, op0
);
10714 if (REG_P (mask
) && REGNO (dest
) == REGNO (mask
)) /* alternative 0 */
10716 gcc_assert (REGNO (op1
) == REGNO (scratch
));
10718 x
= gen_rtx_AND (vmode
, scratch
, mask
);
10719 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, x
));
10722 op0
= simplify_gen_subreg (vmode
, op0
, mode
, 0);
10723 x
= gen_rtx_NOT (vmode
, dest
);
10724 x
= gen_rtx_AND (vmode
, x
, op0
);
10725 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
10729 if (REGNO (op1
) == REGNO (scratch
)) /* alternative 1,3 */
10731 x
= gen_rtx_AND (vmode
, scratch
, mask
);
10733 else /* alternative 2,4 */
10735 gcc_assert (REGNO (mask
) == REGNO (scratch
));
10736 op1
= simplify_gen_subreg (vmode
, op1
, mode
, 0);
10737 x
= gen_rtx_AND (vmode
, scratch
, op1
);
10739 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, x
));
10741 if (REGNO (op0
) == REGNO (dest
)) /* alternative 1,2 */
10743 dest
= simplify_gen_subreg (vmode
, op0
, mode
, 0);
10744 x
= gen_rtx_AND (vmode
, dest
, nmask
);
10746 else /* alternative 3,4 */
10748 gcc_assert (REGNO (nmask
) == REGNO (dest
));
10750 op0
= simplify_gen_subreg (vmode
, op0
, mode
, 0);
10751 x
= gen_rtx_AND (vmode
, dest
, op0
);
10753 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
10756 x
= gen_rtx_IOR (vmode
, dest
, scratch
);
10757 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
10760 /* Return TRUE or FALSE depending on whether the first SET in INSN
10761 has source and destination with matching CC modes, and that the
10762 CC mode is at least as constrained as REQ_MODE. */
10765 ix86_match_ccmode (rtx insn
, enum machine_mode req_mode
)
10768 enum machine_mode set_mode
;
10770 set
= PATTERN (insn
);
10771 if (GET_CODE (set
) == PARALLEL
)
10772 set
= XVECEXP (set
, 0, 0);
10773 gcc_assert (GET_CODE (set
) == SET
);
10774 gcc_assert (GET_CODE (SET_SRC (set
)) == COMPARE
);
10776 set_mode
= GET_MODE (SET_DEST (set
));
10780 if (req_mode
!= CCNOmode
10781 && (req_mode
!= CCmode
10782 || XEXP (SET_SRC (set
), 1) != const0_rtx
))
10786 if (req_mode
== CCGCmode
)
10790 if (req_mode
== CCGOCmode
|| req_mode
== CCNOmode
)
10794 if (req_mode
== CCZmode
)
10801 gcc_unreachable ();
10804 return (GET_MODE (SET_SRC (set
)) == set_mode
);
10807 /* Generate insn patterns to do an integer compare of OPERANDS. */
10810 ix86_expand_int_compare (enum rtx_code code
, rtx op0
, rtx op1
)
10812 enum machine_mode cmpmode
;
10815 cmpmode
= SELECT_CC_MODE (code
, op0
, op1
);
10816 flags
= gen_rtx_REG (cmpmode
, FLAGS_REG
);
10818 /* This is very simple, but making the interface the same as in the
10819 FP case makes the rest of the code easier. */
10820 tmp
= gen_rtx_COMPARE (cmpmode
, op0
, op1
);
10821 emit_insn (gen_rtx_SET (VOIDmode
, flags
, tmp
));
10823 /* Return the test that should be put into the flags user, i.e.
10824 the bcc, scc, or cmov instruction. */
10825 return gen_rtx_fmt_ee (code
, VOIDmode
, flags
, const0_rtx
);
10828 /* Figure out whether to use ordered or unordered fp comparisons.
10829 Return the appropriate mode to use. */
10832 ix86_fp_compare_mode (enum rtx_code code ATTRIBUTE_UNUSED
)
10834 /* ??? In order to make all comparisons reversible, we do all comparisons
10835 non-trapping when compiling for IEEE. Once gcc is able to distinguish
10836 all forms trapping and nontrapping comparisons, we can make inequality
10837 comparisons trapping again, since it results in better code when using
10838 FCOM based compares. */
10839 return TARGET_IEEE_FP
? CCFPUmode
: CCFPmode
;
10843 ix86_cc_mode (enum rtx_code code
, rtx op0
, rtx op1
)
10845 enum machine_mode mode
= GET_MODE (op0
);
10847 if (SCALAR_FLOAT_MODE_P (mode
))
10849 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode
));
10850 return ix86_fp_compare_mode (code
);
10855 /* Only zero flag is needed. */
10856 case EQ
: /* ZF=0 */
10857 case NE
: /* ZF!=0 */
10859 /* Codes needing carry flag. */
10860 case GEU
: /* CF=0 */
10861 case GTU
: /* CF=0 & ZF=0 */
10862 case LTU
: /* CF=1 */
10863 case LEU
: /* CF=1 | ZF=1 */
10865 /* Codes possibly doable only with sign flag when
10866 comparing against zero. */
10867 case GE
: /* SF=OF or SF=0 */
10868 case LT
: /* SF<>OF or SF=1 */
10869 if (op1
== const0_rtx
)
10872 /* For other cases Carry flag is not required. */
10874 /* Codes doable only with sign flag when comparing
10875 against zero, but we miss jump instruction for it
10876 so we need to use relational tests against overflow
10877 that thus needs to be zero. */
10878 case GT
: /* ZF=0 & SF=OF */
10879 case LE
: /* ZF=1 | SF<>OF */
10880 if (op1
== const0_rtx
)
10884 /* strcmp pattern do (use flags) and combine may ask us for proper
10889 gcc_unreachable ();
10893 /* Return the fixed registers used for condition codes. */
10896 ix86_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
10903 /* If two condition code modes are compatible, return a condition code
10904 mode which is compatible with both. Otherwise, return
10907 static enum machine_mode
10908 ix86_cc_modes_compatible (enum machine_mode m1
, enum machine_mode m2
)
10913 if (GET_MODE_CLASS (m1
) != MODE_CC
|| GET_MODE_CLASS (m2
) != MODE_CC
)
10916 if ((m1
== CCGCmode
&& m2
== CCGOCmode
)
10917 || (m1
== CCGOCmode
&& m2
== CCGCmode
))
10923 gcc_unreachable ();
10945 /* These are only compatible with themselves, which we already
10951 /* Split comparison code CODE into comparisons we can do using branch
10952 instructions. BYPASS_CODE is comparison code for branch that will
10953 branch around FIRST_CODE and SECOND_CODE. If some of branches
10954 is not required, set value to UNKNOWN.
10955 We never require more than two branches. */
10958 ix86_fp_comparison_codes (enum rtx_code code
, enum rtx_code
*bypass_code
,
10959 enum rtx_code
*first_code
,
10960 enum rtx_code
*second_code
)
10962 *first_code
= code
;
10963 *bypass_code
= UNKNOWN
;
10964 *second_code
= UNKNOWN
;
10966 /* The fcomi comparison sets flags as follows:
10976 case GT
: /* GTU - CF=0 & ZF=0 */
10977 case GE
: /* GEU - CF=0 */
10978 case ORDERED
: /* PF=0 */
10979 case UNORDERED
: /* PF=1 */
10980 case UNEQ
: /* EQ - ZF=1 */
10981 case UNLT
: /* LTU - CF=1 */
10982 case UNLE
: /* LEU - CF=1 | ZF=1 */
10983 case LTGT
: /* EQ - ZF=0 */
10985 case LT
: /* LTU - CF=1 - fails on unordered */
10986 *first_code
= UNLT
;
10987 *bypass_code
= UNORDERED
;
10989 case LE
: /* LEU - CF=1 | ZF=1 - fails on unordered */
10990 *first_code
= UNLE
;
10991 *bypass_code
= UNORDERED
;
10993 case EQ
: /* EQ - ZF=1 - fails on unordered */
10994 *first_code
= UNEQ
;
10995 *bypass_code
= UNORDERED
;
10997 case NE
: /* NE - ZF=0 - fails on unordered */
10998 *first_code
= LTGT
;
10999 *second_code
= UNORDERED
;
11001 case UNGE
: /* GEU - CF=0 - fails on unordered */
11003 *second_code
= UNORDERED
;
11005 case UNGT
: /* GTU - CF=0 & ZF=0 - fails on unordered */
11007 *second_code
= UNORDERED
;
11010 gcc_unreachable ();
11012 if (!TARGET_IEEE_FP
)
11014 *second_code
= UNKNOWN
;
11015 *bypass_code
= UNKNOWN
;
11019 /* Return cost of comparison done fcom + arithmetics operations on AX.
11020 All following functions do use number of instructions as a cost metrics.
11021 In future this should be tweaked to compute bytes for optimize_size and
11022 take into account performance of various instructions on various CPUs. */
11024 ix86_fp_comparison_arithmetics_cost (enum rtx_code code
)
11026 if (!TARGET_IEEE_FP
)
11028 /* The cost of code output by ix86_expand_fp_compare. */
11052 gcc_unreachable ();
11056 /* Return cost of comparison done using fcomi operation.
11057 See ix86_fp_comparison_arithmetics_cost for the metrics. */
11059 ix86_fp_comparison_fcomi_cost (enum rtx_code code
)
11061 enum rtx_code bypass_code
, first_code
, second_code
;
11062 /* Return arbitrarily high cost when instruction is not supported - this
11063 prevents gcc from using it. */
11066 ix86_fp_comparison_codes (code
, &bypass_code
, &first_code
, &second_code
);
11067 return (bypass_code
!= UNKNOWN
|| second_code
!= UNKNOWN
) + 2;
11070 /* Return cost of comparison done using sahf operation.
11071 See ix86_fp_comparison_arithmetics_cost for the metrics. */
11073 ix86_fp_comparison_sahf_cost (enum rtx_code code
)
11075 enum rtx_code bypass_code
, first_code
, second_code
;
11076 /* Return arbitrarily high cost when instruction is not preferred - this
11077 avoids gcc from using it. */
11078 if (!(TARGET_SAHF
&& (TARGET_USE_SAHF
|| optimize_size
)))
11080 ix86_fp_comparison_codes (code
, &bypass_code
, &first_code
, &second_code
);
11081 return (bypass_code
!= UNKNOWN
|| second_code
!= UNKNOWN
) + 3;
11084 /* Compute cost of the comparison done using any method.
11085 See ix86_fp_comparison_arithmetics_cost for the metrics. */
11087 ix86_fp_comparison_cost (enum rtx_code code
)
11089 int fcomi_cost
, sahf_cost
, arithmetics_cost
= 1024;
11092 fcomi_cost
= ix86_fp_comparison_fcomi_cost (code
);
11093 sahf_cost
= ix86_fp_comparison_sahf_cost (code
);
11095 min
= arithmetics_cost
= ix86_fp_comparison_arithmetics_cost (code
);
11096 if (min
> sahf_cost
)
11098 if (min
> fcomi_cost
)
11103 /* Return true if we should use an FCOMI instruction for this
11107 ix86_use_fcomi_compare (enum rtx_code code ATTRIBUTE_UNUSED
)
11109 enum rtx_code swapped_code
= swap_condition (code
);
11111 return ((ix86_fp_comparison_cost (code
)
11112 == ix86_fp_comparison_fcomi_cost (code
))
11113 || (ix86_fp_comparison_cost (swapped_code
)
11114 == ix86_fp_comparison_fcomi_cost (swapped_code
)));
11117 /* Swap, force into registers, or otherwise massage the two operands
11118 to a fp comparison. The operands are updated in place; the new
11119 comparison code is returned. */
11121 static enum rtx_code
11122 ix86_prepare_fp_compare_args (enum rtx_code code
, rtx
*pop0
, rtx
*pop1
)
11124 enum machine_mode fpcmp_mode
= ix86_fp_compare_mode (code
);
11125 rtx op0
= *pop0
, op1
= *pop1
;
11126 enum machine_mode op_mode
= GET_MODE (op0
);
11127 int is_sse
= TARGET_SSE_MATH
&& SSE_FLOAT_MODE_P (op_mode
);
11129 /* All of the unordered compare instructions only work on registers.
11130 The same is true of the fcomi compare instructions. The XFmode
11131 compare instructions require registers except when comparing
11132 against zero or when converting operand 1 from fixed point to
11136 && (fpcmp_mode
== CCFPUmode
11137 || (op_mode
== XFmode
11138 && ! (standard_80387_constant_p (op0
) == 1
11139 || standard_80387_constant_p (op1
) == 1)
11140 && GET_CODE (op1
) != FLOAT
)
11141 || ix86_use_fcomi_compare (code
)))
11143 op0
= force_reg (op_mode
, op0
);
11144 op1
= force_reg (op_mode
, op1
);
11148 /* %%% We only allow op1 in memory; op0 must be st(0). So swap
11149 things around if they appear profitable, otherwise force op0
11150 into a register. */
11152 if (standard_80387_constant_p (op0
) == 0
11154 && ! (standard_80387_constant_p (op1
) == 0
11158 tmp
= op0
, op0
= op1
, op1
= tmp
;
11159 code
= swap_condition (code
);
11163 op0
= force_reg (op_mode
, op0
);
11165 if (CONSTANT_P (op1
))
11167 int tmp
= standard_80387_constant_p (op1
);
11169 op1
= validize_mem (force_const_mem (op_mode
, op1
));
11173 op1
= force_reg (op_mode
, op1
);
11176 op1
= force_reg (op_mode
, op1
);
11180 /* Try to rearrange the comparison to make it cheaper. */
11181 if (ix86_fp_comparison_cost (code
)
11182 > ix86_fp_comparison_cost (swap_condition (code
))
11183 && (REG_P (op1
) || !no_new_pseudos
))
11186 tmp
= op0
, op0
= op1
, op1
= tmp
;
11187 code
= swap_condition (code
);
11189 op0
= force_reg (op_mode
, op0
);
11197 /* Convert comparison codes we use to represent FP comparison to integer
11198 code that will result in proper branch. Return UNKNOWN if no such code
11202 ix86_fp_compare_code_to_integer (enum rtx_code code
)
11231 /* Generate insn patterns to do a floating point compare of OPERANDS. */
11234 ix86_expand_fp_compare (enum rtx_code code
, rtx op0
, rtx op1
, rtx scratch
,
11235 rtx
*second_test
, rtx
*bypass_test
)
11237 enum machine_mode fpcmp_mode
, intcmp_mode
;
11239 int cost
= ix86_fp_comparison_cost (code
);
11240 enum rtx_code bypass_code
, first_code
, second_code
;
11242 fpcmp_mode
= ix86_fp_compare_mode (code
);
11243 code
= ix86_prepare_fp_compare_args (code
, &op0
, &op1
);
11246 *second_test
= NULL_RTX
;
11248 *bypass_test
= NULL_RTX
;
11250 ix86_fp_comparison_codes (code
, &bypass_code
, &first_code
, &second_code
);
11252 /* Do fcomi/sahf based test when profitable. */
11253 if ((TARGET_CMOVE
|| TARGET_SAHF
)
11254 && (bypass_code
== UNKNOWN
|| bypass_test
)
11255 && (second_code
== UNKNOWN
|| second_test
)
11256 && ix86_fp_comparison_arithmetics_cost (code
) > cost
)
11260 tmp
= gen_rtx_COMPARE (fpcmp_mode
, op0
, op1
);
11261 tmp
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (fpcmp_mode
, FLAGS_REG
),
11267 tmp
= gen_rtx_COMPARE (fpcmp_mode
, op0
, op1
);
11268 tmp2
= gen_rtx_UNSPEC (HImode
, gen_rtvec (1, tmp
), UNSPEC_FNSTSW
);
11270 scratch
= gen_reg_rtx (HImode
);
11271 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, tmp2
));
11272 emit_insn (gen_x86_sahf_1 (scratch
));
11275 /* The FP codes work out to act like unsigned. */
11276 intcmp_mode
= fpcmp_mode
;
11278 if (bypass_code
!= UNKNOWN
)
11279 *bypass_test
= gen_rtx_fmt_ee (bypass_code
, VOIDmode
,
11280 gen_rtx_REG (intcmp_mode
, FLAGS_REG
),
11282 if (second_code
!= UNKNOWN
)
11283 *second_test
= gen_rtx_fmt_ee (second_code
, VOIDmode
,
11284 gen_rtx_REG (intcmp_mode
, FLAGS_REG
),
11289 /* Sadness wrt reg-stack pops killing fpsr -- gotta get fnstsw first. */
11290 tmp
= gen_rtx_COMPARE (fpcmp_mode
, op0
, op1
);
11291 tmp2
= gen_rtx_UNSPEC (HImode
, gen_rtvec (1, tmp
), UNSPEC_FNSTSW
);
11293 scratch
= gen_reg_rtx (HImode
);
11294 emit_insn (gen_rtx_SET (VOIDmode
, scratch
, tmp2
));
11296 /* In the unordered case, we have to check C2 for NaN's, which
11297 doesn't happen to work out to anything nice combination-wise.
11298 So do some bit twiddling on the value we've got in AH to come
11299 up with an appropriate set of condition codes. */
11301 intcmp_mode
= CCNOmode
;
11306 if (code
== GT
|| !TARGET_IEEE_FP
)
11308 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x45)));
11313 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11314 emit_insn (gen_addqi_ext_1 (scratch
, scratch
, constm1_rtx
));
11315 emit_insn (gen_cmpqi_ext_3 (scratch
, GEN_INT (0x44)));
11316 intcmp_mode
= CCmode
;
11322 if (code
== LT
&& TARGET_IEEE_FP
)
11324 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11325 emit_insn (gen_cmpqi_ext_3 (scratch
, GEN_INT (0x01)));
11326 intcmp_mode
= CCmode
;
11331 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x01)));
11337 if (code
== GE
|| !TARGET_IEEE_FP
)
11339 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x05)));
11344 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11345 emit_insn (gen_xorqi_cc_ext_1 (scratch
, scratch
,
11352 if (code
== LE
&& TARGET_IEEE_FP
)
11354 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11355 emit_insn (gen_addqi_ext_1 (scratch
, scratch
, constm1_rtx
));
11356 emit_insn (gen_cmpqi_ext_3 (scratch
, GEN_INT (0x40)));
11357 intcmp_mode
= CCmode
;
11362 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x45)));
11368 if (code
== EQ
&& TARGET_IEEE_FP
)
11370 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11371 emit_insn (gen_cmpqi_ext_3 (scratch
, GEN_INT (0x40)));
11372 intcmp_mode
= CCmode
;
11377 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x40)));
11384 if (code
== NE
&& TARGET_IEEE_FP
)
11386 emit_insn (gen_andqi_ext_0 (scratch
, scratch
, GEN_INT (0x45)));
11387 emit_insn (gen_xorqi_cc_ext_1 (scratch
, scratch
,
11393 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x40)));
11399 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x04)));
11403 emit_insn (gen_testqi_ext_ccno_0 (scratch
, GEN_INT (0x04)));
11408 gcc_unreachable ();
11412 /* Return the test that should be put into the flags user, i.e.
11413 the bcc, scc, or cmov instruction. */
11414 return gen_rtx_fmt_ee (code
, VOIDmode
,
11415 gen_rtx_REG (intcmp_mode
, FLAGS_REG
),
11420 ix86_expand_compare (enum rtx_code code
, rtx
*second_test
, rtx
*bypass_test
)
11423 op0
= ix86_compare_op0
;
11424 op1
= ix86_compare_op1
;
11427 *second_test
= NULL_RTX
;
11429 *bypass_test
= NULL_RTX
;
11431 if (ix86_compare_emitted
)
11433 ret
= gen_rtx_fmt_ee (code
, VOIDmode
, ix86_compare_emitted
, const0_rtx
);
11434 ix86_compare_emitted
= NULL_RTX
;
11436 else if (SCALAR_FLOAT_MODE_P (GET_MODE (op0
)))
11438 gcc_assert (!DECIMAL_FLOAT_MODE_P (GET_MODE (op0
)));
11439 ret
= ix86_expand_fp_compare (code
, op0
, op1
, NULL_RTX
,
11440 second_test
, bypass_test
);
11443 ret
= ix86_expand_int_compare (code
, op0
, op1
);
11448 /* Return true if the CODE will result in nontrivial jump sequence. */
11450 ix86_fp_jump_nontrivial_p (enum rtx_code code
)
11452 enum rtx_code bypass_code
, first_code
, second_code
;
11455 ix86_fp_comparison_codes (code
, &bypass_code
, &first_code
, &second_code
);
11456 return bypass_code
!= UNKNOWN
|| second_code
!= UNKNOWN
;
11460 ix86_expand_branch (enum rtx_code code
, rtx label
)
11464 /* If we have emitted a compare insn, go straight to simple.
11465 ix86_expand_compare won't emit anything if ix86_compare_emitted
11467 if (ix86_compare_emitted
)
11470 switch (GET_MODE (ix86_compare_op0
))
11476 tmp
= ix86_expand_compare (code
, NULL
, NULL
);
11477 tmp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, tmp
,
11478 gen_rtx_LABEL_REF (VOIDmode
, label
),
11480 emit_jump_insn (gen_rtx_SET (VOIDmode
, pc_rtx
, tmp
));
11489 enum rtx_code bypass_code
, first_code
, second_code
;
11491 code
= ix86_prepare_fp_compare_args (code
, &ix86_compare_op0
,
11492 &ix86_compare_op1
);
11494 ix86_fp_comparison_codes (code
, &bypass_code
, &first_code
, &second_code
);
11496 /* Check whether we will use the natural sequence with one jump. If
11497 so, we can expand jump early. Otherwise delay expansion by
11498 creating compound insn to not confuse optimizers. */
11499 if (bypass_code
== UNKNOWN
&& second_code
== UNKNOWN
11502 ix86_split_fp_branch (code
, ix86_compare_op0
, ix86_compare_op1
,
11503 gen_rtx_LABEL_REF (VOIDmode
, label
),
11504 pc_rtx
, NULL_RTX
, NULL_RTX
);
11508 tmp
= gen_rtx_fmt_ee (code
, VOIDmode
,
11509 ix86_compare_op0
, ix86_compare_op1
);
11510 tmp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, tmp
,
11511 gen_rtx_LABEL_REF (VOIDmode
, label
),
11513 tmp
= gen_rtx_SET (VOIDmode
, pc_rtx
, tmp
);
11515 use_fcomi
= ix86_use_fcomi_compare (code
);
11516 vec
= rtvec_alloc (3 + !use_fcomi
);
11517 RTVEC_ELT (vec
, 0) = tmp
;
11519 = gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCFPmode
, 18));
11521 = gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCFPmode
, 17));
11524 = gen_rtx_CLOBBER (VOIDmode
, gen_rtx_SCRATCH (HImode
));
11526 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, vec
));
11535 /* Expand DImode branch into multiple compare+branch. */
11537 rtx lo
[2], hi
[2], label2
;
11538 enum rtx_code code1
, code2
, code3
;
11539 enum machine_mode submode
;
11541 if (CONSTANT_P (ix86_compare_op0
) && ! CONSTANT_P (ix86_compare_op1
))
11543 tmp
= ix86_compare_op0
;
11544 ix86_compare_op0
= ix86_compare_op1
;
11545 ix86_compare_op1
= tmp
;
11546 code
= swap_condition (code
);
11548 if (GET_MODE (ix86_compare_op0
) == DImode
)
11550 split_di (&ix86_compare_op0
, 1, lo
+0, hi
+0);
11551 split_di (&ix86_compare_op1
, 1, lo
+1, hi
+1);
11556 split_ti (&ix86_compare_op0
, 1, lo
+0, hi
+0);
11557 split_ti (&ix86_compare_op1
, 1, lo
+1, hi
+1);
11561 /* When comparing for equality, we can use (hi0^hi1)|(lo0^lo1) to
11562 avoid two branches. This costs one extra insn, so disable when
11563 optimizing for size. */
11565 if ((code
== EQ
|| code
== NE
)
11567 || hi
[1] == const0_rtx
|| lo
[1] == const0_rtx
))
11572 if (hi
[1] != const0_rtx
)
11573 xor1
= expand_binop (submode
, xor_optab
, xor1
, hi
[1],
11574 NULL_RTX
, 0, OPTAB_WIDEN
);
11577 if (lo
[1] != const0_rtx
)
11578 xor0
= expand_binop (submode
, xor_optab
, xor0
, lo
[1],
11579 NULL_RTX
, 0, OPTAB_WIDEN
);
11581 tmp
= expand_binop (submode
, ior_optab
, xor1
, xor0
,
11582 NULL_RTX
, 0, OPTAB_WIDEN
);
11584 ix86_compare_op0
= tmp
;
11585 ix86_compare_op1
= const0_rtx
;
11586 ix86_expand_branch (code
, label
);
11590 /* Otherwise, if we are doing less-than or greater-or-equal-than,
11591 op1 is a constant and the low word is zero, then we can just
11592 examine the high word. */
11594 if (CONST_INT_P (hi
[1]) && lo
[1] == const0_rtx
)
11597 case LT
: case LTU
: case GE
: case GEU
:
11598 ix86_compare_op0
= hi
[0];
11599 ix86_compare_op1
= hi
[1];
11600 ix86_expand_branch (code
, label
);
11606 /* Otherwise, we need two or three jumps. */
11608 label2
= gen_label_rtx ();
11611 code2
= swap_condition (code
);
11612 code3
= unsigned_condition (code
);
11616 case LT
: case GT
: case LTU
: case GTU
:
11619 case LE
: code1
= LT
; code2
= GT
; break;
11620 case GE
: code1
= GT
; code2
= LT
; break;
11621 case LEU
: code1
= LTU
; code2
= GTU
; break;
11622 case GEU
: code1
= GTU
; code2
= LTU
; break;
11624 case EQ
: code1
= UNKNOWN
; code2
= NE
; break;
11625 case NE
: code2
= UNKNOWN
; break;
11628 gcc_unreachable ();
11633 * if (hi(a) < hi(b)) goto true;
11634 * if (hi(a) > hi(b)) goto false;
11635 * if (lo(a) < lo(b)) goto true;
11639 ix86_compare_op0
= hi
[0];
11640 ix86_compare_op1
= hi
[1];
11642 if (code1
!= UNKNOWN
)
11643 ix86_expand_branch (code1
, label
);
11644 if (code2
!= UNKNOWN
)
11645 ix86_expand_branch (code2
, label2
);
11647 ix86_compare_op0
= lo
[0];
11648 ix86_compare_op1
= lo
[1];
11649 ix86_expand_branch (code3
, label
);
11651 if (code2
!= UNKNOWN
)
11652 emit_label (label2
);
11657 gcc_unreachable ();
11661 /* Split branch based on floating point condition. */
11663 ix86_split_fp_branch (enum rtx_code code
, rtx op1
, rtx op2
,
11664 rtx target1
, rtx target2
, rtx tmp
, rtx pushed
)
11666 rtx second
, bypass
;
11667 rtx label
= NULL_RTX
;
11669 int bypass_probability
= -1, second_probability
= -1, probability
= -1;
11672 if (target2
!= pc_rtx
)
11675 code
= reverse_condition_maybe_unordered (code
);
11680 condition
= ix86_expand_fp_compare (code
, op1
, op2
,
11681 tmp
, &second
, &bypass
);
11683 /* Remove pushed operand from stack. */
11685 ix86_free_from_memory (GET_MODE (pushed
));
11687 if (split_branch_probability
>= 0)
11689 /* Distribute the probabilities across the jumps.
11690 Assume the BYPASS and SECOND to be always test
11692 probability
= split_branch_probability
;
11694 /* Value of 1 is low enough to make no need for probability
11695 to be updated. Later we may run some experiments and see
11696 if unordered values are more frequent in practice. */
11698 bypass_probability
= 1;
11700 second_probability
= 1;
11702 if (bypass
!= NULL_RTX
)
11704 label
= gen_label_rtx ();
11705 i
= emit_jump_insn (gen_rtx_SET
11707 gen_rtx_IF_THEN_ELSE (VOIDmode
,
11709 gen_rtx_LABEL_REF (VOIDmode
,
11712 if (bypass_probability
>= 0)
11714 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
11715 GEN_INT (bypass_probability
),
11718 i
= emit_jump_insn (gen_rtx_SET
11720 gen_rtx_IF_THEN_ELSE (VOIDmode
,
11721 condition
, target1
, target2
)));
11722 if (probability
>= 0)
11724 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
11725 GEN_INT (probability
),
11727 if (second
!= NULL_RTX
)
11729 i
= emit_jump_insn (gen_rtx_SET
11731 gen_rtx_IF_THEN_ELSE (VOIDmode
, second
, target1
,
11733 if (second_probability
>= 0)
11735 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
11736 GEN_INT (second_probability
),
11739 if (label
!= NULL_RTX
)
11740 emit_label (label
);
11744 ix86_expand_setcc (enum rtx_code code
, rtx dest
)
11746 rtx ret
, tmp
, tmpreg
, equiv
;
11747 rtx second_test
, bypass_test
;
11749 if (GET_MODE (ix86_compare_op0
) == (TARGET_64BIT
? TImode
: DImode
))
11750 return 0; /* FAIL */
11752 gcc_assert (GET_MODE (dest
) == QImode
);
11754 ret
= ix86_expand_compare (code
, &second_test
, &bypass_test
);
11755 PUT_MODE (ret
, QImode
);
11760 emit_insn (gen_rtx_SET (VOIDmode
, tmp
, ret
));
11761 if (bypass_test
|| second_test
)
11763 rtx test
= second_test
;
11765 rtx tmp2
= gen_reg_rtx (QImode
);
11768 gcc_assert (!second_test
);
11769 test
= bypass_test
;
11771 PUT_CODE (test
, reverse_condition_maybe_unordered (GET_CODE (test
)));
11773 PUT_MODE (test
, QImode
);
11774 emit_insn (gen_rtx_SET (VOIDmode
, tmp2
, test
));
11777 emit_insn (gen_andqi3 (tmp
, tmpreg
, tmp2
));
11779 emit_insn (gen_iorqi3 (tmp
, tmpreg
, tmp2
));
11782 /* Attach a REG_EQUAL note describing the comparison result. */
11783 if (ix86_compare_op0
&& ix86_compare_op1
)
11785 equiv
= simplify_gen_relational (code
, QImode
,
11786 GET_MODE (ix86_compare_op0
),
11787 ix86_compare_op0
, ix86_compare_op1
);
11788 set_unique_reg_note (get_last_insn (), REG_EQUAL
, equiv
);
11791 return 1; /* DONE */
11794 /* Expand comparison setting or clearing carry flag. Return true when
11795 successful and set pop for the operation. */
11797 ix86_expand_carry_flag_compare (enum rtx_code code
, rtx op0
, rtx op1
, rtx
*pop
)
11799 enum machine_mode mode
=
11800 GET_MODE (op0
) != VOIDmode
? GET_MODE (op0
) : GET_MODE (op1
);
11802 /* Do not handle DImode compares that go through special path.
11803 Also we can't deal with FP compares yet. This is possible to add. */
11804 if (mode
== (TARGET_64BIT
? TImode
: DImode
))
11807 if (SCALAR_FLOAT_MODE_P (mode
))
11809 rtx second_test
= NULL
, bypass_test
= NULL
;
11810 rtx compare_op
, compare_seq
;
11812 gcc_assert (!DECIMAL_FLOAT_MODE_P (mode
));
11814 /* Shortcut: following common codes never translate
11815 into carry flag compares. */
11816 if (code
== EQ
|| code
== NE
|| code
== UNEQ
|| code
== LTGT
11817 || code
== ORDERED
|| code
== UNORDERED
)
11820 /* These comparisons require zero flag; swap operands so they won't. */
11821 if ((code
== GT
|| code
== UNLE
|| code
== LE
|| code
== UNGT
)
11822 && !TARGET_IEEE_FP
)
11827 code
= swap_condition (code
);
11830 /* Try to expand the comparison and verify that we end up with carry flag
11831 based comparison. This is fails to be true only when we decide to expand
11832 comparison using arithmetic that is not too common scenario. */
11834 compare_op
= ix86_expand_fp_compare (code
, op0
, op1
, NULL_RTX
,
11835 &second_test
, &bypass_test
);
11836 compare_seq
= get_insns ();
11839 if (second_test
|| bypass_test
)
11841 if (GET_MODE (XEXP (compare_op
, 0)) == CCFPmode
11842 || GET_MODE (XEXP (compare_op
, 0)) == CCFPUmode
)
11843 code
= ix86_fp_compare_code_to_integer (GET_CODE (compare_op
));
11845 code
= GET_CODE (compare_op
);
11846 if (code
!= LTU
&& code
!= GEU
)
11848 emit_insn (compare_seq
);
11852 if (!INTEGRAL_MODE_P (mode
))
11860 /* Convert a==0 into (unsigned)a<1. */
11863 if (op1
!= const0_rtx
)
11866 code
= (code
== EQ
? LTU
: GEU
);
11869 /* Convert a>b into b<a or a>=b-1. */
11872 if (CONST_INT_P (op1
))
11874 op1
= gen_int_mode (INTVAL (op1
) + 1, GET_MODE (op0
));
11875 /* Bail out on overflow. We still can swap operands but that
11876 would force loading of the constant into register. */
11877 if (op1
== const0_rtx
11878 || !x86_64_immediate_operand (op1
, GET_MODE (op1
)))
11880 code
= (code
== GTU
? GEU
: LTU
);
11887 code
= (code
== GTU
? LTU
: GEU
);
11891 /* Convert a>=0 into (unsigned)a<0x80000000. */
11894 if (mode
== DImode
|| op1
!= const0_rtx
)
11896 op1
= gen_int_mode (1 << (GET_MODE_BITSIZE (mode
) - 1), mode
);
11897 code
= (code
== LT
? GEU
: LTU
);
11901 if (mode
== DImode
|| op1
!= constm1_rtx
)
11903 op1
= gen_int_mode (1 << (GET_MODE_BITSIZE (mode
) - 1), mode
);
11904 code
= (code
== LE
? GEU
: LTU
);
11910 /* Swapping operands may cause constant to appear as first operand. */
11911 if (!nonimmediate_operand (op0
, VOIDmode
))
11913 if (no_new_pseudos
)
11915 op0
= force_reg (mode
, op0
);
11917 ix86_compare_op0
= op0
;
11918 ix86_compare_op1
= op1
;
11919 *pop
= ix86_expand_compare (code
, NULL
, NULL
);
11920 gcc_assert (GET_CODE (*pop
) == LTU
|| GET_CODE (*pop
) == GEU
);
11925 ix86_expand_int_movcc (rtx operands
[])
11927 enum rtx_code code
= GET_CODE (operands
[1]), compare_code
;
11928 rtx compare_seq
, compare_op
;
11929 rtx second_test
, bypass_test
;
11930 enum machine_mode mode
= GET_MODE (operands
[0]);
11931 bool sign_bit_compare_p
= false;;
11934 compare_op
= ix86_expand_compare (code
, &second_test
, &bypass_test
);
11935 compare_seq
= get_insns ();
11938 compare_code
= GET_CODE (compare_op
);
11940 if ((ix86_compare_op1
== const0_rtx
&& (code
== GE
|| code
== LT
))
11941 || (ix86_compare_op1
== constm1_rtx
&& (code
== GT
|| code
== LE
)))
11942 sign_bit_compare_p
= true;
11944 /* Don't attempt mode expansion here -- if we had to expand 5 or 6
11945 HImode insns, we'd be swallowed in word prefix ops. */
11947 if ((mode
!= HImode
|| TARGET_FAST_PREFIX
)
11948 && (mode
!= (TARGET_64BIT
? TImode
: DImode
))
11949 && CONST_INT_P (operands
[2])
11950 && CONST_INT_P (operands
[3]))
11952 rtx out
= operands
[0];
11953 HOST_WIDE_INT ct
= INTVAL (operands
[2]);
11954 HOST_WIDE_INT cf
= INTVAL (operands
[3]);
11955 HOST_WIDE_INT diff
;
11958 /* Sign bit compares are better done using shifts than we do by using
11960 if (sign_bit_compare_p
11961 || ix86_expand_carry_flag_compare (code
, ix86_compare_op0
,
11962 ix86_compare_op1
, &compare_op
))
11964 /* Detect overlap between destination and compare sources. */
11967 if (!sign_bit_compare_p
)
11969 bool fpcmp
= false;
11971 compare_code
= GET_CODE (compare_op
);
11973 if (GET_MODE (XEXP (compare_op
, 0)) == CCFPmode
11974 || GET_MODE (XEXP (compare_op
, 0)) == CCFPUmode
)
11977 compare_code
= ix86_fp_compare_code_to_integer (compare_code
);
11980 /* To simplify rest of code, restrict to the GEU case. */
11981 if (compare_code
== LTU
)
11983 HOST_WIDE_INT tmp
= ct
;
11986 compare_code
= reverse_condition (compare_code
);
11987 code
= reverse_condition (code
);
11992 PUT_CODE (compare_op
,
11993 reverse_condition_maybe_unordered
11994 (GET_CODE (compare_op
)));
11996 PUT_CODE (compare_op
, reverse_condition (GET_CODE (compare_op
)));
12000 if (reg_overlap_mentioned_p (out
, ix86_compare_op0
)
12001 || reg_overlap_mentioned_p (out
, ix86_compare_op1
))
12002 tmp
= gen_reg_rtx (mode
);
12004 if (mode
== DImode
)
12005 emit_insn (gen_x86_movdicc_0_m1_rex64 (tmp
, compare_op
));
12007 emit_insn (gen_x86_movsicc_0_m1 (gen_lowpart (SImode
, tmp
), compare_op
));
12011 if (code
== GT
|| code
== GE
)
12012 code
= reverse_condition (code
);
12015 HOST_WIDE_INT tmp
= ct
;
12020 tmp
= emit_store_flag (tmp
, code
, ix86_compare_op0
,
12021 ix86_compare_op1
, VOIDmode
, 0, -1);
12034 tmp
= expand_simple_binop (mode
, PLUS
,
12036 copy_rtx (tmp
), 1, OPTAB_DIRECT
);
12047 tmp
= expand_simple_binop (mode
, IOR
,
12049 copy_rtx (tmp
), 1, OPTAB_DIRECT
);
12051 else if (diff
== -1 && ct
)
12061 tmp
= expand_simple_unop (mode
, NOT
, tmp
, copy_rtx (tmp
), 1);
12063 tmp
= expand_simple_binop (mode
, PLUS
,
12064 copy_rtx (tmp
), GEN_INT (cf
),
12065 copy_rtx (tmp
), 1, OPTAB_DIRECT
);
12073 * andl cf - ct, dest
12083 tmp
= expand_simple_unop (mode
, NOT
, tmp
, copy_rtx (tmp
), 1);
12086 tmp
= expand_simple_binop (mode
, AND
,
12088 gen_int_mode (cf
- ct
, mode
),
12089 copy_rtx (tmp
), 1, OPTAB_DIRECT
);
12091 tmp
= expand_simple_binop (mode
, PLUS
,
12092 copy_rtx (tmp
), GEN_INT (ct
),
12093 copy_rtx (tmp
), 1, OPTAB_DIRECT
);
12096 if (!rtx_equal_p (tmp
, out
))
12097 emit_move_insn (copy_rtx (out
), copy_rtx (tmp
));
12099 return 1; /* DONE */
12104 enum machine_mode cmp_mode
= GET_MODE (ix86_compare_op0
);
12107 tmp
= ct
, ct
= cf
, cf
= tmp
;
12110 if (SCALAR_FLOAT_MODE_P (cmp_mode
))
12112 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode
));
12114 /* We may be reversing unordered compare to normal compare, that
12115 is not valid in general (we may convert non-trapping condition
12116 to trapping one), however on i386 we currently emit all
12117 comparisons unordered. */
12118 compare_code
= reverse_condition_maybe_unordered (compare_code
);
12119 code
= reverse_condition_maybe_unordered (code
);
12123 compare_code
= reverse_condition (compare_code
);
12124 code
= reverse_condition (code
);
12128 compare_code
= UNKNOWN
;
12129 if (GET_MODE_CLASS (GET_MODE (ix86_compare_op0
)) == MODE_INT
12130 && CONST_INT_P (ix86_compare_op1
))
12132 if (ix86_compare_op1
== const0_rtx
12133 && (code
== LT
|| code
== GE
))
12134 compare_code
= code
;
12135 else if (ix86_compare_op1
== constm1_rtx
)
12139 else if (code
== GT
)
12144 /* Optimize dest = (op0 < 0) ? -1 : cf. */
12145 if (compare_code
!= UNKNOWN
12146 && GET_MODE (ix86_compare_op0
) == GET_MODE (out
)
12147 && (cf
== -1 || ct
== -1))
12149 /* If lea code below could be used, only optimize
12150 if it results in a 2 insn sequence. */
12152 if (! (diff
== 1 || diff
== 2 || diff
== 4 || diff
== 8
12153 || diff
== 3 || diff
== 5 || diff
== 9)
12154 || (compare_code
== LT
&& ct
== -1)
12155 || (compare_code
== GE
&& cf
== -1))
12158 * notl op1 (if necessary)
12166 code
= reverse_condition (code
);
12169 out
= emit_store_flag (out
, code
, ix86_compare_op0
,
12170 ix86_compare_op1
, VOIDmode
, 0, -1);
12172 out
= expand_simple_binop (mode
, IOR
,
12174 out
, 1, OPTAB_DIRECT
);
12175 if (out
!= operands
[0])
12176 emit_move_insn (operands
[0], out
);
12178 return 1; /* DONE */
12183 if ((diff
== 1 || diff
== 2 || diff
== 4 || diff
== 8
12184 || diff
== 3 || diff
== 5 || diff
== 9)
12185 && ((mode
!= QImode
&& mode
!= HImode
) || !TARGET_PARTIAL_REG_STALL
)
12187 || x86_64_immediate_operand (GEN_INT (cf
), VOIDmode
)))
12193 * lea cf(dest*(ct-cf)),dest
12197 * This also catches the degenerate setcc-only case.
12203 out
= emit_store_flag (out
, code
, ix86_compare_op0
,
12204 ix86_compare_op1
, VOIDmode
, 0, 1);
12207 /* On x86_64 the lea instruction operates on Pmode, so we need
12208 to get arithmetics done in proper mode to match. */
12210 tmp
= copy_rtx (out
);
12214 out1
= copy_rtx (out
);
12215 tmp
= gen_rtx_MULT (mode
, out1
, GEN_INT (diff
& ~1));
12219 tmp
= gen_rtx_PLUS (mode
, tmp
, out1
);
12225 tmp
= gen_rtx_PLUS (mode
, tmp
, GEN_INT (cf
));
12228 if (!rtx_equal_p (tmp
, out
))
12231 out
= force_operand (tmp
, copy_rtx (out
));
12233 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (out
), copy_rtx (tmp
)));
12235 if (!rtx_equal_p (out
, operands
[0]))
12236 emit_move_insn (operands
[0], copy_rtx (out
));
12238 return 1; /* DONE */
12242 * General case: Jumpful:
12243 * xorl dest,dest cmpl op1, op2
12244 * cmpl op1, op2 movl ct, dest
12245 * setcc dest jcc 1f
12246 * decl dest movl cf, dest
12247 * andl (cf-ct),dest 1:
12250 * Size 20. Size 14.
12252 * This is reasonably steep, but branch mispredict costs are
12253 * high on modern cpus, so consider failing only if optimizing
12257 if ((!TARGET_CMOVE
|| (mode
== QImode
&& TARGET_PARTIAL_REG_STALL
))
12258 && BRANCH_COST
>= 2)
12262 enum machine_mode cmp_mode
= GET_MODE (ix86_compare_op0
);
12267 if (SCALAR_FLOAT_MODE_P (cmp_mode
))
12269 gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode
));
12271 /* We may be reversing unordered compare to normal compare,
12272 that is not valid in general (we may convert non-trapping
12273 condition to trapping one), however on i386 we currently
12274 emit all comparisons unordered. */
12275 code
= reverse_condition_maybe_unordered (code
);
12279 code
= reverse_condition (code
);
12280 if (compare_code
!= UNKNOWN
)
12281 compare_code
= reverse_condition (compare_code
);
12285 if (compare_code
!= UNKNOWN
)
12287 /* notl op1 (if needed)
12292 For x < 0 (resp. x <= -1) there will be no notl,
12293 so if possible swap the constants to get rid of the
12295 True/false will be -1/0 while code below (store flag
12296 followed by decrement) is 0/-1, so the constants need
12297 to be exchanged once more. */
12299 if (compare_code
== GE
|| !cf
)
12301 code
= reverse_condition (code
);
12306 HOST_WIDE_INT tmp
= cf
;
12311 out
= emit_store_flag (out
, code
, ix86_compare_op0
,
12312 ix86_compare_op1
, VOIDmode
, 0, -1);
12316 out
= emit_store_flag (out
, code
, ix86_compare_op0
,
12317 ix86_compare_op1
, VOIDmode
, 0, 1);
12319 out
= expand_simple_binop (mode
, PLUS
, copy_rtx (out
), constm1_rtx
,
12320 copy_rtx (out
), 1, OPTAB_DIRECT
);
12323 out
= expand_simple_binop (mode
, AND
, copy_rtx (out
),
12324 gen_int_mode (cf
- ct
, mode
),
12325 copy_rtx (out
), 1, OPTAB_DIRECT
);
12327 out
= expand_simple_binop (mode
, PLUS
, copy_rtx (out
), GEN_INT (ct
),
12328 copy_rtx (out
), 1, OPTAB_DIRECT
);
12329 if (!rtx_equal_p (out
, operands
[0]))
12330 emit_move_insn (operands
[0], copy_rtx (out
));
12332 return 1; /* DONE */
12336 if (!TARGET_CMOVE
|| (mode
== QImode
&& TARGET_PARTIAL_REG_STALL
))
12338 /* Try a few things more with specific constants and a variable. */
12341 rtx var
, orig_out
, out
, tmp
;
12343 if (BRANCH_COST
<= 2)
12344 return 0; /* FAIL */
12346 /* If one of the two operands is an interesting constant, load a
12347 constant with the above and mask it in with a logical operation. */
12349 if (CONST_INT_P (operands
[2]))
12352 if (INTVAL (operands
[2]) == 0 && operands
[3] != constm1_rtx
)
12353 operands
[3] = constm1_rtx
, op
= and_optab
;
12354 else if (INTVAL (operands
[2]) == -1 && operands
[3] != const0_rtx
)
12355 operands
[3] = const0_rtx
, op
= ior_optab
;
12357 return 0; /* FAIL */
12359 else if (CONST_INT_P (operands
[3]))
12362 if (INTVAL (operands
[3]) == 0 && operands
[2] != constm1_rtx
)
12363 operands
[2] = constm1_rtx
, op
= and_optab
;
12364 else if (INTVAL (operands
[3]) == -1 && operands
[3] != const0_rtx
)
12365 operands
[2] = const0_rtx
, op
= ior_optab
;
12367 return 0; /* FAIL */
12370 return 0; /* FAIL */
12372 orig_out
= operands
[0];
12373 tmp
= gen_reg_rtx (mode
);
12376 /* Recurse to get the constant loaded. */
12377 if (ix86_expand_int_movcc (operands
) == 0)
12378 return 0; /* FAIL */
12380 /* Mask in the interesting variable. */
12381 out
= expand_binop (mode
, op
, var
, tmp
, orig_out
, 0,
12383 if (!rtx_equal_p (out
, orig_out
))
12384 emit_move_insn (copy_rtx (orig_out
), copy_rtx (out
));
12386 return 1; /* DONE */
12390 * For comparison with above,
12400 if (! nonimmediate_operand (operands
[2], mode
))
12401 operands
[2] = force_reg (mode
, operands
[2]);
12402 if (! nonimmediate_operand (operands
[3], mode
))
12403 operands
[3] = force_reg (mode
, operands
[3]);
12405 if (bypass_test
&& reg_overlap_mentioned_p (operands
[0], operands
[3]))
12407 rtx tmp
= gen_reg_rtx (mode
);
12408 emit_move_insn (tmp
, operands
[3]);
12411 if (second_test
&& reg_overlap_mentioned_p (operands
[0], operands
[2]))
12413 rtx tmp
= gen_reg_rtx (mode
);
12414 emit_move_insn (tmp
, operands
[2]);
12418 if (! register_operand (operands
[2], VOIDmode
)
12420 || ! register_operand (operands
[3], VOIDmode
)))
12421 operands
[2] = force_reg (mode
, operands
[2]);
12424 && ! register_operand (operands
[3], VOIDmode
))
12425 operands
[3] = force_reg (mode
, operands
[3]);
12427 emit_insn (compare_seq
);
12428 emit_insn (gen_rtx_SET (VOIDmode
, operands
[0],
12429 gen_rtx_IF_THEN_ELSE (mode
,
12430 compare_op
, operands
[2],
12433 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (operands
[0]),
12434 gen_rtx_IF_THEN_ELSE (mode
,
12436 copy_rtx (operands
[3]),
12437 copy_rtx (operands
[0]))));
12439 emit_insn (gen_rtx_SET (VOIDmode
, copy_rtx (operands
[0]),
12440 gen_rtx_IF_THEN_ELSE (mode
,
12442 copy_rtx (operands
[2]),
12443 copy_rtx (operands
[0]))));
12445 return 1; /* DONE */
12448 /* Swap, force into registers, or otherwise massage the two operands
12449 to an sse comparison with a mask result. Thus we differ a bit from
12450 ix86_prepare_fp_compare_args which expects to produce a flags result.
12452 The DEST operand exists to help determine whether to commute commutative
12453 operators. The POP0/POP1 operands are updated in place. The new
12454 comparison code is returned, or UNKNOWN if not implementable. */
12456 static enum rtx_code
12457 ix86_prepare_sse_fp_compare_args (rtx dest
, enum rtx_code code
,
12458 rtx
*pop0
, rtx
*pop1
)
12466 /* We have no LTGT as an operator. We could implement it with
12467 NE & ORDERED, but this requires an extra temporary. It's
12468 not clear that it's worth it. */
12475 /* These are supported directly. */
12482 /* For commutative operators, try to canonicalize the destination
12483 operand to be first in the comparison - this helps reload to
12484 avoid extra moves. */
12485 if (!dest
|| !rtx_equal_p (dest
, *pop1
))
12493 /* These are not supported directly. Swap the comparison operands
12494 to transform into something that is supported. */
12498 code
= swap_condition (code
);
12502 gcc_unreachable ();
12508 /* Detect conditional moves that exactly match min/max operational
12509 semantics. Note that this is IEEE safe, as long as we don't
12510 interchange the operands.
12512 Returns FALSE if this conditional move doesn't match a MIN/MAX,
12513 and TRUE if the operation is successful and instructions are emitted. */
12516 ix86_expand_sse_fp_minmax (rtx dest
, enum rtx_code code
, rtx cmp_op0
,
12517 rtx cmp_op1
, rtx if_true
, rtx if_false
)
12519 enum machine_mode mode
;
12525 else if (code
== UNGE
)
12528 if_true
= if_false
;
12534 if (rtx_equal_p (cmp_op0
, if_true
) && rtx_equal_p (cmp_op1
, if_false
))
12536 else if (rtx_equal_p (cmp_op1
, if_true
) && rtx_equal_p (cmp_op0
, if_false
))
12541 mode
= GET_MODE (dest
);
12543 /* We want to check HONOR_NANS and HONOR_SIGNED_ZEROS here,
12544 but MODE may be a vector mode and thus not appropriate. */
12545 if (!flag_finite_math_only
|| !flag_unsafe_math_optimizations
)
12547 int u
= is_min
? UNSPEC_IEEE_MIN
: UNSPEC_IEEE_MAX
;
12550 if_true
= force_reg (mode
, if_true
);
12551 v
= gen_rtvec (2, if_true
, if_false
);
12552 tmp
= gen_rtx_UNSPEC (mode
, v
, u
);
12556 code
= is_min
? SMIN
: SMAX
;
12557 tmp
= gen_rtx_fmt_ee (code
, mode
, if_true
, if_false
);
12560 emit_insn (gen_rtx_SET (VOIDmode
, dest
, tmp
));
12564 /* Expand an sse vector comparison. Return the register with the result. */
12567 ix86_expand_sse_cmp (rtx dest
, enum rtx_code code
, rtx cmp_op0
, rtx cmp_op1
,
12568 rtx op_true
, rtx op_false
)
12570 enum machine_mode mode
= GET_MODE (dest
);
12573 cmp_op0
= force_reg (mode
, cmp_op0
);
12574 if (!nonimmediate_operand (cmp_op1
, mode
))
12575 cmp_op1
= force_reg (mode
, cmp_op1
);
12578 || reg_overlap_mentioned_p (dest
, op_true
)
12579 || reg_overlap_mentioned_p (dest
, op_false
))
12580 dest
= gen_reg_rtx (mode
);
12582 x
= gen_rtx_fmt_ee (code
, mode
, cmp_op0
, cmp_op1
);
12583 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
12588 /* Expand DEST = CMP ? OP_TRUE : OP_FALSE into a sequence of logical
12589 operations. This is used for both scalar and vector conditional moves. */
12592 ix86_expand_sse_movcc (rtx dest
, rtx cmp
, rtx op_true
, rtx op_false
)
12594 enum machine_mode mode
= GET_MODE (dest
);
12597 if (op_false
== CONST0_RTX (mode
))
12599 op_true
= force_reg (mode
, op_true
);
12600 x
= gen_rtx_AND (mode
, cmp
, op_true
);
12601 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
12603 else if (op_true
== CONST0_RTX (mode
))
12605 op_false
= force_reg (mode
, op_false
);
12606 x
= gen_rtx_NOT (mode
, cmp
);
12607 x
= gen_rtx_AND (mode
, x
, op_false
);
12608 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
12612 op_true
= force_reg (mode
, op_true
);
12613 op_false
= force_reg (mode
, op_false
);
12615 t2
= gen_reg_rtx (mode
);
12617 t3
= gen_reg_rtx (mode
);
12621 x
= gen_rtx_AND (mode
, op_true
, cmp
);
12622 emit_insn (gen_rtx_SET (VOIDmode
, t2
, x
));
12624 x
= gen_rtx_NOT (mode
, cmp
);
12625 x
= gen_rtx_AND (mode
, x
, op_false
);
12626 emit_insn (gen_rtx_SET (VOIDmode
, t3
, x
));
12628 x
= gen_rtx_IOR (mode
, t3
, t2
);
12629 emit_insn (gen_rtx_SET (VOIDmode
, dest
, x
));
12633 /* Expand a floating-point conditional move. Return true if successful. */
12636 ix86_expand_fp_movcc (rtx operands
[])
12638 enum machine_mode mode
= GET_MODE (operands
[0]);
12639 enum rtx_code code
= GET_CODE (operands
[1]);
12640 rtx tmp
, compare_op
, second_test
, bypass_test
;
12642 if (TARGET_SSE_MATH
&& SSE_FLOAT_MODE_P (mode
))
12644 enum machine_mode cmode
;
12646 /* Since we've no cmove for sse registers, don't force bad register
12647 allocation just to gain access to it. Deny movcc when the
12648 comparison mode doesn't match the move mode. */
12649 cmode
= GET_MODE (ix86_compare_op0
);
12650 if (cmode
== VOIDmode
)
12651 cmode
= GET_MODE (ix86_compare_op1
);
12655 code
= ix86_prepare_sse_fp_compare_args (operands
[0], code
,
12657 &ix86_compare_op1
);
12658 if (code
== UNKNOWN
)
12661 if (ix86_expand_sse_fp_minmax (operands
[0], code
, ix86_compare_op0
,
12662 ix86_compare_op1
, operands
[2],
12666 tmp
= ix86_expand_sse_cmp (operands
[0], code
, ix86_compare_op0
,
12667 ix86_compare_op1
, operands
[2], operands
[3]);
12668 ix86_expand_sse_movcc (operands
[0], tmp
, operands
[2], operands
[3]);
12672 /* The floating point conditional move instructions don't directly
12673 support conditions resulting from a signed integer comparison. */
12675 compare_op
= ix86_expand_compare (code
, &second_test
, &bypass_test
);
12677 /* The floating point conditional move instructions don't directly
12678 support signed integer comparisons. */
12680 if (!fcmov_comparison_operator (compare_op
, VOIDmode
))
12682 gcc_assert (!second_test
&& !bypass_test
);
12683 tmp
= gen_reg_rtx (QImode
);
12684 ix86_expand_setcc (code
, tmp
);
12686 ix86_compare_op0
= tmp
;
12687 ix86_compare_op1
= const0_rtx
;
12688 compare_op
= ix86_expand_compare (code
, &second_test
, &bypass_test
);
12690 if (bypass_test
&& reg_overlap_mentioned_p (operands
[0], operands
[3]))
12692 tmp
= gen_reg_rtx (mode
);
12693 emit_move_insn (tmp
, operands
[3]);
12696 if (second_test
&& reg_overlap_mentioned_p (operands
[0], operands
[2]))
12698 tmp
= gen_reg_rtx (mode
);
12699 emit_move_insn (tmp
, operands
[2]);
12703 emit_insn (gen_rtx_SET (VOIDmode
, operands
[0],
12704 gen_rtx_IF_THEN_ELSE (mode
, compare_op
,
12705 operands
[2], operands
[3])));
12707 emit_insn (gen_rtx_SET (VOIDmode
, operands
[0],
12708 gen_rtx_IF_THEN_ELSE (mode
, bypass_test
,
12709 operands
[3], operands
[0])));
12711 emit_insn (gen_rtx_SET (VOIDmode
, operands
[0],
12712 gen_rtx_IF_THEN_ELSE (mode
, second_test
,
12713 operands
[2], operands
[0])));
12718 /* Expand a floating-point vector conditional move; a vcond operation
12719 rather than a movcc operation. */
12722 ix86_expand_fp_vcond (rtx operands
[])
12724 enum rtx_code code
= GET_CODE (operands
[3]);
12727 code
= ix86_prepare_sse_fp_compare_args (operands
[0], code
,
12728 &operands
[4], &operands
[5]);
12729 if (code
== UNKNOWN
)
12732 if (ix86_expand_sse_fp_minmax (operands
[0], code
, operands
[4],
12733 operands
[5], operands
[1], operands
[2]))
12736 cmp
= ix86_expand_sse_cmp (operands
[0], code
, operands
[4], operands
[5],
12737 operands
[1], operands
[2]);
12738 ix86_expand_sse_movcc (operands
[0], cmp
, operands
[1], operands
[2]);
12742 /* Expand a signed integral vector conditional move. */
12745 ix86_expand_int_vcond (rtx operands
[])
12747 enum machine_mode mode
= GET_MODE (operands
[0]);
12748 enum rtx_code code
= GET_CODE (operands
[3]);
12749 bool negate
= false;
12752 cop0
= operands
[4];
12753 cop1
= operands
[5];
12755 /* Canonicalize the comparison to EQ, GT, GTU. */
12766 code
= reverse_condition (code
);
12772 code
= reverse_condition (code
);
12778 code
= swap_condition (code
);
12779 x
= cop0
, cop0
= cop1
, cop1
= x
;
12783 gcc_unreachable ();
12786 /* Unsigned parallel compare is not supported by the hardware. Play some
12787 tricks to turn this into a signed comparison against 0. */
12790 cop0
= force_reg (mode
, cop0
);
12798 /* Perform a parallel modulo subtraction. */
12799 t1
= gen_reg_rtx (mode
);
12800 emit_insn (gen_subv4si3 (t1
, cop0
, cop1
));
12802 /* Extract the original sign bit of op0. */
12803 mask
= GEN_INT (-0x80000000);
12804 mask
= gen_rtx_CONST_VECTOR (mode
,
12805 gen_rtvec (4, mask
, mask
, mask
, mask
));
12806 mask
= force_reg (mode
, mask
);
12807 t2
= gen_reg_rtx (mode
);
12808 emit_insn (gen_andv4si3 (t2
, cop0
, mask
));
12810 /* XOR it back into the result of the subtraction. This results
12811 in the sign bit set iff we saw unsigned underflow. */
12812 x
= gen_reg_rtx (mode
);
12813 emit_insn (gen_xorv4si3 (x
, t1
, t2
));
12821 /* Perform a parallel unsigned saturating subtraction. */
12822 x
= gen_reg_rtx (mode
);
12823 emit_insn (gen_rtx_SET (VOIDmode
, x
,
12824 gen_rtx_US_MINUS (mode
, cop0
, cop1
)));
12831 gcc_unreachable ();
12835 cop1
= CONST0_RTX (mode
);
12838 x
= ix86_expand_sse_cmp (operands
[0], code
, cop0
, cop1
,
12839 operands
[1+negate
], operands
[2-negate
]);
12841 ix86_expand_sse_movcc (operands
[0], x
, operands
[1+negate
],
12842 operands
[2-negate
]);
12846 /* Unpack OP[1] into the next wider integer vector type. UNSIGNED_P is
12847 true if we should do zero extension, else sign extension. HIGH_P is
12848 true if we want the N/2 high elements, else the low elements. */
12851 ix86_expand_sse_unpack (rtx operands
[2], bool unsigned_p
, bool high_p
)
12853 enum machine_mode imode
= GET_MODE (operands
[1]);
12854 rtx (*unpack
)(rtx
, rtx
, rtx
);
12861 unpack
= gen_vec_interleave_highv16qi
;
12863 unpack
= gen_vec_interleave_lowv16qi
;
12867 unpack
= gen_vec_interleave_highv8hi
;
12869 unpack
= gen_vec_interleave_lowv8hi
;
12873 unpack
= gen_vec_interleave_highv4si
;
12875 unpack
= gen_vec_interleave_lowv4si
;
12878 gcc_unreachable ();
12881 dest
= gen_lowpart (imode
, operands
[0]);
12884 se
= force_reg (imode
, CONST0_RTX (imode
));
12886 se
= ix86_expand_sse_cmp (gen_reg_rtx (imode
), GT
, CONST0_RTX (imode
),
12887 operands
[1], pc_rtx
, pc_rtx
);
12889 emit_insn (unpack (dest
, operands
[1], se
));
12892 /* This function performs the same task as ix86_expand_sse_unpack,
12893 but with SSE4.1 instructions. */
12896 ix86_expand_sse4_unpack (rtx operands
[2], bool unsigned_p
, bool high_p
)
12898 enum machine_mode imode
= GET_MODE (operands
[1]);
12899 rtx (*unpack
)(rtx
, rtx
);
12906 unpack
= gen_sse4_1_zero_extendv8qiv8hi2
;
12908 unpack
= gen_sse4_1_extendv8qiv8hi2
;
12912 unpack
= gen_sse4_1_zero_extendv4hiv4si2
;
12914 unpack
= gen_sse4_1_extendv4hiv4si2
;
12918 unpack
= gen_sse4_1_zero_extendv2siv2di2
;
12920 unpack
= gen_sse4_1_extendv2siv2di2
;
12923 gcc_unreachable ();
12926 dest
= operands
[0];
12929 /* Shift higher 8 bytes to lower 8 bytes. */
12930 src
= gen_reg_rtx (imode
);
12931 emit_insn (gen_sse2_lshrti3 (gen_lowpart (TImode
, src
),
12932 gen_lowpart (TImode
, operands
[1]),
12938 emit_insn (unpack (dest
, src
));
12941 /* Expand conditional increment or decrement using adb/sbb instructions.
12942 The default case using setcc followed by the conditional move can be
12943 done by generic code. */
12945 ix86_expand_int_addcc (rtx operands
[])
12947 enum rtx_code code
= GET_CODE (operands
[1]);
12949 rtx val
= const0_rtx
;
12950 bool fpcmp
= false;
12951 enum machine_mode mode
= GET_MODE (operands
[0]);
12953 if (operands
[3] != const1_rtx
12954 && operands
[3] != constm1_rtx
)
12956 if (!ix86_expand_carry_flag_compare (code
, ix86_compare_op0
,
12957 ix86_compare_op1
, &compare_op
))
12959 code
= GET_CODE (compare_op
);
12961 if (GET_MODE (XEXP (compare_op
, 0)) == CCFPmode
12962 || GET_MODE (XEXP (compare_op
, 0)) == CCFPUmode
)
12965 code
= ix86_fp_compare_code_to_integer (code
);
12972 PUT_CODE (compare_op
,
12973 reverse_condition_maybe_unordered
12974 (GET_CODE (compare_op
)));
12976 PUT_CODE (compare_op
, reverse_condition (GET_CODE (compare_op
)));
12978 PUT_MODE (compare_op
, mode
);
12980 /* Construct either adc or sbb insn. */
12981 if ((code
== LTU
) == (operands
[3] == constm1_rtx
))
12983 switch (GET_MODE (operands
[0]))
12986 emit_insn (gen_subqi3_carry (operands
[0], operands
[2], val
, compare_op
));
12989 emit_insn (gen_subhi3_carry (operands
[0], operands
[2], val
, compare_op
));
12992 emit_insn (gen_subsi3_carry (operands
[0], operands
[2], val
, compare_op
));
12995 emit_insn (gen_subdi3_carry_rex64 (operands
[0], operands
[2], val
, compare_op
));
12998 gcc_unreachable ();
13003 switch (GET_MODE (operands
[0]))
13006 emit_insn (gen_addqi3_carry (operands
[0], operands
[2], val
, compare_op
));
13009 emit_insn (gen_addhi3_carry (operands
[0], operands
[2], val
, compare_op
));
13012 emit_insn (gen_addsi3_carry (operands
[0], operands
[2], val
, compare_op
));
13015 emit_insn (gen_adddi3_carry_rex64 (operands
[0], operands
[2], val
, compare_op
));
13018 gcc_unreachable ();
13021 return 1; /* DONE */
13025 /* Split operands 0 and 1 into SImode parts. Similar to split_di, but
13026 works for floating pointer parameters and nonoffsetable memories.
13027 For pushes, it returns just stack offsets; the values will be saved
13028 in the right order. Maximally three parts are generated. */
13031 ix86_split_to_parts (rtx operand
, rtx
*parts
, enum machine_mode mode
)
13036 size
= mode
==XFmode
? 3 : GET_MODE_SIZE (mode
) / 4;
13038 size
= (GET_MODE_SIZE (mode
) + 4) / 8;
13040 gcc_assert (!REG_P (operand
) || !MMX_REGNO_P (REGNO (operand
)));
13041 gcc_assert (size
>= 2 && size
<= 3);
13043 /* Optimize constant pool reference to immediates. This is used by fp
13044 moves, that force all constants to memory to allow combining. */
13045 if (MEM_P (operand
) && MEM_READONLY_P (operand
))
13047 rtx tmp
= maybe_get_pool_constant (operand
);
13052 if (MEM_P (operand
) && !offsettable_memref_p (operand
))
13054 /* The only non-offsetable memories we handle are pushes. */
13055 int ok
= push_operand (operand
, VOIDmode
);
13059 operand
= copy_rtx (operand
);
13060 PUT_MODE (operand
, Pmode
);
13061 parts
[0] = parts
[1] = parts
[2] = operand
;
13065 if (GET_CODE (operand
) == CONST_VECTOR
)
13067 enum machine_mode imode
= int_mode_for_mode (mode
);
13068 /* Caution: if we looked through a constant pool memory above,
13069 the operand may actually have a different mode now. That's
13070 ok, since we want to pun this all the way back to an integer. */
13071 operand
= simplify_subreg (imode
, operand
, GET_MODE (operand
), 0);
13072 gcc_assert (operand
!= NULL
);
13078 if (mode
== DImode
)
13079 split_di (&operand
, 1, &parts
[0], &parts
[1]);
13082 if (REG_P (operand
))
13084 gcc_assert (reload_completed
);
13085 parts
[0] = gen_rtx_REG (SImode
, REGNO (operand
) + 0);
13086 parts
[1] = gen_rtx_REG (SImode
, REGNO (operand
) + 1);
13088 parts
[2] = gen_rtx_REG (SImode
, REGNO (operand
) + 2);
13090 else if (offsettable_memref_p (operand
))
13092 operand
= adjust_address (operand
, SImode
, 0);
13093 parts
[0] = operand
;
13094 parts
[1] = adjust_address (operand
, SImode
, 4);
13096 parts
[2] = adjust_address (operand
, SImode
, 8);
13098 else if (GET_CODE (operand
) == CONST_DOUBLE
)
13103 REAL_VALUE_FROM_CONST_DOUBLE (r
, operand
);
13107 REAL_VALUE_TO_TARGET_LONG_DOUBLE (r
, l
);
13108 parts
[2] = gen_int_mode (l
[2], SImode
);
13111 REAL_VALUE_TO_TARGET_DOUBLE (r
, l
);
13114 gcc_unreachable ();
13116 parts
[1] = gen_int_mode (l
[1], SImode
);
13117 parts
[0] = gen_int_mode (l
[0], SImode
);
13120 gcc_unreachable ();
13125 if (mode
== TImode
)
13126 split_ti (&operand
, 1, &parts
[0], &parts
[1]);
13127 if (mode
== XFmode
|| mode
== TFmode
)
13129 enum machine_mode upper_mode
= mode
==XFmode
? SImode
: DImode
;
13130 if (REG_P (operand
))
13132 gcc_assert (reload_completed
);
13133 parts
[0] = gen_rtx_REG (DImode
, REGNO (operand
) + 0);
13134 parts
[1] = gen_rtx_REG (upper_mode
, REGNO (operand
) + 1);
13136 else if (offsettable_memref_p (operand
))
13138 operand
= adjust_address (operand
, DImode
, 0);
13139 parts
[0] = operand
;
13140 parts
[1] = adjust_address (operand
, upper_mode
, 8);
13142 else if (GET_CODE (operand
) == CONST_DOUBLE
)
13147 REAL_VALUE_FROM_CONST_DOUBLE (r
, operand
);
13148 real_to_target (l
, &r
, mode
);
13150 /* Do not use shift by 32 to avoid warning on 32bit systems. */
13151 if (HOST_BITS_PER_WIDE_INT
>= 64)
13154 ((l
[0] & (((HOST_WIDE_INT
) 2 << 31) - 1))
13155 + ((((HOST_WIDE_INT
) l
[1]) << 31) << 1),
13158 parts
[0] = immed_double_const (l
[0], l
[1], DImode
);
13160 if (upper_mode
== SImode
)
13161 parts
[1] = gen_int_mode (l
[2], SImode
);
13162 else if (HOST_BITS_PER_WIDE_INT
>= 64)
13165 ((l
[2] & (((HOST_WIDE_INT
) 2 << 31) - 1))
13166 + ((((HOST_WIDE_INT
) l
[3]) << 31) << 1),
13169 parts
[1] = immed_double_const (l
[2], l
[3], DImode
);
13172 gcc_unreachable ();
13179 /* Emit insns to perform a move or push of DI, DF, and XF values.
13180 Return false when normal moves are needed; true when all required
13181 insns have been emitted. Operands 2-4 contain the input values
13182 int the correct order; operands 5-7 contain the output values. */
13185 ix86_split_long_move (rtx operands
[])
13190 int collisions
= 0;
13191 enum machine_mode mode
= GET_MODE (operands
[0]);
13193 /* The DFmode expanders may ask us to move double.
13194 For 64bit target this is single move. By hiding the fact
13195 here we simplify i386.md splitters. */
13196 if (GET_MODE_SIZE (GET_MODE (operands
[0])) == 8 && TARGET_64BIT
)
13198 /* Optimize constant pool reference to immediates. This is used by
13199 fp moves, that force all constants to memory to allow combining. */
13201 if (MEM_P (operands
[1])
13202 && GET_CODE (XEXP (operands
[1], 0)) == SYMBOL_REF
13203 && CONSTANT_POOL_ADDRESS_P (XEXP (operands
[1], 0)))
13204 operands
[1] = get_pool_constant (XEXP (operands
[1], 0));
13205 if (push_operand (operands
[0], VOIDmode
))
13207 operands
[0] = copy_rtx (operands
[0]);
13208 PUT_MODE (operands
[0], Pmode
);
13211 operands
[0] = gen_lowpart (DImode
, operands
[0]);
13212 operands
[1] = gen_lowpart (DImode
, operands
[1]);
13213 emit_move_insn (operands
[0], operands
[1]);
13217 /* The only non-offsettable memory we handle is push. */
13218 if (push_operand (operands
[0], VOIDmode
))
13221 gcc_assert (!MEM_P (operands
[0])
13222 || offsettable_memref_p (operands
[0]));
13224 nparts
= ix86_split_to_parts (operands
[1], part
[1], GET_MODE (operands
[0]));
13225 ix86_split_to_parts (operands
[0], part
[0], GET_MODE (operands
[0]));
13227 /* When emitting push, take care for source operands on the stack. */
13228 if (push
&& MEM_P (operands
[1])
13229 && reg_overlap_mentioned_p (stack_pointer_rtx
, operands
[1]))
13232 part
[1][1] = change_address (part
[1][1], GET_MODE (part
[1][1]),
13233 XEXP (part
[1][2], 0));
13234 part
[1][0] = change_address (part
[1][0], GET_MODE (part
[1][0]),
13235 XEXP (part
[1][1], 0));
13238 /* We need to do copy in the right order in case an address register
13239 of the source overlaps the destination. */
13240 if (REG_P (part
[0][0]) && MEM_P (part
[1][0]))
13242 if (reg_overlap_mentioned_p (part
[0][0], XEXP (part
[1][0], 0)))
13244 if (reg_overlap_mentioned_p (part
[0][1], XEXP (part
[1][0], 0)))
13247 && reg_overlap_mentioned_p (part
[0][2], XEXP (part
[1][0], 0)))
13250 /* Collision in the middle part can be handled by reordering. */
13251 if (collisions
== 1 && nparts
== 3
13252 && reg_overlap_mentioned_p (part
[0][1], XEXP (part
[1][0], 0)))
13255 tmp
= part
[0][1]; part
[0][1] = part
[0][2]; part
[0][2] = tmp
;
13256 tmp
= part
[1][1]; part
[1][1] = part
[1][2]; part
[1][2] = tmp
;
13259 /* If there are more collisions, we can't handle it by reordering.
13260 Do an lea to the last part and use only one colliding move. */
13261 else if (collisions
> 1)
13267 base
= part
[0][nparts
- 1];
13269 /* Handle the case when the last part isn't valid for lea.
13270 Happens in 64-bit mode storing the 12-byte XFmode. */
13271 if (GET_MODE (base
) != Pmode
)
13272 base
= gen_rtx_REG (Pmode
, REGNO (base
));
13274 emit_insn (gen_rtx_SET (VOIDmode
, base
, XEXP (part
[1][0], 0)));
13275 part
[1][0] = replace_equiv_address (part
[1][0], base
);
13276 part
[1][1] = replace_equiv_address (part
[1][1],
13277 plus_constant (base
, UNITS_PER_WORD
));
13279 part
[1][2] = replace_equiv_address (part
[1][2],
13280 plus_constant (base
, 8));
13290 if (TARGET_128BIT_LONG_DOUBLE
&& mode
== XFmode
)
13291 emit_insn (gen_addsi3 (stack_pointer_rtx
, stack_pointer_rtx
, GEN_INT (-4)));
13292 emit_move_insn (part
[0][2], part
[1][2]);
13297 /* In 64bit mode we don't have 32bit push available. In case this is
13298 register, it is OK - we will just use larger counterpart. We also
13299 retype memory - these comes from attempt to avoid REX prefix on
13300 moving of second half of TFmode value. */
13301 if (GET_MODE (part
[1][1]) == SImode
)
13303 switch (GET_CODE (part
[1][1]))
13306 part
[1][1] = adjust_address (part
[1][1], DImode
, 0);
13310 part
[1][1] = gen_rtx_REG (DImode
, REGNO (part
[1][1]));
13314 gcc_unreachable ();
13317 if (GET_MODE (part
[1][0]) == SImode
)
13318 part
[1][0] = part
[1][1];
13321 emit_move_insn (part
[0][1], part
[1][1]);
13322 emit_move_insn (part
[0][0], part
[1][0]);
13326 /* Choose correct order to not overwrite the source before it is copied. */
13327 if ((REG_P (part
[0][0])
13328 && REG_P (part
[1][1])
13329 && (REGNO (part
[0][0]) == REGNO (part
[1][1])
13331 && REGNO (part
[0][0]) == REGNO (part
[1][2]))))
13333 && reg_overlap_mentioned_p (part
[0][0], XEXP (part
[1][0], 0))))
13337 operands
[2] = part
[0][2];
13338 operands
[3] = part
[0][1];
13339 operands
[4] = part
[0][0];
13340 operands
[5] = part
[1][2];
13341 operands
[6] = part
[1][1];
13342 operands
[7] = part
[1][0];
13346 operands
[2] = part
[0][1];
13347 operands
[3] = part
[0][0];
13348 operands
[5] = part
[1][1];
13349 operands
[6] = part
[1][0];
13356 operands
[2] = part
[0][0];
13357 operands
[3] = part
[0][1];
13358 operands
[4] = part
[0][2];
13359 operands
[5] = part
[1][0];
13360 operands
[6] = part
[1][1];
13361 operands
[7] = part
[1][2];
13365 operands
[2] = part
[0][0];
13366 operands
[3] = part
[0][1];
13367 operands
[5] = part
[1][0];
13368 operands
[6] = part
[1][1];
13372 /* If optimizing for size, attempt to locally unCSE nonzero constants. */
13375 if (CONST_INT_P (operands
[5])
13376 && operands
[5] != const0_rtx
13377 && REG_P (operands
[2]))
13379 if (CONST_INT_P (operands
[6])
13380 && INTVAL (operands
[6]) == INTVAL (operands
[5]))
13381 operands
[6] = operands
[2];
13384 && CONST_INT_P (operands
[7])
13385 && INTVAL (operands
[7]) == INTVAL (operands
[5]))
13386 operands
[7] = operands
[2];
13390 && CONST_INT_P (operands
[6])
13391 && operands
[6] != const0_rtx
13392 && REG_P (operands
[3])
13393 && CONST_INT_P (operands
[7])
13394 && INTVAL (operands
[7]) == INTVAL (operands
[6]))
13395 operands
[7] = operands
[3];
13398 emit_move_insn (operands
[2], operands
[5]);
13399 emit_move_insn (operands
[3], operands
[6]);
13401 emit_move_insn (operands
[4], operands
[7]);
13406 /* Helper function of ix86_split_ashl used to generate an SImode/DImode
13407 left shift by a constant, either using a single shift or
13408 a sequence of add instructions. */
13411 ix86_expand_ashl_const (rtx operand
, int count
, enum machine_mode mode
)
13415 emit_insn ((mode
== DImode
13417 : gen_adddi3
) (operand
, operand
, operand
));
13419 else if (!optimize_size
13420 && count
* ix86_cost
->add
<= ix86_cost
->shift_const
)
13423 for (i
=0; i
<count
; i
++)
13425 emit_insn ((mode
== DImode
13427 : gen_adddi3
) (operand
, operand
, operand
));
13431 emit_insn ((mode
== DImode
13433 : gen_ashldi3
) (operand
, operand
, GEN_INT (count
)));
13437 ix86_split_ashl (rtx
*operands
, rtx scratch
, enum machine_mode mode
)
13439 rtx low
[2], high
[2];
13441 const int single_width
= mode
== DImode
? 32 : 64;
13443 if (CONST_INT_P (operands
[2]))
13445 (mode
== DImode
? split_di
: split_ti
) (operands
, 2, low
, high
);
13446 count
= INTVAL (operands
[2]) & (single_width
* 2 - 1);
13448 if (count
>= single_width
)
13450 emit_move_insn (high
[0], low
[1]);
13451 emit_move_insn (low
[0], const0_rtx
);
13453 if (count
> single_width
)
13454 ix86_expand_ashl_const (high
[0], count
- single_width
, mode
);
13458 if (!rtx_equal_p (operands
[0], operands
[1]))
13459 emit_move_insn (operands
[0], operands
[1]);
13460 emit_insn ((mode
== DImode
13462 : gen_x86_64_shld
) (high
[0], low
[0], GEN_INT (count
)));
13463 ix86_expand_ashl_const (low
[0], count
, mode
);
13468 (mode
== DImode
? split_di
: split_ti
) (operands
, 1, low
, high
);
13470 if (operands
[1] == const1_rtx
)
13472 /* Assuming we've chosen a QImode capable registers, then 1 << N
13473 can be done with two 32/64-bit shifts, no branches, no cmoves. */
13474 if (ANY_QI_REG_P (low
[0]) && ANY_QI_REG_P (high
[0]))
13476 rtx s
, d
, flags
= gen_rtx_REG (CCZmode
, FLAGS_REG
);
13478 ix86_expand_clear (low
[0]);
13479 ix86_expand_clear (high
[0]);
13480 emit_insn (gen_testqi_ccz_1 (operands
[2], GEN_INT (single_width
)));
13482 d
= gen_lowpart (QImode
, low
[0]);
13483 d
= gen_rtx_STRICT_LOW_PART (VOIDmode
, d
);
13484 s
= gen_rtx_EQ (QImode
, flags
, const0_rtx
);
13485 emit_insn (gen_rtx_SET (VOIDmode
, d
, s
));
13487 d
= gen_lowpart (QImode
, high
[0]);
13488 d
= gen_rtx_STRICT_LOW_PART (VOIDmode
, d
);
13489 s
= gen_rtx_NE (QImode
, flags
, const0_rtx
);
13490 emit_insn (gen_rtx_SET (VOIDmode
, d
, s
));
13493 /* Otherwise, we can get the same results by manually performing
13494 a bit extract operation on bit 5/6, and then performing the two
13495 shifts. The two methods of getting 0/1 into low/high are exactly
13496 the same size. Avoiding the shift in the bit extract case helps
13497 pentium4 a bit; no one else seems to care much either way. */
13502 if (TARGET_PARTIAL_REG_STALL
&& !optimize_size
)
13503 x
= gen_rtx_ZERO_EXTEND (mode
== DImode
? SImode
: DImode
, operands
[2]);
13505 x
= gen_lowpart (mode
== DImode
? SImode
: DImode
, operands
[2]);
13506 emit_insn (gen_rtx_SET (VOIDmode
, high
[0], x
));
13508 emit_insn ((mode
== DImode
13510 : gen_lshrdi3
) (high
[0], high
[0], GEN_INT (mode
== DImode
? 5 : 6)));
13511 emit_insn ((mode
== DImode
13513 : gen_anddi3
) (high
[0], high
[0], GEN_INT (1)));
13514 emit_move_insn (low
[0], high
[0]);
13515 emit_insn ((mode
== DImode
13517 : gen_xordi3
) (low
[0], low
[0], GEN_INT (1)));
13520 emit_insn ((mode
== DImode
13522 : gen_ashldi3
) (low
[0], low
[0], operands
[2]));
13523 emit_insn ((mode
== DImode
13525 : gen_ashldi3
) (high
[0], high
[0], operands
[2]));
13529 if (operands
[1] == constm1_rtx
)
13531 /* For -1 << N, we can avoid the shld instruction, because we
13532 know that we're shifting 0...31/63 ones into a -1. */
13533 emit_move_insn (low
[0], constm1_rtx
);
13535 emit_move_insn (high
[0], low
[0]);
13537 emit_move_insn (high
[0], constm1_rtx
);
13541 if (!rtx_equal_p (operands
[0], operands
[1]))
13542 emit_move_insn (operands
[0], operands
[1]);
13544 (mode
== DImode
? split_di
: split_ti
) (operands
, 1, low
, high
);
13545 emit_insn ((mode
== DImode
13547 : gen_x86_64_shld
) (high
[0], low
[0], operands
[2]));
13550 emit_insn ((mode
== DImode
? gen_ashlsi3
: gen_ashldi3
) (low
[0], low
[0], operands
[2]));
13552 if (TARGET_CMOVE
&& scratch
)
13554 ix86_expand_clear (scratch
);
13555 emit_insn ((mode
== DImode
13556 ? gen_x86_shift_adj_1
13557 : gen_x86_64_shift_adj
) (high
[0], low
[0], operands
[2], scratch
));
13560 emit_insn (gen_x86_shift_adj_2 (high
[0], low
[0], operands
[2]));
13564 ix86_split_ashr (rtx
*operands
, rtx scratch
, enum machine_mode mode
)
13566 rtx low
[2], high
[2];
13568 const int single_width
= mode
== DImode
? 32 : 64;
13570 if (CONST_INT_P (operands
[2]))
13572 (mode
== DImode
? split_di
: split_ti
) (operands
, 2, low
, high
);
13573 count
= INTVAL (operands
[2]) & (single_width
* 2 - 1);
13575 if (count
== single_width
* 2 - 1)
13577 emit_move_insn (high
[0], high
[1]);
13578 emit_insn ((mode
== DImode
13580 : gen_ashrdi3
) (high
[0], high
[0],
13581 GEN_INT (single_width
- 1)));
13582 emit_move_insn (low
[0], high
[0]);
13585 else if (count
>= single_width
)
13587 emit_move_insn (low
[0], high
[1]);
13588 emit_move_insn (high
[0], low
[0]);
13589 emit_insn ((mode
== DImode
13591 : gen_ashrdi3
) (high
[0], high
[0],
13592 GEN_INT (single_width
- 1)));
13593 if (count
> single_width
)
13594 emit_insn ((mode
== DImode
13596 : gen_ashrdi3
) (low
[0], low
[0],
13597 GEN_INT (count
- single_width
)));
13601 if (!rtx_equal_p (operands
[0], operands
[1]))
13602 emit_move_insn (operands
[0], operands
[1]);
13603 emit_insn ((mode
== DImode
13605 : gen_x86_64_shrd
) (low
[0], high
[0], GEN_INT (count
)));
13606 emit_insn ((mode
== DImode
13608 : gen_ashrdi3
) (high
[0], high
[0], GEN_INT (count
)));
13613 if (!rtx_equal_p (operands
[0], operands
[1]))
13614 emit_move_insn (operands
[0], operands
[1]);
13616 (mode
== DImode
? split_di
: split_ti
) (operands
, 1, low
, high
);
13618 emit_insn ((mode
== DImode
13620 : gen_x86_64_shrd
) (low
[0], high
[0], operands
[2]));
13621 emit_insn ((mode
== DImode
13623 : gen_ashrdi3
) (high
[0], high
[0], operands
[2]));
13625 if (TARGET_CMOVE
&& scratch
)
13627 emit_move_insn (scratch
, high
[0]);
13628 emit_insn ((mode
== DImode
13630 : gen_ashrdi3
) (scratch
, scratch
,
13631 GEN_INT (single_width
- 1)));
13632 emit_insn ((mode
== DImode
13633 ? gen_x86_shift_adj_1
13634 : gen_x86_64_shift_adj
) (low
[0], high
[0], operands
[2],
13638 emit_insn (gen_x86_shift_adj_3 (low
[0], high
[0], operands
[2]));
13643 ix86_split_lshr (rtx
*operands
, rtx scratch
, enum machine_mode mode
)
13645 rtx low
[2], high
[2];
13647 const int single_width
= mode
== DImode
? 32 : 64;
13649 if (CONST_INT_P (operands
[2]))
13651 (mode
== DImode
? split_di
: split_ti
) (operands
, 2, low
, high
);
13652 count
= INTVAL (operands
[2]) & (single_width
* 2 - 1);
13654 if (count
>= single_width
)
13656 emit_move_insn (low
[0], high
[1]);
13657 ix86_expand_clear (high
[0]);
13659 if (count
> single_width
)
13660 emit_insn ((mode
== DImode
13662 : gen_lshrdi3
) (low
[0], low
[0],
13663 GEN_INT (count
- single_width
)));
13667 if (!rtx_equal_p (operands
[0], operands
[1]))
13668 emit_move_insn (operands
[0], operands
[1]);
13669 emit_insn ((mode
== DImode
13671 : gen_x86_64_shrd
) (low
[0], high
[0], GEN_INT (count
)));
13672 emit_insn ((mode
== DImode
13674 : gen_lshrdi3
) (high
[0], high
[0], GEN_INT (count
)));
13679 if (!rtx_equal_p (operands
[0], operands
[1]))
13680 emit_move_insn (operands
[0], operands
[1]);
13682 (mode
== DImode
? split_di
: split_ti
) (operands
, 1, low
, high
);
13684 emit_insn ((mode
== DImode
13686 : gen_x86_64_shrd
) (low
[0], high
[0], operands
[2]));
13687 emit_insn ((mode
== DImode
13689 : gen_lshrdi3
) (high
[0], high
[0], operands
[2]));
13691 /* Heh. By reversing the arguments, we can reuse this pattern. */
13692 if (TARGET_CMOVE
&& scratch
)
13694 ix86_expand_clear (scratch
);
13695 emit_insn ((mode
== DImode
13696 ? gen_x86_shift_adj_1
13697 : gen_x86_64_shift_adj
) (low
[0], high
[0], operands
[2],
13701 emit_insn (gen_x86_shift_adj_2 (low
[0], high
[0], operands
[2]));
13705 /* Predict just emitted jump instruction to be taken with probability PROB. */
13707 predict_jump (int prob
)
13709 rtx insn
= get_last_insn ();
13710 gcc_assert (JUMP_P (insn
));
13712 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
13717 /* Helper function for the string operations below. Dest VARIABLE whether
13718 it is aligned to VALUE bytes. If true, jump to the label. */
13720 ix86_expand_aligntest (rtx variable
, int value
, bool epilogue
)
13722 rtx label
= gen_label_rtx ();
13723 rtx tmpcount
= gen_reg_rtx (GET_MODE (variable
));
13724 if (GET_MODE (variable
) == DImode
)
13725 emit_insn (gen_anddi3 (tmpcount
, variable
, GEN_INT (value
)));
13727 emit_insn (gen_andsi3 (tmpcount
, variable
, GEN_INT (value
)));
13728 emit_cmp_and_jump_insns (tmpcount
, const0_rtx
, EQ
, 0, GET_MODE (variable
),
13731 predict_jump (REG_BR_PROB_BASE
* 50 / 100);
13733 predict_jump (REG_BR_PROB_BASE
* 90 / 100);
13737 /* Adjust COUNTER by the VALUE. */
13739 ix86_adjust_counter (rtx countreg
, HOST_WIDE_INT value
)
13741 if (GET_MODE (countreg
) == DImode
)
13742 emit_insn (gen_adddi3 (countreg
, countreg
, GEN_INT (-value
)));
13744 emit_insn (gen_addsi3 (countreg
, countreg
, GEN_INT (-value
)));
13747 /* Zero extend possibly SImode EXP to Pmode register. */
13749 ix86_zero_extend_to_Pmode (rtx exp
)
13752 if (GET_MODE (exp
) == VOIDmode
)
13753 return force_reg (Pmode
, exp
);
13754 if (GET_MODE (exp
) == Pmode
)
13755 return copy_to_mode_reg (Pmode
, exp
);
13756 r
= gen_reg_rtx (Pmode
);
13757 emit_insn (gen_zero_extendsidi2 (r
, exp
));
13761 /* Divide COUNTREG by SCALE. */
13763 scale_counter (rtx countreg
, int scale
)
13766 rtx piece_size_mask
;
13770 if (CONST_INT_P (countreg
))
13771 return GEN_INT (INTVAL (countreg
) / scale
);
13772 gcc_assert (REG_P (countreg
));
13774 piece_size_mask
= GEN_INT (scale
- 1);
13775 sc
= expand_simple_binop (GET_MODE (countreg
), LSHIFTRT
, countreg
,
13776 GEN_INT (exact_log2 (scale
)),
13777 NULL
, 1, OPTAB_DIRECT
);
13781 /* Return mode for the memcpy/memset loop counter. Prefer SImode over
13782 DImode for constant loop counts. */
13784 static enum machine_mode
13785 counter_mode (rtx count_exp
)
13787 if (GET_MODE (count_exp
) != VOIDmode
)
13788 return GET_MODE (count_exp
);
13789 if (GET_CODE (count_exp
) != CONST_INT
)
13791 if (TARGET_64BIT
&& (INTVAL (count_exp
) & ~0xffffffff))
13796 /* When SRCPTR is non-NULL, output simple loop to move memory
13797 pointer to SRCPTR to DESTPTR via chunks of MODE unrolled UNROLL times,
13798 overall size is COUNT specified in bytes. When SRCPTR is NULL, output the
13799 equivalent loop to set memory by VALUE (supposed to be in MODE).
13801 The size is rounded down to whole number of chunk size moved at once.
13802 SRCMEM and DESTMEM provide MEMrtx to feed proper aliasing info. */
13806 expand_set_or_movmem_via_loop (rtx destmem
, rtx srcmem
,
13807 rtx destptr
, rtx srcptr
, rtx value
,
13808 rtx count
, enum machine_mode mode
, int unroll
,
13811 rtx out_label
, top_label
, iter
, tmp
;
13812 enum machine_mode iter_mode
= counter_mode (count
);
13813 rtx piece_size
= GEN_INT (GET_MODE_SIZE (mode
) * unroll
);
13814 rtx piece_size_mask
= GEN_INT (~((GET_MODE_SIZE (mode
) * unroll
) - 1));
13820 top_label
= gen_label_rtx ();
13821 out_label
= gen_label_rtx ();
13822 iter
= gen_reg_rtx (iter_mode
);
13824 size
= expand_simple_binop (iter_mode
, AND
, count
, piece_size_mask
,
13825 NULL
, 1, OPTAB_DIRECT
);
13826 /* Those two should combine. */
13827 if (piece_size
== const1_rtx
)
13829 emit_cmp_and_jump_insns (size
, const0_rtx
, EQ
, NULL_RTX
, iter_mode
,
13831 predict_jump (REG_BR_PROB_BASE
* 10 / 100);
13833 emit_move_insn (iter
, const0_rtx
);
13835 emit_label (top_label
);
13837 tmp
= convert_modes (Pmode
, iter_mode
, iter
, true);
13838 x_addr
= gen_rtx_PLUS (Pmode
, destptr
, tmp
);
13839 destmem
= change_address (destmem
, mode
, x_addr
);
13843 y_addr
= gen_rtx_PLUS (Pmode
, srcptr
, copy_rtx (tmp
));
13844 srcmem
= change_address (srcmem
, mode
, y_addr
);
13846 /* When unrolling for chips that reorder memory reads and writes,
13847 we can save registers by using single temporary.
13848 Also using 4 temporaries is overkill in 32bit mode. */
13849 if (!TARGET_64BIT
&& 0)
13851 for (i
= 0; i
< unroll
; i
++)
13856 adjust_address (copy_rtx (destmem
), mode
, GET_MODE_SIZE (mode
));
13858 adjust_address (copy_rtx (srcmem
), mode
, GET_MODE_SIZE (mode
));
13860 emit_move_insn (destmem
, srcmem
);
13866 gcc_assert (unroll
<= 4);
13867 for (i
= 0; i
< unroll
; i
++)
13869 tmpreg
[i
] = gen_reg_rtx (mode
);
13873 adjust_address (copy_rtx (srcmem
), mode
, GET_MODE_SIZE (mode
));
13875 emit_move_insn (tmpreg
[i
], srcmem
);
13877 for (i
= 0; i
< unroll
; i
++)
13882 adjust_address (copy_rtx (destmem
), mode
, GET_MODE_SIZE (mode
));
13884 emit_move_insn (destmem
, tmpreg
[i
]);
13889 for (i
= 0; i
< unroll
; i
++)
13893 adjust_address (copy_rtx (destmem
), mode
, GET_MODE_SIZE (mode
));
13894 emit_move_insn (destmem
, value
);
13897 tmp
= expand_simple_binop (iter_mode
, PLUS
, iter
, piece_size
, iter
,
13898 true, OPTAB_LIB_WIDEN
);
13900 emit_move_insn (iter
, tmp
);
13902 emit_cmp_and_jump_insns (iter
, size
, LT
, NULL_RTX
, iter_mode
,
13904 if (expected_size
!= -1)
13906 expected_size
/= GET_MODE_SIZE (mode
) * unroll
;
13907 if (expected_size
== 0)
13909 else if (expected_size
> REG_BR_PROB_BASE
)
13910 predict_jump (REG_BR_PROB_BASE
- 1);
13912 predict_jump (REG_BR_PROB_BASE
- (REG_BR_PROB_BASE
+ expected_size
/ 2) / expected_size
);
13915 predict_jump (REG_BR_PROB_BASE
* 80 / 100);
13916 iter
= ix86_zero_extend_to_Pmode (iter
);
13917 tmp
= expand_simple_binop (Pmode
, PLUS
, destptr
, iter
, destptr
,
13918 true, OPTAB_LIB_WIDEN
);
13919 if (tmp
!= destptr
)
13920 emit_move_insn (destptr
, tmp
);
13923 tmp
= expand_simple_binop (Pmode
, PLUS
, srcptr
, iter
, srcptr
,
13924 true, OPTAB_LIB_WIDEN
);
13926 emit_move_insn (srcptr
, tmp
);
13928 emit_label (out_label
);
13931 /* Output "rep; mov" instruction.
13932 Arguments have same meaning as for previous function */
13934 expand_movmem_via_rep_mov (rtx destmem
, rtx srcmem
,
13935 rtx destptr
, rtx srcptr
,
13937 enum machine_mode mode
)
13943 /* If the size is known, it is shorter to use rep movs. */
13944 if (mode
== QImode
&& CONST_INT_P (count
)
13945 && !(INTVAL (count
) & 3))
13948 if (destptr
!= XEXP (destmem
, 0) || GET_MODE (destmem
) != BLKmode
)
13949 destmem
= adjust_automodify_address_nv (destmem
, BLKmode
, destptr
, 0);
13950 if (srcptr
!= XEXP (srcmem
, 0) || GET_MODE (srcmem
) != BLKmode
)
13951 srcmem
= adjust_automodify_address_nv (srcmem
, BLKmode
, srcptr
, 0);
13952 countreg
= ix86_zero_extend_to_Pmode (scale_counter (count
, GET_MODE_SIZE (mode
)));
13953 if (mode
!= QImode
)
13955 destexp
= gen_rtx_ASHIFT (Pmode
, countreg
,
13956 GEN_INT (exact_log2 (GET_MODE_SIZE (mode
))));
13957 destexp
= gen_rtx_PLUS (Pmode
, destexp
, destptr
);
13958 srcexp
= gen_rtx_ASHIFT (Pmode
, countreg
,
13959 GEN_INT (exact_log2 (GET_MODE_SIZE (mode
))));
13960 srcexp
= gen_rtx_PLUS (Pmode
, srcexp
, srcptr
);
13964 destexp
= gen_rtx_PLUS (Pmode
, destptr
, countreg
);
13965 srcexp
= gen_rtx_PLUS (Pmode
, srcptr
, countreg
);
13967 emit_insn (gen_rep_mov (destptr
, destmem
, srcptr
, srcmem
, countreg
,
13971 /* Output "rep; stos" instruction.
13972 Arguments have same meaning as for previous function */
13974 expand_setmem_via_rep_stos (rtx destmem
, rtx destptr
, rtx value
,
13976 enum machine_mode mode
)
13981 if (destptr
!= XEXP (destmem
, 0) || GET_MODE (destmem
) != BLKmode
)
13982 destmem
= adjust_automodify_address_nv (destmem
, BLKmode
, destptr
, 0);
13983 value
= force_reg (mode
, gen_lowpart (mode
, value
));
13984 countreg
= ix86_zero_extend_to_Pmode (scale_counter (count
, GET_MODE_SIZE (mode
)));
13985 if (mode
!= QImode
)
13987 destexp
= gen_rtx_ASHIFT (Pmode
, countreg
,
13988 GEN_INT (exact_log2 (GET_MODE_SIZE (mode
))));
13989 destexp
= gen_rtx_PLUS (Pmode
, destexp
, destptr
);
13992 destexp
= gen_rtx_PLUS (Pmode
, destptr
, countreg
);
13993 emit_insn (gen_rep_stos (destptr
, countreg
, destmem
, value
, destexp
));
13997 emit_strmov (rtx destmem
, rtx srcmem
,
13998 rtx destptr
, rtx srcptr
, enum machine_mode mode
, int offset
)
14000 rtx src
= adjust_automodify_address_nv (srcmem
, mode
, srcptr
, offset
);
14001 rtx dest
= adjust_automodify_address_nv (destmem
, mode
, destptr
, offset
);
14002 emit_insn (gen_strmov (destptr
, dest
, srcptr
, src
));
14005 /* Output code to copy at most count & (max_size - 1) bytes from SRC to DEST. */
14007 expand_movmem_epilogue (rtx destmem
, rtx srcmem
,
14008 rtx destptr
, rtx srcptr
, rtx count
, int max_size
)
14011 if (CONST_INT_P (count
))
14013 HOST_WIDE_INT countval
= INTVAL (count
);
14016 if ((countval
& 0x10) && max_size
> 16)
14020 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, DImode
, offset
);
14021 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, DImode
, offset
+ 8);
14024 gcc_unreachable ();
14027 if ((countval
& 0x08) && max_size
> 8)
14030 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, DImode
, offset
);
14033 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, SImode
, offset
);
14034 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, SImode
, offset
+ 4);
14038 if ((countval
& 0x04) && max_size
> 4)
14040 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, SImode
, offset
);
14043 if ((countval
& 0x02) && max_size
> 2)
14045 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, HImode
, offset
);
14048 if ((countval
& 0x01) && max_size
> 1)
14050 emit_strmov (destmem
, srcmem
, destptr
, srcptr
, QImode
, offset
);
14057 count
= expand_simple_binop (GET_MODE (count
), AND
, count
, GEN_INT (max_size
- 1),
14058 count
, 1, OPTAB_DIRECT
);
14059 expand_set_or_movmem_via_loop (destmem
, srcmem
, destptr
, srcptr
, NULL
,
14060 count
, QImode
, 1, 4);
14064 /* When there are stringops, we can cheaply increase dest and src pointers.
14065 Otherwise we save code size by maintaining offset (zero is readily
14066 available from preceding rep operation) and using x86 addressing modes.
14068 if (TARGET_SINGLE_STRINGOP
)
14072 rtx label
= ix86_expand_aligntest (count
, 4, true);
14073 src
= change_address (srcmem
, SImode
, srcptr
);
14074 dest
= change_address (destmem
, SImode
, destptr
);
14075 emit_insn (gen_strmov (destptr
, dest
, srcptr
, src
));
14076 emit_label (label
);
14077 LABEL_NUSES (label
) = 1;
14081 rtx label
= ix86_expand_aligntest (count
, 2, true);
14082 src
= change_address (srcmem
, HImode
, srcptr
);
14083 dest
= change_address (destmem
, HImode
, destptr
);
14084 emit_insn (gen_strmov (destptr
, dest
, srcptr
, src
));
14085 emit_label (label
);
14086 LABEL_NUSES (label
) = 1;
14090 rtx label
= ix86_expand_aligntest (count
, 1, true);
14091 src
= change_address (srcmem
, QImode
, srcptr
);
14092 dest
= change_address (destmem
, QImode
, destptr
);
14093 emit_insn (gen_strmov (destptr
, dest
, srcptr
, src
));
14094 emit_label (label
);
14095 LABEL_NUSES (label
) = 1;
14100 rtx offset
= force_reg (Pmode
, const0_rtx
);
14105 rtx label
= ix86_expand_aligntest (count
, 4, true);
14106 src
= change_address (srcmem
, SImode
, srcptr
);
14107 dest
= change_address (destmem
, SImode
, destptr
);
14108 emit_move_insn (dest
, src
);
14109 tmp
= expand_simple_binop (Pmode
, PLUS
, offset
, GEN_INT (4), NULL
,
14110 true, OPTAB_LIB_WIDEN
);
14112 emit_move_insn (offset
, tmp
);
14113 emit_label (label
);
14114 LABEL_NUSES (label
) = 1;
14118 rtx label
= ix86_expand_aligntest (count
, 2, true);
14119 tmp
= gen_rtx_PLUS (Pmode
, srcptr
, offset
);
14120 src
= change_address (srcmem
, HImode
, tmp
);
14121 tmp
= gen_rtx_PLUS (Pmode
, destptr
, offset
);
14122 dest
= change_address (destmem
, HImode
, tmp
);
14123 emit_move_insn (dest
, src
);
14124 tmp
= expand_simple_binop (Pmode
, PLUS
, offset
, GEN_INT (2), tmp
,
14125 true, OPTAB_LIB_WIDEN
);
14127 emit_move_insn (offset
, tmp
);
14128 emit_label (label
);
14129 LABEL_NUSES (label
) = 1;
14133 rtx label
= ix86_expand_aligntest (count
, 1, true);
14134 tmp
= gen_rtx_PLUS (Pmode
, srcptr
, offset
);
14135 src
= change_address (srcmem
, QImode
, tmp
);
14136 tmp
= gen_rtx_PLUS (Pmode
, destptr
, offset
);
14137 dest
= change_address (destmem
, QImode
, tmp
);
14138 emit_move_insn (dest
, src
);
14139 emit_label (label
);
14140 LABEL_NUSES (label
) = 1;
14145 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
14147 expand_setmem_epilogue_via_loop (rtx destmem
, rtx destptr
, rtx value
,
14148 rtx count
, int max_size
)
14151 expand_simple_binop (counter_mode (count
), AND
, count
,
14152 GEN_INT (max_size
- 1), count
, 1, OPTAB_DIRECT
);
14153 expand_set_or_movmem_via_loop (destmem
, NULL
, destptr
, NULL
,
14154 gen_lowpart (QImode
, value
), count
, QImode
,
14158 /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */
14160 expand_setmem_epilogue (rtx destmem
, rtx destptr
, rtx value
, rtx count
, int max_size
)
14164 if (CONST_INT_P (count
))
14166 HOST_WIDE_INT countval
= INTVAL (count
);
14169 if ((countval
& 0x10) && max_size
> 16)
14173 dest
= adjust_automodify_address_nv (destmem
, DImode
, destptr
, offset
);
14174 emit_insn (gen_strset (destptr
, dest
, value
));
14175 dest
= adjust_automodify_address_nv (destmem
, DImode
, destptr
, offset
+ 8);
14176 emit_insn (gen_strset (destptr
, dest
, value
));
14179 gcc_unreachable ();
14182 if ((countval
& 0x08) && max_size
> 8)
14186 dest
= adjust_automodify_address_nv (destmem
, DImode
, destptr
, offset
);
14187 emit_insn (gen_strset (destptr
, dest
, value
));
14191 dest
= adjust_automodify_address_nv (destmem
, SImode
, destptr
, offset
);
14192 emit_insn (gen_strset (destptr
, dest
, value
));
14193 dest
= adjust_automodify_address_nv (destmem
, SImode
, destptr
, offset
+ 4);
14194 emit_insn (gen_strset (destptr
, dest
, value
));
14198 if ((countval
& 0x04) && max_size
> 4)
14200 dest
= adjust_automodify_address_nv (destmem
, SImode
, destptr
, offset
);
14201 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (SImode
, value
)));
14204 if ((countval
& 0x02) && max_size
> 2)
14206 dest
= adjust_automodify_address_nv (destmem
, HImode
, destptr
, offset
);
14207 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (HImode
, value
)));
14210 if ((countval
& 0x01) && max_size
> 1)
14212 dest
= adjust_automodify_address_nv (destmem
, QImode
, destptr
, offset
);
14213 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (QImode
, value
)));
14220 expand_setmem_epilogue_via_loop (destmem
, destptr
, value
, count
, max_size
);
14225 rtx label
= ix86_expand_aligntest (count
, 16, true);
14228 dest
= change_address (destmem
, DImode
, destptr
);
14229 emit_insn (gen_strset (destptr
, dest
, value
));
14230 emit_insn (gen_strset (destptr
, dest
, value
));
14234 dest
= change_address (destmem
, SImode
, destptr
);
14235 emit_insn (gen_strset (destptr
, dest
, value
));
14236 emit_insn (gen_strset (destptr
, dest
, value
));
14237 emit_insn (gen_strset (destptr
, dest
, value
));
14238 emit_insn (gen_strset (destptr
, dest
, value
));
14240 emit_label (label
);
14241 LABEL_NUSES (label
) = 1;
14245 rtx label
= ix86_expand_aligntest (count
, 8, true);
14248 dest
= change_address (destmem
, DImode
, destptr
);
14249 emit_insn (gen_strset (destptr
, dest
, value
));
14253 dest
= change_address (destmem
, SImode
, destptr
);
14254 emit_insn (gen_strset (destptr
, dest
, value
));
14255 emit_insn (gen_strset (destptr
, dest
, value
));
14257 emit_label (label
);
14258 LABEL_NUSES (label
) = 1;
14262 rtx label
= ix86_expand_aligntest (count
, 4, true);
14263 dest
= change_address (destmem
, SImode
, destptr
);
14264 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (SImode
, value
)));
14265 emit_label (label
);
14266 LABEL_NUSES (label
) = 1;
14270 rtx label
= ix86_expand_aligntest (count
, 2, true);
14271 dest
= change_address (destmem
, HImode
, destptr
);
14272 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (HImode
, value
)));
14273 emit_label (label
);
14274 LABEL_NUSES (label
) = 1;
14278 rtx label
= ix86_expand_aligntest (count
, 1, true);
14279 dest
= change_address (destmem
, QImode
, destptr
);
14280 emit_insn (gen_strset (destptr
, dest
, gen_lowpart (QImode
, value
)));
14281 emit_label (label
);
14282 LABEL_NUSES (label
) = 1;
14286 /* Copy enough from DEST to SRC to align DEST known to by aligned by ALIGN to
14287 DESIRED_ALIGNMENT. */
14289 expand_movmem_prologue (rtx destmem
, rtx srcmem
,
14290 rtx destptr
, rtx srcptr
, rtx count
,
14291 int align
, int desired_alignment
)
14293 if (align
<= 1 && desired_alignment
> 1)
14295 rtx label
= ix86_expand_aligntest (destptr
, 1, false);
14296 srcmem
= change_address (srcmem
, QImode
, srcptr
);
14297 destmem
= change_address (destmem
, QImode
, destptr
);
14298 emit_insn (gen_strmov (destptr
, destmem
, srcptr
, srcmem
));
14299 ix86_adjust_counter (count
, 1);
14300 emit_label (label
);
14301 LABEL_NUSES (label
) = 1;
14303 if (align
<= 2 && desired_alignment
> 2)
14305 rtx label
= ix86_expand_aligntest (destptr
, 2, false);
14306 srcmem
= change_address (srcmem
, HImode
, srcptr
);
14307 destmem
= change_address (destmem
, HImode
, destptr
);
14308 emit_insn (gen_strmov (destptr
, destmem
, srcptr
, srcmem
));
14309 ix86_adjust_counter (count
, 2);
14310 emit_label (label
);
14311 LABEL_NUSES (label
) = 1;
14313 if (align
<= 4 && desired_alignment
> 4)
14315 rtx label
= ix86_expand_aligntest (destptr
, 4, false);
14316 srcmem
= change_address (srcmem
, SImode
, srcptr
);
14317 destmem
= change_address (destmem
, SImode
, destptr
);
14318 emit_insn (gen_strmov (destptr
, destmem
, srcptr
, srcmem
));
14319 ix86_adjust_counter (count
, 4);
14320 emit_label (label
);
14321 LABEL_NUSES (label
) = 1;
14323 gcc_assert (desired_alignment
<= 8);
14326 /* Set enough from DEST to align DEST known to by aligned by ALIGN to
14327 DESIRED_ALIGNMENT. */
14329 expand_setmem_prologue (rtx destmem
, rtx destptr
, rtx value
, rtx count
,
14330 int align
, int desired_alignment
)
14332 if (align
<= 1 && desired_alignment
> 1)
14334 rtx label
= ix86_expand_aligntest (destptr
, 1, false);
14335 destmem
= change_address (destmem
, QImode
, destptr
);
14336 emit_insn (gen_strset (destptr
, destmem
, gen_lowpart (QImode
, value
)));
14337 ix86_adjust_counter (count
, 1);
14338 emit_label (label
);
14339 LABEL_NUSES (label
) = 1;
14341 if (align
<= 2 && desired_alignment
> 2)
14343 rtx label
= ix86_expand_aligntest (destptr
, 2, false);
14344 destmem
= change_address (destmem
, HImode
, destptr
);
14345 emit_insn (gen_strset (destptr
, destmem
, gen_lowpart (HImode
, value
)));
14346 ix86_adjust_counter (count
, 2);
14347 emit_label (label
);
14348 LABEL_NUSES (label
) = 1;
14350 if (align
<= 4 && desired_alignment
> 4)
14352 rtx label
= ix86_expand_aligntest (destptr
, 4, false);
14353 destmem
= change_address (destmem
, SImode
, destptr
);
14354 emit_insn (gen_strset (destptr
, destmem
, gen_lowpart (SImode
, value
)));
14355 ix86_adjust_counter (count
, 4);
14356 emit_label (label
);
14357 LABEL_NUSES (label
) = 1;
14359 gcc_assert (desired_alignment
<= 8);
14362 /* Given COUNT and EXPECTED_SIZE, decide on codegen of string operation. */
14363 static enum stringop_alg
14364 decide_alg (HOST_WIDE_INT count
, HOST_WIDE_INT expected_size
, bool memset
,
14365 int *dynamic_check
)
14367 const struct stringop_algs
* algs
;
14369 *dynamic_check
= -1;
14371 algs
= &ix86_cost
->memset
[TARGET_64BIT
!= 0];
14373 algs
= &ix86_cost
->memcpy
[TARGET_64BIT
!= 0];
14374 if (stringop_alg
!= no_stringop
)
14375 return stringop_alg
;
14376 /* rep; movq or rep; movl is the smallest variant. */
14377 else if (optimize_size
)
14379 if (!count
|| (count
& 3))
14380 return rep_prefix_1_byte
;
14382 return rep_prefix_4_byte
;
14384 /* Very tiny blocks are best handled via the loop, REP is expensive to setup.
14386 else if (expected_size
!= -1 && expected_size
< 4)
14387 return loop_1_byte
;
14388 else if (expected_size
!= -1)
14391 enum stringop_alg alg
= libcall
;
14392 for (i
= 0; i
< NAX_STRINGOP_ALGS
; i
++)
14394 gcc_assert (algs
->size
[i
].max
);
14395 if (algs
->size
[i
].max
>= expected_size
|| algs
->size
[i
].max
== -1)
14397 if (algs
->size
[i
].alg
!= libcall
)
14398 alg
= algs
->size
[i
].alg
;
14399 /* Honor TARGET_INLINE_ALL_STRINGOPS by picking
14400 last non-libcall inline algorithm. */
14401 if (TARGET_INLINE_ALL_STRINGOPS
)
14403 /* When the current size is best to be copied by a libcall,
14404 but we are still forced to inline, run the heuristic bellow
14405 that will pick code for medium sized blocks. */
14406 if (alg
!= libcall
)
14411 return algs
->size
[i
].alg
;
14414 gcc_assert (TARGET_INLINE_ALL_STRINGOPS
);
14416 /* When asked to inline the call anyway, try to pick meaningful choice.
14417 We look for maximal size of block that is faster to copy by hand and
14418 take blocks of at most of that size guessing that average size will
14419 be roughly half of the block.
14421 If this turns out to be bad, we might simply specify the preferred
14422 choice in ix86_costs. */
14423 if ((TARGET_INLINE_ALL_STRINGOPS
|| TARGET_INLINE_STRINGOPS_DYNAMICALLY
)
14424 && algs
->unknown_size
== libcall
)
14427 enum stringop_alg alg
;
14430 for (i
= 0; i
< NAX_STRINGOP_ALGS
; i
++)
14431 if (algs
->size
[i
].alg
!= libcall
&& algs
->size
[i
].alg
)
14432 max
= algs
->size
[i
].max
;
14435 alg
= decide_alg (count
, max
/ 2, memset
, dynamic_check
);
14436 gcc_assert (*dynamic_check
== -1);
14437 gcc_assert (alg
!= libcall
);
14438 if (TARGET_INLINE_STRINGOPS_DYNAMICALLY
)
14439 *dynamic_check
= max
;
14442 return algs
->unknown_size
;
14445 /* Decide on alignment. We know that the operand is already aligned to ALIGN
14446 (ALIGN can be based on profile feedback and thus it is not 100% guaranteed). */
14448 decide_alignment (int align
,
14449 enum stringop_alg alg
,
14452 int desired_align
= 0;
14456 gcc_unreachable ();
14458 case unrolled_loop
:
14459 desired_align
= GET_MODE_SIZE (Pmode
);
14461 case rep_prefix_8_byte
:
14464 case rep_prefix_4_byte
:
14465 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
14466 copying whole cacheline at once. */
14467 if (TARGET_PENTIUMPRO
)
14472 case rep_prefix_1_byte
:
14473 /* PentiumPro has special logic triggering for 8 byte aligned blocks.
14474 copying whole cacheline at once. */
14475 if (TARGET_PENTIUMPRO
)
14489 if (desired_align
< align
)
14490 desired_align
= align
;
14491 if (expected_size
!= -1 && expected_size
< 4)
14492 desired_align
= align
;
14493 return desired_align
;
14496 /* Return the smallest power of 2 greater than VAL. */
14498 smallest_pow2_greater_than (int val
)
14506 /* Expand string move (memcpy) operation. Use i386 string operations when
14507 profitable. expand_clrmem contains similar code. The code depends upon
14508 architecture, block size and alignment, but always has the same
14511 1) Prologue guard: Conditional that jumps up to epilogues for small
14512 blocks that can be handled by epilogue alone. This is faster but
14513 also needed for correctness, since prologue assume the block is larger
14514 than the desired alignment.
14516 Optional dynamic check for size and libcall for large
14517 blocks is emitted here too, with -minline-stringops-dynamically.
14519 2) Prologue: copy first few bytes in order to get destination aligned
14520 to DESIRED_ALIGN. It is emitted only when ALIGN is less than
14521 DESIRED_ALIGN and and up to DESIRED_ALIGN - ALIGN bytes can be copied.
14522 We emit either a jump tree on power of two sized blocks, or a byte loop.
14524 3) Main body: the copying loop itself, copying in SIZE_NEEDED chunks
14525 with specified algorithm.
14527 4) Epilogue: code copying tail of the block that is too small to be
14528 handled by main body (or up to size guarded by prologue guard). */
14531 ix86_expand_movmem (rtx dst
, rtx src
, rtx count_exp
, rtx align_exp
,
14532 rtx expected_align_exp
, rtx expected_size_exp
)
14538 rtx jump_around_label
= NULL
;
14539 HOST_WIDE_INT align
= 1;
14540 unsigned HOST_WIDE_INT count
= 0;
14541 HOST_WIDE_INT expected_size
= -1;
14542 int size_needed
= 0, epilogue_size_needed
;
14543 int desired_align
= 0;
14544 enum stringop_alg alg
;
14547 if (CONST_INT_P (align_exp
))
14548 align
= INTVAL (align_exp
);
14549 /* i386 can do misaligned access on reasonably increased cost. */
14550 if (CONST_INT_P (expected_align_exp
)
14551 && INTVAL (expected_align_exp
) > align
)
14552 align
= INTVAL (expected_align_exp
);
14553 if (CONST_INT_P (count_exp
))
14554 count
= expected_size
= INTVAL (count_exp
);
14555 if (CONST_INT_P (expected_size_exp
) && count
== 0)
14556 expected_size
= INTVAL (expected_size_exp
);
14558 /* Step 0: Decide on preferred algorithm, desired alignment and
14559 size of chunks to be copied by main loop. */
14561 alg
= decide_alg (count
, expected_size
, false, &dynamic_check
);
14562 desired_align
= decide_alignment (align
, alg
, expected_size
);
14564 if (!TARGET_ALIGN_STRINGOPS
)
14565 align
= desired_align
;
14567 if (alg
== libcall
)
14569 gcc_assert (alg
!= no_stringop
);
14571 count_exp
= copy_to_mode_reg (GET_MODE (count_exp
), count_exp
);
14572 destreg
= copy_to_mode_reg (Pmode
, XEXP (dst
, 0));
14573 srcreg
= copy_to_mode_reg (Pmode
, XEXP (src
, 0));
14578 gcc_unreachable ();
14580 size_needed
= GET_MODE_SIZE (Pmode
);
14582 case unrolled_loop
:
14583 size_needed
= GET_MODE_SIZE (Pmode
) * (TARGET_64BIT
? 4 : 2);
14585 case rep_prefix_8_byte
:
14588 case rep_prefix_4_byte
:
14591 case rep_prefix_1_byte
:
14597 epilogue_size_needed
= size_needed
;
14599 /* Step 1: Prologue guard. */
14601 /* Alignment code needs count to be in register. */
14602 if (CONST_INT_P (count_exp
) && desired_align
> align
)
14604 enum machine_mode mode
= SImode
;
14605 if (TARGET_64BIT
&& (count
& ~0xffffffff))
14607 count_exp
= force_reg (mode
, count_exp
);
14609 gcc_assert (desired_align
>= 1 && align
>= 1);
14611 /* Ensure that alignment prologue won't copy past end of block. */
14612 if (size_needed
> 1 || (desired_align
> 1 && desired_align
> align
))
14614 epilogue_size_needed
= MAX (size_needed
- 1, desired_align
- align
);
14615 /* Epilogue always copies COUNT_EXP & EPILOGUE_SIZE_NEEDED bytes.
14616 Make sure it is power of 2. */
14617 epilogue_size_needed
= smallest_pow2_greater_than (epilogue_size_needed
);
14619 label
= gen_label_rtx ();
14620 emit_cmp_and_jump_insns (count_exp
,
14621 GEN_INT (epilogue_size_needed
),
14622 LTU
, 0, counter_mode (count_exp
), 1, label
);
14623 if (GET_CODE (count_exp
) == CONST_INT
)
14625 else if (expected_size
== -1 || expected_size
< epilogue_size_needed
)
14626 predict_jump (REG_BR_PROB_BASE
* 60 / 100);
14628 predict_jump (REG_BR_PROB_BASE
* 20 / 100);
14630 /* Emit code to decide on runtime whether library call or inline should be
14632 if (dynamic_check
!= -1)
14634 rtx hot_label
= gen_label_rtx ();
14635 jump_around_label
= gen_label_rtx ();
14636 emit_cmp_and_jump_insns (count_exp
, GEN_INT (dynamic_check
- 1),
14637 LEU
, 0, GET_MODE (count_exp
), 1, hot_label
);
14638 predict_jump (REG_BR_PROB_BASE
* 90 / 100);
14639 emit_block_move_via_libcall (dst
, src
, count_exp
, false);
14640 emit_jump (jump_around_label
);
14641 emit_label (hot_label
);
14644 /* Step 2: Alignment prologue. */
14646 if (desired_align
> align
)
14648 /* Except for the first move in epilogue, we no longer know
14649 constant offset in aliasing info. It don't seems to worth
14650 the pain to maintain it for the first move, so throw away
14652 src
= change_address (src
, BLKmode
, srcreg
);
14653 dst
= change_address (dst
, BLKmode
, destreg
);
14654 expand_movmem_prologue (dst
, src
, destreg
, srcreg
, count_exp
, align
,
14657 if (label
&& size_needed
== 1)
14659 emit_label (label
);
14660 LABEL_NUSES (label
) = 1;
14664 /* Step 3: Main loop. */
14670 gcc_unreachable ();
14672 expand_set_or_movmem_via_loop (dst
, src
, destreg
, srcreg
, NULL
,
14673 count_exp
, QImode
, 1, expected_size
);
14676 expand_set_or_movmem_via_loop (dst
, src
, destreg
, srcreg
, NULL
,
14677 count_exp
, Pmode
, 1, expected_size
);
14679 case unrolled_loop
:
14680 /* Unroll only by factor of 2 in 32bit mode, since we don't have enough
14681 registers for 4 temporaries anyway. */
14682 expand_set_or_movmem_via_loop (dst
, src
, destreg
, srcreg
, NULL
,
14683 count_exp
, Pmode
, TARGET_64BIT
? 4 : 2,
14686 case rep_prefix_8_byte
:
14687 expand_movmem_via_rep_mov (dst
, src
, destreg
, srcreg
, count_exp
,
14690 case rep_prefix_4_byte
:
14691 expand_movmem_via_rep_mov (dst
, src
, destreg
, srcreg
, count_exp
,
14694 case rep_prefix_1_byte
:
14695 expand_movmem_via_rep_mov (dst
, src
, destreg
, srcreg
, count_exp
,
14699 /* Adjust properly the offset of src and dest memory for aliasing. */
14700 if (CONST_INT_P (count_exp
))
14702 src
= adjust_automodify_address_nv (src
, BLKmode
, srcreg
,
14703 (count
/ size_needed
) * size_needed
);
14704 dst
= adjust_automodify_address_nv (dst
, BLKmode
, destreg
,
14705 (count
/ size_needed
) * size_needed
);
14709 src
= change_address (src
, BLKmode
, srcreg
);
14710 dst
= change_address (dst
, BLKmode
, destreg
);
14713 /* Step 4: Epilogue to copy the remaining bytes. */
14717 /* When the main loop is done, COUNT_EXP might hold original count,
14718 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
14719 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
14720 bytes. Compensate if needed. */
14722 if (size_needed
< epilogue_size_needed
)
14725 expand_simple_binop (counter_mode (count_exp
), AND
, count_exp
,
14726 GEN_INT (size_needed
- 1), count_exp
, 1,
14728 if (tmp
!= count_exp
)
14729 emit_move_insn (count_exp
, tmp
);
14731 emit_label (label
);
14732 LABEL_NUSES (label
) = 1;
14735 if (count_exp
!= const0_rtx
&& epilogue_size_needed
> 1)
14736 expand_movmem_epilogue (dst
, src
, destreg
, srcreg
, count_exp
,
14737 epilogue_size_needed
);
14738 if (jump_around_label
)
14739 emit_label (jump_around_label
);
14743 /* Helper function for memcpy. For QImode value 0xXY produce
14744 0xXYXYXYXY of wide specified by MODE. This is essentially
14745 a * 0x10101010, but we can do slightly better than
14746 synth_mult by unwinding the sequence by hand on CPUs with
14749 promote_duplicated_reg (enum machine_mode mode
, rtx val
)
14751 enum machine_mode valmode
= GET_MODE (val
);
14753 int nops
= mode
== DImode
? 3 : 2;
14755 gcc_assert (mode
== SImode
|| mode
== DImode
);
14756 if (val
== const0_rtx
)
14757 return copy_to_mode_reg (mode
, const0_rtx
);
14758 if (CONST_INT_P (val
))
14760 HOST_WIDE_INT v
= INTVAL (val
) & 255;
14764 if (mode
== DImode
)
14765 v
|= (v
<< 16) << 16;
14766 return copy_to_mode_reg (mode
, gen_int_mode (v
, mode
));
14769 if (valmode
== VOIDmode
)
14771 if (valmode
!= QImode
)
14772 val
= gen_lowpart (QImode
, val
);
14773 if (mode
== QImode
)
14775 if (!TARGET_PARTIAL_REG_STALL
)
14777 if (ix86_cost
->mult_init
[mode
== DImode
? 3 : 2]
14778 + ix86_cost
->mult_bit
* (mode
== DImode
? 8 : 4)
14779 <= (ix86_cost
->shift_const
+ ix86_cost
->add
) * nops
14780 + (COSTS_N_INSNS (TARGET_PARTIAL_REG_STALL
== 0)))
14782 rtx reg
= convert_modes (mode
, QImode
, val
, true);
14783 tmp
= promote_duplicated_reg (mode
, const1_rtx
);
14784 return expand_simple_binop (mode
, MULT
, reg
, tmp
, NULL
, 1,
14789 rtx reg
= convert_modes (mode
, QImode
, val
, true);
14791 if (!TARGET_PARTIAL_REG_STALL
)
14792 if (mode
== SImode
)
14793 emit_insn (gen_movsi_insv_1 (reg
, reg
));
14795 emit_insn (gen_movdi_insv_1_rex64 (reg
, reg
));
14798 tmp
= expand_simple_binop (mode
, ASHIFT
, reg
, GEN_INT (8),
14799 NULL
, 1, OPTAB_DIRECT
);
14801 expand_simple_binop (mode
, IOR
, reg
, tmp
, reg
, 1, OPTAB_DIRECT
);
14803 tmp
= expand_simple_binop (mode
, ASHIFT
, reg
, GEN_INT (16),
14804 NULL
, 1, OPTAB_DIRECT
);
14805 reg
= expand_simple_binop (mode
, IOR
, reg
, tmp
, reg
, 1, OPTAB_DIRECT
);
14806 if (mode
== SImode
)
14808 tmp
= expand_simple_binop (mode
, ASHIFT
, reg
, GEN_INT (32),
14809 NULL
, 1, OPTAB_DIRECT
);
14810 reg
= expand_simple_binop (mode
, IOR
, reg
, tmp
, reg
, 1, OPTAB_DIRECT
);
14815 /* Duplicate value VAL using promote_duplicated_reg into maximal size that will
14816 be needed by main loop copying SIZE_NEEDED chunks and prologue getting
14817 alignment from ALIGN to DESIRED_ALIGN. */
14819 promote_duplicated_reg_to_size (rtx val
, int size_needed
, int desired_align
, int align
)
14824 && (size_needed
> 4 || (desired_align
> align
&& desired_align
> 4)))
14825 promoted_val
= promote_duplicated_reg (DImode
, val
);
14826 else if (size_needed
> 2 || (desired_align
> align
&& desired_align
> 2))
14827 promoted_val
= promote_duplicated_reg (SImode
, val
);
14828 else if (size_needed
> 1 || (desired_align
> align
&& desired_align
> 1))
14829 promoted_val
= promote_duplicated_reg (HImode
, val
);
14831 promoted_val
= val
;
14833 return promoted_val
;
14836 /* Expand string clear operation (bzero). Use i386 string operations when
14837 profitable. See expand_movmem comment for explanation of individual
14838 steps performed. */
14840 ix86_expand_setmem (rtx dst
, rtx count_exp
, rtx val_exp
, rtx align_exp
,
14841 rtx expected_align_exp
, rtx expected_size_exp
)
14846 rtx jump_around_label
= NULL
;
14847 HOST_WIDE_INT align
= 1;
14848 unsigned HOST_WIDE_INT count
= 0;
14849 HOST_WIDE_INT expected_size
= -1;
14850 int size_needed
= 0, epilogue_size_needed
;
14851 int desired_align
= 0;
14852 enum stringop_alg alg
;
14853 rtx promoted_val
= NULL
;
14854 bool force_loopy_epilogue
= false;
14857 if (CONST_INT_P (align_exp
))
14858 align
= INTVAL (align_exp
);
14859 /* i386 can do misaligned access on reasonably increased cost. */
14860 if (CONST_INT_P (expected_align_exp
)
14861 && INTVAL (expected_align_exp
) > align
)
14862 align
= INTVAL (expected_align_exp
);
14863 if (CONST_INT_P (count_exp
))
14864 count
= expected_size
= INTVAL (count_exp
);
14865 if (CONST_INT_P (expected_size_exp
) && count
== 0)
14866 expected_size
= INTVAL (expected_size_exp
);
14868 /* Step 0: Decide on preferred algorithm, desired alignment and
14869 size of chunks to be copied by main loop. */
14871 alg
= decide_alg (count
, expected_size
, true, &dynamic_check
);
14872 desired_align
= decide_alignment (align
, alg
, expected_size
);
14874 if (!TARGET_ALIGN_STRINGOPS
)
14875 align
= desired_align
;
14877 if (alg
== libcall
)
14879 gcc_assert (alg
!= no_stringop
);
14881 count_exp
= copy_to_mode_reg (counter_mode (count_exp
), count_exp
);
14882 destreg
= copy_to_mode_reg (Pmode
, XEXP (dst
, 0));
14887 gcc_unreachable ();
14889 size_needed
= GET_MODE_SIZE (Pmode
);
14891 case unrolled_loop
:
14892 size_needed
= GET_MODE_SIZE (Pmode
) * 4;
14894 case rep_prefix_8_byte
:
14897 case rep_prefix_4_byte
:
14900 case rep_prefix_1_byte
:
14905 epilogue_size_needed
= size_needed
;
14907 /* Step 1: Prologue guard. */
14909 /* Alignment code needs count to be in register. */
14910 if (CONST_INT_P (count_exp
) && desired_align
> align
)
14912 enum machine_mode mode
= SImode
;
14913 if (TARGET_64BIT
&& (count
& ~0xffffffff))
14915 count_exp
= force_reg (mode
, count_exp
);
14917 /* Do the cheap promotion to allow better CSE across the
14918 main loop and epilogue (ie one load of the big constant in the
14919 front of all code. */
14920 if (CONST_INT_P (val_exp
))
14921 promoted_val
= promote_duplicated_reg_to_size (val_exp
, size_needed
,
14922 desired_align
, align
);
14923 /* Ensure that alignment prologue won't copy past end of block. */
14924 if (size_needed
> 1 || (desired_align
> 1 && desired_align
> align
))
14926 epilogue_size_needed
= MAX (size_needed
- 1, desired_align
- align
);
14927 /* Epilogue always copies COUNT_EXP & EPILOGUE_SIZE_NEEDED bytes.
14928 Make sure it is power of 2. */
14929 epilogue_size_needed
= smallest_pow2_greater_than (epilogue_size_needed
);
14931 /* To improve performance of small blocks, we jump around the VAL
14932 promoting mode. This mean that if the promoted VAL is not constant,
14933 we might not use it in the epilogue and have to use byte
14935 if (epilogue_size_needed
> 2 && !promoted_val
)
14936 force_loopy_epilogue
= true;
14937 label
= gen_label_rtx ();
14938 emit_cmp_and_jump_insns (count_exp
,
14939 GEN_INT (epilogue_size_needed
),
14940 LTU
, 0, counter_mode (count_exp
), 1, label
);
14941 if (GET_CODE (count_exp
) == CONST_INT
)
14943 else if (expected_size
== -1 || expected_size
<= epilogue_size_needed
)
14944 predict_jump (REG_BR_PROB_BASE
* 60 / 100);
14946 predict_jump (REG_BR_PROB_BASE
* 20 / 100);
14948 if (dynamic_check
!= -1)
14950 rtx hot_label
= gen_label_rtx ();
14951 jump_around_label
= gen_label_rtx ();
14952 emit_cmp_and_jump_insns (count_exp
, GEN_INT (dynamic_check
- 1),
14953 LEU
, 0, counter_mode (count_exp
), 1, hot_label
);
14954 predict_jump (REG_BR_PROB_BASE
* 90 / 100);
14955 set_storage_via_libcall (dst
, count_exp
, val_exp
, false);
14956 emit_jump (jump_around_label
);
14957 emit_label (hot_label
);
14960 /* Step 2: Alignment prologue. */
14962 /* Do the expensive promotion once we branched off the small blocks. */
14964 promoted_val
= promote_duplicated_reg_to_size (val_exp
, size_needed
,
14965 desired_align
, align
);
14966 gcc_assert (desired_align
>= 1 && align
>= 1);
14968 if (desired_align
> align
)
14970 /* Except for the first move in epilogue, we no longer know
14971 constant offset in aliasing info. It don't seems to worth
14972 the pain to maintain it for the first move, so throw away
14974 dst
= change_address (dst
, BLKmode
, destreg
);
14975 expand_setmem_prologue (dst
, destreg
, promoted_val
, count_exp
, align
,
14978 if (label
&& size_needed
== 1)
14980 emit_label (label
);
14981 LABEL_NUSES (label
) = 1;
14985 /* Step 3: Main loop. */
14991 gcc_unreachable ();
14993 expand_set_or_movmem_via_loop (dst
, NULL
, destreg
, NULL
, promoted_val
,
14994 count_exp
, QImode
, 1, expected_size
);
14997 expand_set_or_movmem_via_loop (dst
, NULL
, destreg
, NULL
, promoted_val
,
14998 count_exp
, Pmode
, 1, expected_size
);
15000 case unrolled_loop
:
15001 expand_set_or_movmem_via_loop (dst
, NULL
, destreg
, NULL
, promoted_val
,
15002 count_exp
, Pmode
, 4, expected_size
);
15004 case rep_prefix_8_byte
:
15005 expand_setmem_via_rep_stos (dst
, destreg
, promoted_val
, count_exp
,
15008 case rep_prefix_4_byte
:
15009 expand_setmem_via_rep_stos (dst
, destreg
, promoted_val
, count_exp
,
15012 case rep_prefix_1_byte
:
15013 expand_setmem_via_rep_stos (dst
, destreg
, promoted_val
, count_exp
,
15017 /* Adjust properly the offset of src and dest memory for aliasing. */
15018 if (CONST_INT_P (count_exp
))
15019 dst
= adjust_automodify_address_nv (dst
, BLKmode
, destreg
,
15020 (count
/ size_needed
) * size_needed
);
15022 dst
= change_address (dst
, BLKmode
, destreg
);
15024 /* Step 4: Epilogue to copy the remaining bytes. */
15028 /* When the main loop is done, COUNT_EXP might hold original count,
15029 while we want to copy only COUNT_EXP & SIZE_NEEDED bytes.
15030 Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED
15031 bytes. Compensate if needed. */
15033 if (size_needed
< desired_align
- align
)
15036 expand_simple_binop (counter_mode (count_exp
), AND
, count_exp
,
15037 GEN_INT (size_needed
- 1), count_exp
, 1,
15039 size_needed
= desired_align
- align
+ 1;
15040 if (tmp
!= count_exp
)
15041 emit_move_insn (count_exp
, tmp
);
15043 emit_label (label
);
15044 LABEL_NUSES (label
) = 1;
15046 if (count_exp
!= const0_rtx
&& epilogue_size_needed
> 1)
15048 if (force_loopy_epilogue
)
15049 expand_setmem_epilogue_via_loop (dst
, destreg
, val_exp
, count_exp
,
15052 expand_setmem_epilogue (dst
, destreg
, promoted_val
, count_exp
,
15055 if (jump_around_label
)
15056 emit_label (jump_around_label
);
15060 /* Expand the appropriate insns for doing strlen if not just doing
15063 out = result, initialized with the start address
15064 align_rtx = alignment of the address.
15065 scratch = scratch register, initialized with the startaddress when
15066 not aligned, otherwise undefined
15068 This is just the body. It needs the initializations mentioned above and
15069 some address computing at the end. These things are done in i386.md. */
15072 ix86_expand_strlensi_unroll_1 (rtx out
, rtx src
, rtx align_rtx
)
15076 rtx align_2_label
= NULL_RTX
;
15077 rtx align_3_label
= NULL_RTX
;
15078 rtx align_4_label
= gen_label_rtx ();
15079 rtx end_0_label
= gen_label_rtx ();
15081 rtx tmpreg
= gen_reg_rtx (SImode
);
15082 rtx scratch
= gen_reg_rtx (SImode
);
15086 if (CONST_INT_P (align_rtx
))
15087 align
= INTVAL (align_rtx
);
15089 /* Loop to check 1..3 bytes for null to get an aligned pointer. */
15091 /* Is there a known alignment and is it less than 4? */
15094 rtx scratch1
= gen_reg_rtx (Pmode
);
15095 emit_move_insn (scratch1
, out
);
15096 /* Is there a known alignment and is it not 2? */
15099 align_3_label
= gen_label_rtx (); /* Label when aligned to 3-byte */
15100 align_2_label
= gen_label_rtx (); /* Label when aligned to 2-byte */
15102 /* Leave just the 3 lower bits. */
15103 align_rtx
= expand_binop (Pmode
, and_optab
, scratch1
, GEN_INT (3),
15104 NULL_RTX
, 0, OPTAB_WIDEN
);
15106 emit_cmp_and_jump_insns (align_rtx
, const0_rtx
, EQ
, NULL
,
15107 Pmode
, 1, align_4_label
);
15108 emit_cmp_and_jump_insns (align_rtx
, const2_rtx
, EQ
, NULL
,
15109 Pmode
, 1, align_2_label
);
15110 emit_cmp_and_jump_insns (align_rtx
, const2_rtx
, GTU
, NULL
,
15111 Pmode
, 1, align_3_label
);
15115 /* Since the alignment is 2, we have to check 2 or 0 bytes;
15116 check if is aligned to 4 - byte. */
15118 align_rtx
= expand_binop (Pmode
, and_optab
, scratch1
, const2_rtx
,
15119 NULL_RTX
, 0, OPTAB_WIDEN
);
15121 emit_cmp_and_jump_insns (align_rtx
, const0_rtx
, EQ
, NULL
,
15122 Pmode
, 1, align_4_label
);
15125 mem
= change_address (src
, QImode
, out
);
15127 /* Now compare the bytes. */
15129 /* Compare the first n unaligned byte on a byte per byte basis. */
15130 emit_cmp_and_jump_insns (mem
, const0_rtx
, EQ
, NULL
,
15131 QImode
, 1, end_0_label
);
15133 /* Increment the address. */
15135 emit_insn (gen_adddi3 (out
, out
, const1_rtx
));
15137 emit_insn (gen_addsi3 (out
, out
, const1_rtx
));
15139 /* Not needed with an alignment of 2 */
15142 emit_label (align_2_label
);
15144 emit_cmp_and_jump_insns (mem
, const0_rtx
, EQ
, NULL
, QImode
, 1,
15148 emit_insn (gen_adddi3 (out
, out
, const1_rtx
));
15150 emit_insn (gen_addsi3 (out
, out
, const1_rtx
));
15152 emit_label (align_3_label
);
15155 emit_cmp_and_jump_insns (mem
, const0_rtx
, EQ
, NULL
, QImode
, 1,
15159 emit_insn (gen_adddi3 (out
, out
, const1_rtx
));
15161 emit_insn (gen_addsi3 (out
, out
, const1_rtx
));
15164 /* Generate loop to check 4 bytes at a time. It is not a good idea to
15165 align this loop. It gives only huge programs, but does not help to
15167 emit_label (align_4_label
);
15169 mem
= change_address (src
, SImode
, out
);
15170 emit_move_insn (scratch
, mem
);
15172 emit_insn (gen_adddi3 (out
, out
, GEN_INT (4)));
15174 emit_insn (gen_addsi3 (out
, out
, GEN_INT (4)));
15176 /* This formula yields a nonzero result iff one of the bytes is zero.
15177 This saves three branches inside loop and many cycles. */
15179 emit_insn (gen_addsi3 (tmpreg
, scratch
, GEN_INT (-0x01010101)));
15180 emit_insn (gen_one_cmplsi2 (scratch
, scratch
));
15181 emit_insn (gen_andsi3 (tmpreg
, tmpreg
, scratch
));
15182 emit_insn (gen_andsi3 (tmpreg
, tmpreg
,
15183 gen_int_mode (0x80808080, SImode
)));
15184 emit_cmp_and_jump_insns (tmpreg
, const0_rtx
, EQ
, 0, SImode
, 1,
15189 rtx reg
= gen_reg_rtx (SImode
);
15190 rtx reg2
= gen_reg_rtx (Pmode
);
15191 emit_move_insn (reg
, tmpreg
);
15192 emit_insn (gen_lshrsi3 (reg
, reg
, GEN_INT (16)));
15194 /* If zero is not in the first two bytes, move two bytes forward. */
15195 emit_insn (gen_testsi_ccno_1 (tmpreg
, GEN_INT (0x8080)));
15196 tmp
= gen_rtx_REG (CCNOmode
, FLAGS_REG
);
15197 tmp
= gen_rtx_EQ (VOIDmode
, tmp
, const0_rtx
);
15198 emit_insn (gen_rtx_SET (VOIDmode
, tmpreg
,
15199 gen_rtx_IF_THEN_ELSE (SImode
, tmp
,
15202 /* Emit lea manually to avoid clobbering of flags. */
15203 emit_insn (gen_rtx_SET (SImode
, reg2
,
15204 gen_rtx_PLUS (Pmode
, out
, const2_rtx
)));
15206 tmp
= gen_rtx_REG (CCNOmode
, FLAGS_REG
);
15207 tmp
= gen_rtx_EQ (VOIDmode
, tmp
, const0_rtx
);
15208 emit_insn (gen_rtx_SET (VOIDmode
, out
,
15209 gen_rtx_IF_THEN_ELSE (Pmode
, tmp
,
15216 rtx end_2_label
= gen_label_rtx ();
15217 /* Is zero in the first two bytes? */
15219 emit_insn (gen_testsi_ccno_1 (tmpreg
, GEN_INT (0x8080)));
15220 tmp
= gen_rtx_REG (CCNOmode
, FLAGS_REG
);
15221 tmp
= gen_rtx_NE (VOIDmode
, tmp
, const0_rtx
);
15222 tmp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, tmp
,
15223 gen_rtx_LABEL_REF (VOIDmode
, end_2_label
),
15225 tmp
= emit_jump_insn (gen_rtx_SET (VOIDmode
, pc_rtx
, tmp
));
15226 JUMP_LABEL (tmp
) = end_2_label
;
15228 /* Not in the first two. Move two bytes forward. */
15229 emit_insn (gen_lshrsi3 (tmpreg
, tmpreg
, GEN_INT (16)));
15231 emit_insn (gen_adddi3 (out
, out
, const2_rtx
));
15233 emit_insn (gen_addsi3 (out
, out
, const2_rtx
));
15235 emit_label (end_2_label
);
15239 /* Avoid branch in fixing the byte. */
15240 tmpreg
= gen_lowpart (QImode
, tmpreg
);
15241 emit_insn (gen_addqi3_cc (tmpreg
, tmpreg
, tmpreg
));
15242 cmp
= gen_rtx_LTU (Pmode
, gen_rtx_REG (CCmode
, 17), const0_rtx
);
15244 emit_insn (gen_subdi3_carry_rex64 (out
, out
, GEN_INT (3), cmp
));
15246 emit_insn (gen_subsi3_carry (out
, out
, GEN_INT (3), cmp
));
15248 emit_label (end_0_label
);
15251 /* Expand strlen. */
15254 ix86_expand_strlen (rtx out
, rtx src
, rtx eoschar
, rtx align
)
15256 rtx addr
, scratch1
, scratch2
, scratch3
, scratch4
;
15258 /* The generic case of strlen expander is long. Avoid it's
15259 expanding unless TARGET_INLINE_ALL_STRINGOPS. */
15261 if (TARGET_UNROLL_STRLEN
&& eoschar
== const0_rtx
&& optimize
> 1
15262 && !TARGET_INLINE_ALL_STRINGOPS
15264 && (!CONST_INT_P (align
) || INTVAL (align
) < 4))
15267 addr
= force_reg (Pmode
, XEXP (src
, 0));
15268 scratch1
= gen_reg_rtx (Pmode
);
15270 if (TARGET_UNROLL_STRLEN
&& eoschar
== const0_rtx
&& optimize
> 1
15273 /* Well it seems that some optimizer does not combine a call like
15274 foo(strlen(bar), strlen(bar));
15275 when the move and the subtraction is done here. It does calculate
15276 the length just once when these instructions are done inside of
15277 output_strlen_unroll(). But I think since &bar[strlen(bar)] is
15278 often used and I use one fewer register for the lifetime of
15279 output_strlen_unroll() this is better. */
15281 emit_move_insn (out
, addr
);
15283 ix86_expand_strlensi_unroll_1 (out
, src
, align
);
15285 /* strlensi_unroll_1 returns the address of the zero at the end of
15286 the string, like memchr(), so compute the length by subtracting
15287 the start address. */
15289 emit_insn (gen_subdi3 (out
, out
, addr
));
15291 emit_insn (gen_subsi3 (out
, out
, addr
));
15296 scratch2
= gen_reg_rtx (Pmode
);
15297 scratch3
= gen_reg_rtx (Pmode
);
15298 scratch4
= force_reg (Pmode
, constm1_rtx
);
15300 emit_move_insn (scratch3
, addr
);
15301 eoschar
= force_reg (QImode
, eoschar
);
15303 src
= replace_equiv_address_nv (src
, scratch3
);
15305 /* If .md starts supporting :P, this can be done in .md. */
15306 unspec
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (4, src
, eoschar
, align
,
15307 scratch4
), UNSPEC_SCAS
);
15308 emit_insn (gen_strlenqi_1 (scratch1
, scratch3
, unspec
));
15311 emit_insn (gen_one_cmpldi2 (scratch2
, scratch1
));
15312 emit_insn (gen_adddi3 (out
, scratch2
, constm1_rtx
));
15316 emit_insn (gen_one_cmplsi2 (scratch2
, scratch1
));
15317 emit_insn (gen_addsi3 (out
, scratch2
, constm1_rtx
));
15323 /* For given symbol (function) construct code to compute address of it's PLT
15324 entry in large x86-64 PIC model. */
15326 construct_plt_address (rtx symbol
)
15328 rtx tmp
= gen_reg_rtx (Pmode
);
15329 rtx unspec
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, symbol
), UNSPEC_PLTOFF
);
15331 gcc_assert (GET_CODE (symbol
) == SYMBOL_REF
);
15332 gcc_assert (ix86_cmodel
== CM_LARGE_PIC
);
15334 emit_move_insn (tmp
, gen_rtx_CONST (Pmode
, unspec
));
15335 emit_insn (gen_adddi3 (tmp
, tmp
, pic_offset_table_rtx
));
15340 ix86_expand_call (rtx retval
, rtx fnaddr
, rtx callarg1
,
15341 rtx callarg2 ATTRIBUTE_UNUSED
,
15342 rtx pop
, int sibcall
)
15344 rtx use
= NULL
, call
;
15346 if (pop
== const0_rtx
)
15348 gcc_assert (!TARGET_64BIT
|| !pop
);
15350 if (TARGET_MACHO
&& !TARGET_64BIT
)
15353 if (flag_pic
&& GET_CODE (XEXP (fnaddr
, 0)) == SYMBOL_REF
)
15354 fnaddr
= machopic_indirect_call_target (fnaddr
);
15359 /* Static functions and indirect calls don't need the pic register. */
15360 if (flag_pic
&& (!TARGET_64BIT
|| ix86_cmodel
== CM_LARGE_PIC
)
15361 && GET_CODE (XEXP (fnaddr
, 0)) == SYMBOL_REF
15362 && ! SYMBOL_REF_LOCAL_P (XEXP (fnaddr
, 0)))
15363 use_reg (&use
, pic_offset_table_rtx
);
15366 if (TARGET_64BIT
&& INTVAL (callarg2
) >= 0)
15368 rtx al
= gen_rtx_REG (QImode
, 0);
15369 emit_move_insn (al
, callarg2
);
15370 use_reg (&use
, al
);
15373 if (ix86_cmodel
== CM_LARGE_PIC
15374 && GET_CODE (fnaddr
) == MEM
15375 && GET_CODE (XEXP (fnaddr
, 0)) == SYMBOL_REF
15376 && !local_symbolic_operand (XEXP (fnaddr
, 0), VOIDmode
))
15377 fnaddr
= gen_rtx_MEM (QImode
, construct_plt_address (XEXP (fnaddr
, 0)));
15378 else if (! call_insn_operand (XEXP (fnaddr
, 0), Pmode
))
15380 fnaddr
= copy_to_mode_reg (Pmode
, XEXP (fnaddr
, 0));
15381 fnaddr
= gen_rtx_MEM (QImode
, fnaddr
);
15383 if (sibcall
&& TARGET_64BIT
15384 && !constant_call_address_operand (XEXP (fnaddr
, 0), Pmode
))
15387 addr
= copy_to_mode_reg (Pmode
, XEXP (fnaddr
, 0));
15388 fnaddr
= gen_rtx_REG (Pmode
, R11_REG
);
15389 emit_move_insn (fnaddr
, addr
);
15390 fnaddr
= gen_rtx_MEM (QImode
, fnaddr
);
15393 call
= gen_rtx_CALL (VOIDmode
, fnaddr
, callarg1
);
15395 call
= gen_rtx_SET (VOIDmode
, retval
, call
);
15398 pop
= gen_rtx_PLUS (Pmode
, stack_pointer_rtx
, pop
);
15399 pop
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
, pop
);
15400 call
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, call
, pop
));
15403 call
= emit_call_insn (call
);
15405 CALL_INSN_FUNCTION_USAGE (call
) = use
;
15409 /* Clear stack slot assignments remembered from previous functions.
15410 This is called from INIT_EXPANDERS once before RTL is emitted for each
15413 static struct machine_function
*
15414 ix86_init_machine_status (void)
15416 struct machine_function
*f
;
15418 f
= ggc_alloc_cleared (sizeof (struct machine_function
));
15419 f
->use_fast_prologue_epilogue_nregs
= -1;
15420 f
->tls_descriptor_call_expanded_p
= 0;
15425 /* Return a MEM corresponding to a stack slot with mode MODE.
15426 Allocate a new slot if necessary.
15428 The RTL for a function can have several slots available: N is
15429 which slot to use. */
15432 assign_386_stack_local (enum machine_mode mode
, enum ix86_stack_slot n
)
15434 struct stack_local_entry
*s
;
15436 gcc_assert (n
< MAX_386_STACK_LOCALS
);
15438 for (s
= ix86_stack_locals
; s
; s
= s
->next
)
15439 if (s
->mode
== mode
&& s
->n
== n
)
15440 return copy_rtx (s
->rtl
);
15442 s
= (struct stack_local_entry
*)
15443 ggc_alloc (sizeof (struct stack_local_entry
));
15446 s
->rtl
= assign_stack_local (mode
, GET_MODE_SIZE (mode
), 0);
15448 s
->next
= ix86_stack_locals
;
15449 ix86_stack_locals
= s
;
15453 /* Construct the SYMBOL_REF for the tls_get_addr function. */
15455 static GTY(()) rtx ix86_tls_symbol
;
15457 ix86_tls_get_addr (void)
15460 if (!ix86_tls_symbol
)
15462 ix86_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
,
15463 (TARGET_ANY_GNU_TLS
15465 ? "___tls_get_addr"
15466 : "__tls_get_addr");
15469 return ix86_tls_symbol
;
15472 /* Construct the SYMBOL_REF for the _TLS_MODULE_BASE_ symbol. */
15474 static GTY(()) rtx ix86_tls_module_base_symbol
;
15476 ix86_tls_module_base (void)
15479 if (!ix86_tls_module_base_symbol
)
15481 ix86_tls_module_base_symbol
= gen_rtx_SYMBOL_REF (Pmode
,
15482 "_TLS_MODULE_BASE_");
15483 SYMBOL_REF_FLAGS (ix86_tls_module_base_symbol
)
15484 |= TLS_MODEL_GLOBAL_DYNAMIC
<< SYMBOL_FLAG_TLS_SHIFT
;
15487 return ix86_tls_module_base_symbol
;
15490 /* Calculate the length of the memory address in the instruction
15491 encoding. Does not include the one-byte modrm, opcode, or prefix. */
15494 memory_address_length (rtx addr
)
15496 struct ix86_address parts
;
15497 rtx base
, index
, disp
;
15501 if (GET_CODE (addr
) == PRE_DEC
15502 || GET_CODE (addr
) == POST_INC
15503 || GET_CODE (addr
) == PRE_MODIFY
15504 || GET_CODE (addr
) == POST_MODIFY
)
15507 ok
= ix86_decompose_address (addr
, &parts
);
15510 if (parts
.base
&& GET_CODE (parts
.base
) == SUBREG
)
15511 parts
.base
= SUBREG_REG (parts
.base
);
15512 if (parts
.index
&& GET_CODE (parts
.index
) == SUBREG
)
15513 parts
.index
= SUBREG_REG (parts
.index
);
15516 index
= parts
.index
;
15521 - esp as the base always wants an index,
15522 - ebp as the base always wants a displacement. */
15524 /* Register Indirect. */
15525 if (base
&& !index
&& !disp
)
15527 /* esp (for its index) and ebp (for its displacement) need
15528 the two-byte modrm form. */
15529 if (addr
== stack_pointer_rtx
15530 || addr
== arg_pointer_rtx
15531 || addr
== frame_pointer_rtx
15532 || addr
== hard_frame_pointer_rtx
)
15536 /* Direct Addressing. */
15537 else if (disp
&& !base
&& !index
)
15542 /* Find the length of the displacement constant. */
15545 if (base
&& satisfies_constraint_K (disp
))
15550 /* ebp always wants a displacement. */
15551 else if (base
== hard_frame_pointer_rtx
)
15554 /* An index requires the two-byte modrm form.... */
15556 /* ...like esp, which always wants an index. */
15557 || base
== stack_pointer_rtx
15558 || base
== arg_pointer_rtx
15559 || base
== frame_pointer_rtx
)
15566 /* Compute default value for "length_immediate" attribute. When SHORTFORM
15567 is set, expect that insn have 8bit immediate alternative. */
15569 ix86_attr_length_immediate_default (rtx insn
, int shortform
)
15573 extract_insn_cached (insn
);
15574 for (i
= recog_data
.n_operands
- 1; i
>= 0; --i
)
15575 if (CONSTANT_P (recog_data
.operand
[i
]))
15578 if (shortform
&& satisfies_constraint_K (recog_data
.operand
[i
]))
15582 switch (get_attr_mode (insn
))
15593 /* Immediates for DImode instructions are encoded as 32bit sign extended values. */
15598 fatal_insn ("unknown insn mode", insn
);
15604 /* Compute default value for "length_address" attribute. */
15606 ix86_attr_length_address_default (rtx insn
)
15610 if (get_attr_type (insn
) == TYPE_LEA
)
15612 rtx set
= PATTERN (insn
);
15614 if (GET_CODE (set
) == PARALLEL
)
15615 set
= XVECEXP (set
, 0, 0);
15617 gcc_assert (GET_CODE (set
) == SET
);
15619 return memory_address_length (SET_SRC (set
));
15622 extract_insn_cached (insn
);
15623 for (i
= recog_data
.n_operands
- 1; i
>= 0; --i
)
15624 if (MEM_P (recog_data
.operand
[i
]))
15626 return memory_address_length (XEXP (recog_data
.operand
[i
], 0));
15632 /* Return the maximum number of instructions a cpu can issue. */
15635 ix86_issue_rate (void)
15639 case PROCESSOR_PENTIUM
:
15643 case PROCESSOR_PENTIUMPRO
:
15644 case PROCESSOR_PENTIUM4
:
15645 case PROCESSOR_ATHLON
:
15647 case PROCESSOR_AMDFAM10
:
15648 case PROCESSOR_NOCONA
:
15649 case PROCESSOR_GENERIC32
:
15650 case PROCESSOR_GENERIC64
:
15653 case PROCESSOR_CORE2
:
15661 /* A subroutine of ix86_adjust_cost -- return true iff INSN reads flags set
15662 by DEP_INSN and nothing set by DEP_INSN. */
15665 ix86_flags_dependent (rtx insn
, rtx dep_insn
, enum attr_type insn_type
)
15669 /* Simplify the test for uninteresting insns. */
15670 if (insn_type
!= TYPE_SETCC
15671 && insn_type
!= TYPE_ICMOV
15672 && insn_type
!= TYPE_FCMOV
15673 && insn_type
!= TYPE_IBR
)
15676 if ((set
= single_set (dep_insn
)) != 0)
15678 set
= SET_DEST (set
);
15681 else if (GET_CODE (PATTERN (dep_insn
)) == PARALLEL
15682 && XVECLEN (PATTERN (dep_insn
), 0) == 2
15683 && GET_CODE (XVECEXP (PATTERN (dep_insn
), 0, 0)) == SET
15684 && GET_CODE (XVECEXP (PATTERN (dep_insn
), 0, 1)) == SET
)
15686 set
= SET_DEST (XVECEXP (PATTERN (dep_insn
), 0, 0));
15687 set2
= SET_DEST (XVECEXP (PATTERN (dep_insn
), 0, 0));
15692 if (!REG_P (set
) || REGNO (set
) != FLAGS_REG
)
15695 /* This test is true if the dependent insn reads the flags but
15696 not any other potentially set register. */
15697 if (!reg_overlap_mentioned_p (set
, PATTERN (insn
)))
15700 if (set2
&& reg_overlap_mentioned_p (set2
, PATTERN (insn
)))
15706 /* A subroutine of ix86_adjust_cost -- return true iff INSN has a memory
15707 address with operands set by DEP_INSN. */
15710 ix86_agi_dependent (rtx insn
, rtx dep_insn
, enum attr_type insn_type
)
15714 if (insn_type
== TYPE_LEA
15717 addr
= PATTERN (insn
);
15719 if (GET_CODE (addr
) == PARALLEL
)
15720 addr
= XVECEXP (addr
, 0, 0);
15722 gcc_assert (GET_CODE (addr
) == SET
);
15724 addr
= SET_SRC (addr
);
15729 extract_insn_cached (insn
);
15730 for (i
= recog_data
.n_operands
- 1; i
>= 0; --i
)
15731 if (MEM_P (recog_data
.operand
[i
]))
15733 addr
= XEXP (recog_data
.operand
[i
], 0);
15740 return modified_in_p (addr
, dep_insn
);
15744 ix86_adjust_cost (rtx insn
, rtx link
, rtx dep_insn
, int cost
)
15746 enum attr_type insn_type
, dep_insn_type
;
15747 enum attr_memory memory
;
15749 int dep_insn_code_number
;
15751 /* Anti and output dependencies have zero cost on all CPUs. */
15752 if (REG_NOTE_KIND (link
) != 0)
15755 dep_insn_code_number
= recog_memoized (dep_insn
);
15757 /* If we can't recognize the insns, we can't really do anything. */
15758 if (dep_insn_code_number
< 0 || recog_memoized (insn
) < 0)
15761 insn_type
= get_attr_type (insn
);
15762 dep_insn_type
= get_attr_type (dep_insn
);
15766 case PROCESSOR_PENTIUM
:
15767 /* Address Generation Interlock adds a cycle of latency. */
15768 if (ix86_agi_dependent (insn
, dep_insn
, insn_type
))
15771 /* ??? Compares pair with jump/setcc. */
15772 if (ix86_flags_dependent (insn
, dep_insn
, insn_type
))
15775 /* Floating point stores require value to be ready one cycle earlier. */
15776 if (insn_type
== TYPE_FMOV
15777 && get_attr_memory (insn
) == MEMORY_STORE
15778 && !ix86_agi_dependent (insn
, dep_insn
, insn_type
))
15782 case PROCESSOR_PENTIUMPRO
:
15783 memory
= get_attr_memory (insn
);
15785 /* INT->FP conversion is expensive. */
15786 if (get_attr_fp_int_src (dep_insn
))
15789 /* There is one cycle extra latency between an FP op and a store. */
15790 if (insn_type
== TYPE_FMOV
15791 && (set
= single_set (dep_insn
)) != NULL_RTX
15792 && (set2
= single_set (insn
)) != NULL_RTX
15793 && rtx_equal_p (SET_DEST (set
), SET_SRC (set2
))
15794 && MEM_P (SET_DEST (set2
)))
15797 /* Show ability of reorder buffer to hide latency of load by executing
15798 in parallel with previous instruction in case
15799 previous instruction is not needed to compute the address. */
15800 if ((memory
== MEMORY_LOAD
|| memory
== MEMORY_BOTH
)
15801 && !ix86_agi_dependent (insn
, dep_insn
, insn_type
))
15803 /* Claim moves to take one cycle, as core can issue one load
15804 at time and the next load can start cycle later. */
15805 if (dep_insn_type
== TYPE_IMOV
15806 || dep_insn_type
== TYPE_FMOV
)
15814 memory
= get_attr_memory (insn
);
15816 /* The esp dependency is resolved before the instruction is really
15818 if ((insn_type
== TYPE_PUSH
|| insn_type
== TYPE_POP
)
15819 && (dep_insn_type
== TYPE_PUSH
|| dep_insn_type
== TYPE_POP
))
15822 /* INT->FP conversion is expensive. */
15823 if (get_attr_fp_int_src (dep_insn
))
15826 /* Show ability of reorder buffer to hide latency of load by executing
15827 in parallel with previous instruction in case
15828 previous instruction is not needed to compute the address. */
15829 if ((memory
== MEMORY_LOAD
|| memory
== MEMORY_BOTH
)
15830 && !ix86_agi_dependent (insn
, dep_insn
, insn_type
))
15832 /* Claim moves to take one cycle, as core can issue one load
15833 at time and the next load can start cycle later. */
15834 if (dep_insn_type
== TYPE_IMOV
15835 || dep_insn_type
== TYPE_FMOV
)
15844 case PROCESSOR_ATHLON
:
15846 case PROCESSOR_AMDFAM10
:
15847 case PROCESSOR_GENERIC32
:
15848 case PROCESSOR_GENERIC64
:
15849 memory
= get_attr_memory (insn
);
15851 /* Show ability of reorder buffer to hide latency of load by executing
15852 in parallel with previous instruction in case
15853 previous instruction is not needed to compute the address. */
15854 if ((memory
== MEMORY_LOAD
|| memory
== MEMORY_BOTH
)
15855 && !ix86_agi_dependent (insn
, dep_insn
, insn_type
))
15857 enum attr_unit unit
= get_attr_unit (insn
);
15860 /* Because of the difference between the length of integer and
15861 floating unit pipeline preparation stages, the memory operands
15862 for floating point are cheaper.
15864 ??? For Athlon it the difference is most probably 2. */
15865 if (unit
== UNIT_INTEGER
|| unit
== UNIT_UNKNOWN
)
15868 loadcost
= TARGET_ATHLON
? 2 : 0;
15870 if (cost
>= loadcost
)
15883 /* How many alternative schedules to try. This should be as wide as the
15884 scheduling freedom in the DFA, but no wider. Making this value too
15885 large results extra work for the scheduler. */
15888 ia32_multipass_dfa_lookahead (void)
15890 if (ix86_tune
== PROCESSOR_PENTIUM
)
15893 if (ix86_tune
== PROCESSOR_PENTIUMPRO
15894 || ix86_tune
== PROCESSOR_K6
)
15902 /* Compute the alignment given to a constant that is being placed in memory.
15903 EXP is the constant and ALIGN is the alignment that the object would
15905 The value of this function is used instead of that alignment to align
15909 ix86_constant_alignment (tree exp
, int align
)
15911 if (TREE_CODE (exp
) == REAL_CST
)
15913 if (TYPE_MODE (TREE_TYPE (exp
)) == DFmode
&& align
< 64)
15915 else if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (exp
))) && align
< 128)
15918 else if (!optimize_size
&& TREE_CODE (exp
) == STRING_CST
15919 && TREE_STRING_LENGTH (exp
) >= 31 && align
< BITS_PER_WORD
)
15920 return BITS_PER_WORD
;
15925 /* Compute the alignment for a static variable.
15926 TYPE is the data type, and ALIGN is the alignment that
15927 the object would ordinarily have. The value of this function is used
15928 instead of that alignment to align the object. */
15931 ix86_data_alignment (tree type
, int align
)
15933 int max_align
= optimize_size
? BITS_PER_WORD
: MIN (256, MAX_OFILE_ALIGNMENT
);
15935 if (AGGREGATE_TYPE_P (type
)
15936 && TYPE_SIZE (type
)
15937 && TREE_CODE (TYPE_SIZE (type
)) == INTEGER_CST
15938 && (TREE_INT_CST_LOW (TYPE_SIZE (type
)) >= (unsigned) max_align
15939 || TREE_INT_CST_HIGH (TYPE_SIZE (type
)))
15940 && align
< max_align
)
15943 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
15944 to 16byte boundary. */
15947 if (AGGREGATE_TYPE_P (type
)
15948 && TYPE_SIZE (type
)
15949 && TREE_CODE (TYPE_SIZE (type
)) == INTEGER_CST
15950 && (TREE_INT_CST_LOW (TYPE_SIZE (type
)) >= 128
15951 || TREE_INT_CST_HIGH (TYPE_SIZE (type
))) && align
< 128)
15955 if (TREE_CODE (type
) == ARRAY_TYPE
)
15957 if (TYPE_MODE (TREE_TYPE (type
)) == DFmode
&& align
< 64)
15959 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type
))) && align
< 128)
15962 else if (TREE_CODE (type
) == COMPLEX_TYPE
)
15965 if (TYPE_MODE (type
) == DCmode
&& align
< 64)
15967 if (TYPE_MODE (type
) == XCmode
&& align
< 128)
15970 else if ((TREE_CODE (type
) == RECORD_TYPE
15971 || TREE_CODE (type
) == UNION_TYPE
15972 || TREE_CODE (type
) == QUAL_UNION_TYPE
)
15973 && TYPE_FIELDS (type
))
15975 if (DECL_MODE (TYPE_FIELDS (type
)) == DFmode
&& align
< 64)
15977 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type
))) && align
< 128)
15980 else if (TREE_CODE (type
) == REAL_TYPE
|| TREE_CODE (type
) == VECTOR_TYPE
15981 || TREE_CODE (type
) == INTEGER_TYPE
)
15983 if (TYPE_MODE (type
) == DFmode
&& align
< 64)
15985 if (ALIGN_MODE_128 (TYPE_MODE (type
)) && align
< 128)
15992 /* Compute the alignment for a local variable.
15993 TYPE is the data type, and ALIGN is the alignment that
15994 the object would ordinarily have. The value of this macro is used
15995 instead of that alignment to align the object. */
15998 ix86_local_alignment (tree type
, int align
)
16000 /* x86-64 ABI requires arrays greater than 16 bytes to be aligned
16001 to 16byte boundary. */
16004 if (AGGREGATE_TYPE_P (type
)
16005 && TYPE_SIZE (type
)
16006 && TREE_CODE (TYPE_SIZE (type
)) == INTEGER_CST
16007 && (TREE_INT_CST_LOW (TYPE_SIZE (type
)) >= 16
16008 || TREE_INT_CST_HIGH (TYPE_SIZE (type
))) && align
< 128)
16011 if (TREE_CODE (type
) == ARRAY_TYPE
)
16013 if (TYPE_MODE (TREE_TYPE (type
)) == DFmode
&& align
< 64)
16015 if (ALIGN_MODE_128 (TYPE_MODE (TREE_TYPE (type
))) && align
< 128)
16018 else if (TREE_CODE (type
) == COMPLEX_TYPE
)
16020 if (TYPE_MODE (type
) == DCmode
&& align
< 64)
16022 if (TYPE_MODE (type
) == XCmode
&& align
< 128)
16025 else if ((TREE_CODE (type
) == RECORD_TYPE
16026 || TREE_CODE (type
) == UNION_TYPE
16027 || TREE_CODE (type
) == QUAL_UNION_TYPE
)
16028 && TYPE_FIELDS (type
))
16030 if (DECL_MODE (TYPE_FIELDS (type
)) == DFmode
&& align
< 64)
16032 if (ALIGN_MODE_128 (DECL_MODE (TYPE_FIELDS (type
))) && align
< 128)
16035 else if (TREE_CODE (type
) == REAL_TYPE
|| TREE_CODE (type
) == VECTOR_TYPE
16036 || TREE_CODE (type
) == INTEGER_TYPE
)
16039 if (TYPE_MODE (type
) == DFmode
&& align
< 64)
16041 if (ALIGN_MODE_128 (TYPE_MODE (type
)) && align
< 128)
16047 /* Emit RTL insns to initialize the variable parts of a trampoline.
16048 FNADDR is an RTX for the address of the function's pure code.
16049 CXT is an RTX for the static chain value for the function. */
16051 x86_initialize_trampoline (rtx tramp
, rtx fnaddr
, rtx cxt
)
16055 /* Compute offset from the end of the jmp to the target function. */
16056 rtx disp
= expand_binop (SImode
, sub_optab
, fnaddr
,
16057 plus_constant (tramp
, 10),
16058 NULL_RTX
, 1, OPTAB_DIRECT
);
16059 emit_move_insn (gen_rtx_MEM (QImode
, tramp
),
16060 gen_int_mode (0xb9, QImode
));
16061 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (tramp
, 1)), cxt
);
16062 emit_move_insn (gen_rtx_MEM (QImode
, plus_constant (tramp
, 5)),
16063 gen_int_mode (0xe9, QImode
));
16064 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (tramp
, 6)), disp
);
16069 /* Try to load address using shorter movl instead of movabs.
16070 We may want to support movq for kernel mode, but kernel does not use
16071 trampolines at the moment. */
16072 if (x86_64_zext_immediate_operand (fnaddr
, VOIDmode
))
16074 fnaddr
= copy_to_mode_reg (DImode
, fnaddr
);
16075 emit_move_insn (gen_rtx_MEM (HImode
, plus_constant (tramp
, offset
)),
16076 gen_int_mode (0xbb41, HImode
));
16077 emit_move_insn (gen_rtx_MEM (SImode
, plus_constant (tramp
, offset
+ 2)),
16078 gen_lowpart (SImode
, fnaddr
));
16083 emit_move_insn (gen_rtx_MEM (HImode
, plus_constant (tramp
, offset
)),
16084 gen_int_mode (0xbb49, HImode
));
16085 emit_move_insn (gen_rtx_MEM (DImode
, plus_constant (tramp
, offset
+ 2)),
16089 /* Load static chain using movabs to r10. */
16090 emit_move_insn (gen_rtx_MEM (HImode
, plus_constant (tramp
, offset
)),
16091 gen_int_mode (0xba49, HImode
));
16092 emit_move_insn (gen_rtx_MEM (DImode
, plus_constant (tramp
, offset
+ 2)),
16095 /* Jump to the r11 */
16096 emit_move_insn (gen_rtx_MEM (HImode
, plus_constant (tramp
, offset
)),
16097 gen_int_mode (0xff49, HImode
));
16098 emit_move_insn (gen_rtx_MEM (QImode
, plus_constant (tramp
, offset
+2)),
16099 gen_int_mode (0xe3, QImode
));
16101 gcc_assert (offset
<= TRAMPOLINE_SIZE
);
16104 #ifdef ENABLE_EXECUTE_STACK
16105 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "__enable_execute_stack"),
16106 LCT_NORMAL
, VOIDmode
, 1, tramp
, Pmode
);
16110 /* Codes for all the SSE/MMX builtins. */
16113 IX86_BUILTIN_ADDPS
,
16114 IX86_BUILTIN_ADDSS
,
16115 IX86_BUILTIN_DIVPS
,
16116 IX86_BUILTIN_DIVSS
,
16117 IX86_BUILTIN_MULPS
,
16118 IX86_BUILTIN_MULSS
,
16119 IX86_BUILTIN_SUBPS
,
16120 IX86_BUILTIN_SUBSS
,
16122 IX86_BUILTIN_CMPEQPS
,
16123 IX86_BUILTIN_CMPLTPS
,
16124 IX86_BUILTIN_CMPLEPS
,
16125 IX86_BUILTIN_CMPGTPS
,
16126 IX86_BUILTIN_CMPGEPS
,
16127 IX86_BUILTIN_CMPNEQPS
,
16128 IX86_BUILTIN_CMPNLTPS
,
16129 IX86_BUILTIN_CMPNLEPS
,
16130 IX86_BUILTIN_CMPNGTPS
,
16131 IX86_BUILTIN_CMPNGEPS
,
16132 IX86_BUILTIN_CMPORDPS
,
16133 IX86_BUILTIN_CMPUNORDPS
,
16134 IX86_BUILTIN_CMPEQSS
,
16135 IX86_BUILTIN_CMPLTSS
,
16136 IX86_BUILTIN_CMPLESS
,
16137 IX86_BUILTIN_CMPNEQSS
,
16138 IX86_BUILTIN_CMPNLTSS
,
16139 IX86_BUILTIN_CMPNLESS
,
16140 IX86_BUILTIN_CMPNGTSS
,
16141 IX86_BUILTIN_CMPNGESS
,
16142 IX86_BUILTIN_CMPORDSS
,
16143 IX86_BUILTIN_CMPUNORDSS
,
16145 IX86_BUILTIN_COMIEQSS
,
16146 IX86_BUILTIN_COMILTSS
,
16147 IX86_BUILTIN_COMILESS
,
16148 IX86_BUILTIN_COMIGTSS
,
16149 IX86_BUILTIN_COMIGESS
,
16150 IX86_BUILTIN_COMINEQSS
,
16151 IX86_BUILTIN_UCOMIEQSS
,
16152 IX86_BUILTIN_UCOMILTSS
,
16153 IX86_BUILTIN_UCOMILESS
,
16154 IX86_BUILTIN_UCOMIGTSS
,
16155 IX86_BUILTIN_UCOMIGESS
,
16156 IX86_BUILTIN_UCOMINEQSS
,
16158 IX86_BUILTIN_CVTPI2PS
,
16159 IX86_BUILTIN_CVTPS2PI
,
16160 IX86_BUILTIN_CVTSI2SS
,
16161 IX86_BUILTIN_CVTSI642SS
,
16162 IX86_BUILTIN_CVTSS2SI
,
16163 IX86_BUILTIN_CVTSS2SI64
,
16164 IX86_BUILTIN_CVTTPS2PI
,
16165 IX86_BUILTIN_CVTTSS2SI
,
16166 IX86_BUILTIN_CVTTSS2SI64
,
16168 IX86_BUILTIN_MAXPS
,
16169 IX86_BUILTIN_MAXSS
,
16170 IX86_BUILTIN_MINPS
,
16171 IX86_BUILTIN_MINSS
,
16173 IX86_BUILTIN_LOADUPS
,
16174 IX86_BUILTIN_STOREUPS
,
16175 IX86_BUILTIN_MOVSS
,
16177 IX86_BUILTIN_MOVHLPS
,
16178 IX86_BUILTIN_MOVLHPS
,
16179 IX86_BUILTIN_LOADHPS
,
16180 IX86_BUILTIN_LOADLPS
,
16181 IX86_BUILTIN_STOREHPS
,
16182 IX86_BUILTIN_STORELPS
,
16184 IX86_BUILTIN_MASKMOVQ
,
16185 IX86_BUILTIN_MOVMSKPS
,
16186 IX86_BUILTIN_PMOVMSKB
,
16188 IX86_BUILTIN_MOVNTPS
,
16189 IX86_BUILTIN_MOVNTQ
,
16191 IX86_BUILTIN_LOADDQU
,
16192 IX86_BUILTIN_STOREDQU
,
16194 IX86_BUILTIN_PACKSSWB
,
16195 IX86_BUILTIN_PACKSSDW
,
16196 IX86_BUILTIN_PACKUSWB
,
16198 IX86_BUILTIN_PADDB
,
16199 IX86_BUILTIN_PADDW
,
16200 IX86_BUILTIN_PADDD
,
16201 IX86_BUILTIN_PADDQ
,
16202 IX86_BUILTIN_PADDSB
,
16203 IX86_BUILTIN_PADDSW
,
16204 IX86_BUILTIN_PADDUSB
,
16205 IX86_BUILTIN_PADDUSW
,
16206 IX86_BUILTIN_PSUBB
,
16207 IX86_BUILTIN_PSUBW
,
16208 IX86_BUILTIN_PSUBD
,
16209 IX86_BUILTIN_PSUBQ
,
16210 IX86_BUILTIN_PSUBSB
,
16211 IX86_BUILTIN_PSUBSW
,
16212 IX86_BUILTIN_PSUBUSB
,
16213 IX86_BUILTIN_PSUBUSW
,
16216 IX86_BUILTIN_PANDN
,
16220 IX86_BUILTIN_PAVGB
,
16221 IX86_BUILTIN_PAVGW
,
16223 IX86_BUILTIN_PCMPEQB
,
16224 IX86_BUILTIN_PCMPEQW
,
16225 IX86_BUILTIN_PCMPEQD
,
16226 IX86_BUILTIN_PCMPGTB
,
16227 IX86_BUILTIN_PCMPGTW
,
16228 IX86_BUILTIN_PCMPGTD
,
16230 IX86_BUILTIN_PMADDWD
,
16232 IX86_BUILTIN_PMAXSW
,
16233 IX86_BUILTIN_PMAXUB
,
16234 IX86_BUILTIN_PMINSW
,
16235 IX86_BUILTIN_PMINUB
,
16237 IX86_BUILTIN_PMULHUW
,
16238 IX86_BUILTIN_PMULHW
,
16239 IX86_BUILTIN_PMULLW
,
16241 IX86_BUILTIN_PSADBW
,
16242 IX86_BUILTIN_PSHUFW
,
16244 IX86_BUILTIN_PSLLW
,
16245 IX86_BUILTIN_PSLLD
,
16246 IX86_BUILTIN_PSLLQ
,
16247 IX86_BUILTIN_PSRAW
,
16248 IX86_BUILTIN_PSRAD
,
16249 IX86_BUILTIN_PSRLW
,
16250 IX86_BUILTIN_PSRLD
,
16251 IX86_BUILTIN_PSRLQ
,
16252 IX86_BUILTIN_PSLLWI
,
16253 IX86_BUILTIN_PSLLDI
,
16254 IX86_BUILTIN_PSLLQI
,
16255 IX86_BUILTIN_PSRAWI
,
16256 IX86_BUILTIN_PSRADI
,
16257 IX86_BUILTIN_PSRLWI
,
16258 IX86_BUILTIN_PSRLDI
,
16259 IX86_BUILTIN_PSRLQI
,
16261 IX86_BUILTIN_PUNPCKHBW
,
16262 IX86_BUILTIN_PUNPCKHWD
,
16263 IX86_BUILTIN_PUNPCKHDQ
,
16264 IX86_BUILTIN_PUNPCKLBW
,
16265 IX86_BUILTIN_PUNPCKLWD
,
16266 IX86_BUILTIN_PUNPCKLDQ
,
16268 IX86_BUILTIN_SHUFPS
,
16270 IX86_BUILTIN_RCPPS
,
16271 IX86_BUILTIN_RCPSS
,
16272 IX86_BUILTIN_RSQRTPS
,
16273 IX86_BUILTIN_RSQRTSS
,
16274 IX86_BUILTIN_SQRTPS
,
16275 IX86_BUILTIN_SQRTSS
,
16277 IX86_BUILTIN_UNPCKHPS
,
16278 IX86_BUILTIN_UNPCKLPS
,
16280 IX86_BUILTIN_ANDPS
,
16281 IX86_BUILTIN_ANDNPS
,
16283 IX86_BUILTIN_XORPS
,
16286 IX86_BUILTIN_LDMXCSR
,
16287 IX86_BUILTIN_STMXCSR
,
16288 IX86_BUILTIN_SFENCE
,
16290 /* 3DNow! Original */
16291 IX86_BUILTIN_FEMMS
,
16292 IX86_BUILTIN_PAVGUSB
,
16293 IX86_BUILTIN_PF2ID
,
16294 IX86_BUILTIN_PFACC
,
16295 IX86_BUILTIN_PFADD
,
16296 IX86_BUILTIN_PFCMPEQ
,
16297 IX86_BUILTIN_PFCMPGE
,
16298 IX86_BUILTIN_PFCMPGT
,
16299 IX86_BUILTIN_PFMAX
,
16300 IX86_BUILTIN_PFMIN
,
16301 IX86_BUILTIN_PFMUL
,
16302 IX86_BUILTIN_PFRCP
,
16303 IX86_BUILTIN_PFRCPIT1
,
16304 IX86_BUILTIN_PFRCPIT2
,
16305 IX86_BUILTIN_PFRSQIT1
,
16306 IX86_BUILTIN_PFRSQRT
,
16307 IX86_BUILTIN_PFSUB
,
16308 IX86_BUILTIN_PFSUBR
,
16309 IX86_BUILTIN_PI2FD
,
16310 IX86_BUILTIN_PMULHRW
,
16312 /* 3DNow! Athlon Extensions */
16313 IX86_BUILTIN_PF2IW
,
16314 IX86_BUILTIN_PFNACC
,
16315 IX86_BUILTIN_PFPNACC
,
16316 IX86_BUILTIN_PI2FW
,
16317 IX86_BUILTIN_PSWAPDSI
,
16318 IX86_BUILTIN_PSWAPDSF
,
16321 IX86_BUILTIN_ADDPD
,
16322 IX86_BUILTIN_ADDSD
,
16323 IX86_BUILTIN_DIVPD
,
16324 IX86_BUILTIN_DIVSD
,
16325 IX86_BUILTIN_MULPD
,
16326 IX86_BUILTIN_MULSD
,
16327 IX86_BUILTIN_SUBPD
,
16328 IX86_BUILTIN_SUBSD
,
16330 IX86_BUILTIN_CMPEQPD
,
16331 IX86_BUILTIN_CMPLTPD
,
16332 IX86_BUILTIN_CMPLEPD
,
16333 IX86_BUILTIN_CMPGTPD
,
16334 IX86_BUILTIN_CMPGEPD
,
16335 IX86_BUILTIN_CMPNEQPD
,
16336 IX86_BUILTIN_CMPNLTPD
,
16337 IX86_BUILTIN_CMPNLEPD
,
16338 IX86_BUILTIN_CMPNGTPD
,
16339 IX86_BUILTIN_CMPNGEPD
,
16340 IX86_BUILTIN_CMPORDPD
,
16341 IX86_BUILTIN_CMPUNORDPD
,
16342 IX86_BUILTIN_CMPEQSD
,
16343 IX86_BUILTIN_CMPLTSD
,
16344 IX86_BUILTIN_CMPLESD
,
16345 IX86_BUILTIN_CMPNEQSD
,
16346 IX86_BUILTIN_CMPNLTSD
,
16347 IX86_BUILTIN_CMPNLESD
,
16348 IX86_BUILTIN_CMPORDSD
,
16349 IX86_BUILTIN_CMPUNORDSD
,
16351 IX86_BUILTIN_COMIEQSD
,
16352 IX86_BUILTIN_COMILTSD
,
16353 IX86_BUILTIN_COMILESD
,
16354 IX86_BUILTIN_COMIGTSD
,
16355 IX86_BUILTIN_COMIGESD
,
16356 IX86_BUILTIN_COMINEQSD
,
16357 IX86_BUILTIN_UCOMIEQSD
,
16358 IX86_BUILTIN_UCOMILTSD
,
16359 IX86_BUILTIN_UCOMILESD
,
16360 IX86_BUILTIN_UCOMIGTSD
,
16361 IX86_BUILTIN_UCOMIGESD
,
16362 IX86_BUILTIN_UCOMINEQSD
,
16364 IX86_BUILTIN_MAXPD
,
16365 IX86_BUILTIN_MAXSD
,
16366 IX86_BUILTIN_MINPD
,
16367 IX86_BUILTIN_MINSD
,
16369 IX86_BUILTIN_ANDPD
,
16370 IX86_BUILTIN_ANDNPD
,
16372 IX86_BUILTIN_XORPD
,
16374 IX86_BUILTIN_SQRTPD
,
16375 IX86_BUILTIN_SQRTSD
,
16377 IX86_BUILTIN_UNPCKHPD
,
16378 IX86_BUILTIN_UNPCKLPD
,
16380 IX86_BUILTIN_SHUFPD
,
16382 IX86_BUILTIN_LOADUPD
,
16383 IX86_BUILTIN_STOREUPD
,
16384 IX86_BUILTIN_MOVSD
,
16386 IX86_BUILTIN_LOADHPD
,
16387 IX86_BUILTIN_LOADLPD
,
16389 IX86_BUILTIN_CVTDQ2PD
,
16390 IX86_BUILTIN_CVTDQ2PS
,
16392 IX86_BUILTIN_CVTPD2DQ
,
16393 IX86_BUILTIN_CVTPD2PI
,
16394 IX86_BUILTIN_CVTPD2PS
,
16395 IX86_BUILTIN_CVTTPD2DQ
,
16396 IX86_BUILTIN_CVTTPD2PI
,
16398 IX86_BUILTIN_CVTPI2PD
,
16399 IX86_BUILTIN_CVTSI2SD
,
16400 IX86_BUILTIN_CVTSI642SD
,
16402 IX86_BUILTIN_CVTSD2SI
,
16403 IX86_BUILTIN_CVTSD2SI64
,
16404 IX86_BUILTIN_CVTSD2SS
,
16405 IX86_BUILTIN_CVTSS2SD
,
16406 IX86_BUILTIN_CVTTSD2SI
,
16407 IX86_BUILTIN_CVTTSD2SI64
,
16409 IX86_BUILTIN_CVTPS2DQ
,
16410 IX86_BUILTIN_CVTPS2PD
,
16411 IX86_BUILTIN_CVTTPS2DQ
,
16413 IX86_BUILTIN_MOVNTI
,
16414 IX86_BUILTIN_MOVNTPD
,
16415 IX86_BUILTIN_MOVNTDQ
,
16418 IX86_BUILTIN_MASKMOVDQU
,
16419 IX86_BUILTIN_MOVMSKPD
,
16420 IX86_BUILTIN_PMOVMSKB128
,
16422 IX86_BUILTIN_PACKSSWB128
,
16423 IX86_BUILTIN_PACKSSDW128
,
16424 IX86_BUILTIN_PACKUSWB128
,
16426 IX86_BUILTIN_PADDB128
,
16427 IX86_BUILTIN_PADDW128
,
16428 IX86_BUILTIN_PADDD128
,
16429 IX86_BUILTIN_PADDQ128
,
16430 IX86_BUILTIN_PADDSB128
,
16431 IX86_BUILTIN_PADDSW128
,
16432 IX86_BUILTIN_PADDUSB128
,
16433 IX86_BUILTIN_PADDUSW128
,
16434 IX86_BUILTIN_PSUBB128
,
16435 IX86_BUILTIN_PSUBW128
,
16436 IX86_BUILTIN_PSUBD128
,
16437 IX86_BUILTIN_PSUBQ128
,
16438 IX86_BUILTIN_PSUBSB128
,
16439 IX86_BUILTIN_PSUBSW128
,
16440 IX86_BUILTIN_PSUBUSB128
,
16441 IX86_BUILTIN_PSUBUSW128
,
16443 IX86_BUILTIN_PAND128
,
16444 IX86_BUILTIN_PANDN128
,
16445 IX86_BUILTIN_POR128
,
16446 IX86_BUILTIN_PXOR128
,
16448 IX86_BUILTIN_PAVGB128
,
16449 IX86_BUILTIN_PAVGW128
,
16451 IX86_BUILTIN_PCMPEQB128
,
16452 IX86_BUILTIN_PCMPEQW128
,
16453 IX86_BUILTIN_PCMPEQD128
,
16454 IX86_BUILTIN_PCMPGTB128
,
16455 IX86_BUILTIN_PCMPGTW128
,
16456 IX86_BUILTIN_PCMPGTD128
,
16458 IX86_BUILTIN_PMADDWD128
,
16460 IX86_BUILTIN_PMAXSW128
,
16461 IX86_BUILTIN_PMAXUB128
,
16462 IX86_BUILTIN_PMINSW128
,
16463 IX86_BUILTIN_PMINUB128
,
16465 IX86_BUILTIN_PMULUDQ
,
16466 IX86_BUILTIN_PMULUDQ128
,
16467 IX86_BUILTIN_PMULHUW128
,
16468 IX86_BUILTIN_PMULHW128
,
16469 IX86_BUILTIN_PMULLW128
,
16471 IX86_BUILTIN_PSADBW128
,
16472 IX86_BUILTIN_PSHUFHW
,
16473 IX86_BUILTIN_PSHUFLW
,
16474 IX86_BUILTIN_PSHUFD
,
16476 IX86_BUILTIN_PSLLDQI128
,
16477 IX86_BUILTIN_PSLLWI128
,
16478 IX86_BUILTIN_PSLLDI128
,
16479 IX86_BUILTIN_PSLLQI128
,
16480 IX86_BUILTIN_PSRAWI128
,
16481 IX86_BUILTIN_PSRADI128
,
16482 IX86_BUILTIN_PSRLDQI128
,
16483 IX86_BUILTIN_PSRLWI128
,
16484 IX86_BUILTIN_PSRLDI128
,
16485 IX86_BUILTIN_PSRLQI128
,
16487 IX86_BUILTIN_PSLLDQ128
,
16488 IX86_BUILTIN_PSLLW128
,
16489 IX86_BUILTIN_PSLLD128
,
16490 IX86_BUILTIN_PSLLQ128
,
16491 IX86_BUILTIN_PSRAW128
,
16492 IX86_BUILTIN_PSRAD128
,
16493 IX86_BUILTIN_PSRLW128
,
16494 IX86_BUILTIN_PSRLD128
,
16495 IX86_BUILTIN_PSRLQ128
,
16497 IX86_BUILTIN_PUNPCKHBW128
,
16498 IX86_BUILTIN_PUNPCKHWD128
,
16499 IX86_BUILTIN_PUNPCKHDQ128
,
16500 IX86_BUILTIN_PUNPCKHQDQ128
,
16501 IX86_BUILTIN_PUNPCKLBW128
,
16502 IX86_BUILTIN_PUNPCKLWD128
,
16503 IX86_BUILTIN_PUNPCKLDQ128
,
16504 IX86_BUILTIN_PUNPCKLQDQ128
,
16506 IX86_BUILTIN_CLFLUSH
,
16507 IX86_BUILTIN_MFENCE
,
16508 IX86_BUILTIN_LFENCE
,
16510 /* Prescott New Instructions. */
16511 IX86_BUILTIN_ADDSUBPS
,
16512 IX86_BUILTIN_HADDPS
,
16513 IX86_BUILTIN_HSUBPS
,
16514 IX86_BUILTIN_MOVSHDUP
,
16515 IX86_BUILTIN_MOVSLDUP
,
16516 IX86_BUILTIN_ADDSUBPD
,
16517 IX86_BUILTIN_HADDPD
,
16518 IX86_BUILTIN_HSUBPD
,
16519 IX86_BUILTIN_LDDQU
,
16521 IX86_BUILTIN_MONITOR
,
16522 IX86_BUILTIN_MWAIT
,
16525 IX86_BUILTIN_PHADDW
,
16526 IX86_BUILTIN_PHADDD
,
16527 IX86_BUILTIN_PHADDSW
,
16528 IX86_BUILTIN_PHSUBW
,
16529 IX86_BUILTIN_PHSUBD
,
16530 IX86_BUILTIN_PHSUBSW
,
16531 IX86_BUILTIN_PMADDUBSW
,
16532 IX86_BUILTIN_PMULHRSW
,
16533 IX86_BUILTIN_PSHUFB
,
16534 IX86_BUILTIN_PSIGNB
,
16535 IX86_BUILTIN_PSIGNW
,
16536 IX86_BUILTIN_PSIGND
,
16537 IX86_BUILTIN_PALIGNR
,
16538 IX86_BUILTIN_PABSB
,
16539 IX86_BUILTIN_PABSW
,
16540 IX86_BUILTIN_PABSD
,
16542 IX86_BUILTIN_PHADDW128
,
16543 IX86_BUILTIN_PHADDD128
,
16544 IX86_BUILTIN_PHADDSW128
,
16545 IX86_BUILTIN_PHSUBW128
,
16546 IX86_BUILTIN_PHSUBD128
,
16547 IX86_BUILTIN_PHSUBSW128
,
16548 IX86_BUILTIN_PMADDUBSW128
,
16549 IX86_BUILTIN_PMULHRSW128
,
16550 IX86_BUILTIN_PSHUFB128
,
16551 IX86_BUILTIN_PSIGNB128
,
16552 IX86_BUILTIN_PSIGNW128
,
16553 IX86_BUILTIN_PSIGND128
,
16554 IX86_BUILTIN_PALIGNR128
,
16555 IX86_BUILTIN_PABSB128
,
16556 IX86_BUILTIN_PABSW128
,
16557 IX86_BUILTIN_PABSD128
,
16559 /* AMDFAM10 - SSE4A New Instructions. */
16560 IX86_BUILTIN_MOVNTSD
,
16561 IX86_BUILTIN_MOVNTSS
,
16562 IX86_BUILTIN_EXTRQI
,
16563 IX86_BUILTIN_EXTRQ
,
16564 IX86_BUILTIN_INSERTQI
,
16565 IX86_BUILTIN_INSERTQ
,
16568 IX86_BUILTIN_BLENDPD
,
16569 IX86_BUILTIN_BLENDPS
,
16570 IX86_BUILTIN_BLENDVPD
,
16571 IX86_BUILTIN_BLENDVPS
,
16572 IX86_BUILTIN_PBLENDVB128
,
16573 IX86_BUILTIN_PBLENDW128
,
16578 IX86_BUILTIN_INSERTPS128
,
16580 IX86_BUILTIN_MOVNTDQA
,
16581 IX86_BUILTIN_MPSADBW128
,
16582 IX86_BUILTIN_PACKUSDW128
,
16583 IX86_BUILTIN_PCMPEQQ
,
16584 IX86_BUILTIN_PHMINPOSUW128
,
16586 IX86_BUILTIN_PMAXSB128
,
16587 IX86_BUILTIN_PMAXSD128
,
16588 IX86_BUILTIN_PMAXUD128
,
16589 IX86_BUILTIN_PMAXUW128
,
16591 IX86_BUILTIN_PMINSB128
,
16592 IX86_BUILTIN_PMINSD128
,
16593 IX86_BUILTIN_PMINUD128
,
16594 IX86_BUILTIN_PMINUW128
,
16596 IX86_BUILTIN_PMOVSXBW128
,
16597 IX86_BUILTIN_PMOVSXBD128
,
16598 IX86_BUILTIN_PMOVSXBQ128
,
16599 IX86_BUILTIN_PMOVSXWD128
,
16600 IX86_BUILTIN_PMOVSXWQ128
,
16601 IX86_BUILTIN_PMOVSXDQ128
,
16603 IX86_BUILTIN_PMOVZXBW128
,
16604 IX86_BUILTIN_PMOVZXBD128
,
16605 IX86_BUILTIN_PMOVZXBQ128
,
16606 IX86_BUILTIN_PMOVZXWD128
,
16607 IX86_BUILTIN_PMOVZXWQ128
,
16608 IX86_BUILTIN_PMOVZXDQ128
,
16610 IX86_BUILTIN_PMULDQ128
,
16611 IX86_BUILTIN_PMULLD128
,
16613 IX86_BUILTIN_ROUNDPD
,
16614 IX86_BUILTIN_ROUNDPS
,
16615 IX86_BUILTIN_ROUNDSD
,
16616 IX86_BUILTIN_ROUNDSS
,
16618 IX86_BUILTIN_PTESTZ
,
16619 IX86_BUILTIN_PTESTC
,
16620 IX86_BUILTIN_PTESTNZC
,
16622 IX86_BUILTIN_VEC_INIT_V2SI
,
16623 IX86_BUILTIN_VEC_INIT_V4HI
,
16624 IX86_BUILTIN_VEC_INIT_V8QI
,
16625 IX86_BUILTIN_VEC_EXT_V2DF
,
16626 IX86_BUILTIN_VEC_EXT_V2DI
,
16627 IX86_BUILTIN_VEC_EXT_V4SF
,
16628 IX86_BUILTIN_VEC_EXT_V4SI
,
16629 IX86_BUILTIN_VEC_EXT_V8HI
,
16630 IX86_BUILTIN_VEC_EXT_V2SI
,
16631 IX86_BUILTIN_VEC_EXT_V4HI
,
16632 IX86_BUILTIN_VEC_EXT_V16QI
,
16633 IX86_BUILTIN_VEC_SET_V2DI
,
16634 IX86_BUILTIN_VEC_SET_V4SF
,
16635 IX86_BUILTIN_VEC_SET_V4SI
,
16636 IX86_BUILTIN_VEC_SET_V8HI
,
16637 IX86_BUILTIN_VEC_SET_V4HI
,
16638 IX86_BUILTIN_VEC_SET_V16QI
,
16643 /* Table for the ix86 builtin decls. */
16644 static GTY(()) tree ix86_builtins
[(int) IX86_BUILTIN_MAX
];
16646 /* Add an ix86 target builtin function with CODE, NAME and TYPE. Do so,
16647 * if the target_flags include one of MASK. Stores the function decl
16648 * in the ix86_builtins array.
16649 * Returns the function decl or NULL_TREE, if the builtin was not added. */
16652 def_builtin (int mask
, const char *name
, tree type
, enum ix86_builtins code
)
16654 tree decl
= NULL_TREE
;
16656 if (mask
& ix86_isa_flags
16657 && (!(mask
& OPTION_MASK_ISA_64BIT
) || TARGET_64BIT
))
16659 decl
= add_builtin_function (name
, type
, code
, BUILT_IN_MD
,
16661 ix86_builtins
[(int) code
] = decl
;
16667 /* Like def_builtin, but also marks the function decl "const". */
16670 def_builtin_const (int mask
, const char *name
, tree type
,
16671 enum ix86_builtins code
)
16673 tree decl
= def_builtin (mask
, name
, type
, code
);
16675 TREE_READONLY (decl
) = 1;
16679 /* Bits for builtin_description.flag. */
16681 /* Set when we don't support the comparison natively, and should
16682 swap_comparison in order to support it. */
16683 #define BUILTIN_DESC_SWAP_OPERANDS 1
16685 struct builtin_description
16687 const unsigned int mask
;
16688 const enum insn_code icode
;
16689 const char *const name
;
16690 const enum ix86_builtins code
;
16691 const enum rtx_code comparison
;
16692 const unsigned int flag
;
16695 static const struct builtin_description bdesc_comi
[] =
16697 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comieq", IX86_BUILTIN_COMIEQSS
, UNEQ
, 0 },
16698 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comilt", IX86_BUILTIN_COMILTSS
, UNLT
, 0 },
16699 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comile", IX86_BUILTIN_COMILESS
, UNLE
, 0 },
16700 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comigt", IX86_BUILTIN_COMIGTSS
, GT
, 0 },
16701 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comige", IX86_BUILTIN_COMIGESS
, GE
, 0 },
16702 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_comi
, "__builtin_ia32_comineq", IX86_BUILTIN_COMINEQSS
, LTGT
, 0 },
16703 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomieq", IX86_BUILTIN_UCOMIEQSS
, UNEQ
, 0 },
16704 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomilt", IX86_BUILTIN_UCOMILTSS
, UNLT
, 0 },
16705 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomile", IX86_BUILTIN_UCOMILESS
, UNLE
, 0 },
16706 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomigt", IX86_BUILTIN_UCOMIGTSS
, GT
, 0 },
16707 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomige", IX86_BUILTIN_UCOMIGESS
, GE
, 0 },
16708 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_ucomi
, "__builtin_ia32_ucomineq", IX86_BUILTIN_UCOMINEQSS
, LTGT
, 0 },
16709 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdeq", IX86_BUILTIN_COMIEQSD
, UNEQ
, 0 },
16710 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdlt", IX86_BUILTIN_COMILTSD
, UNLT
, 0 },
16711 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdle", IX86_BUILTIN_COMILESD
, UNLE
, 0 },
16712 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdgt", IX86_BUILTIN_COMIGTSD
, GT
, 0 },
16713 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdge", IX86_BUILTIN_COMIGESD
, GE
, 0 },
16714 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_comi
, "__builtin_ia32_comisdneq", IX86_BUILTIN_COMINEQSD
, LTGT
, 0 },
16715 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdeq", IX86_BUILTIN_UCOMIEQSD
, UNEQ
, 0 },
16716 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdlt", IX86_BUILTIN_UCOMILTSD
, UNLT
, 0 },
16717 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdle", IX86_BUILTIN_UCOMILESD
, UNLE
, 0 },
16718 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdgt", IX86_BUILTIN_UCOMIGTSD
, GT
, 0 },
16719 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdge", IX86_BUILTIN_UCOMIGESD
, GE
, 0 },
16720 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_ucomi
, "__builtin_ia32_ucomisdneq", IX86_BUILTIN_UCOMINEQSD
, LTGT
, 0 },
16723 static const struct builtin_description bdesc_ptest
[] =
16726 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_ptest
, "__builtin_ia32_ptestz128", IX86_BUILTIN_PTESTZ
, EQ
, 0 },
16727 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_ptest
, "__builtin_ia32_ptestc128", IX86_BUILTIN_PTESTC
, LTU
, 0 },
16728 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_ptest
, "__builtin_ia32_ptestnzc128", IX86_BUILTIN_PTESTNZC
, GTU
, 0 },
16731 /* SSE builtins with 3 arguments and the last argument must be a 8 bit
16732 constant or xmm0. */
16733 static const struct builtin_description bdesc_sse_3arg
[] =
16736 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_blendpd
, "__builtin_ia32_blendpd", IX86_BUILTIN_BLENDPD
, 0, 0 },
16737 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_blendps
, "__builtin_ia32_blendps", IX86_BUILTIN_BLENDPS
, 0, 0 },
16738 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_blendvpd
, "__builtin_ia32_blendvpd", IX86_BUILTIN_BLENDVPD
, 0, 0 },
16739 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_blendvps
, "__builtin_ia32_blendvps", IX86_BUILTIN_BLENDVPS
, 0, 0 },
16740 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_dppd
, "__builtin_ia32_dppd", IX86_BUILTIN_DPPD
, 0, 0 },
16741 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_dpps
, "__builtin_ia32_dpps", IX86_BUILTIN_DPPS
, 0, 0 },
16742 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_insertps
, "__builtin_ia32_insertps128", IX86_BUILTIN_INSERTPS128
, 0, 0 },
16743 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_mpsadbw
, "__builtin_ia32_mpsadbw128", IX86_BUILTIN_MPSADBW128
, 0, 0 },
16744 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_pblendvb
, "__builtin_ia32_pblendvb128", IX86_BUILTIN_PBLENDVB128
, 0, 0 },
16745 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_pblendw
, "__builtin_ia32_pblendw128", IX86_BUILTIN_PBLENDW128
, 0, 0 },
16746 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_roundsd
, 0, IX86_BUILTIN_ROUNDSD
, 0, 0 },
16747 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_roundss
, 0, IX86_BUILTIN_ROUNDSS
, 0, 0 },
16750 static const struct builtin_description bdesc_2arg
[] =
16753 { OPTION_MASK_ISA_SSE
, CODE_FOR_addv4sf3
, "__builtin_ia32_addps", IX86_BUILTIN_ADDPS
, 0, 0 },
16754 { OPTION_MASK_ISA_SSE
, CODE_FOR_subv4sf3
, "__builtin_ia32_subps", IX86_BUILTIN_SUBPS
, 0, 0 },
16755 { OPTION_MASK_ISA_SSE
, CODE_FOR_mulv4sf3
, "__builtin_ia32_mulps", IX86_BUILTIN_MULPS
, 0, 0 },
16756 { OPTION_MASK_ISA_SSE
, CODE_FOR_divv4sf3
, "__builtin_ia32_divps", IX86_BUILTIN_DIVPS
, 0, 0 },
16757 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmaddv4sf3
, "__builtin_ia32_addss", IX86_BUILTIN_ADDSS
, 0, 0 },
16758 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmsubv4sf3
, "__builtin_ia32_subss", IX86_BUILTIN_SUBSS
, 0, 0 },
16759 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmulv4sf3
, "__builtin_ia32_mulss", IX86_BUILTIN_MULSS
, 0, 0 },
16760 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmdivv4sf3
, "__builtin_ia32_divss", IX86_BUILTIN_DIVSS
, 0, 0 },
16762 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpeqps", IX86_BUILTIN_CMPEQPS
, EQ
, 0 },
16763 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpltps", IX86_BUILTIN_CMPLTPS
, LT
, 0 },
16764 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpleps", IX86_BUILTIN_CMPLEPS
, LE
, 0 },
16765 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpgtps", IX86_BUILTIN_CMPGTPS
, LT
, BUILTIN_DESC_SWAP_OPERANDS
},
16766 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpgeps", IX86_BUILTIN_CMPGEPS
, LE
, BUILTIN_DESC_SWAP_OPERANDS
},
16767 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpunordps", IX86_BUILTIN_CMPUNORDPS
, UNORDERED
, 0 },
16768 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpneqps", IX86_BUILTIN_CMPNEQPS
, NE
, 0 },
16769 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpnltps", IX86_BUILTIN_CMPNLTPS
, UNGE
, 0 },
16770 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpnleps", IX86_BUILTIN_CMPNLEPS
, UNGT
, 0 },
16771 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpngtps", IX86_BUILTIN_CMPNGTPS
, UNGE
, BUILTIN_DESC_SWAP_OPERANDS
},
16772 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpngeps", IX86_BUILTIN_CMPNGEPS
, UNGT
, BUILTIN_DESC_SWAP_OPERANDS
},
16773 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_maskcmpv4sf3
, "__builtin_ia32_cmpordps", IX86_BUILTIN_CMPORDPS
, ORDERED
, 0 },
16774 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpeqss", IX86_BUILTIN_CMPEQSS
, EQ
, 0 },
16775 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpltss", IX86_BUILTIN_CMPLTSS
, LT
, 0 },
16776 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpless", IX86_BUILTIN_CMPLESS
, LE
, 0 },
16777 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpunordss", IX86_BUILTIN_CMPUNORDSS
, UNORDERED
, 0 },
16778 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpneqss", IX86_BUILTIN_CMPNEQSS
, NE
, 0 },
16779 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpnltss", IX86_BUILTIN_CMPNLTSS
, UNGE
, 0 },
16780 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpnless", IX86_BUILTIN_CMPNLESS
, UNGT
, 0 },
16781 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpngtss", IX86_BUILTIN_CMPNGTSS
, UNGE
, BUILTIN_DESC_SWAP_OPERANDS
},
16782 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpngess", IX86_BUILTIN_CMPNGESS
, UNGT
, BUILTIN_DESC_SWAP_OPERANDS
},
16783 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmmaskcmpv4sf3
, "__builtin_ia32_cmpordss", IX86_BUILTIN_CMPORDSS
, ORDERED
, 0 },
16785 { OPTION_MASK_ISA_SSE
, CODE_FOR_sminv4sf3
, "__builtin_ia32_minps", IX86_BUILTIN_MINPS
, 0, 0 },
16786 { OPTION_MASK_ISA_SSE
, CODE_FOR_smaxv4sf3
, "__builtin_ia32_maxps", IX86_BUILTIN_MAXPS
, 0, 0 },
16787 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmsminv4sf3
, "__builtin_ia32_minss", IX86_BUILTIN_MINSS
, 0, 0 },
16788 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_vmsmaxv4sf3
, "__builtin_ia32_maxss", IX86_BUILTIN_MAXSS
, 0, 0 },
16790 { OPTION_MASK_ISA_SSE
, CODE_FOR_andv4sf3
, "__builtin_ia32_andps", IX86_BUILTIN_ANDPS
, 0, 0 },
16791 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_nandv4sf3
, "__builtin_ia32_andnps", IX86_BUILTIN_ANDNPS
, 0, 0 },
16792 { OPTION_MASK_ISA_SSE
, CODE_FOR_iorv4sf3
, "__builtin_ia32_orps", IX86_BUILTIN_ORPS
, 0, 0 },
16793 { OPTION_MASK_ISA_SSE
, CODE_FOR_xorv4sf3
, "__builtin_ia32_xorps", IX86_BUILTIN_XORPS
, 0, 0 },
16795 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_movss
, "__builtin_ia32_movss", IX86_BUILTIN_MOVSS
, 0, 0 },
16796 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_movhlps
, "__builtin_ia32_movhlps", IX86_BUILTIN_MOVHLPS
, 0, 0 },
16797 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_movlhps
, "__builtin_ia32_movlhps", IX86_BUILTIN_MOVLHPS
, 0, 0 },
16798 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_unpckhps
, "__builtin_ia32_unpckhps", IX86_BUILTIN_UNPCKHPS
, 0, 0 },
16799 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_unpcklps
, "__builtin_ia32_unpcklps", IX86_BUILTIN_UNPCKLPS
, 0, 0 },
16802 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_addv8qi3
, "__builtin_ia32_paddb", IX86_BUILTIN_PADDB
, 0, 0 },
16803 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_addv4hi3
, "__builtin_ia32_paddw", IX86_BUILTIN_PADDW
, 0, 0 },
16804 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_addv2si3
, "__builtin_ia32_paddd", IX86_BUILTIN_PADDD
, 0, 0 },
16805 { OPTION_MASK_ISA_SSE2
, CODE_FOR_mmx_adddi3
, "__builtin_ia32_paddq", IX86_BUILTIN_PADDQ
, 0, 0 },
16806 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_subv8qi3
, "__builtin_ia32_psubb", IX86_BUILTIN_PSUBB
, 0, 0 },
16807 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_subv4hi3
, "__builtin_ia32_psubw", IX86_BUILTIN_PSUBW
, 0, 0 },
16808 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_subv2si3
, "__builtin_ia32_psubd", IX86_BUILTIN_PSUBD
, 0, 0 },
16809 { OPTION_MASK_ISA_SSE2
, CODE_FOR_mmx_subdi3
, "__builtin_ia32_psubq", IX86_BUILTIN_PSUBQ
, 0, 0 },
16811 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ssaddv8qi3
, "__builtin_ia32_paddsb", IX86_BUILTIN_PADDSB
, 0, 0 },
16812 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ssaddv4hi3
, "__builtin_ia32_paddsw", IX86_BUILTIN_PADDSW
, 0, 0 },
16813 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_sssubv8qi3
, "__builtin_ia32_psubsb", IX86_BUILTIN_PSUBSB
, 0, 0 },
16814 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_sssubv4hi3
, "__builtin_ia32_psubsw", IX86_BUILTIN_PSUBSW
, 0, 0 },
16815 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_usaddv8qi3
, "__builtin_ia32_paddusb", IX86_BUILTIN_PADDUSB
, 0, 0 },
16816 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_usaddv4hi3
, "__builtin_ia32_paddusw", IX86_BUILTIN_PADDUSW
, 0, 0 },
16817 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ussubv8qi3
, "__builtin_ia32_psubusb", IX86_BUILTIN_PSUBUSB
, 0, 0 },
16818 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ussubv4hi3
, "__builtin_ia32_psubusw", IX86_BUILTIN_PSUBUSW
, 0, 0 },
16820 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_mulv4hi3
, "__builtin_ia32_pmullw", IX86_BUILTIN_PMULLW
, 0, 0 },
16821 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_smulv4hi3_highpart
, "__builtin_ia32_pmulhw", IX86_BUILTIN_PMULHW
, 0, 0 },
16822 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_umulv4hi3_highpart
, "__builtin_ia32_pmulhuw", IX86_BUILTIN_PMULHUW
, 0, 0 },
16824 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_andv2si3
, "__builtin_ia32_pand", IX86_BUILTIN_PAND
, 0, 0 },
16825 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_nandv2si3
, "__builtin_ia32_pandn", IX86_BUILTIN_PANDN
, 0, 0 },
16826 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_iorv2si3
, "__builtin_ia32_por", IX86_BUILTIN_POR
, 0, 0 },
16827 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_xorv2si3
, "__builtin_ia32_pxor", IX86_BUILTIN_PXOR
, 0, 0 },
16829 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_uavgv8qi3
, "__builtin_ia32_pavgb", IX86_BUILTIN_PAVGB
, 0, 0 },
16830 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_uavgv4hi3
, "__builtin_ia32_pavgw", IX86_BUILTIN_PAVGW
, 0, 0 },
16832 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_eqv8qi3
, "__builtin_ia32_pcmpeqb", IX86_BUILTIN_PCMPEQB
, 0, 0 },
16833 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_eqv4hi3
, "__builtin_ia32_pcmpeqw", IX86_BUILTIN_PCMPEQW
, 0, 0 },
16834 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_eqv2si3
, "__builtin_ia32_pcmpeqd", IX86_BUILTIN_PCMPEQD
, 0, 0 },
16835 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_gtv8qi3
, "__builtin_ia32_pcmpgtb", IX86_BUILTIN_PCMPGTB
, 0, 0 },
16836 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_gtv4hi3
, "__builtin_ia32_pcmpgtw", IX86_BUILTIN_PCMPGTW
, 0, 0 },
16837 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_gtv2si3
, "__builtin_ia32_pcmpgtd", IX86_BUILTIN_PCMPGTD
, 0, 0 },
16839 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_umaxv8qi3
, "__builtin_ia32_pmaxub", IX86_BUILTIN_PMAXUB
, 0, 0 },
16840 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_smaxv4hi3
, "__builtin_ia32_pmaxsw", IX86_BUILTIN_PMAXSW
, 0, 0 },
16841 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_uminv8qi3
, "__builtin_ia32_pminub", IX86_BUILTIN_PMINUB
, 0, 0 },
16842 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_sminv4hi3
, "__builtin_ia32_pminsw", IX86_BUILTIN_PMINSW
, 0, 0 },
16844 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpckhbw
, "__builtin_ia32_punpckhbw", IX86_BUILTIN_PUNPCKHBW
, 0, 0 },
16845 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpckhwd
, "__builtin_ia32_punpckhwd", IX86_BUILTIN_PUNPCKHWD
, 0, 0 },
16846 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpckhdq
, "__builtin_ia32_punpckhdq", IX86_BUILTIN_PUNPCKHDQ
, 0, 0 },
16847 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpcklbw
, "__builtin_ia32_punpcklbw", IX86_BUILTIN_PUNPCKLBW
, 0, 0 },
16848 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpcklwd
, "__builtin_ia32_punpcklwd", IX86_BUILTIN_PUNPCKLWD
, 0, 0 },
16849 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_punpckldq
, "__builtin_ia32_punpckldq", IX86_BUILTIN_PUNPCKLDQ
, 0, 0 },
16852 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_packsswb
, 0, IX86_BUILTIN_PACKSSWB
, 0, 0 },
16853 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_packssdw
, 0, IX86_BUILTIN_PACKSSDW
, 0, 0 },
16854 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_packuswb
, 0, IX86_BUILTIN_PACKUSWB
, 0, 0 },
16856 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvtpi2ps
, 0, IX86_BUILTIN_CVTPI2PS
, 0, 0 },
16857 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvtsi2ss
, 0, IX86_BUILTIN_CVTSI2SS
, 0, 0 },
16858 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse_cvtsi2ssq
, 0, IX86_BUILTIN_CVTSI642SS
, 0, 0 },
16860 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashlv4hi3
, 0, IX86_BUILTIN_PSLLW
, 0, 0 },
16861 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashlv4hi3
, 0, IX86_BUILTIN_PSLLWI
, 0, 0 },
16862 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashlv2si3
, 0, IX86_BUILTIN_PSLLD
, 0, 0 },
16863 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashlv2si3
, 0, IX86_BUILTIN_PSLLDI
, 0, 0 },
16864 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashldi3
, 0, IX86_BUILTIN_PSLLQ
, 0, 0 },
16865 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashldi3
, 0, IX86_BUILTIN_PSLLQI
, 0, 0 },
16867 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrv4hi3
, 0, IX86_BUILTIN_PSRLW
, 0, 0 },
16868 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrv4hi3
, 0, IX86_BUILTIN_PSRLWI
, 0, 0 },
16869 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrv2si3
, 0, IX86_BUILTIN_PSRLD
, 0, 0 },
16870 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrv2si3
, 0, IX86_BUILTIN_PSRLDI
, 0, 0 },
16871 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrdi3
, 0, IX86_BUILTIN_PSRLQ
, 0, 0 },
16872 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_lshrdi3
, 0, IX86_BUILTIN_PSRLQI
, 0, 0 },
16874 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashrv4hi3
, 0, IX86_BUILTIN_PSRAW
, 0, 0 },
16875 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashrv4hi3
, 0, IX86_BUILTIN_PSRAWI
, 0, 0 },
16876 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashrv2si3
, 0, IX86_BUILTIN_PSRAD
, 0, 0 },
16877 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_ashrv2si3
, 0, IX86_BUILTIN_PSRADI
, 0, 0 },
16879 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_psadbw
, 0, IX86_BUILTIN_PSADBW
, 0, 0 },
16880 { OPTION_MASK_ISA_MMX
, CODE_FOR_mmx_pmaddwd
, 0, IX86_BUILTIN_PMADDWD
, 0, 0 },
16883 { OPTION_MASK_ISA_SSE2
, CODE_FOR_addv2df3
, "__builtin_ia32_addpd", IX86_BUILTIN_ADDPD
, 0, 0 },
16884 { OPTION_MASK_ISA_SSE2
, CODE_FOR_subv2df3
, "__builtin_ia32_subpd", IX86_BUILTIN_SUBPD
, 0, 0 },
16885 { OPTION_MASK_ISA_SSE2
, CODE_FOR_mulv2df3
, "__builtin_ia32_mulpd", IX86_BUILTIN_MULPD
, 0, 0 },
16886 { OPTION_MASK_ISA_SSE2
, CODE_FOR_divv2df3
, "__builtin_ia32_divpd", IX86_BUILTIN_DIVPD
, 0, 0 },
16887 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmaddv2df3
, "__builtin_ia32_addsd", IX86_BUILTIN_ADDSD
, 0, 0 },
16888 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmsubv2df3
, "__builtin_ia32_subsd", IX86_BUILTIN_SUBSD
, 0, 0 },
16889 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmulv2df3
, "__builtin_ia32_mulsd", IX86_BUILTIN_MULSD
, 0, 0 },
16890 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmdivv2df3
, "__builtin_ia32_divsd", IX86_BUILTIN_DIVSD
, 0, 0 },
16892 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpeqpd", IX86_BUILTIN_CMPEQPD
, EQ
, 0 },
16893 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpltpd", IX86_BUILTIN_CMPLTPD
, LT
, 0 },
16894 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmplepd", IX86_BUILTIN_CMPLEPD
, LE
, 0 },
16895 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpgtpd", IX86_BUILTIN_CMPGTPD
, LT
, BUILTIN_DESC_SWAP_OPERANDS
},
16896 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpgepd", IX86_BUILTIN_CMPGEPD
, LE
, BUILTIN_DESC_SWAP_OPERANDS
},
16897 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpunordpd", IX86_BUILTIN_CMPUNORDPD
, UNORDERED
, 0 },
16898 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpneqpd", IX86_BUILTIN_CMPNEQPD
, NE
, 0 },
16899 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpnltpd", IX86_BUILTIN_CMPNLTPD
, UNGE
, 0 },
16900 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpnlepd", IX86_BUILTIN_CMPNLEPD
, UNGT
, 0 },
16901 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpngtpd", IX86_BUILTIN_CMPNGTPD
, UNGE
, BUILTIN_DESC_SWAP_OPERANDS
},
16902 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpngepd", IX86_BUILTIN_CMPNGEPD
, UNGT
, BUILTIN_DESC_SWAP_OPERANDS
},
16903 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_maskcmpv2df3
, "__builtin_ia32_cmpordpd", IX86_BUILTIN_CMPORDPD
, ORDERED
, 0 },
16904 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpeqsd", IX86_BUILTIN_CMPEQSD
, EQ
, 0 },
16905 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpltsd", IX86_BUILTIN_CMPLTSD
, LT
, 0 },
16906 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmplesd", IX86_BUILTIN_CMPLESD
, LE
, 0 },
16907 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpunordsd", IX86_BUILTIN_CMPUNORDSD
, UNORDERED
, 0 },
16908 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpneqsd", IX86_BUILTIN_CMPNEQSD
, NE
, 0 },
16909 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpnltsd", IX86_BUILTIN_CMPNLTSD
, UNGE
, 0 },
16910 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpnlesd", IX86_BUILTIN_CMPNLESD
, UNGT
, 0 },
16911 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmmaskcmpv2df3
, "__builtin_ia32_cmpordsd", IX86_BUILTIN_CMPORDSD
, ORDERED
, 0 },
16913 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sminv2df3
, "__builtin_ia32_minpd", IX86_BUILTIN_MINPD
, 0, 0 },
16914 { OPTION_MASK_ISA_SSE2
, CODE_FOR_smaxv2df3
, "__builtin_ia32_maxpd", IX86_BUILTIN_MAXPD
, 0, 0 },
16915 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmsminv2df3
, "__builtin_ia32_minsd", IX86_BUILTIN_MINSD
, 0, 0 },
16916 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_vmsmaxv2df3
, "__builtin_ia32_maxsd", IX86_BUILTIN_MAXSD
, 0, 0 },
16918 { OPTION_MASK_ISA_SSE2
, CODE_FOR_andv2df3
, "__builtin_ia32_andpd", IX86_BUILTIN_ANDPD
, 0, 0 },
16919 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_nandv2df3
, "__builtin_ia32_andnpd", IX86_BUILTIN_ANDNPD
, 0, 0 },
16920 { OPTION_MASK_ISA_SSE2
, CODE_FOR_iorv2df3
, "__builtin_ia32_orpd", IX86_BUILTIN_ORPD
, 0, 0 },
16921 { OPTION_MASK_ISA_SSE2
, CODE_FOR_xorv2df3
, "__builtin_ia32_xorpd", IX86_BUILTIN_XORPD
, 0, 0 },
16923 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_movsd
, "__builtin_ia32_movsd", IX86_BUILTIN_MOVSD
, 0, 0 },
16924 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_unpckhpd
, "__builtin_ia32_unpckhpd", IX86_BUILTIN_UNPCKHPD
, 0, 0 },
16925 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_unpcklpd
, "__builtin_ia32_unpcklpd", IX86_BUILTIN_UNPCKLPD
, 0, 0 },
16928 { OPTION_MASK_ISA_SSE2
, CODE_FOR_addv16qi3
, "__builtin_ia32_paddb128", IX86_BUILTIN_PADDB128
, 0, 0 },
16929 { OPTION_MASK_ISA_SSE2
, CODE_FOR_addv8hi3
, "__builtin_ia32_paddw128", IX86_BUILTIN_PADDW128
, 0, 0 },
16930 { OPTION_MASK_ISA_SSE2
, CODE_FOR_addv4si3
, "__builtin_ia32_paddd128", IX86_BUILTIN_PADDD128
, 0, 0 },
16931 { OPTION_MASK_ISA_SSE2
, CODE_FOR_addv2di3
, "__builtin_ia32_paddq128", IX86_BUILTIN_PADDQ128
, 0, 0 },
16932 { OPTION_MASK_ISA_SSE2
, CODE_FOR_subv16qi3
, "__builtin_ia32_psubb128", IX86_BUILTIN_PSUBB128
, 0, 0 },
16933 { OPTION_MASK_ISA_SSE2
, CODE_FOR_subv8hi3
, "__builtin_ia32_psubw128", IX86_BUILTIN_PSUBW128
, 0, 0 },
16934 { OPTION_MASK_ISA_SSE2
, CODE_FOR_subv4si3
, "__builtin_ia32_psubd128", IX86_BUILTIN_PSUBD128
, 0, 0 },
16935 { OPTION_MASK_ISA_SSE2
, CODE_FOR_subv2di3
, "__builtin_ia32_psubq128", IX86_BUILTIN_PSUBQ128
, 0, 0 },
16937 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_ssaddv16qi3
, "__builtin_ia32_paddsb128", IX86_BUILTIN_PADDSB128
, 0, 0 },
16938 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_ssaddv8hi3
, "__builtin_ia32_paddsw128", IX86_BUILTIN_PADDSW128
, 0, 0 },
16939 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_sssubv16qi3
, "__builtin_ia32_psubsb128", IX86_BUILTIN_PSUBSB128
, 0, 0 },
16940 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_sssubv8hi3
, "__builtin_ia32_psubsw128", IX86_BUILTIN_PSUBSW128
, 0, 0 },
16941 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_usaddv16qi3
, "__builtin_ia32_paddusb128", IX86_BUILTIN_PADDUSB128
, 0, 0 },
16942 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_usaddv8hi3
, "__builtin_ia32_paddusw128", IX86_BUILTIN_PADDUSW128
, 0, 0 },
16943 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_ussubv16qi3
, "__builtin_ia32_psubusb128", IX86_BUILTIN_PSUBUSB128
, 0, 0 },
16944 { OPTION_MASK_ISA_MMX
, CODE_FOR_sse2_ussubv8hi3
, "__builtin_ia32_psubusw128", IX86_BUILTIN_PSUBUSW128
, 0, 0 },
16946 { OPTION_MASK_ISA_SSE2
, CODE_FOR_mulv8hi3
, "__builtin_ia32_pmullw128", IX86_BUILTIN_PMULLW128
, 0, 0 },
16947 { OPTION_MASK_ISA_SSE2
, CODE_FOR_smulv8hi3_highpart
, "__builtin_ia32_pmulhw128", IX86_BUILTIN_PMULHW128
, 0, 0 },
16949 { OPTION_MASK_ISA_SSE2
, CODE_FOR_andv2di3
, "__builtin_ia32_pand128", IX86_BUILTIN_PAND128
, 0, 0 },
16950 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_nandv2di3
, "__builtin_ia32_pandn128", IX86_BUILTIN_PANDN128
, 0, 0 },
16951 { OPTION_MASK_ISA_SSE2
, CODE_FOR_iorv2di3
, "__builtin_ia32_por128", IX86_BUILTIN_POR128
, 0, 0 },
16952 { OPTION_MASK_ISA_SSE2
, CODE_FOR_xorv2di3
, "__builtin_ia32_pxor128", IX86_BUILTIN_PXOR128
, 0, 0 },
16954 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_uavgv16qi3
, "__builtin_ia32_pavgb128", IX86_BUILTIN_PAVGB128
, 0, 0 },
16955 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_uavgv8hi3
, "__builtin_ia32_pavgw128", IX86_BUILTIN_PAVGW128
, 0, 0 },
16957 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_eqv16qi3
, "__builtin_ia32_pcmpeqb128", IX86_BUILTIN_PCMPEQB128
, 0, 0 },
16958 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_eqv8hi3
, "__builtin_ia32_pcmpeqw128", IX86_BUILTIN_PCMPEQW128
, 0, 0 },
16959 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_eqv4si3
, "__builtin_ia32_pcmpeqd128", IX86_BUILTIN_PCMPEQD128
, 0, 0 },
16960 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_gtv16qi3
, "__builtin_ia32_pcmpgtb128", IX86_BUILTIN_PCMPGTB128
, 0, 0 },
16961 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_gtv8hi3
, "__builtin_ia32_pcmpgtw128", IX86_BUILTIN_PCMPGTW128
, 0, 0 },
16962 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_gtv4si3
, "__builtin_ia32_pcmpgtd128", IX86_BUILTIN_PCMPGTD128
, 0, 0 },
16964 { OPTION_MASK_ISA_SSE2
, CODE_FOR_umaxv16qi3
, "__builtin_ia32_pmaxub128", IX86_BUILTIN_PMAXUB128
, 0, 0 },
16965 { OPTION_MASK_ISA_SSE2
, CODE_FOR_smaxv8hi3
, "__builtin_ia32_pmaxsw128", IX86_BUILTIN_PMAXSW128
, 0, 0 },
16966 { OPTION_MASK_ISA_SSE2
, CODE_FOR_uminv16qi3
, "__builtin_ia32_pminub128", IX86_BUILTIN_PMINUB128
, 0, 0 },
16967 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sminv8hi3
, "__builtin_ia32_pminsw128", IX86_BUILTIN_PMINSW128
, 0, 0 },
16969 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpckhbw
, "__builtin_ia32_punpckhbw128", IX86_BUILTIN_PUNPCKHBW128
, 0, 0 },
16970 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpckhwd
, "__builtin_ia32_punpckhwd128", IX86_BUILTIN_PUNPCKHWD128
, 0, 0 },
16971 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpckhdq
, "__builtin_ia32_punpckhdq128", IX86_BUILTIN_PUNPCKHDQ128
, 0, 0 },
16972 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpckhqdq
, "__builtin_ia32_punpckhqdq128", IX86_BUILTIN_PUNPCKHQDQ128
, 0, 0 },
16973 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpcklbw
, "__builtin_ia32_punpcklbw128", IX86_BUILTIN_PUNPCKLBW128
, 0, 0 },
16974 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpcklwd
, "__builtin_ia32_punpcklwd128", IX86_BUILTIN_PUNPCKLWD128
, 0, 0 },
16975 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpckldq
, "__builtin_ia32_punpckldq128", IX86_BUILTIN_PUNPCKLDQ128
, 0, 0 },
16976 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_punpcklqdq
, "__builtin_ia32_punpcklqdq128", IX86_BUILTIN_PUNPCKLQDQ128
, 0, 0 },
16978 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_packsswb
, "__builtin_ia32_packsswb128", IX86_BUILTIN_PACKSSWB128
, 0, 0 },
16979 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_packssdw
, "__builtin_ia32_packssdw128", IX86_BUILTIN_PACKSSDW128
, 0, 0 },
16980 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_packuswb
, "__builtin_ia32_packuswb128", IX86_BUILTIN_PACKUSWB128
, 0, 0 },
16982 { OPTION_MASK_ISA_SSE2
, CODE_FOR_umulv8hi3_highpart
, "__builtin_ia32_pmulhuw128", IX86_BUILTIN_PMULHUW128
, 0, 0 },
16983 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_psadbw
, 0, IX86_BUILTIN_PSADBW128
, 0, 0 },
16985 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_umulsidi3
, 0, IX86_BUILTIN_PMULUDQ
, 0, 0 },
16986 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_umulv2siv2di3
, 0, IX86_BUILTIN_PMULUDQ128
, 0, 0 },
16988 { OPTION_MASK_ISA_SSE2
, CODE_FOR_ashlv8hi3
, 0, IX86_BUILTIN_PSLLWI128
, 0, 0 },
16989 { OPTION_MASK_ISA_SSE2
, CODE_FOR_ashlv4si3
, 0, IX86_BUILTIN_PSLLDI128
, 0, 0 },
16990 { OPTION_MASK_ISA_SSE2
, CODE_FOR_ashlv2di3
, 0, IX86_BUILTIN_PSLLQI128
, 0, 0 },
16992 { OPTION_MASK_ISA_SSE2
, CODE_FOR_lshrv8hi3
, 0, IX86_BUILTIN_PSRLWI128
, 0, 0 },
16993 { OPTION_MASK_ISA_SSE2
, CODE_FOR_lshrv4si3
, 0, IX86_BUILTIN_PSRLDI128
, 0, 0 },
16994 { OPTION_MASK_ISA_SSE2
, CODE_FOR_lshrv2di3
, 0, IX86_BUILTIN_PSRLQI128
, 0, 0 },
16996 { OPTION_MASK_ISA_SSE2
, CODE_FOR_ashrv8hi3
, 0, IX86_BUILTIN_PSRAWI128
, 0, 0 },
16997 { OPTION_MASK_ISA_SSE2
, CODE_FOR_ashrv4si3
, 0, IX86_BUILTIN_PSRADI128
, 0, 0 },
16999 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_pmaddwd
, 0, IX86_BUILTIN_PMADDWD128
, 0, 0 },
17001 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtsi2sd
, 0, IX86_BUILTIN_CVTSI2SD
, 0, 0 },
17002 { OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse2_cvtsi2sdq
, 0, IX86_BUILTIN_CVTSI642SD
, 0, 0 },
17003 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtsd2ss
, 0, IX86_BUILTIN_CVTSD2SS
, 0, 0 },
17004 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtss2sd
, 0, IX86_BUILTIN_CVTSS2SD
, 0, 0 },
17007 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_addsubv4sf3
, "__builtin_ia32_addsubps", IX86_BUILTIN_ADDSUBPS
, 0, 0 },
17008 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_addsubv2df3
, "__builtin_ia32_addsubpd", IX86_BUILTIN_ADDSUBPD
, 0, 0 },
17009 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_haddv4sf3
, "__builtin_ia32_haddps", IX86_BUILTIN_HADDPS
, 0, 0 },
17010 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_haddv2df3
, "__builtin_ia32_haddpd", IX86_BUILTIN_HADDPD
, 0, 0 },
17011 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_hsubv4sf3
, "__builtin_ia32_hsubps", IX86_BUILTIN_HSUBPS
, 0, 0 },
17012 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_hsubv2df3
, "__builtin_ia32_hsubpd", IX86_BUILTIN_HSUBPD
, 0, 0 },
17015 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phaddwv8hi3
, "__builtin_ia32_phaddw128", IX86_BUILTIN_PHADDW128
, 0, 0 },
17016 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phaddwv4hi3
, "__builtin_ia32_phaddw", IX86_BUILTIN_PHADDW
, 0, 0 },
17017 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phadddv4si3
, "__builtin_ia32_phaddd128", IX86_BUILTIN_PHADDD128
, 0, 0 },
17018 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phadddv2si3
, "__builtin_ia32_phaddd", IX86_BUILTIN_PHADDD
, 0, 0 },
17019 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phaddswv8hi3
, "__builtin_ia32_phaddsw128", IX86_BUILTIN_PHADDSW128
, 0, 0 },
17020 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phaddswv4hi3
, "__builtin_ia32_phaddsw", IX86_BUILTIN_PHADDSW
, 0, 0 },
17021 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubwv8hi3
, "__builtin_ia32_phsubw128", IX86_BUILTIN_PHSUBW128
, 0, 0 },
17022 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubwv4hi3
, "__builtin_ia32_phsubw", IX86_BUILTIN_PHSUBW
, 0, 0 },
17023 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubdv4si3
, "__builtin_ia32_phsubd128", IX86_BUILTIN_PHSUBD128
, 0, 0 },
17024 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubdv2si3
, "__builtin_ia32_phsubd", IX86_BUILTIN_PHSUBD
, 0, 0 },
17025 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubswv8hi3
, "__builtin_ia32_phsubsw128", IX86_BUILTIN_PHSUBSW128
, 0, 0 },
17026 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_phsubswv4hi3
, "__builtin_ia32_phsubsw", IX86_BUILTIN_PHSUBSW
, 0, 0 },
17027 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pmaddubswv8hi3
, "__builtin_ia32_pmaddubsw128", IX86_BUILTIN_PMADDUBSW128
, 0, 0 },
17028 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pmaddubswv4hi3
, "__builtin_ia32_pmaddubsw", IX86_BUILTIN_PMADDUBSW
, 0, 0 },
17029 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pmulhrswv8hi3
, "__builtin_ia32_pmulhrsw128", IX86_BUILTIN_PMULHRSW128
, 0, 0 },
17030 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pmulhrswv4hi3
, "__builtin_ia32_pmulhrsw", IX86_BUILTIN_PMULHRSW
, 0, 0 },
17031 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pshufbv16qi3
, "__builtin_ia32_pshufb128", IX86_BUILTIN_PSHUFB128
, 0, 0 },
17032 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_pshufbv8qi3
, "__builtin_ia32_pshufb", IX86_BUILTIN_PSHUFB
, 0, 0 },
17033 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv16qi3
, "__builtin_ia32_psignb128", IX86_BUILTIN_PSIGNB128
, 0, 0 },
17034 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv8qi3
, "__builtin_ia32_psignb", IX86_BUILTIN_PSIGNB
, 0, 0 },
17035 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv8hi3
, "__builtin_ia32_psignw128", IX86_BUILTIN_PSIGNW128
, 0, 0 },
17036 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv4hi3
, "__builtin_ia32_psignw", IX86_BUILTIN_PSIGNW
, 0, 0 },
17037 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv4si3
, "__builtin_ia32_psignd128", IX86_BUILTIN_PSIGND128
, 0, 0 },
17038 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_ssse3_psignv2si3
, "__builtin_ia32_psignd", IX86_BUILTIN_PSIGND
, 0, 0 },
17041 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_packusdw
, "__builtin_ia32_packusdw128", IX86_BUILTIN_PACKUSDW128
, 0, 0 },
17042 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_eqv2di3
, "__builtin_ia32_pcmpeqq", IX86_BUILTIN_PCMPEQQ
, 0, 0 },
17043 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_smaxv16qi3
, "__builtin_ia32_pmaxsb128", IX86_BUILTIN_PMAXSB128
, 0, 0 },
17044 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_smaxv4si3
, "__builtin_ia32_pmaxsd128", IX86_BUILTIN_PMAXSD128
, 0, 0 },
17045 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_umaxv4si3
, "__builtin_ia32_pmaxud128", IX86_BUILTIN_PMAXUD128
, 0, 0 },
17046 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_umaxv8hi3
, "__builtin_ia32_pmaxuw128", IX86_BUILTIN_PMAXUW128
, 0, 0 },
17047 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sminv16qi3
, "__builtin_ia32_pminsb128", IX86_BUILTIN_PMINSB128
, 0, 0 },
17048 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sminv4si3
, "__builtin_ia32_pminsd128", IX86_BUILTIN_PMINSD128
, 0, 0 },
17049 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_uminv4si3
, "__builtin_ia32_pminud128", IX86_BUILTIN_PMINUD128
, 0, 0 },
17050 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_uminv8hi3
, "__builtin_ia32_pminuw128", IX86_BUILTIN_PMINUW128
, 0, 0 },
17051 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_mulv2siv2di3
, 0, IX86_BUILTIN_PMULDQ128
, 0, 0 },
17052 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_mulv4si3
, "__builtin_ia32_pmulld128", IX86_BUILTIN_PMULLD128
, 0, 0 },
17055 static const struct builtin_description bdesc_1arg
[] =
17057 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, CODE_FOR_mmx_pmovmskb
, 0, IX86_BUILTIN_PMOVMSKB
, 0, 0 },
17058 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_movmskps
, 0, IX86_BUILTIN_MOVMSKPS
, 0, 0 },
17060 { OPTION_MASK_ISA_SSE
, CODE_FOR_sqrtv4sf2
, 0, IX86_BUILTIN_SQRTPS
, 0, 0 },
17061 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_rsqrtv4sf2
, 0, IX86_BUILTIN_RSQRTPS
, 0, 0 },
17062 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_rcpv4sf2
, 0, IX86_BUILTIN_RCPPS
, 0, 0 },
17064 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvtps2pi
, 0, IX86_BUILTIN_CVTPS2PI
, 0, 0 },
17065 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvtss2si
, 0, IX86_BUILTIN_CVTSS2SI
, 0, 0 },
17066 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse_cvtss2siq
, 0, IX86_BUILTIN_CVTSS2SI64
, 0, 0 },
17067 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvttps2pi
, 0, IX86_BUILTIN_CVTTPS2PI
, 0, 0 },
17068 { OPTION_MASK_ISA_SSE
, CODE_FOR_sse_cvttss2si
, 0, IX86_BUILTIN_CVTTSS2SI
, 0, 0 },
17069 { OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse_cvttss2siq
, 0, IX86_BUILTIN_CVTTSS2SI64
, 0, 0 },
17071 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_pmovmskb
, 0, IX86_BUILTIN_PMOVMSKB128
, 0, 0 },
17072 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_movmskpd
, 0, IX86_BUILTIN_MOVMSKPD
, 0, 0 },
17074 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sqrtv2df2
, 0, IX86_BUILTIN_SQRTPD
, 0, 0 },
17076 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtdq2pd
, 0, IX86_BUILTIN_CVTDQ2PD
, 0, 0 },
17077 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtdq2ps
, 0, IX86_BUILTIN_CVTDQ2PS
, 0, 0 },
17079 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtpd2dq
, 0, IX86_BUILTIN_CVTPD2DQ
, 0, 0 },
17080 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtpd2pi
, 0, IX86_BUILTIN_CVTPD2PI
, 0, 0 },
17081 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtpd2ps
, 0, IX86_BUILTIN_CVTPD2PS
, 0, 0 },
17082 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvttpd2dq
, 0, IX86_BUILTIN_CVTTPD2DQ
, 0, 0 },
17083 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvttpd2pi
, 0, IX86_BUILTIN_CVTTPD2PI
, 0, 0 },
17085 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtpi2pd
, 0, IX86_BUILTIN_CVTPI2PD
, 0, 0 },
17087 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtsd2si
, 0, IX86_BUILTIN_CVTSD2SI
, 0, 0 },
17088 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvttsd2si
, 0, IX86_BUILTIN_CVTTSD2SI
, 0, 0 },
17089 { OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse2_cvtsd2siq
, 0, IX86_BUILTIN_CVTSD2SI64
, 0, 0 },
17090 { OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, CODE_FOR_sse2_cvttsd2siq
, 0, IX86_BUILTIN_CVTTSD2SI64
, 0, 0 },
17092 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtps2dq
, 0, IX86_BUILTIN_CVTPS2DQ
, 0, 0 },
17093 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvtps2pd
, 0, IX86_BUILTIN_CVTPS2PD
, 0, 0 },
17094 { OPTION_MASK_ISA_SSE2
, CODE_FOR_sse2_cvttps2dq
, 0, IX86_BUILTIN_CVTTPS2DQ
, 0, 0 },
17097 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_movshdup
, "__builtin_ia32_movshdup", IX86_BUILTIN_MOVSHDUP
, 0, 0 },
17098 { OPTION_MASK_ISA_SSE3
, CODE_FOR_sse3_movsldup
, "__builtin_ia32_movsldup", IX86_BUILTIN_MOVSLDUP
, 0, 0 },
17101 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv16qi2
, "__builtin_ia32_pabsb128", IX86_BUILTIN_PABSB128
, 0, 0 },
17102 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv8qi2
, "__builtin_ia32_pabsb", IX86_BUILTIN_PABSB
, 0, 0 },
17103 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv8hi2
, "__builtin_ia32_pabsw128", IX86_BUILTIN_PABSW128
, 0, 0 },
17104 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv4hi2
, "__builtin_ia32_pabsw", IX86_BUILTIN_PABSW
, 0, 0 },
17105 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv4si2
, "__builtin_ia32_pabsd128", IX86_BUILTIN_PABSD128
, 0, 0 },
17106 { OPTION_MASK_ISA_SSSE3
, CODE_FOR_absv2si2
, "__builtin_ia32_pabsd", IX86_BUILTIN_PABSD
, 0, 0 },
17109 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv8qiv8hi2
, 0, IX86_BUILTIN_PMOVSXBW128
, 0, 0 },
17110 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv4qiv4si2
, 0, IX86_BUILTIN_PMOVSXBD128
, 0, 0 },
17111 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv2qiv2di2
, 0, IX86_BUILTIN_PMOVSXBQ128
, 0, 0 },
17112 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv4hiv4si2
, 0, IX86_BUILTIN_PMOVSXWD128
, 0, 0 },
17113 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv2hiv2di2
, 0, IX86_BUILTIN_PMOVSXWQ128
, 0, 0 },
17114 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_extendv2siv2di2
, 0, IX86_BUILTIN_PMOVSXDQ128
, 0, 0 },
17115 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv8qiv8hi2
, 0, IX86_BUILTIN_PMOVZXBW128
, 0, 0 },
17116 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv4qiv4si2
, 0, IX86_BUILTIN_PMOVZXBD128
, 0, 0 },
17117 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv2qiv2di2
, 0, IX86_BUILTIN_PMOVZXBQ128
, 0, 0 },
17118 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv4hiv4si2
, 0, IX86_BUILTIN_PMOVZXWD128
, 0, 0 },
17119 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv2hiv2di2
, 0, IX86_BUILTIN_PMOVZXWQ128
, 0, 0 },
17120 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_zero_extendv2siv2di2
, 0, IX86_BUILTIN_PMOVZXDQ128
, 0, 0 },
17121 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_phminposuw
, "__builtin_ia32_phminposuw128", IX86_BUILTIN_PHMINPOSUW128
, 0, 0 },
17123 /* Fake 1 arg builtins with a constant smaller than 8 bits as the 2nd arg. */
17124 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_roundpd
, 0, IX86_BUILTIN_ROUNDPD
, 0, 0 },
17125 { OPTION_MASK_ISA_SSE4_1
, CODE_FOR_sse4_1_roundps
, 0, IX86_BUILTIN_ROUNDPS
, 0, 0 },
17128 /* Set up all the MMX/SSE builtins. This is not called if TARGET_MMX
17129 is zero. Otherwise, if TARGET_SSE is not set, only expand the MMX
17132 ix86_init_mmx_sse_builtins (void)
17134 const struct builtin_description
* d
;
17137 tree V16QI_type_node
= build_vector_type_for_mode (char_type_node
, V16QImode
);
17138 tree V2SI_type_node
= build_vector_type_for_mode (intSI_type_node
, V2SImode
);
17139 tree V2SF_type_node
= build_vector_type_for_mode (float_type_node
, V2SFmode
);
17140 tree V2DI_type_node
17141 = build_vector_type_for_mode (long_long_integer_type_node
, V2DImode
);
17142 tree V2DF_type_node
= build_vector_type_for_mode (double_type_node
, V2DFmode
);
17143 tree V4SF_type_node
= build_vector_type_for_mode (float_type_node
, V4SFmode
);
17144 tree V4SI_type_node
= build_vector_type_for_mode (intSI_type_node
, V4SImode
);
17145 tree V4HI_type_node
= build_vector_type_for_mode (intHI_type_node
, V4HImode
);
17146 tree V8QI_type_node
= build_vector_type_for_mode (char_type_node
, V8QImode
);
17147 tree V8HI_type_node
= build_vector_type_for_mode (intHI_type_node
, V8HImode
);
17149 tree pchar_type_node
= build_pointer_type (char_type_node
);
17150 tree pcchar_type_node
= build_pointer_type (
17151 build_type_variant (char_type_node
, 1, 0));
17152 tree pfloat_type_node
= build_pointer_type (float_type_node
);
17153 tree pcfloat_type_node
= build_pointer_type (
17154 build_type_variant (float_type_node
, 1, 0));
17155 tree pv2si_type_node
= build_pointer_type (V2SI_type_node
);
17156 tree pv2di_type_node
= build_pointer_type (V2DI_type_node
);
17157 tree pdi_type_node
= build_pointer_type (long_long_unsigned_type_node
);
17160 tree int_ftype_v4sf_v4sf
17161 = build_function_type_list (integer_type_node
,
17162 V4SF_type_node
, V4SF_type_node
, NULL_TREE
);
17163 tree v4si_ftype_v4sf_v4sf
17164 = build_function_type_list (V4SI_type_node
,
17165 V4SF_type_node
, V4SF_type_node
, NULL_TREE
);
17166 /* MMX/SSE/integer conversions. */
17167 tree int_ftype_v4sf
17168 = build_function_type_list (integer_type_node
,
17169 V4SF_type_node
, NULL_TREE
);
17170 tree int64_ftype_v4sf
17171 = build_function_type_list (long_long_integer_type_node
,
17172 V4SF_type_node
, NULL_TREE
);
17173 tree int_ftype_v8qi
17174 = build_function_type_list (integer_type_node
, V8QI_type_node
, NULL_TREE
);
17175 tree v4sf_ftype_v4sf_int
17176 = build_function_type_list (V4SF_type_node
,
17177 V4SF_type_node
, integer_type_node
, NULL_TREE
);
17178 tree v4sf_ftype_v4sf_int64
17179 = build_function_type_list (V4SF_type_node
,
17180 V4SF_type_node
, long_long_integer_type_node
,
17182 tree v4sf_ftype_v4sf_v2si
17183 = build_function_type_list (V4SF_type_node
,
17184 V4SF_type_node
, V2SI_type_node
, NULL_TREE
);
17186 /* Miscellaneous. */
17187 tree v8qi_ftype_v4hi_v4hi
17188 = build_function_type_list (V8QI_type_node
,
17189 V4HI_type_node
, V4HI_type_node
, NULL_TREE
);
17190 tree v4hi_ftype_v2si_v2si
17191 = build_function_type_list (V4HI_type_node
,
17192 V2SI_type_node
, V2SI_type_node
, NULL_TREE
);
17193 tree v4sf_ftype_v4sf_v4sf_int
17194 = build_function_type_list (V4SF_type_node
,
17195 V4SF_type_node
, V4SF_type_node
,
17196 integer_type_node
, NULL_TREE
);
17197 tree v2si_ftype_v4hi_v4hi
17198 = build_function_type_list (V2SI_type_node
,
17199 V4HI_type_node
, V4HI_type_node
, NULL_TREE
);
17200 tree v4hi_ftype_v4hi_int
17201 = build_function_type_list (V4HI_type_node
,
17202 V4HI_type_node
, integer_type_node
, NULL_TREE
);
17203 tree v4hi_ftype_v4hi_di
17204 = build_function_type_list (V4HI_type_node
,
17205 V4HI_type_node
, long_long_unsigned_type_node
,
17207 tree v2si_ftype_v2si_di
17208 = build_function_type_list (V2SI_type_node
,
17209 V2SI_type_node
, long_long_unsigned_type_node
,
17211 tree void_ftype_void
17212 = build_function_type (void_type_node
, void_list_node
);
17213 tree void_ftype_unsigned
17214 = build_function_type_list (void_type_node
, unsigned_type_node
, NULL_TREE
);
17215 tree void_ftype_unsigned_unsigned
17216 = build_function_type_list (void_type_node
, unsigned_type_node
,
17217 unsigned_type_node
, NULL_TREE
);
17218 tree void_ftype_pcvoid_unsigned_unsigned
17219 = build_function_type_list (void_type_node
, const_ptr_type_node
,
17220 unsigned_type_node
, unsigned_type_node
,
17222 tree unsigned_ftype_void
17223 = build_function_type (unsigned_type_node
, void_list_node
);
17224 tree v2si_ftype_v4sf
17225 = build_function_type_list (V2SI_type_node
, V4SF_type_node
, NULL_TREE
);
17226 /* Loads/stores. */
17227 tree void_ftype_v8qi_v8qi_pchar
17228 = build_function_type_list (void_type_node
,
17229 V8QI_type_node
, V8QI_type_node
,
17230 pchar_type_node
, NULL_TREE
);
17231 tree v4sf_ftype_pcfloat
17232 = build_function_type_list (V4SF_type_node
, pcfloat_type_node
, NULL_TREE
);
17233 /* @@@ the type is bogus */
17234 tree v4sf_ftype_v4sf_pv2si
17235 = build_function_type_list (V4SF_type_node
,
17236 V4SF_type_node
, pv2si_type_node
, NULL_TREE
);
17237 tree void_ftype_pv2si_v4sf
17238 = build_function_type_list (void_type_node
,
17239 pv2si_type_node
, V4SF_type_node
, NULL_TREE
);
17240 tree void_ftype_pfloat_v4sf
17241 = build_function_type_list (void_type_node
,
17242 pfloat_type_node
, V4SF_type_node
, NULL_TREE
);
17243 tree void_ftype_pdi_di
17244 = build_function_type_list (void_type_node
,
17245 pdi_type_node
, long_long_unsigned_type_node
,
17247 tree void_ftype_pv2di_v2di
17248 = build_function_type_list (void_type_node
,
17249 pv2di_type_node
, V2DI_type_node
, NULL_TREE
);
17250 /* Normal vector unops. */
17251 tree v4sf_ftype_v4sf
17252 = build_function_type_list (V4SF_type_node
, V4SF_type_node
, NULL_TREE
);
17253 tree v16qi_ftype_v16qi
17254 = build_function_type_list (V16QI_type_node
, V16QI_type_node
, NULL_TREE
);
17255 tree v8hi_ftype_v8hi
17256 = build_function_type_list (V8HI_type_node
, V8HI_type_node
, NULL_TREE
);
17257 tree v4si_ftype_v4si
17258 = build_function_type_list (V4SI_type_node
, V4SI_type_node
, NULL_TREE
);
17259 tree v8qi_ftype_v8qi
17260 = build_function_type_list (V8QI_type_node
, V8QI_type_node
, NULL_TREE
);
17261 tree v4hi_ftype_v4hi
17262 = build_function_type_list (V4HI_type_node
, V4HI_type_node
, NULL_TREE
);
17264 /* Normal vector binops. */
17265 tree v4sf_ftype_v4sf_v4sf
17266 = build_function_type_list (V4SF_type_node
,
17267 V4SF_type_node
, V4SF_type_node
, NULL_TREE
);
17268 tree v8qi_ftype_v8qi_v8qi
17269 = build_function_type_list (V8QI_type_node
,
17270 V8QI_type_node
, V8QI_type_node
, NULL_TREE
);
17271 tree v4hi_ftype_v4hi_v4hi
17272 = build_function_type_list (V4HI_type_node
,
17273 V4HI_type_node
, V4HI_type_node
, NULL_TREE
);
17274 tree v2si_ftype_v2si_v2si
17275 = build_function_type_list (V2SI_type_node
,
17276 V2SI_type_node
, V2SI_type_node
, NULL_TREE
);
17277 tree di_ftype_di_di
17278 = build_function_type_list (long_long_unsigned_type_node
,
17279 long_long_unsigned_type_node
,
17280 long_long_unsigned_type_node
, NULL_TREE
);
17282 tree di_ftype_di_di_int
17283 = build_function_type_list (long_long_unsigned_type_node
,
17284 long_long_unsigned_type_node
,
17285 long_long_unsigned_type_node
,
17286 integer_type_node
, NULL_TREE
);
17288 tree v2si_ftype_v2sf
17289 = build_function_type_list (V2SI_type_node
, V2SF_type_node
, NULL_TREE
);
17290 tree v2sf_ftype_v2si
17291 = build_function_type_list (V2SF_type_node
, V2SI_type_node
, NULL_TREE
);
17292 tree v2si_ftype_v2si
17293 = build_function_type_list (V2SI_type_node
, V2SI_type_node
, NULL_TREE
);
17294 tree v2sf_ftype_v2sf
17295 = build_function_type_list (V2SF_type_node
, V2SF_type_node
, NULL_TREE
);
17296 tree v2sf_ftype_v2sf_v2sf
17297 = build_function_type_list (V2SF_type_node
,
17298 V2SF_type_node
, V2SF_type_node
, NULL_TREE
);
17299 tree v2si_ftype_v2sf_v2sf
17300 = build_function_type_list (V2SI_type_node
,
17301 V2SF_type_node
, V2SF_type_node
, NULL_TREE
);
17302 tree pint_type_node
= build_pointer_type (integer_type_node
);
17303 tree pdouble_type_node
= build_pointer_type (double_type_node
);
17304 tree pcdouble_type_node
= build_pointer_type (
17305 build_type_variant (double_type_node
, 1, 0));
17306 tree int_ftype_v2df_v2df
17307 = build_function_type_list (integer_type_node
,
17308 V2DF_type_node
, V2DF_type_node
, NULL_TREE
);
17310 tree void_ftype_pcvoid
17311 = build_function_type_list (void_type_node
, const_ptr_type_node
, NULL_TREE
);
17312 tree v4sf_ftype_v4si
17313 = build_function_type_list (V4SF_type_node
, V4SI_type_node
, NULL_TREE
);
17314 tree v4si_ftype_v4sf
17315 = build_function_type_list (V4SI_type_node
, V4SF_type_node
, NULL_TREE
);
17316 tree v2df_ftype_v4si
17317 = build_function_type_list (V2DF_type_node
, V4SI_type_node
, NULL_TREE
);
17318 tree v4si_ftype_v2df
17319 = build_function_type_list (V4SI_type_node
, V2DF_type_node
, NULL_TREE
);
17320 tree v2si_ftype_v2df
17321 = build_function_type_list (V2SI_type_node
, V2DF_type_node
, NULL_TREE
);
17322 tree v4sf_ftype_v2df
17323 = build_function_type_list (V4SF_type_node
, V2DF_type_node
, NULL_TREE
);
17324 tree v2df_ftype_v2si
17325 = build_function_type_list (V2DF_type_node
, V2SI_type_node
, NULL_TREE
);
17326 tree v2df_ftype_v4sf
17327 = build_function_type_list (V2DF_type_node
, V4SF_type_node
, NULL_TREE
);
17328 tree int_ftype_v2df
17329 = build_function_type_list (integer_type_node
, V2DF_type_node
, NULL_TREE
);
17330 tree int64_ftype_v2df
17331 = build_function_type_list (long_long_integer_type_node
,
17332 V2DF_type_node
, NULL_TREE
);
17333 tree v2df_ftype_v2df_int
17334 = build_function_type_list (V2DF_type_node
,
17335 V2DF_type_node
, integer_type_node
, NULL_TREE
);
17336 tree v2df_ftype_v2df_int64
17337 = build_function_type_list (V2DF_type_node
,
17338 V2DF_type_node
, long_long_integer_type_node
,
17340 tree v4sf_ftype_v4sf_v2df
17341 = build_function_type_list (V4SF_type_node
,
17342 V4SF_type_node
, V2DF_type_node
, NULL_TREE
);
17343 tree v2df_ftype_v2df_v4sf
17344 = build_function_type_list (V2DF_type_node
,
17345 V2DF_type_node
, V4SF_type_node
, NULL_TREE
);
17346 tree v2df_ftype_v2df_v2df_int
17347 = build_function_type_list (V2DF_type_node
,
17348 V2DF_type_node
, V2DF_type_node
,
17351 tree v2df_ftype_v2df_pcdouble
17352 = build_function_type_list (V2DF_type_node
,
17353 V2DF_type_node
, pcdouble_type_node
, NULL_TREE
);
17354 tree void_ftype_pdouble_v2df
17355 = build_function_type_list (void_type_node
,
17356 pdouble_type_node
, V2DF_type_node
, NULL_TREE
);
17357 tree void_ftype_pint_int
17358 = build_function_type_list (void_type_node
,
17359 pint_type_node
, integer_type_node
, NULL_TREE
);
17360 tree void_ftype_v16qi_v16qi_pchar
17361 = build_function_type_list (void_type_node
,
17362 V16QI_type_node
, V16QI_type_node
,
17363 pchar_type_node
, NULL_TREE
);
17364 tree v2df_ftype_pcdouble
17365 = build_function_type_list (V2DF_type_node
, pcdouble_type_node
, NULL_TREE
);
17366 tree v2df_ftype_v2df_v2df
17367 = build_function_type_list (V2DF_type_node
,
17368 V2DF_type_node
, V2DF_type_node
, NULL_TREE
);
17369 tree v16qi_ftype_v16qi_v16qi
17370 = build_function_type_list (V16QI_type_node
,
17371 V16QI_type_node
, V16QI_type_node
, NULL_TREE
);
17372 tree v8hi_ftype_v8hi_v8hi
17373 = build_function_type_list (V8HI_type_node
,
17374 V8HI_type_node
, V8HI_type_node
, NULL_TREE
);
17375 tree v4si_ftype_v4si_v4si
17376 = build_function_type_list (V4SI_type_node
,
17377 V4SI_type_node
, V4SI_type_node
, NULL_TREE
);
17378 tree v2di_ftype_v2di_v2di
17379 = build_function_type_list (V2DI_type_node
,
17380 V2DI_type_node
, V2DI_type_node
, NULL_TREE
);
17381 tree v2di_ftype_v2df_v2df
17382 = build_function_type_list (V2DI_type_node
,
17383 V2DF_type_node
, V2DF_type_node
, NULL_TREE
);
17384 tree v2df_ftype_v2df
17385 = build_function_type_list (V2DF_type_node
, V2DF_type_node
, NULL_TREE
);
17386 tree v2di_ftype_v2di_int
17387 = build_function_type_list (V2DI_type_node
,
17388 V2DI_type_node
, integer_type_node
, NULL_TREE
);
17389 tree v2di_ftype_v2di_v2di_int
17390 = build_function_type_list (V2DI_type_node
, V2DI_type_node
,
17391 V2DI_type_node
, integer_type_node
, NULL_TREE
);
17392 tree v4si_ftype_v4si_int
17393 = build_function_type_list (V4SI_type_node
,
17394 V4SI_type_node
, integer_type_node
, NULL_TREE
);
17395 tree v8hi_ftype_v8hi_int
17396 = build_function_type_list (V8HI_type_node
,
17397 V8HI_type_node
, integer_type_node
, NULL_TREE
);
17398 tree v4si_ftype_v8hi_v8hi
17399 = build_function_type_list (V4SI_type_node
,
17400 V8HI_type_node
, V8HI_type_node
, NULL_TREE
);
17401 tree di_ftype_v8qi_v8qi
17402 = build_function_type_list (long_long_unsigned_type_node
,
17403 V8QI_type_node
, V8QI_type_node
, NULL_TREE
);
17404 tree di_ftype_v2si_v2si
17405 = build_function_type_list (long_long_unsigned_type_node
,
17406 V2SI_type_node
, V2SI_type_node
, NULL_TREE
);
17407 tree v2di_ftype_v16qi_v16qi
17408 = build_function_type_list (V2DI_type_node
,
17409 V16QI_type_node
, V16QI_type_node
, NULL_TREE
);
17410 tree v2di_ftype_v4si_v4si
17411 = build_function_type_list (V2DI_type_node
,
17412 V4SI_type_node
, V4SI_type_node
, NULL_TREE
);
17413 tree int_ftype_v16qi
17414 = build_function_type_list (integer_type_node
, V16QI_type_node
, NULL_TREE
);
17415 tree v16qi_ftype_pcchar
17416 = build_function_type_list (V16QI_type_node
, pcchar_type_node
, NULL_TREE
);
17417 tree void_ftype_pchar_v16qi
17418 = build_function_type_list (void_type_node
,
17419 pchar_type_node
, V16QI_type_node
, NULL_TREE
);
17421 tree v2di_ftype_v2di_unsigned_unsigned
17422 = build_function_type_list (V2DI_type_node
, V2DI_type_node
,
17423 unsigned_type_node
, unsigned_type_node
,
17425 tree v2di_ftype_v2di_v2di_unsigned_unsigned
17426 = build_function_type_list (V2DI_type_node
, V2DI_type_node
, V2DI_type_node
,
17427 unsigned_type_node
, unsigned_type_node
,
17429 tree v2di_ftype_v2di_v16qi
17430 = build_function_type_list (V2DI_type_node
, V2DI_type_node
, V16QI_type_node
,
17432 tree v2df_ftype_v2df_v2df_v2df
17433 = build_function_type_list (V2DF_type_node
,
17434 V2DF_type_node
, V2DF_type_node
,
17435 V2DF_type_node
, NULL_TREE
);
17436 tree v4sf_ftype_v4sf_v4sf_v4sf
17437 = build_function_type_list (V4SF_type_node
,
17438 V4SF_type_node
, V4SF_type_node
,
17439 V4SF_type_node
, NULL_TREE
);
17440 tree v8hi_ftype_v16qi
17441 = build_function_type_list (V8HI_type_node
, V16QI_type_node
,
17443 tree v4si_ftype_v16qi
17444 = build_function_type_list (V4SI_type_node
, V16QI_type_node
,
17446 tree v2di_ftype_v16qi
17447 = build_function_type_list (V2DI_type_node
, V16QI_type_node
,
17449 tree v4si_ftype_v8hi
17450 = build_function_type_list (V4SI_type_node
, V8HI_type_node
,
17452 tree v2di_ftype_v8hi
17453 = build_function_type_list (V2DI_type_node
, V8HI_type_node
,
17455 tree v2di_ftype_v4si
17456 = build_function_type_list (V2DI_type_node
, V4SI_type_node
,
17458 tree v2di_ftype_pv2di
17459 = build_function_type_list (V2DI_type_node
, pv2di_type_node
,
17461 tree v16qi_ftype_v16qi_v16qi_int
17462 = build_function_type_list (V16QI_type_node
, V16QI_type_node
,
17463 V16QI_type_node
, integer_type_node
,
17465 tree v16qi_ftype_v16qi_v16qi_v16qi
17466 = build_function_type_list (V16QI_type_node
, V16QI_type_node
,
17467 V16QI_type_node
, V16QI_type_node
,
17469 tree v8hi_ftype_v8hi_v8hi_int
17470 = build_function_type_list (V8HI_type_node
, V8HI_type_node
,
17471 V8HI_type_node
, integer_type_node
,
17473 tree v4si_ftype_v4si_v4si_int
17474 = build_function_type_list (V4SI_type_node
, V4SI_type_node
,
17475 V4SI_type_node
, integer_type_node
,
17477 tree int_ftype_v2di_v2di
17478 = build_function_type_list (integer_type_node
,
17479 V2DI_type_node
, V2DI_type_node
,
17483 tree float128_type
;
17486 /* The __float80 type. */
17487 if (TYPE_MODE (long_double_type_node
) == XFmode
)
17488 (*lang_hooks
.types
.register_builtin_type
) (long_double_type_node
,
17492 /* The __float80 type. */
17493 float80_type
= make_node (REAL_TYPE
);
17494 TYPE_PRECISION (float80_type
) = 80;
17495 layout_type (float80_type
);
17496 (*lang_hooks
.types
.register_builtin_type
) (float80_type
, "__float80");
17501 float128_type
= make_node (REAL_TYPE
);
17502 TYPE_PRECISION (float128_type
) = 128;
17503 layout_type (float128_type
);
17504 (*lang_hooks
.types
.register_builtin_type
) (float128_type
, "__float128");
17507 /* Add all SSE builtins that are more or less simple operations on
17509 for (i
= 0, d
= bdesc_sse_3arg
;
17510 i
< ARRAY_SIZE (bdesc_sse_3arg
);
17513 /* Use one of the operands; the target can have a different mode for
17514 mask-generating compares. */
17515 enum machine_mode mode
;
17520 mode
= insn_data
[d
->icode
].operand
[1].mode
;
17525 type
= v16qi_ftype_v16qi_v16qi_int
;
17528 type
= v8hi_ftype_v8hi_v8hi_int
;
17531 type
= v4si_ftype_v4si_v4si_int
;
17534 type
= v2di_ftype_v2di_v2di_int
;
17537 type
= v2df_ftype_v2df_v2df_int
;
17540 type
= v4sf_ftype_v4sf_v4sf_int
;
17543 gcc_unreachable ();
17546 /* Override for variable blends. */
17549 case CODE_FOR_sse4_1_blendvpd
:
17550 type
= v2df_ftype_v2df_v2df_v2df
;
17552 case CODE_FOR_sse4_1_blendvps
:
17553 type
= v4sf_ftype_v4sf_v4sf_v4sf
;
17555 case CODE_FOR_sse4_1_pblendvb
:
17556 type
= v16qi_ftype_v16qi_v16qi_v16qi
;
17562 def_builtin (d
->mask
, d
->name
, type
, d
->code
);
17565 /* Add all builtins that are more or less simple operations on two
17567 for (i
= 0, d
= bdesc_2arg
; i
< ARRAY_SIZE (bdesc_2arg
); i
++, d
++)
17569 /* Use one of the operands; the target can have a different mode for
17570 mask-generating compares. */
17571 enum machine_mode mode
;
17576 mode
= insn_data
[d
->icode
].operand
[1].mode
;
17581 type
= v16qi_ftype_v16qi_v16qi
;
17584 type
= v8hi_ftype_v8hi_v8hi
;
17587 type
= v4si_ftype_v4si_v4si
;
17590 type
= v2di_ftype_v2di_v2di
;
17593 type
= v2df_ftype_v2df_v2df
;
17596 type
= v4sf_ftype_v4sf_v4sf
;
17599 type
= v8qi_ftype_v8qi_v8qi
;
17602 type
= v4hi_ftype_v4hi_v4hi
;
17605 type
= v2si_ftype_v2si_v2si
;
17608 type
= di_ftype_di_di
;
17612 gcc_unreachable ();
17615 /* Override for comparisons. */
17616 if (d
->icode
== CODE_FOR_sse_maskcmpv4sf3
17617 || d
->icode
== CODE_FOR_sse_vmmaskcmpv4sf3
)
17618 type
= v4si_ftype_v4sf_v4sf
;
17620 if (d
->icode
== CODE_FOR_sse2_maskcmpv2df3
17621 || d
->icode
== CODE_FOR_sse2_vmmaskcmpv2df3
)
17622 type
= v2di_ftype_v2df_v2df
;
17624 def_builtin (d
->mask
, d
->name
, type
, d
->code
);
17627 /* Add all builtins that are more or less simple operations on 1 operand. */
17628 for (i
= 0, d
= bdesc_1arg
; i
< ARRAY_SIZE (bdesc_1arg
); i
++, d
++)
17630 enum machine_mode mode
;
17635 mode
= insn_data
[d
->icode
].operand
[1].mode
;
17640 type
= v16qi_ftype_v16qi
;
17643 type
= v8hi_ftype_v8hi
;
17646 type
= v4si_ftype_v4si
;
17649 type
= v2df_ftype_v2df
;
17652 type
= v4sf_ftype_v4sf
;
17655 type
= v8qi_ftype_v8qi
;
17658 type
= v4hi_ftype_v4hi
;
17661 type
= v2si_ftype_v2si
;
17668 def_builtin (d
->mask
, d
->name
, type
, d
->code
);
17671 /* Add the remaining MMX insns with somewhat more complicated types. */
17672 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_emms", void_ftype_void
, IX86_BUILTIN_EMMS
);
17673 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psllw", v4hi_ftype_v4hi_di
, IX86_BUILTIN_PSLLW
);
17674 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_pslld", v2si_ftype_v2si_di
, IX86_BUILTIN_PSLLD
);
17675 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psllq", di_ftype_di_di
, IX86_BUILTIN_PSLLQ
);
17677 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psrlw", v4hi_ftype_v4hi_di
, IX86_BUILTIN_PSRLW
);
17678 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psrld", v2si_ftype_v2si_di
, IX86_BUILTIN_PSRLD
);
17679 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psrlq", di_ftype_di_di
, IX86_BUILTIN_PSRLQ
);
17681 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psraw", v4hi_ftype_v4hi_di
, IX86_BUILTIN_PSRAW
);
17682 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_psrad", v2si_ftype_v2si_di
, IX86_BUILTIN_PSRAD
);
17684 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pshufw", v4hi_ftype_v4hi_int
, IX86_BUILTIN_PSHUFW
);
17685 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_pmaddwd", v2si_ftype_v4hi_v4hi
, IX86_BUILTIN_PMADDWD
);
17687 /* comi/ucomi insns. */
17688 for (i
= 0, d
= bdesc_comi
; i
< ARRAY_SIZE (bdesc_comi
); i
++, d
++)
17689 if (d
->mask
== OPTION_MASK_ISA_SSE2
)
17690 def_builtin (d
->mask
, d
->name
, int_ftype_v2df_v2df
, d
->code
);
17692 def_builtin (d
->mask
, d
->name
, int_ftype_v4sf_v4sf
, d
->code
);
17695 for (i
= 0, d
= bdesc_ptest
; i
< ARRAY_SIZE (bdesc_ptest
); i
++, d
++)
17696 def_builtin (d
->mask
, d
->name
, int_ftype_v2di_v2di
, d
->code
);
17698 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_packsswb", v8qi_ftype_v4hi_v4hi
, IX86_BUILTIN_PACKSSWB
);
17699 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_packssdw", v4hi_ftype_v2si_v2si
, IX86_BUILTIN_PACKSSDW
);
17700 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_packuswb", v8qi_ftype_v4hi_v4hi
, IX86_BUILTIN_PACKUSWB
);
17702 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_ldmxcsr", void_ftype_unsigned
, IX86_BUILTIN_LDMXCSR
);
17703 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_stmxcsr", unsigned_ftype_void
, IX86_BUILTIN_STMXCSR
);
17704 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvtpi2ps", v4sf_ftype_v4sf_v2si
, IX86_BUILTIN_CVTPI2PS
);
17705 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvtps2pi", v2si_ftype_v4sf
, IX86_BUILTIN_CVTPS2PI
);
17706 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvtsi2ss", v4sf_ftype_v4sf_int
, IX86_BUILTIN_CVTSI2SS
);
17707 def_builtin_const (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvtsi642ss", v4sf_ftype_v4sf_int64
, IX86_BUILTIN_CVTSI642SS
);
17708 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvtss2si", int_ftype_v4sf
, IX86_BUILTIN_CVTSS2SI
);
17709 def_builtin_const (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvtss2si64", int64_ftype_v4sf
, IX86_BUILTIN_CVTSS2SI64
);
17710 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvttps2pi", v2si_ftype_v4sf
, IX86_BUILTIN_CVTTPS2PI
);
17711 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_cvttss2si", int_ftype_v4sf
, IX86_BUILTIN_CVTTSS2SI
);
17712 def_builtin_const (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvttss2si64", int64_ftype_v4sf
, IX86_BUILTIN_CVTTSS2SI64
);
17714 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_maskmovq", void_ftype_v8qi_v8qi_pchar
, IX86_BUILTIN_MASKMOVQ
);
17716 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_loadups", v4sf_ftype_pcfloat
, IX86_BUILTIN_LOADUPS
);
17717 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_storeups", void_ftype_pfloat_v4sf
, IX86_BUILTIN_STOREUPS
);
17719 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_loadhps", v4sf_ftype_v4sf_pv2si
, IX86_BUILTIN_LOADHPS
);
17720 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_loadlps", v4sf_ftype_v4sf_pv2si
, IX86_BUILTIN_LOADLPS
);
17721 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_storehps", void_ftype_pv2si_v4sf
, IX86_BUILTIN_STOREHPS
);
17722 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_storelps", void_ftype_pv2si_v4sf
, IX86_BUILTIN_STORELPS
);
17724 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_movmskps", int_ftype_v4sf
, IX86_BUILTIN_MOVMSKPS
);
17725 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pmovmskb", int_ftype_v8qi
, IX86_BUILTIN_PMOVMSKB
);
17726 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_movntps", void_ftype_pfloat_v4sf
, IX86_BUILTIN_MOVNTPS
);
17727 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_movntq", void_ftype_pdi_di
, IX86_BUILTIN_MOVNTQ
);
17729 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_sfence", void_ftype_void
, IX86_BUILTIN_SFENCE
);
17731 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_psadbw", di_ftype_v8qi_v8qi
, IX86_BUILTIN_PSADBW
);
17733 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_rcpps", v4sf_ftype_v4sf
, IX86_BUILTIN_RCPPS
);
17734 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_rcpss", v4sf_ftype_v4sf
, IX86_BUILTIN_RCPSS
);
17735 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_rsqrtps", v4sf_ftype_v4sf
, IX86_BUILTIN_RSQRTPS
);
17736 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_rsqrtss", v4sf_ftype_v4sf
, IX86_BUILTIN_RSQRTSS
);
17737 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_sqrtps", v4sf_ftype_v4sf
, IX86_BUILTIN_SQRTPS
);
17738 def_builtin_const (OPTION_MASK_ISA_SSE
, "__builtin_ia32_sqrtss", v4sf_ftype_v4sf
, IX86_BUILTIN_SQRTSS
);
17740 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_shufps", v4sf_ftype_v4sf_v4sf_int
, IX86_BUILTIN_SHUFPS
);
17742 /* Original 3DNow! */
17743 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_femms", void_ftype_void
, IX86_BUILTIN_FEMMS
);
17744 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pavgusb", v8qi_ftype_v8qi_v8qi
, IX86_BUILTIN_PAVGUSB
);
17745 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pf2id", v2si_ftype_v2sf
, IX86_BUILTIN_PF2ID
);
17746 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfacc", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFACC
);
17747 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfadd", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFADD
);
17748 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfcmpeq", v2si_ftype_v2sf_v2sf
, IX86_BUILTIN_PFCMPEQ
);
17749 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfcmpge", v2si_ftype_v2sf_v2sf
, IX86_BUILTIN_PFCMPGE
);
17750 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfcmpgt", v2si_ftype_v2sf_v2sf
, IX86_BUILTIN_PFCMPGT
);
17751 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfmax", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFMAX
);
17752 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfmin", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFMIN
);
17753 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfmul", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFMUL
);
17754 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfrcp", v2sf_ftype_v2sf
, IX86_BUILTIN_PFRCP
);
17755 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfrcpit1", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFRCPIT1
);
17756 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfrcpit2", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFRCPIT2
);
17757 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfrsqrt", v2sf_ftype_v2sf
, IX86_BUILTIN_PFRSQRT
);
17758 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfrsqit1", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFRSQIT1
);
17759 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfsub", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFSUB
);
17760 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pfsubr", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFSUBR
);
17761 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pi2fd", v2sf_ftype_v2si
, IX86_BUILTIN_PI2FD
);
17762 def_builtin (OPTION_MASK_ISA_3DNOW
, "__builtin_ia32_pmulhrw", v4hi_ftype_v4hi_v4hi
, IX86_BUILTIN_PMULHRW
);
17764 /* 3DNow! extension as used in the Athlon CPU. */
17765 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pf2iw", v2si_ftype_v2sf
, IX86_BUILTIN_PF2IW
);
17766 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pfnacc", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFNACC
);
17767 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pfpnacc", v2sf_ftype_v2sf_v2sf
, IX86_BUILTIN_PFPNACC
);
17768 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pi2fw", v2sf_ftype_v2si
, IX86_BUILTIN_PI2FW
);
17769 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pswapdsf", v2sf_ftype_v2sf
, IX86_BUILTIN_PSWAPDSF
);
17770 def_builtin (OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_pswapdsi", v2si_ftype_v2si
, IX86_BUILTIN_PSWAPDSI
);
17773 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_maskmovdqu", void_ftype_v16qi_v16qi_pchar
, IX86_BUILTIN_MASKMOVDQU
);
17775 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_loadupd", v2df_ftype_pcdouble
, IX86_BUILTIN_LOADUPD
);
17776 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_storeupd", void_ftype_pdouble_v2df
, IX86_BUILTIN_STOREUPD
);
17778 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_loadhpd", v2df_ftype_v2df_pcdouble
, IX86_BUILTIN_LOADHPD
);
17779 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_loadlpd", v2df_ftype_v2df_pcdouble
, IX86_BUILTIN_LOADLPD
);
17781 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_movmskpd", int_ftype_v2df
, IX86_BUILTIN_MOVMSKPD
);
17782 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pmovmskb128", int_ftype_v16qi
, IX86_BUILTIN_PMOVMSKB128
);
17783 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_movnti", void_ftype_pint_int
, IX86_BUILTIN_MOVNTI
);
17784 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_movntpd", void_ftype_pdouble_v2df
, IX86_BUILTIN_MOVNTPD
);
17785 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_movntdq", void_ftype_pv2di_v2di
, IX86_BUILTIN_MOVNTDQ
);
17787 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pshufd", v4si_ftype_v4si_int
, IX86_BUILTIN_PSHUFD
);
17788 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pshuflw", v8hi_ftype_v8hi_int
, IX86_BUILTIN_PSHUFLW
);
17789 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pshufhw", v8hi_ftype_v8hi_int
, IX86_BUILTIN_PSHUFHW
);
17790 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psadbw128", v2di_ftype_v16qi_v16qi
, IX86_BUILTIN_PSADBW128
);
17792 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_sqrtpd", v2df_ftype_v2df
, IX86_BUILTIN_SQRTPD
);
17793 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_sqrtsd", v2df_ftype_v2df
, IX86_BUILTIN_SQRTSD
);
17795 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_shufpd", v2df_ftype_v2df_v2df_int
, IX86_BUILTIN_SHUFPD
);
17797 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtdq2pd", v2df_ftype_v4si
, IX86_BUILTIN_CVTDQ2PD
);
17798 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtdq2ps", v4sf_ftype_v4si
, IX86_BUILTIN_CVTDQ2PS
);
17800 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtpd2dq", v4si_ftype_v2df
, IX86_BUILTIN_CVTPD2DQ
);
17801 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtpd2pi", v2si_ftype_v2df
, IX86_BUILTIN_CVTPD2PI
);
17802 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtpd2ps", v4sf_ftype_v2df
, IX86_BUILTIN_CVTPD2PS
);
17803 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvttpd2dq", v4si_ftype_v2df
, IX86_BUILTIN_CVTTPD2DQ
);
17804 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvttpd2pi", v2si_ftype_v2df
, IX86_BUILTIN_CVTTPD2PI
);
17806 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtpi2pd", v2df_ftype_v2si
, IX86_BUILTIN_CVTPI2PD
);
17808 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtsd2si", int_ftype_v2df
, IX86_BUILTIN_CVTSD2SI
);
17809 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvttsd2si", int_ftype_v2df
, IX86_BUILTIN_CVTTSD2SI
);
17810 def_builtin_const (OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvtsd2si64", int64_ftype_v2df
, IX86_BUILTIN_CVTSD2SI64
);
17811 def_builtin_const (OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvttsd2si64", int64_ftype_v2df
, IX86_BUILTIN_CVTTSD2SI64
);
17813 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtps2dq", v4si_ftype_v4sf
, IX86_BUILTIN_CVTPS2DQ
);
17814 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtps2pd", v2df_ftype_v4sf
, IX86_BUILTIN_CVTPS2PD
);
17815 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvttps2dq", v4si_ftype_v4sf
, IX86_BUILTIN_CVTTPS2DQ
);
17817 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtsi2sd", v2df_ftype_v2df_int
, IX86_BUILTIN_CVTSI2SD
);
17818 def_builtin_const (OPTION_MASK_ISA_SSE2
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_cvtsi642sd", v2df_ftype_v2df_int64
, IX86_BUILTIN_CVTSI642SD
);
17819 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtsd2ss", v4sf_ftype_v4sf_v2df
, IX86_BUILTIN_CVTSD2SS
);
17820 def_builtin_const (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_cvtss2sd", v2df_ftype_v2df_v4sf
, IX86_BUILTIN_CVTSS2SD
);
17822 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_clflush", void_ftype_pcvoid
, IX86_BUILTIN_CLFLUSH
);
17823 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_lfence", void_ftype_void
, IX86_BUILTIN_LFENCE
);
17824 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_mfence", void_ftype_void
, IX86_BUILTIN_MFENCE
);
17826 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_loaddqu", v16qi_ftype_pcchar
, IX86_BUILTIN_LOADDQU
);
17827 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_storedqu", void_ftype_pchar_v16qi
, IX86_BUILTIN_STOREDQU
);
17829 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pmuludq", di_ftype_v2si_v2si
, IX86_BUILTIN_PMULUDQ
);
17830 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pmuludq128", v2di_ftype_v4si_v4si
, IX86_BUILTIN_PMULUDQ128
);
17832 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pslldqi128", v2di_ftype_v2di_int
, IX86_BUILTIN_PSLLDQI128
);
17833 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psllwi128", v8hi_ftype_v8hi_int
, IX86_BUILTIN_PSLLWI128
);
17834 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pslldi128", v4si_ftype_v4si_int
, IX86_BUILTIN_PSLLDI128
);
17835 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psllqi128", v2di_ftype_v2di_int
, IX86_BUILTIN_PSLLQI128
);
17836 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psllw128", v8hi_ftype_v8hi_v8hi
, IX86_BUILTIN_PSLLW128
);
17837 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pslld128", v4si_ftype_v4si_v4si
, IX86_BUILTIN_PSLLD128
);
17838 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psllq128", v2di_ftype_v2di_v2di
, IX86_BUILTIN_PSLLQ128
);
17840 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrldqi128", v2di_ftype_v2di_int
, IX86_BUILTIN_PSRLDQI128
);
17841 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrlwi128", v8hi_ftype_v8hi_int
, IX86_BUILTIN_PSRLWI128
);
17842 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrldi128", v4si_ftype_v4si_int
, IX86_BUILTIN_PSRLDI128
);
17843 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrlqi128", v2di_ftype_v2di_int
, IX86_BUILTIN_PSRLQI128
);
17844 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrlw128", v8hi_ftype_v8hi_v8hi
, IX86_BUILTIN_PSRLW128
);
17845 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrld128", v4si_ftype_v4si_v4si
, IX86_BUILTIN_PSRLD128
);
17846 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrlq128", v2di_ftype_v2di_v2di
, IX86_BUILTIN_PSRLQ128
);
17848 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrawi128", v8hi_ftype_v8hi_int
, IX86_BUILTIN_PSRAWI128
);
17849 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psradi128", v4si_ftype_v4si_int
, IX86_BUILTIN_PSRADI128
);
17850 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psraw128", v8hi_ftype_v8hi_v8hi
, IX86_BUILTIN_PSRAW128
);
17851 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_psrad128", v4si_ftype_v4si_v4si
, IX86_BUILTIN_PSRAD128
);
17853 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_pmaddwd128", v4si_ftype_v8hi_v8hi
, IX86_BUILTIN_PMADDWD128
);
17855 /* Prescott New Instructions. */
17856 def_builtin (OPTION_MASK_ISA_SSE3
, "__builtin_ia32_monitor", void_ftype_pcvoid_unsigned_unsigned
, IX86_BUILTIN_MONITOR
);
17857 def_builtin (OPTION_MASK_ISA_SSE3
, "__builtin_ia32_mwait", void_ftype_unsigned_unsigned
, IX86_BUILTIN_MWAIT
);
17858 def_builtin (OPTION_MASK_ISA_SSE3
, "__builtin_ia32_lddqu", v16qi_ftype_pcchar
, IX86_BUILTIN_LDDQU
);
17861 def_builtin (OPTION_MASK_ISA_SSSE3
, "__builtin_ia32_palignr128", v2di_ftype_v2di_v2di_int
, IX86_BUILTIN_PALIGNR128
);
17862 def_builtin (OPTION_MASK_ISA_SSSE3
, "__builtin_ia32_palignr", di_ftype_di_di_int
, IX86_BUILTIN_PALIGNR
);
17865 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_movntdqa", v2di_ftype_pv2di
, IX86_BUILTIN_MOVNTDQA
);
17866 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxbw128", v8hi_ftype_v16qi
, IX86_BUILTIN_PMOVSXBW128
);
17867 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxbd128", v4si_ftype_v16qi
, IX86_BUILTIN_PMOVSXBD128
);
17868 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxbq128", v2di_ftype_v16qi
, IX86_BUILTIN_PMOVSXBQ128
);
17869 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxwd128", v4si_ftype_v8hi
, IX86_BUILTIN_PMOVSXWD128
);
17870 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxwq128", v2di_ftype_v8hi
, IX86_BUILTIN_PMOVSXWQ128
);
17871 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovsxdq128", v2di_ftype_v4si
, IX86_BUILTIN_PMOVSXDQ128
);
17872 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxbw128", v8hi_ftype_v16qi
, IX86_BUILTIN_PMOVZXBW128
);
17873 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxbd128", v4si_ftype_v16qi
, IX86_BUILTIN_PMOVZXBD128
);
17874 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxbq128", v2di_ftype_v16qi
, IX86_BUILTIN_PMOVZXBQ128
);
17875 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxwd128", v4si_ftype_v8hi
, IX86_BUILTIN_PMOVZXWD128
);
17876 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxwq128", v2di_ftype_v8hi
, IX86_BUILTIN_PMOVZXWQ128
);
17877 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmovzxdq128", v2di_ftype_v4si
, IX86_BUILTIN_PMOVZXDQ128
);
17878 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_pmuldq128", v2di_ftype_v4si_v4si
, IX86_BUILTIN_PMULDQ128
);
17879 def_builtin_const (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_roundpd", v2df_ftype_v2df_int
, IX86_BUILTIN_ROUNDPD
);
17880 def_builtin_const (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_roundps", v4sf_ftype_v4sf_int
, IX86_BUILTIN_ROUNDPS
);
17881 def_builtin_const (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_roundsd", v2df_ftype_v2df_v2df_int
, IX86_BUILTIN_ROUNDSD
);
17882 def_builtin_const (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_roundss", v4sf_ftype_v4sf_v4sf_int
, IX86_BUILTIN_ROUNDSS
);
17884 /* AMDFAM10 SSE4A New built-ins */
17885 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_movntsd", void_ftype_pdouble_v2df
, IX86_BUILTIN_MOVNTSD
);
17886 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_movntss", void_ftype_pfloat_v4sf
, IX86_BUILTIN_MOVNTSS
);
17887 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_extrqi", v2di_ftype_v2di_unsigned_unsigned
, IX86_BUILTIN_EXTRQI
);
17888 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_extrq", v2di_ftype_v2di_v16qi
, IX86_BUILTIN_EXTRQ
);
17889 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_insertqi", v2di_ftype_v2di_v2di_unsigned_unsigned
, IX86_BUILTIN_INSERTQI
);
17890 def_builtin (OPTION_MASK_ISA_SSE4A
, "__builtin_ia32_insertq", v2di_ftype_v2di_v2di
, IX86_BUILTIN_INSERTQ
);
17892 /* Access to the vec_init patterns. */
17893 ftype
= build_function_type_list (V2SI_type_node
, integer_type_node
,
17894 integer_type_node
, NULL_TREE
);
17895 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_vec_init_v2si", ftype
, IX86_BUILTIN_VEC_INIT_V2SI
);
17897 ftype
= build_function_type_list (V4HI_type_node
, short_integer_type_node
,
17898 short_integer_type_node
,
17899 short_integer_type_node
,
17900 short_integer_type_node
, NULL_TREE
);
17901 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_vec_init_v4hi", ftype
, IX86_BUILTIN_VEC_INIT_V4HI
);
17903 ftype
= build_function_type_list (V8QI_type_node
, char_type_node
,
17904 char_type_node
, char_type_node
,
17905 char_type_node
, char_type_node
,
17906 char_type_node
, char_type_node
,
17907 char_type_node
, NULL_TREE
);
17908 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_vec_init_v8qi", ftype
, IX86_BUILTIN_VEC_INIT_V8QI
);
17910 /* Access to the vec_extract patterns. */
17911 ftype
= build_function_type_list (double_type_node
, V2DF_type_node
,
17912 integer_type_node
, NULL_TREE
);
17913 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_ext_v2df", ftype
, IX86_BUILTIN_VEC_EXT_V2DF
);
17915 ftype
= build_function_type_list (long_long_integer_type_node
,
17916 V2DI_type_node
, integer_type_node
,
17918 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_ext_v2di", ftype
, IX86_BUILTIN_VEC_EXT_V2DI
);
17920 ftype
= build_function_type_list (float_type_node
, V4SF_type_node
,
17921 integer_type_node
, NULL_TREE
);
17922 def_builtin (OPTION_MASK_ISA_SSE
, "__builtin_ia32_vec_ext_v4sf", ftype
, IX86_BUILTIN_VEC_EXT_V4SF
);
17924 ftype
= build_function_type_list (intSI_type_node
, V4SI_type_node
,
17925 integer_type_node
, NULL_TREE
);
17926 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_ext_v4si", ftype
, IX86_BUILTIN_VEC_EXT_V4SI
);
17928 ftype
= build_function_type_list (intHI_type_node
, V8HI_type_node
,
17929 integer_type_node
, NULL_TREE
);
17930 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_ext_v8hi", ftype
, IX86_BUILTIN_VEC_EXT_V8HI
);
17932 ftype
= build_function_type_list (intHI_type_node
, V4HI_type_node
,
17933 integer_type_node
, NULL_TREE
);
17934 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_vec_ext_v4hi", ftype
, IX86_BUILTIN_VEC_EXT_V4HI
);
17936 ftype
= build_function_type_list (intSI_type_node
, V2SI_type_node
,
17937 integer_type_node
, NULL_TREE
);
17938 def_builtin (OPTION_MASK_ISA_MMX
, "__builtin_ia32_vec_ext_v2si", ftype
, IX86_BUILTIN_VEC_EXT_V2SI
);
17940 ftype
= build_function_type_list (intQI_type_node
, V16QI_type_node
,
17941 integer_type_node
, NULL_TREE
);
17942 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_ext_v16qi", ftype
, IX86_BUILTIN_VEC_EXT_V16QI
);
17944 /* Access to the vec_set patterns. */
17945 ftype
= build_function_type_list (V2DI_type_node
, V2DI_type_node
,
17947 integer_type_node
, NULL_TREE
);
17948 def_builtin (OPTION_MASK_ISA_SSE4_1
| OPTION_MASK_ISA_64BIT
, "__builtin_ia32_vec_set_v2di", ftype
, IX86_BUILTIN_VEC_SET_V2DI
);
17950 ftype
= build_function_type_list (V4SF_type_node
, V4SF_type_node
,
17952 integer_type_node
, NULL_TREE
);
17953 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_vec_set_v4sf", ftype
, IX86_BUILTIN_VEC_SET_V4SF
);
17955 ftype
= build_function_type_list (V4SI_type_node
, V4SI_type_node
,
17957 integer_type_node
, NULL_TREE
);
17958 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_vec_set_v4si", ftype
, IX86_BUILTIN_VEC_SET_V4SI
);
17960 ftype
= build_function_type_list (V8HI_type_node
, V8HI_type_node
,
17962 integer_type_node
, NULL_TREE
);
17963 def_builtin (OPTION_MASK_ISA_SSE2
, "__builtin_ia32_vec_set_v8hi", ftype
, IX86_BUILTIN_VEC_SET_V8HI
);
17965 ftype
= build_function_type_list (V4HI_type_node
, V4HI_type_node
,
17967 integer_type_node
, NULL_TREE
);
17968 def_builtin (OPTION_MASK_ISA_SSE
| OPTION_MASK_ISA_3DNOW_A
, "__builtin_ia32_vec_set_v4hi", ftype
, IX86_BUILTIN_VEC_SET_V4HI
);
17970 ftype
= build_function_type_list (V16QI_type_node
, V16QI_type_node
,
17972 integer_type_node
, NULL_TREE
);
17973 def_builtin (OPTION_MASK_ISA_SSE4_1
, "__builtin_ia32_vec_set_v16qi", ftype
, IX86_BUILTIN_VEC_SET_V16QI
);
17977 ix86_init_builtins (void)
17980 ix86_init_mmx_sse_builtins ();
17983 /* Errors in the source file can cause expand_expr to return const0_rtx
17984 where we expect a vector. To avoid crashing, use one of the vector
17985 clear instructions. */
17987 safe_vector_operand (rtx x
, enum machine_mode mode
)
17989 if (x
== const0_rtx
)
17990 x
= CONST0_RTX (mode
);
17994 /* Subroutine of ix86_expand_builtin to take care of SSE insns with
17995 4 operands. The third argument must be a constant smaller than 8
17999 ix86_expand_sse_4_operands_builtin (enum insn_code icode
, tree exp
,
18003 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18004 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18005 tree arg2
= CALL_EXPR_ARG (exp
, 2);
18006 rtx op0
= expand_normal (arg0
);
18007 rtx op1
= expand_normal (arg1
);
18008 rtx op2
= expand_normal (arg2
);
18009 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
18010 enum machine_mode mode0
= insn_data
[icode
].operand
[1].mode
;
18011 enum machine_mode mode1
= insn_data
[icode
].operand
[2].mode
;
18012 enum machine_mode mode2
;
18015 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode0
))
18016 op0
= copy_to_mode_reg (mode0
, op0
);
18017 if ((optimize
&& !register_operand (op1
, mode1
))
18018 || !(*insn_data
[icode
].operand
[2].predicate
) (op1
, mode1
))
18019 op1
= copy_to_mode_reg (mode1
, op1
);
18023 case CODE_FOR_sse4_1_blendvpd
:
18024 case CODE_FOR_sse4_1_blendvps
:
18025 case CODE_FOR_sse4_1_pblendvb
:
18026 /* The third argument of variable blends must be xmm0. */
18027 xmm0
= gen_rtx_REG (tmode
, FIRST_SSE_REG
);
18028 emit_move_insn (xmm0
, op2
);
18032 mode2
= insn_data
[icode
].operand
[2].mode
;
18033 if (! (*insn_data
[icode
].operand
[3].predicate
) (op2
, mode2
))
18037 case CODE_FOR_sse4_1_roundsd
:
18038 case CODE_FOR_sse4_1_roundss
:
18039 error ("the third argument must be a 4-bit immediate");
18042 error ("the third argument must be a 8-bit immediate");
18052 || GET_MODE (target
) != tmode
18053 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18054 target
= gen_reg_rtx (tmode
);
18055 pat
= GEN_FCN (icode
) (target
, op0
, op1
, op2
);
18062 /* Subroutine of ix86_expand_builtin to take care of binop insns. */
18065 ix86_expand_binop_builtin (enum insn_code icode
, tree exp
, rtx target
)
18068 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18069 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18070 rtx op0
= expand_normal (arg0
);
18071 rtx op1
= expand_normal (arg1
);
18072 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
18073 enum machine_mode mode0
= insn_data
[icode
].operand
[1].mode
;
18074 enum machine_mode mode1
= insn_data
[icode
].operand
[2].mode
;
18076 if (VECTOR_MODE_P (mode0
))
18077 op0
= safe_vector_operand (op0
, mode0
);
18078 if (VECTOR_MODE_P (mode1
))
18079 op1
= safe_vector_operand (op1
, mode1
);
18081 if (optimize
|| !target
18082 || GET_MODE (target
) != tmode
18083 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18084 target
= gen_reg_rtx (tmode
);
18086 if (GET_MODE (op1
) == SImode
&& mode1
== TImode
)
18088 rtx x
= gen_reg_rtx (V4SImode
);
18089 emit_insn (gen_sse2_loadd (x
, op1
));
18090 op1
= gen_lowpart (TImode
, x
);
18093 /* The insn must want input operands in the same modes as the
18095 gcc_assert ((GET_MODE (op0
) == mode0
|| GET_MODE (op0
) == VOIDmode
)
18096 && (GET_MODE (op1
) == mode1
|| GET_MODE (op1
) == VOIDmode
));
18098 if (!(*insn_data
[icode
].operand
[1].predicate
) (op0
, mode0
))
18099 op0
= copy_to_mode_reg (mode0
, op0
);
18100 if (!(*insn_data
[icode
].operand
[2].predicate
) (op1
, mode1
))
18101 op1
= copy_to_mode_reg (mode1
, op1
);
18103 /* ??? Using ix86_fixup_binary_operands is problematic when
18104 we've got mismatched modes. Fake it. */
18110 if (tmode
== mode0
&& tmode
== mode1
)
18112 target
= ix86_fixup_binary_operands (UNKNOWN
, tmode
, xops
);
18116 else if (optimize
|| !ix86_binary_operator_ok (UNKNOWN
, tmode
, xops
))
18118 op0
= force_reg (mode0
, op0
);
18119 op1
= force_reg (mode1
, op1
);
18120 target
= gen_reg_rtx (tmode
);
18123 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18130 /* Subroutine of ix86_expand_builtin to take care of stores. */
18133 ix86_expand_store_builtin (enum insn_code icode
, tree exp
)
18136 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18137 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18138 rtx op0
= expand_normal (arg0
);
18139 rtx op1
= expand_normal (arg1
);
18140 enum machine_mode mode0
= insn_data
[icode
].operand
[0].mode
;
18141 enum machine_mode mode1
= insn_data
[icode
].operand
[1].mode
;
18143 if (VECTOR_MODE_P (mode1
))
18144 op1
= safe_vector_operand (op1
, mode1
);
18146 op0
= gen_rtx_MEM (mode0
, copy_to_mode_reg (Pmode
, op0
));
18147 op1
= copy_to_mode_reg (mode1
, op1
);
18149 pat
= GEN_FCN (icode
) (op0
, op1
);
18155 /* Subroutine of ix86_expand_builtin to take care of unop insns. */
18158 ix86_expand_unop_builtin (enum insn_code icode
, tree exp
,
18159 rtx target
, int do_load
)
18162 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18163 rtx op0
= expand_normal (arg0
);
18164 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
18165 enum machine_mode mode0
= insn_data
[icode
].operand
[1].mode
;
18167 if (optimize
|| !target
18168 || GET_MODE (target
) != tmode
18169 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18170 target
= gen_reg_rtx (tmode
);
18172 op0
= gen_rtx_MEM (mode0
, copy_to_mode_reg (Pmode
, op0
));
18175 if (VECTOR_MODE_P (mode0
))
18176 op0
= safe_vector_operand (op0
, mode0
);
18178 if ((optimize
&& !register_operand (op0
, mode0
))
18179 || ! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode0
))
18180 op0
= copy_to_mode_reg (mode0
, op0
);
18185 case CODE_FOR_sse4_1_roundpd
:
18186 case CODE_FOR_sse4_1_roundps
:
18188 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18189 rtx op1
= expand_normal (arg1
);
18190 enum machine_mode mode1
= insn_data
[icode
].operand
[2].mode
;
18192 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode1
))
18194 error ("the second argument must be a 4-bit immediate");
18197 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18201 pat
= GEN_FCN (icode
) (target
, op0
);
18211 /* Subroutine of ix86_expand_builtin to take care of three special unop insns:
18212 sqrtss, rsqrtss, rcpss. */
18215 ix86_expand_unop1_builtin (enum insn_code icode
, tree exp
, rtx target
)
18218 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18219 rtx op1
, op0
= expand_normal (arg0
);
18220 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
18221 enum machine_mode mode0
= insn_data
[icode
].operand
[1].mode
;
18223 if (optimize
|| !target
18224 || GET_MODE (target
) != tmode
18225 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18226 target
= gen_reg_rtx (tmode
);
18228 if (VECTOR_MODE_P (mode0
))
18229 op0
= safe_vector_operand (op0
, mode0
);
18231 if ((optimize
&& !register_operand (op0
, mode0
))
18232 || ! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode0
))
18233 op0
= copy_to_mode_reg (mode0
, op0
);
18236 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode0
))
18237 op1
= copy_to_mode_reg (mode0
, op1
);
18239 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18246 /* Subroutine of ix86_expand_builtin to take care of comparison insns. */
18249 ix86_expand_sse_compare (const struct builtin_description
*d
, tree exp
,
18253 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18254 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18255 rtx op0
= expand_normal (arg0
);
18256 rtx op1
= expand_normal (arg1
);
18258 enum machine_mode tmode
= insn_data
[d
->icode
].operand
[0].mode
;
18259 enum machine_mode mode0
= insn_data
[d
->icode
].operand
[1].mode
;
18260 enum machine_mode mode1
= insn_data
[d
->icode
].operand
[2].mode
;
18261 enum rtx_code comparison
= d
->comparison
;
18263 if (VECTOR_MODE_P (mode0
))
18264 op0
= safe_vector_operand (op0
, mode0
);
18265 if (VECTOR_MODE_P (mode1
))
18266 op1
= safe_vector_operand (op1
, mode1
);
18268 /* Swap operands if we have a comparison that isn't available in
18270 if (d
->flag
& BUILTIN_DESC_SWAP_OPERANDS
)
18272 rtx tmp
= gen_reg_rtx (mode1
);
18273 emit_move_insn (tmp
, op1
);
18278 if (optimize
|| !target
18279 || GET_MODE (target
) != tmode
18280 || ! (*insn_data
[d
->icode
].operand
[0].predicate
) (target
, tmode
))
18281 target
= gen_reg_rtx (tmode
);
18283 if ((optimize
&& !register_operand (op0
, mode0
))
18284 || ! (*insn_data
[d
->icode
].operand
[1].predicate
) (op0
, mode0
))
18285 op0
= copy_to_mode_reg (mode0
, op0
);
18286 if ((optimize
&& !register_operand (op1
, mode1
))
18287 || ! (*insn_data
[d
->icode
].operand
[2].predicate
) (op1
, mode1
))
18288 op1
= copy_to_mode_reg (mode1
, op1
);
18290 op2
= gen_rtx_fmt_ee (comparison
, mode0
, op0
, op1
);
18291 pat
= GEN_FCN (d
->icode
) (target
, op0
, op1
, op2
);
18298 /* Subroutine of ix86_expand_builtin to take care of comi insns. */
18301 ix86_expand_sse_comi (const struct builtin_description
*d
, tree exp
,
18305 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18306 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18307 rtx op0
= expand_normal (arg0
);
18308 rtx op1
= expand_normal (arg1
);
18309 enum machine_mode mode0
= insn_data
[d
->icode
].operand
[0].mode
;
18310 enum machine_mode mode1
= insn_data
[d
->icode
].operand
[1].mode
;
18311 enum rtx_code comparison
= d
->comparison
;
18313 if (VECTOR_MODE_P (mode0
))
18314 op0
= safe_vector_operand (op0
, mode0
);
18315 if (VECTOR_MODE_P (mode1
))
18316 op1
= safe_vector_operand (op1
, mode1
);
18318 /* Swap operands if we have a comparison that isn't available in
18320 if (d
->flag
& BUILTIN_DESC_SWAP_OPERANDS
)
18327 target
= gen_reg_rtx (SImode
);
18328 emit_move_insn (target
, const0_rtx
);
18329 target
= gen_rtx_SUBREG (QImode
, target
, 0);
18331 if ((optimize
&& !register_operand (op0
, mode0
))
18332 || !(*insn_data
[d
->icode
].operand
[0].predicate
) (op0
, mode0
))
18333 op0
= copy_to_mode_reg (mode0
, op0
);
18334 if ((optimize
&& !register_operand (op1
, mode1
))
18335 || !(*insn_data
[d
->icode
].operand
[1].predicate
) (op1
, mode1
))
18336 op1
= copy_to_mode_reg (mode1
, op1
);
18338 pat
= GEN_FCN (d
->icode
) (op0
, op1
);
18342 emit_insn (gen_rtx_SET (VOIDmode
,
18343 gen_rtx_STRICT_LOW_PART (VOIDmode
, target
),
18344 gen_rtx_fmt_ee (comparison
, QImode
,
18348 return SUBREG_REG (target
);
18351 /* Subroutine of ix86_expand_builtin to take care of ptest insns. */
18354 ix86_expand_sse_ptest (const struct builtin_description
*d
, tree exp
,
18358 tree arg0
= CALL_EXPR_ARG (exp
, 0);
18359 tree arg1
= CALL_EXPR_ARG (exp
, 1);
18360 rtx op0
= expand_normal (arg0
);
18361 rtx op1
= expand_normal (arg1
);
18362 enum machine_mode mode0
= insn_data
[d
->icode
].operand
[0].mode
;
18363 enum machine_mode mode1
= insn_data
[d
->icode
].operand
[1].mode
;
18364 enum rtx_code comparison
= d
->comparison
;
18366 if (VECTOR_MODE_P (mode0
))
18367 op0
= safe_vector_operand (op0
, mode0
);
18368 if (VECTOR_MODE_P (mode1
))
18369 op1
= safe_vector_operand (op1
, mode1
);
18371 target
= gen_reg_rtx (SImode
);
18372 emit_move_insn (target
, const0_rtx
);
18373 target
= gen_rtx_SUBREG (QImode
, target
, 0);
18375 if ((optimize
&& !register_operand (op0
, mode0
))
18376 || !(*insn_data
[d
->icode
].operand
[0].predicate
) (op0
, mode0
))
18377 op0
= copy_to_mode_reg (mode0
, op0
);
18378 if ((optimize
&& !register_operand (op1
, mode1
))
18379 || !(*insn_data
[d
->icode
].operand
[1].predicate
) (op1
, mode1
))
18380 op1
= copy_to_mode_reg (mode1
, op1
);
18382 pat
= GEN_FCN (d
->icode
) (op0
, op1
);
18386 emit_insn (gen_rtx_SET (VOIDmode
,
18387 gen_rtx_STRICT_LOW_PART (VOIDmode
, target
),
18388 gen_rtx_fmt_ee (comparison
, QImode
,
18392 return SUBREG_REG (target
);
18395 /* Return the integer constant in ARG. Constrain it to be in the range
18396 of the subparts of VEC_TYPE; issue an error if not. */
18399 get_element_number (tree vec_type
, tree arg
)
18401 unsigned HOST_WIDE_INT elt
, max
= TYPE_VECTOR_SUBPARTS (vec_type
) - 1;
18403 if (!host_integerp (arg
, 1)
18404 || (elt
= tree_low_cst (arg
, 1), elt
> max
))
18406 error ("selector must be an integer constant in the range 0..%wi", max
);
18413 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
18414 ix86_expand_vector_init. We DO have language-level syntax for this, in
18415 the form of (type){ init-list }. Except that since we can't place emms
18416 instructions from inside the compiler, we can't allow the use of MMX
18417 registers unless the user explicitly asks for it. So we do *not* define
18418 vec_set/vec_extract/vec_init patterns for MMX modes in mmx.md. Instead
18419 we have builtins invoked by mmintrin.h that gives us license to emit
18420 these sorts of instructions. */
18423 ix86_expand_vec_init_builtin (tree type
, tree exp
, rtx target
)
18425 enum machine_mode tmode
= TYPE_MODE (type
);
18426 enum machine_mode inner_mode
= GET_MODE_INNER (tmode
);
18427 int i
, n_elt
= GET_MODE_NUNITS (tmode
);
18428 rtvec v
= rtvec_alloc (n_elt
);
18430 gcc_assert (VECTOR_MODE_P (tmode
));
18431 gcc_assert (call_expr_nargs (exp
) == n_elt
);
18433 for (i
= 0; i
< n_elt
; ++i
)
18435 rtx x
= expand_normal (CALL_EXPR_ARG (exp
, i
));
18436 RTVEC_ELT (v
, i
) = gen_lowpart (inner_mode
, x
);
18439 if (!target
|| !register_operand (target
, tmode
))
18440 target
= gen_reg_rtx (tmode
);
18442 ix86_expand_vector_init (true, target
, gen_rtx_PARALLEL (tmode
, v
));
18446 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
18447 ix86_expand_vector_extract. They would be redundant (for non-MMX) if we
18448 had a language-level syntax for referencing vector elements. */
18451 ix86_expand_vec_ext_builtin (tree exp
, rtx target
)
18453 enum machine_mode tmode
, mode0
;
18458 arg0
= CALL_EXPR_ARG (exp
, 0);
18459 arg1
= CALL_EXPR_ARG (exp
, 1);
18461 op0
= expand_normal (arg0
);
18462 elt
= get_element_number (TREE_TYPE (arg0
), arg1
);
18464 tmode
= TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0
)));
18465 mode0
= TYPE_MODE (TREE_TYPE (arg0
));
18466 gcc_assert (VECTOR_MODE_P (mode0
));
18468 op0
= force_reg (mode0
, op0
);
18470 if (optimize
|| !target
|| !register_operand (target
, tmode
))
18471 target
= gen_reg_rtx (tmode
);
18473 ix86_expand_vector_extract (true, target
, op0
, elt
);
18478 /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around
18479 ix86_expand_vector_set. They would be redundant (for non-MMX) if we had
18480 a language-level syntax for referencing vector elements. */
18483 ix86_expand_vec_set_builtin (tree exp
)
18485 enum machine_mode tmode
, mode1
;
18486 tree arg0
, arg1
, arg2
;
18488 rtx op0
, op1
, target
;
18490 arg0
= CALL_EXPR_ARG (exp
, 0);
18491 arg1
= CALL_EXPR_ARG (exp
, 1);
18492 arg2
= CALL_EXPR_ARG (exp
, 2);
18494 tmode
= TYPE_MODE (TREE_TYPE (arg0
));
18495 mode1
= TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0
)));
18496 gcc_assert (VECTOR_MODE_P (tmode
));
18498 op0
= expand_expr (arg0
, NULL_RTX
, tmode
, 0);
18499 op1
= expand_expr (arg1
, NULL_RTX
, mode1
, 0);
18500 elt
= get_element_number (TREE_TYPE (arg0
), arg2
);
18502 if (GET_MODE (op1
) != mode1
&& GET_MODE (op1
) != VOIDmode
)
18503 op1
= convert_modes (mode1
, GET_MODE (op1
), op1
, true);
18505 op0
= force_reg (tmode
, op0
);
18506 op1
= force_reg (mode1
, op1
);
18508 /* OP0 is the source of these builtin functions and shouldn't be
18509 modified. Create a copy, use it and return it as target. */
18510 target
= gen_reg_rtx (tmode
);
18511 emit_move_insn (target
, op0
);
18512 ix86_expand_vector_set (true, target
, op1
, elt
);
18517 /* Expand an expression EXP that calls a built-in function,
18518 with result going to TARGET if that's convenient
18519 (and in mode MODE if that's convenient).
18520 SUBTARGET may be used as the target for computing one of EXP's operands.
18521 IGNORE is nonzero if the value is to be ignored. */
18524 ix86_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
18525 enum machine_mode mode ATTRIBUTE_UNUSED
,
18526 int ignore ATTRIBUTE_UNUSED
)
18528 const struct builtin_description
*d
;
18530 enum insn_code icode
;
18531 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
18532 tree arg0
, arg1
, arg2
, arg3
;
18533 rtx op0
, op1
, op2
, op3
, pat
;
18534 enum machine_mode tmode
, mode0
, mode1
, mode2
, mode3
, mode4
;
18535 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
18539 case IX86_BUILTIN_EMMS
:
18540 emit_insn (gen_mmx_emms ());
18543 case IX86_BUILTIN_SFENCE
:
18544 emit_insn (gen_sse_sfence ());
18547 case IX86_BUILTIN_MASKMOVQ
:
18548 case IX86_BUILTIN_MASKMOVDQU
:
18549 icode
= (fcode
== IX86_BUILTIN_MASKMOVQ
18550 ? CODE_FOR_mmx_maskmovq
18551 : CODE_FOR_sse2_maskmovdqu
);
18552 /* Note the arg order is different from the operand order. */
18553 arg1
= CALL_EXPR_ARG (exp
, 0);
18554 arg2
= CALL_EXPR_ARG (exp
, 1);
18555 arg0
= CALL_EXPR_ARG (exp
, 2);
18556 op0
= expand_normal (arg0
);
18557 op1
= expand_normal (arg1
);
18558 op2
= expand_normal (arg2
);
18559 mode0
= insn_data
[icode
].operand
[0].mode
;
18560 mode1
= insn_data
[icode
].operand
[1].mode
;
18561 mode2
= insn_data
[icode
].operand
[2].mode
;
18563 op0
= force_reg (Pmode
, op0
);
18564 op0
= gen_rtx_MEM (mode1
, op0
);
18566 if (! (*insn_data
[icode
].operand
[0].predicate
) (op0
, mode0
))
18567 op0
= copy_to_mode_reg (mode0
, op0
);
18568 if (! (*insn_data
[icode
].operand
[1].predicate
) (op1
, mode1
))
18569 op1
= copy_to_mode_reg (mode1
, op1
);
18570 if (! (*insn_data
[icode
].operand
[2].predicate
) (op2
, mode2
))
18571 op2
= copy_to_mode_reg (mode2
, op2
);
18572 pat
= GEN_FCN (icode
) (op0
, op1
, op2
);
18578 case IX86_BUILTIN_SQRTSS
:
18579 return ix86_expand_unop1_builtin (CODE_FOR_sse_vmsqrtv4sf2
, exp
, target
);
18580 case IX86_BUILTIN_RSQRTSS
:
18581 return ix86_expand_unop1_builtin (CODE_FOR_sse_vmrsqrtv4sf2
, exp
, target
);
18582 case IX86_BUILTIN_RCPSS
:
18583 return ix86_expand_unop1_builtin (CODE_FOR_sse_vmrcpv4sf2
, exp
, target
);
18585 case IX86_BUILTIN_LOADUPS
:
18586 return ix86_expand_unop_builtin (CODE_FOR_sse_movups
, exp
, target
, 1);
18588 case IX86_BUILTIN_STOREUPS
:
18589 return ix86_expand_store_builtin (CODE_FOR_sse_movups
, exp
);
18591 case IX86_BUILTIN_LOADHPS
:
18592 case IX86_BUILTIN_LOADLPS
:
18593 case IX86_BUILTIN_LOADHPD
:
18594 case IX86_BUILTIN_LOADLPD
:
18595 icode
= (fcode
== IX86_BUILTIN_LOADHPS
? CODE_FOR_sse_loadhps
18596 : fcode
== IX86_BUILTIN_LOADLPS
? CODE_FOR_sse_loadlps
18597 : fcode
== IX86_BUILTIN_LOADHPD
? CODE_FOR_sse2_loadhpd
18598 : CODE_FOR_sse2_loadlpd
);
18599 arg0
= CALL_EXPR_ARG (exp
, 0);
18600 arg1
= CALL_EXPR_ARG (exp
, 1);
18601 op0
= expand_normal (arg0
);
18602 op1
= expand_normal (arg1
);
18603 tmode
= insn_data
[icode
].operand
[0].mode
;
18604 mode0
= insn_data
[icode
].operand
[1].mode
;
18605 mode1
= insn_data
[icode
].operand
[2].mode
;
18607 op0
= force_reg (mode0
, op0
);
18608 op1
= gen_rtx_MEM (mode1
, copy_to_mode_reg (Pmode
, op1
));
18609 if (optimize
|| target
== 0
18610 || GET_MODE (target
) != tmode
18611 || !register_operand (target
, tmode
))
18612 target
= gen_reg_rtx (tmode
);
18613 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18619 case IX86_BUILTIN_STOREHPS
:
18620 case IX86_BUILTIN_STORELPS
:
18621 icode
= (fcode
== IX86_BUILTIN_STOREHPS
? CODE_FOR_sse_storehps
18622 : CODE_FOR_sse_storelps
);
18623 arg0
= CALL_EXPR_ARG (exp
, 0);
18624 arg1
= CALL_EXPR_ARG (exp
, 1);
18625 op0
= expand_normal (arg0
);
18626 op1
= expand_normal (arg1
);
18627 mode0
= insn_data
[icode
].operand
[0].mode
;
18628 mode1
= insn_data
[icode
].operand
[1].mode
;
18630 op0
= gen_rtx_MEM (mode0
, copy_to_mode_reg (Pmode
, op0
));
18631 op1
= force_reg (mode1
, op1
);
18633 pat
= GEN_FCN (icode
) (op0
, op1
);
18639 case IX86_BUILTIN_MOVNTPS
:
18640 return ix86_expand_store_builtin (CODE_FOR_sse_movntv4sf
, exp
);
18641 case IX86_BUILTIN_MOVNTQ
:
18642 return ix86_expand_store_builtin (CODE_FOR_sse_movntdi
, exp
);
18644 case IX86_BUILTIN_LDMXCSR
:
18645 op0
= expand_normal (CALL_EXPR_ARG (exp
, 0));
18646 target
= assign_386_stack_local (SImode
, SLOT_TEMP
);
18647 emit_move_insn (target
, op0
);
18648 emit_insn (gen_sse_ldmxcsr (target
));
18651 case IX86_BUILTIN_STMXCSR
:
18652 target
= assign_386_stack_local (SImode
, SLOT_TEMP
);
18653 emit_insn (gen_sse_stmxcsr (target
));
18654 return copy_to_mode_reg (SImode
, target
);
18656 case IX86_BUILTIN_SHUFPS
:
18657 case IX86_BUILTIN_SHUFPD
:
18658 icode
= (fcode
== IX86_BUILTIN_SHUFPS
18659 ? CODE_FOR_sse_shufps
18660 : CODE_FOR_sse2_shufpd
);
18661 arg0
= CALL_EXPR_ARG (exp
, 0);
18662 arg1
= CALL_EXPR_ARG (exp
, 1);
18663 arg2
= CALL_EXPR_ARG (exp
, 2);
18664 op0
= expand_normal (arg0
);
18665 op1
= expand_normal (arg1
);
18666 op2
= expand_normal (arg2
);
18667 tmode
= insn_data
[icode
].operand
[0].mode
;
18668 mode0
= insn_data
[icode
].operand
[1].mode
;
18669 mode1
= insn_data
[icode
].operand
[2].mode
;
18670 mode2
= insn_data
[icode
].operand
[3].mode
;
18672 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode0
))
18673 op0
= copy_to_mode_reg (mode0
, op0
);
18674 if ((optimize
&& !register_operand (op1
, mode1
))
18675 || !(*insn_data
[icode
].operand
[2].predicate
) (op1
, mode1
))
18676 op1
= copy_to_mode_reg (mode1
, op1
);
18677 if (! (*insn_data
[icode
].operand
[3].predicate
) (op2
, mode2
))
18679 /* @@@ better error message */
18680 error ("mask must be an immediate");
18681 return gen_reg_rtx (tmode
);
18683 if (optimize
|| target
== 0
18684 || GET_MODE (target
) != tmode
18685 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18686 target
= gen_reg_rtx (tmode
);
18687 pat
= GEN_FCN (icode
) (target
, op0
, op1
, op2
);
18693 case IX86_BUILTIN_PSHUFW
:
18694 case IX86_BUILTIN_PSHUFD
:
18695 case IX86_BUILTIN_PSHUFHW
:
18696 case IX86_BUILTIN_PSHUFLW
:
18697 icode
= ( fcode
== IX86_BUILTIN_PSHUFHW
? CODE_FOR_sse2_pshufhw
18698 : fcode
== IX86_BUILTIN_PSHUFLW
? CODE_FOR_sse2_pshuflw
18699 : fcode
== IX86_BUILTIN_PSHUFD
? CODE_FOR_sse2_pshufd
18700 : CODE_FOR_mmx_pshufw
);
18701 arg0
= CALL_EXPR_ARG (exp
, 0);
18702 arg1
= CALL_EXPR_ARG (exp
, 1);
18703 op0
= expand_normal (arg0
);
18704 op1
= expand_normal (arg1
);
18705 tmode
= insn_data
[icode
].operand
[0].mode
;
18706 mode1
= insn_data
[icode
].operand
[1].mode
;
18707 mode2
= insn_data
[icode
].operand
[2].mode
;
18709 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
18710 op0
= copy_to_mode_reg (mode1
, op0
);
18711 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
18713 /* @@@ better error message */
18714 error ("mask must be an immediate");
18718 || GET_MODE (target
) != tmode
18719 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
18720 target
= gen_reg_rtx (tmode
);
18721 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18727 case IX86_BUILTIN_PSLLWI128
:
18728 icode
= CODE_FOR_ashlv8hi3
;
18730 case IX86_BUILTIN_PSLLDI128
:
18731 icode
= CODE_FOR_ashlv4si3
;
18733 case IX86_BUILTIN_PSLLQI128
:
18734 icode
= CODE_FOR_ashlv2di3
;
18736 case IX86_BUILTIN_PSRAWI128
:
18737 icode
= CODE_FOR_ashrv8hi3
;
18739 case IX86_BUILTIN_PSRADI128
:
18740 icode
= CODE_FOR_ashrv4si3
;
18742 case IX86_BUILTIN_PSRLWI128
:
18743 icode
= CODE_FOR_lshrv8hi3
;
18745 case IX86_BUILTIN_PSRLDI128
:
18746 icode
= CODE_FOR_lshrv4si3
;
18748 case IX86_BUILTIN_PSRLQI128
:
18749 icode
= CODE_FOR_lshrv2di3
;
18752 arg0
= CALL_EXPR_ARG (exp
, 0);
18753 arg1
= CALL_EXPR_ARG (exp
, 1);
18754 op0
= expand_normal (arg0
);
18755 op1
= expand_normal (arg1
);
18757 if (!CONST_INT_P (op1
))
18759 error ("shift must be an immediate");
18762 if (INTVAL (op1
) < 0 || INTVAL (op1
) > 255)
18763 op1
= GEN_INT (255);
18765 tmode
= insn_data
[icode
].operand
[0].mode
;
18766 mode1
= insn_data
[icode
].operand
[1].mode
;
18767 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
18768 op0
= copy_to_reg (op0
);
18770 target
= gen_reg_rtx (tmode
);
18771 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18777 case IX86_BUILTIN_PSLLW128
:
18778 icode
= CODE_FOR_ashlv8hi3
;
18780 case IX86_BUILTIN_PSLLD128
:
18781 icode
= CODE_FOR_ashlv4si3
;
18783 case IX86_BUILTIN_PSLLQ128
:
18784 icode
= CODE_FOR_ashlv2di3
;
18786 case IX86_BUILTIN_PSRAW128
:
18787 icode
= CODE_FOR_ashrv8hi3
;
18789 case IX86_BUILTIN_PSRAD128
:
18790 icode
= CODE_FOR_ashrv4si3
;
18792 case IX86_BUILTIN_PSRLW128
:
18793 icode
= CODE_FOR_lshrv8hi3
;
18795 case IX86_BUILTIN_PSRLD128
:
18796 icode
= CODE_FOR_lshrv4si3
;
18798 case IX86_BUILTIN_PSRLQ128
:
18799 icode
= CODE_FOR_lshrv2di3
;
18802 arg0
= CALL_EXPR_ARG (exp
, 0);
18803 arg1
= CALL_EXPR_ARG (exp
, 1);
18804 op0
= expand_normal (arg0
);
18805 op1
= expand_normal (arg1
);
18807 tmode
= insn_data
[icode
].operand
[0].mode
;
18808 mode1
= insn_data
[icode
].operand
[1].mode
;
18810 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
18811 op0
= copy_to_reg (op0
);
18813 op1
= simplify_gen_subreg (TImode
, op1
, GET_MODE (op1
), 0);
18814 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, TImode
))
18815 op1
= copy_to_reg (op1
);
18817 target
= gen_reg_rtx (tmode
);
18818 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
18824 case IX86_BUILTIN_PSLLDQI128
:
18825 case IX86_BUILTIN_PSRLDQI128
:
18826 icode
= (fcode
== IX86_BUILTIN_PSLLDQI128
? CODE_FOR_sse2_ashlti3
18827 : CODE_FOR_sse2_lshrti3
);
18828 arg0
= CALL_EXPR_ARG (exp
, 0);
18829 arg1
= CALL_EXPR_ARG (exp
, 1);
18830 op0
= expand_normal (arg0
);
18831 op1
= expand_normal (arg1
);
18832 tmode
= insn_data
[icode
].operand
[0].mode
;
18833 mode1
= insn_data
[icode
].operand
[1].mode
;
18834 mode2
= insn_data
[icode
].operand
[2].mode
;
18836 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
18838 op0
= copy_to_reg (op0
);
18839 op0
= simplify_gen_subreg (mode1
, op0
, GET_MODE (op0
), 0);
18841 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
18843 error ("shift must be an immediate");
18846 target
= gen_reg_rtx (V2DImode
);
18847 pat
= GEN_FCN (icode
) (simplify_gen_subreg (tmode
, target
, V2DImode
, 0),
18854 case IX86_BUILTIN_FEMMS
:
18855 emit_insn (gen_mmx_femms ());
18858 case IX86_BUILTIN_PAVGUSB
:
18859 return ix86_expand_binop_builtin (CODE_FOR_mmx_uavgv8qi3
, exp
, target
);
18861 case IX86_BUILTIN_PF2ID
:
18862 return ix86_expand_unop_builtin (CODE_FOR_mmx_pf2id
, exp
, target
, 0);
18864 case IX86_BUILTIN_PFACC
:
18865 return ix86_expand_binop_builtin (CODE_FOR_mmx_haddv2sf3
, exp
, target
);
18867 case IX86_BUILTIN_PFADD
:
18868 return ix86_expand_binop_builtin (CODE_FOR_mmx_addv2sf3
, exp
, target
);
18870 case IX86_BUILTIN_PFCMPEQ
:
18871 return ix86_expand_binop_builtin (CODE_FOR_mmx_eqv2sf3
, exp
, target
);
18873 case IX86_BUILTIN_PFCMPGE
:
18874 return ix86_expand_binop_builtin (CODE_FOR_mmx_gev2sf3
, exp
, target
);
18876 case IX86_BUILTIN_PFCMPGT
:
18877 return ix86_expand_binop_builtin (CODE_FOR_mmx_gtv2sf3
, exp
, target
);
18879 case IX86_BUILTIN_PFMAX
:
18880 return ix86_expand_binop_builtin (CODE_FOR_mmx_smaxv2sf3
, exp
, target
);
18882 case IX86_BUILTIN_PFMIN
:
18883 return ix86_expand_binop_builtin (CODE_FOR_mmx_sminv2sf3
, exp
, target
);
18885 case IX86_BUILTIN_PFMUL
:
18886 return ix86_expand_binop_builtin (CODE_FOR_mmx_mulv2sf3
, exp
, target
);
18888 case IX86_BUILTIN_PFRCP
:
18889 return ix86_expand_unop_builtin (CODE_FOR_mmx_rcpv2sf2
, exp
, target
, 0);
18891 case IX86_BUILTIN_PFRCPIT1
:
18892 return ix86_expand_binop_builtin (CODE_FOR_mmx_rcpit1v2sf3
, exp
, target
);
18894 case IX86_BUILTIN_PFRCPIT2
:
18895 return ix86_expand_binop_builtin (CODE_FOR_mmx_rcpit2v2sf3
, exp
, target
);
18897 case IX86_BUILTIN_PFRSQIT1
:
18898 return ix86_expand_binop_builtin (CODE_FOR_mmx_rsqit1v2sf3
, exp
, target
);
18900 case IX86_BUILTIN_PFRSQRT
:
18901 return ix86_expand_unop_builtin (CODE_FOR_mmx_rsqrtv2sf2
, exp
, target
, 0);
18903 case IX86_BUILTIN_PFSUB
:
18904 return ix86_expand_binop_builtin (CODE_FOR_mmx_subv2sf3
, exp
, target
);
18906 case IX86_BUILTIN_PFSUBR
:
18907 return ix86_expand_binop_builtin (CODE_FOR_mmx_subrv2sf3
, exp
, target
);
18909 case IX86_BUILTIN_PI2FD
:
18910 return ix86_expand_unop_builtin (CODE_FOR_mmx_floatv2si2
, exp
, target
, 0);
18912 case IX86_BUILTIN_PMULHRW
:
18913 return ix86_expand_binop_builtin (CODE_FOR_mmx_pmulhrwv4hi3
, exp
, target
);
18915 case IX86_BUILTIN_PF2IW
:
18916 return ix86_expand_unop_builtin (CODE_FOR_mmx_pf2iw
, exp
, target
, 0);
18918 case IX86_BUILTIN_PFNACC
:
18919 return ix86_expand_binop_builtin (CODE_FOR_mmx_hsubv2sf3
, exp
, target
);
18921 case IX86_BUILTIN_PFPNACC
:
18922 return ix86_expand_binop_builtin (CODE_FOR_mmx_addsubv2sf3
, exp
, target
);
18924 case IX86_BUILTIN_PI2FW
:
18925 return ix86_expand_unop_builtin (CODE_FOR_mmx_pi2fw
, exp
, target
, 0);
18927 case IX86_BUILTIN_PSWAPDSI
:
18928 return ix86_expand_unop_builtin (CODE_FOR_mmx_pswapdv2si2
, exp
, target
, 0);
18930 case IX86_BUILTIN_PSWAPDSF
:
18931 return ix86_expand_unop_builtin (CODE_FOR_mmx_pswapdv2sf2
, exp
, target
, 0);
18933 case IX86_BUILTIN_SQRTSD
:
18934 return ix86_expand_unop1_builtin (CODE_FOR_sse2_vmsqrtv2df2
, exp
, target
);
18935 case IX86_BUILTIN_LOADUPD
:
18936 return ix86_expand_unop_builtin (CODE_FOR_sse2_movupd
, exp
, target
, 1);
18937 case IX86_BUILTIN_STOREUPD
:
18938 return ix86_expand_store_builtin (CODE_FOR_sse2_movupd
, exp
);
18940 case IX86_BUILTIN_MFENCE
:
18941 emit_insn (gen_sse2_mfence ());
18943 case IX86_BUILTIN_LFENCE
:
18944 emit_insn (gen_sse2_lfence ());
18947 case IX86_BUILTIN_CLFLUSH
:
18948 arg0
= CALL_EXPR_ARG (exp
, 0);
18949 op0
= expand_normal (arg0
);
18950 icode
= CODE_FOR_sse2_clflush
;
18951 if (! (*insn_data
[icode
].operand
[0].predicate
) (op0
, Pmode
))
18952 op0
= copy_to_mode_reg (Pmode
, op0
);
18954 emit_insn (gen_sse2_clflush (op0
));
18957 case IX86_BUILTIN_MOVNTPD
:
18958 return ix86_expand_store_builtin (CODE_FOR_sse2_movntv2df
, exp
);
18959 case IX86_BUILTIN_MOVNTDQ
:
18960 return ix86_expand_store_builtin (CODE_FOR_sse2_movntv2di
, exp
);
18961 case IX86_BUILTIN_MOVNTI
:
18962 return ix86_expand_store_builtin (CODE_FOR_sse2_movntsi
, exp
);
18964 case IX86_BUILTIN_LOADDQU
:
18965 return ix86_expand_unop_builtin (CODE_FOR_sse2_movdqu
, exp
, target
, 1);
18966 case IX86_BUILTIN_STOREDQU
:
18967 return ix86_expand_store_builtin (CODE_FOR_sse2_movdqu
, exp
);
18969 case IX86_BUILTIN_MONITOR
:
18970 arg0
= CALL_EXPR_ARG (exp
, 0);
18971 arg1
= CALL_EXPR_ARG (exp
, 1);
18972 arg2
= CALL_EXPR_ARG (exp
, 2);
18973 op0
= expand_normal (arg0
);
18974 op1
= expand_normal (arg1
);
18975 op2
= expand_normal (arg2
);
18977 op0
= copy_to_mode_reg (Pmode
, op0
);
18979 op1
= copy_to_mode_reg (SImode
, op1
);
18981 op2
= copy_to_mode_reg (SImode
, op2
);
18983 emit_insn (gen_sse3_monitor (op0
, op1
, op2
));
18985 emit_insn (gen_sse3_monitor64 (op0
, op1
, op2
));
18988 case IX86_BUILTIN_MWAIT
:
18989 arg0
= CALL_EXPR_ARG (exp
, 0);
18990 arg1
= CALL_EXPR_ARG (exp
, 1);
18991 op0
= expand_normal (arg0
);
18992 op1
= expand_normal (arg1
);
18994 op0
= copy_to_mode_reg (SImode
, op0
);
18996 op1
= copy_to_mode_reg (SImode
, op1
);
18997 emit_insn (gen_sse3_mwait (op0
, op1
));
19000 case IX86_BUILTIN_LDDQU
:
19001 return ix86_expand_unop_builtin (CODE_FOR_sse3_lddqu
, exp
,
19004 case IX86_BUILTIN_PALIGNR
:
19005 case IX86_BUILTIN_PALIGNR128
:
19006 if (fcode
== IX86_BUILTIN_PALIGNR
)
19008 icode
= CODE_FOR_ssse3_palignrdi
;
19013 icode
= CODE_FOR_ssse3_palignrti
;
19016 arg0
= CALL_EXPR_ARG (exp
, 0);
19017 arg1
= CALL_EXPR_ARG (exp
, 1);
19018 arg2
= CALL_EXPR_ARG (exp
, 2);
19019 op0
= expand_expr (arg0
, NULL_RTX
, VOIDmode
, 0);
19020 op1
= expand_expr (arg1
, NULL_RTX
, VOIDmode
, 0);
19021 op2
= expand_expr (arg2
, NULL_RTX
, VOIDmode
, 0);
19022 tmode
= insn_data
[icode
].operand
[0].mode
;
19023 mode1
= insn_data
[icode
].operand
[1].mode
;
19024 mode2
= insn_data
[icode
].operand
[2].mode
;
19025 mode3
= insn_data
[icode
].operand
[3].mode
;
19027 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
19029 op0
= copy_to_reg (op0
);
19030 op0
= simplify_gen_subreg (mode1
, op0
, GET_MODE (op0
), 0);
19032 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
19034 op1
= copy_to_reg (op1
);
19035 op1
= simplify_gen_subreg (mode2
, op1
, GET_MODE (op1
), 0);
19037 if (! (*insn_data
[icode
].operand
[3].predicate
) (op2
, mode3
))
19039 error ("shift must be an immediate");
19042 target
= gen_reg_rtx (mode
);
19043 pat
= GEN_FCN (icode
) (simplify_gen_subreg (tmode
, target
, mode
, 0),
19050 case IX86_BUILTIN_MOVNTDQA
:
19051 return ix86_expand_unop_builtin (CODE_FOR_sse4_1_movntdqa
, exp
,
19054 case IX86_BUILTIN_MOVNTSD
:
19055 return ix86_expand_store_builtin (CODE_FOR_sse4a_vmmovntv2df
, exp
);
19057 case IX86_BUILTIN_MOVNTSS
:
19058 return ix86_expand_store_builtin (CODE_FOR_sse4a_vmmovntv4sf
, exp
);
19060 case IX86_BUILTIN_INSERTQ
:
19061 case IX86_BUILTIN_EXTRQ
:
19062 icode
= (fcode
== IX86_BUILTIN_EXTRQ
19063 ? CODE_FOR_sse4a_extrq
19064 : CODE_FOR_sse4a_insertq
);
19065 arg0
= CALL_EXPR_ARG (exp
, 0);
19066 arg1
= CALL_EXPR_ARG (exp
, 1);
19067 op0
= expand_normal (arg0
);
19068 op1
= expand_normal (arg1
);
19069 tmode
= insn_data
[icode
].operand
[0].mode
;
19070 mode1
= insn_data
[icode
].operand
[1].mode
;
19071 mode2
= insn_data
[icode
].operand
[2].mode
;
19072 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
19073 op0
= copy_to_mode_reg (mode1
, op0
);
19074 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
19075 op1
= copy_to_mode_reg (mode2
, op1
);
19076 if (optimize
|| target
== 0
19077 || GET_MODE (target
) != tmode
19078 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
19079 target
= gen_reg_rtx (tmode
);
19080 pat
= GEN_FCN (icode
) (target
, op0
, op1
);
19086 case IX86_BUILTIN_EXTRQI
:
19087 icode
= CODE_FOR_sse4a_extrqi
;
19088 arg0
= CALL_EXPR_ARG (exp
, 0);
19089 arg1
= CALL_EXPR_ARG (exp
, 1);
19090 arg2
= CALL_EXPR_ARG (exp
, 2);
19091 op0
= expand_normal (arg0
);
19092 op1
= expand_normal (arg1
);
19093 op2
= expand_normal (arg2
);
19094 tmode
= insn_data
[icode
].operand
[0].mode
;
19095 mode1
= insn_data
[icode
].operand
[1].mode
;
19096 mode2
= insn_data
[icode
].operand
[2].mode
;
19097 mode3
= insn_data
[icode
].operand
[3].mode
;
19098 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
19099 op0
= copy_to_mode_reg (mode1
, op0
);
19100 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
19102 error ("index mask must be an immediate");
19103 return gen_reg_rtx (tmode
);
19105 if (! (*insn_data
[icode
].operand
[3].predicate
) (op2
, mode3
))
19107 error ("length mask must be an immediate");
19108 return gen_reg_rtx (tmode
);
19110 if (optimize
|| target
== 0
19111 || GET_MODE (target
) != tmode
19112 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
19113 target
= gen_reg_rtx (tmode
);
19114 pat
= GEN_FCN (icode
) (target
, op0
, op1
, op2
);
19120 case IX86_BUILTIN_INSERTQI
:
19121 icode
= CODE_FOR_sse4a_insertqi
;
19122 arg0
= CALL_EXPR_ARG (exp
, 0);
19123 arg1
= CALL_EXPR_ARG (exp
, 1);
19124 arg2
= CALL_EXPR_ARG (exp
, 2);
19125 arg3
= CALL_EXPR_ARG (exp
, 3);
19126 op0
= expand_normal (arg0
);
19127 op1
= expand_normal (arg1
);
19128 op2
= expand_normal (arg2
);
19129 op3
= expand_normal (arg3
);
19130 tmode
= insn_data
[icode
].operand
[0].mode
;
19131 mode1
= insn_data
[icode
].operand
[1].mode
;
19132 mode2
= insn_data
[icode
].operand
[2].mode
;
19133 mode3
= insn_data
[icode
].operand
[3].mode
;
19134 mode4
= insn_data
[icode
].operand
[4].mode
;
19136 if (! (*insn_data
[icode
].operand
[1].predicate
) (op0
, mode1
))
19137 op0
= copy_to_mode_reg (mode1
, op0
);
19139 if (! (*insn_data
[icode
].operand
[2].predicate
) (op1
, mode2
))
19140 op1
= copy_to_mode_reg (mode2
, op1
);
19142 if (! (*insn_data
[icode
].operand
[3].predicate
) (op2
, mode3
))
19144 error ("index mask must be an immediate");
19145 return gen_reg_rtx (tmode
);
19147 if (! (*insn_data
[icode
].operand
[4].predicate
) (op3
, mode4
))
19149 error ("length mask must be an immediate");
19150 return gen_reg_rtx (tmode
);
19152 if (optimize
|| target
== 0
19153 || GET_MODE (target
) != tmode
19154 || ! (*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
19155 target
= gen_reg_rtx (tmode
);
19156 pat
= GEN_FCN (icode
) (target
, op0
, op1
, op2
, op3
);
19162 case IX86_BUILTIN_VEC_INIT_V2SI
:
19163 case IX86_BUILTIN_VEC_INIT_V4HI
:
19164 case IX86_BUILTIN_VEC_INIT_V8QI
:
19165 return ix86_expand_vec_init_builtin (TREE_TYPE (exp
), exp
, target
);
19167 case IX86_BUILTIN_VEC_EXT_V2DF
:
19168 case IX86_BUILTIN_VEC_EXT_V2DI
:
19169 case IX86_BUILTIN_VEC_EXT_V4SF
:
19170 case IX86_BUILTIN_VEC_EXT_V4SI
:
19171 case IX86_BUILTIN_VEC_EXT_V8HI
:
19172 case IX86_BUILTIN_VEC_EXT_V2SI
:
19173 case IX86_BUILTIN_VEC_EXT_V4HI
:
19174 case IX86_BUILTIN_VEC_EXT_V16QI
:
19175 return ix86_expand_vec_ext_builtin (exp
, target
);
19177 case IX86_BUILTIN_VEC_SET_V2DI
:
19178 case IX86_BUILTIN_VEC_SET_V4SF
:
19179 case IX86_BUILTIN_VEC_SET_V4SI
:
19180 case IX86_BUILTIN_VEC_SET_V8HI
:
19181 case IX86_BUILTIN_VEC_SET_V4HI
:
19182 case IX86_BUILTIN_VEC_SET_V16QI
:
19183 return ix86_expand_vec_set_builtin (exp
);
19189 for (i
= 0, d
= bdesc_sse_3arg
;
19190 i
< ARRAY_SIZE (bdesc_sse_3arg
);
19192 if (d
->code
== fcode
)
19193 return ix86_expand_sse_4_operands_builtin (d
->icode
, exp
,
19196 for (i
= 0, d
= bdesc_2arg
; i
< ARRAY_SIZE (bdesc_2arg
); i
++, d
++)
19197 if (d
->code
== fcode
)
19199 /* Compares are treated specially. */
19200 if (d
->icode
== CODE_FOR_sse_maskcmpv4sf3
19201 || d
->icode
== CODE_FOR_sse_vmmaskcmpv4sf3
19202 || d
->icode
== CODE_FOR_sse2_maskcmpv2df3
19203 || d
->icode
== CODE_FOR_sse2_vmmaskcmpv2df3
)
19204 return ix86_expand_sse_compare (d
, exp
, target
);
19206 return ix86_expand_binop_builtin (d
->icode
, exp
, target
);
19209 for (i
= 0, d
= bdesc_1arg
; i
< ARRAY_SIZE (bdesc_1arg
); i
++, d
++)
19210 if (d
->code
== fcode
)
19211 return ix86_expand_unop_builtin (d
->icode
, exp
, target
, 0);
19213 for (i
= 0, d
= bdesc_comi
; i
< ARRAY_SIZE (bdesc_comi
); i
++, d
++)
19214 if (d
->code
== fcode
)
19215 return ix86_expand_sse_comi (d
, exp
, target
);
19217 for (i
= 0, d
= bdesc_ptest
; i
< ARRAY_SIZE (bdesc_ptest
); i
++, d
++)
19218 if (d
->code
== fcode
)
19219 return ix86_expand_sse_ptest (d
, exp
, target
);
19221 gcc_unreachable ();
19224 /* Returns a function decl for a vectorized version of the builtin function
19225 with builtin function code FN and the result vector type TYPE, or NULL_TREE
19226 if it is not available. */
19229 ix86_builtin_vectorized_function (enum built_in_function fn
, tree type_out
,
19232 enum machine_mode in_mode
, out_mode
;
19235 if (TREE_CODE (type_out
) != VECTOR_TYPE
19236 || TREE_CODE (type_in
) != VECTOR_TYPE
)
19239 out_mode
= TYPE_MODE (TREE_TYPE (type_out
));
19240 out_n
= TYPE_VECTOR_SUBPARTS (type_out
);
19241 in_mode
= TYPE_MODE (TREE_TYPE (type_in
));
19242 in_n
= TYPE_VECTOR_SUBPARTS (type_in
);
19246 case BUILT_IN_SQRT
:
19247 if (out_mode
== DFmode
&& out_n
== 2
19248 && in_mode
== DFmode
&& in_n
== 2)
19249 return ix86_builtins
[IX86_BUILTIN_SQRTPD
];
19252 case BUILT_IN_SQRTF
:
19253 if (out_mode
== SFmode
&& out_n
== 4
19254 && in_mode
== SFmode
&& in_n
== 4)
19255 return ix86_builtins
[IX86_BUILTIN_SQRTPS
];
19258 case BUILT_IN_LRINTF
:
19259 if (out_mode
== SImode
&& out_n
== 4
19260 && in_mode
== SFmode
&& in_n
== 4)
19261 return ix86_builtins
[IX86_BUILTIN_CVTPS2DQ
];
19271 /* Returns a decl of a function that implements conversion of the
19272 input vector of type TYPE, or NULL_TREE if it is not available. */
19275 ix86_builtin_conversion (enum tree_code code
, tree type
)
19277 if (TREE_CODE (type
) != VECTOR_TYPE
)
19283 switch (TYPE_MODE (type
))
19286 return ix86_builtins
[IX86_BUILTIN_CVTDQ2PS
];
19291 case FIX_TRUNC_EXPR
:
19292 switch (TYPE_MODE (type
))
19295 return ix86_builtins
[IX86_BUILTIN_CVTTPS2DQ
];
19305 /* Store OPERAND to the memory after reload is completed. This means
19306 that we can't easily use assign_stack_local. */
19308 ix86_force_to_memory (enum machine_mode mode
, rtx operand
)
19312 gcc_assert (reload_completed
);
19313 if (TARGET_RED_ZONE
)
19315 result
= gen_rtx_MEM (mode
,
19316 gen_rtx_PLUS (Pmode
,
19318 GEN_INT (-RED_ZONE_SIZE
)));
19319 emit_move_insn (result
, operand
);
19321 else if (!TARGET_RED_ZONE
&& TARGET_64BIT
)
19327 operand
= gen_lowpart (DImode
, operand
);
19331 gen_rtx_SET (VOIDmode
,
19332 gen_rtx_MEM (DImode
,
19333 gen_rtx_PRE_DEC (DImode
,
19334 stack_pointer_rtx
)),
19338 gcc_unreachable ();
19340 result
= gen_rtx_MEM (mode
, stack_pointer_rtx
);
19349 split_di (&operand
, 1, operands
, operands
+ 1);
19351 gen_rtx_SET (VOIDmode
,
19352 gen_rtx_MEM (SImode
,
19353 gen_rtx_PRE_DEC (Pmode
,
19354 stack_pointer_rtx
)),
19357 gen_rtx_SET (VOIDmode
,
19358 gen_rtx_MEM (SImode
,
19359 gen_rtx_PRE_DEC (Pmode
,
19360 stack_pointer_rtx
)),
19365 /* Store HImodes as SImodes. */
19366 operand
= gen_lowpart (SImode
, operand
);
19370 gen_rtx_SET (VOIDmode
,
19371 gen_rtx_MEM (GET_MODE (operand
),
19372 gen_rtx_PRE_DEC (SImode
,
19373 stack_pointer_rtx
)),
19377 gcc_unreachable ();
19379 result
= gen_rtx_MEM (mode
, stack_pointer_rtx
);
19384 /* Free operand from the memory. */
19386 ix86_free_from_memory (enum machine_mode mode
)
19388 if (!TARGET_RED_ZONE
)
19392 if (mode
== DImode
|| TARGET_64BIT
)
19396 /* Use LEA to deallocate stack space. In peephole2 it will be converted
19397 to pop or add instruction if registers are available. */
19398 emit_insn (gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
19399 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
19404 /* Put float CONST_DOUBLE in the constant pool instead of fp regs.
19405 QImode must go into class Q_REGS.
19406 Narrow ALL_REGS to GENERAL_REGS. This supports allowing movsf and
19407 movdf to do mem-to-mem moves through integer regs. */
19409 ix86_preferred_reload_class (rtx x
, enum reg_class
class)
19411 enum machine_mode mode
= GET_MODE (x
);
19413 /* We're only allowed to return a subclass of CLASS. Many of the
19414 following checks fail for NO_REGS, so eliminate that early. */
19415 if (class == NO_REGS
)
19418 /* All classes can load zeros. */
19419 if (x
== CONST0_RTX (mode
))
19422 /* Force constants into memory if we are loading a (nonzero) constant into
19423 an MMX or SSE register. This is because there are no MMX/SSE instructions
19424 to load from a constant. */
19426 && (MAYBE_MMX_CLASS_P (class) || MAYBE_SSE_CLASS_P (class)))
19429 /* Prefer SSE regs only, if we can use them for math. */
19430 if (TARGET_SSE_MATH
&& !TARGET_MIX_SSE_I387
&& SSE_FLOAT_MODE_P (mode
))
19431 return SSE_CLASS_P (class) ? class : NO_REGS
;
19433 /* Floating-point constants need more complex checks. */
19434 if (GET_CODE (x
) == CONST_DOUBLE
&& GET_MODE (x
) != VOIDmode
)
19436 /* General regs can load everything. */
19437 if (reg_class_subset_p (class, GENERAL_REGS
))
19440 /* Floats can load 0 and 1 plus some others. Note that we eliminated
19441 zero above. We only want to wind up preferring 80387 registers if
19442 we plan on doing computation with them. */
19444 && standard_80387_constant_p (x
))
19446 /* Limit class to non-sse. */
19447 if (class == FLOAT_SSE_REGS
)
19449 if (class == FP_TOP_SSE_REGS
)
19451 if (class == FP_SECOND_SSE_REGS
)
19452 return FP_SECOND_REG
;
19453 if (class == FLOAT_INT_REGS
|| class == FLOAT_REGS
)
19460 /* Generally when we see PLUS here, it's the function invariant
19461 (plus soft-fp const_int). Which can only be computed into general
19463 if (GET_CODE (x
) == PLUS
)
19464 return reg_class_subset_p (class, GENERAL_REGS
) ? class : NO_REGS
;
19466 /* QImode constants are easy to load, but non-constant QImode data
19467 must go into Q_REGS. */
19468 if (GET_MODE (x
) == QImode
&& !CONSTANT_P (x
))
19470 if (reg_class_subset_p (class, Q_REGS
))
19472 if (reg_class_subset_p (Q_REGS
, class))
19480 /* Discourage putting floating-point values in SSE registers unless
19481 SSE math is being used, and likewise for the 387 registers. */
19483 ix86_preferred_output_reload_class (rtx x
, enum reg_class
class)
19485 enum machine_mode mode
= GET_MODE (x
);
19487 /* Restrict the output reload class to the register bank that we are doing
19488 math on. If we would like not to return a subset of CLASS, reject this
19489 alternative: if reload cannot do this, it will still use its choice. */
19490 mode
= GET_MODE (x
);
19491 if (TARGET_SSE_MATH
&& SSE_FLOAT_MODE_P (mode
))
19492 return MAYBE_SSE_CLASS_P (class) ? SSE_REGS
: NO_REGS
;
19494 if (X87_FLOAT_MODE_P (mode
))
19496 if (class == FP_TOP_SSE_REGS
)
19498 else if (class == FP_SECOND_SSE_REGS
)
19499 return FP_SECOND_REG
;
19501 return FLOAT_CLASS_P (class) ? class : NO_REGS
;
19507 /* If we are copying between general and FP registers, we need a memory
19508 location. The same is true for SSE and MMX registers.
19510 The macro can't work reliably when one of the CLASSES is class containing
19511 registers from multiple units (SSE, MMX, integer). We avoid this by never
19512 combining those units in single alternative in the machine description.
19513 Ensure that this constraint holds to avoid unexpected surprises.
19515 When STRICT is false, we are being called from REGISTER_MOVE_COST, so do not
19516 enforce these sanity checks. */
19519 ix86_secondary_memory_needed (enum reg_class class1
, enum reg_class class2
,
19520 enum machine_mode mode
, int strict
)
19522 if (MAYBE_FLOAT_CLASS_P (class1
) != FLOAT_CLASS_P (class1
)
19523 || MAYBE_FLOAT_CLASS_P (class2
) != FLOAT_CLASS_P (class2
)
19524 || MAYBE_SSE_CLASS_P (class1
) != SSE_CLASS_P (class1
)
19525 || MAYBE_SSE_CLASS_P (class2
) != SSE_CLASS_P (class2
)
19526 || MAYBE_MMX_CLASS_P (class1
) != MMX_CLASS_P (class1
)
19527 || MAYBE_MMX_CLASS_P (class2
) != MMX_CLASS_P (class2
))
19529 gcc_assert (!strict
);
19533 if (FLOAT_CLASS_P (class1
) != FLOAT_CLASS_P (class2
))
19536 /* ??? This is a lie. We do have moves between mmx/general, and for
19537 mmx/sse2. But by saying we need secondary memory we discourage the
19538 register allocator from using the mmx registers unless needed. */
19539 if (MMX_CLASS_P (class1
) != MMX_CLASS_P (class2
))
19542 if (SSE_CLASS_P (class1
) != SSE_CLASS_P (class2
))
19544 /* SSE1 doesn't have any direct moves from other classes. */
19548 /* If the target says that inter-unit moves are more expensive
19549 than moving through memory, then don't generate them. */
19550 if (!TARGET_INTER_UNIT_MOVES
)
19553 /* Between SSE and general, we have moves no larger than word size. */
19554 if (GET_MODE_SIZE (mode
) > UNITS_PER_WORD
)
19561 /* Return true if the registers in CLASS cannot represent the change from
19562 modes FROM to TO. */
19565 ix86_cannot_change_mode_class (enum machine_mode from
, enum machine_mode to
,
19566 enum reg_class
class)
19571 /* x87 registers can't do subreg at all, as all values are reformatted
19572 to extended precision. */
19573 if (MAYBE_FLOAT_CLASS_P (class))
19576 if (MAYBE_SSE_CLASS_P (class) || MAYBE_MMX_CLASS_P (class))
19578 /* Vector registers do not support QI or HImode loads. If we don't
19579 disallow a change to these modes, reload will assume it's ok to
19580 drop the subreg from (subreg:SI (reg:HI 100) 0). This affects
19581 the vec_dupv4hi pattern. */
19582 if (GET_MODE_SIZE (from
) < 4)
19585 /* Vector registers do not support subreg with nonzero offsets, which
19586 are otherwise valid for integer registers. Since we can't see
19587 whether we have a nonzero offset from here, prohibit all
19588 nonparadoxical subregs changing size. */
19589 if (GET_MODE_SIZE (to
) < GET_MODE_SIZE (from
))
19596 /* Return the cost of moving data from a register in class CLASS1 to
19597 one in class CLASS2.
19599 It is not required that the cost always equal 2 when FROM is the same as TO;
19600 on some machines it is expensive to move between registers if they are not
19601 general registers. */
19604 ix86_register_move_cost (enum machine_mode mode
, enum reg_class class1
,
19605 enum reg_class class2
)
19607 /* In case we require secondary memory, compute cost of the store followed
19608 by load. In order to avoid bad register allocation choices, we need
19609 for this to be *at least* as high as the symmetric MEMORY_MOVE_COST. */
19611 if (ix86_secondary_memory_needed (class1
, class2
, mode
, 0))
19615 cost
+= MAX (MEMORY_MOVE_COST (mode
, class1
, 0),
19616 MEMORY_MOVE_COST (mode
, class1
, 1));
19617 cost
+= MAX (MEMORY_MOVE_COST (mode
, class2
, 0),
19618 MEMORY_MOVE_COST (mode
, class2
, 1));
19620 /* In case of copying from general_purpose_register we may emit multiple
19621 stores followed by single load causing memory size mismatch stall.
19622 Count this as arbitrarily high cost of 20. */
19623 if (CLASS_MAX_NREGS (class1
, mode
) > CLASS_MAX_NREGS (class2
, mode
))
19626 /* In the case of FP/MMX moves, the registers actually overlap, and we
19627 have to switch modes in order to treat them differently. */
19628 if ((MMX_CLASS_P (class1
) && MAYBE_FLOAT_CLASS_P (class2
))
19629 || (MMX_CLASS_P (class2
) && MAYBE_FLOAT_CLASS_P (class1
)))
19635 /* Moves between SSE/MMX and integer unit are expensive. */
19636 if (MMX_CLASS_P (class1
) != MMX_CLASS_P (class2
)
19637 || SSE_CLASS_P (class1
) != SSE_CLASS_P (class2
))
19638 return ix86_cost
->mmxsse_to_integer
;
19639 if (MAYBE_FLOAT_CLASS_P (class1
))
19640 return ix86_cost
->fp_move
;
19641 if (MAYBE_SSE_CLASS_P (class1
))
19642 return ix86_cost
->sse_move
;
19643 if (MAYBE_MMX_CLASS_P (class1
))
19644 return ix86_cost
->mmx_move
;
19648 /* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
19651 ix86_hard_regno_mode_ok (int regno
, enum machine_mode mode
)
19653 /* Flags and only flags can only hold CCmode values. */
19654 if (CC_REGNO_P (regno
))
19655 return GET_MODE_CLASS (mode
) == MODE_CC
;
19656 if (GET_MODE_CLASS (mode
) == MODE_CC
19657 || GET_MODE_CLASS (mode
) == MODE_RANDOM
19658 || GET_MODE_CLASS (mode
) == MODE_PARTIAL_INT
)
19660 if (FP_REGNO_P (regno
))
19661 return VALID_FP_MODE_P (mode
);
19662 if (SSE_REGNO_P (regno
))
19664 /* We implement the move patterns for all vector modes into and
19665 out of SSE registers, even when no operation instructions
19667 return (VALID_SSE_REG_MODE (mode
)
19668 || VALID_SSE2_REG_MODE (mode
)
19669 || VALID_MMX_REG_MODE (mode
)
19670 || VALID_MMX_REG_MODE_3DNOW (mode
));
19672 if (MMX_REGNO_P (regno
))
19674 /* We implement the move patterns for 3DNOW modes even in MMX mode,
19675 so if the register is available at all, then we can move data of
19676 the given mode into or out of it. */
19677 return (VALID_MMX_REG_MODE (mode
)
19678 || VALID_MMX_REG_MODE_3DNOW (mode
));
19681 if (mode
== QImode
)
19683 /* Take care for QImode values - they can be in non-QI regs,
19684 but then they do cause partial register stalls. */
19685 if (regno
< 4 || TARGET_64BIT
)
19687 if (!TARGET_PARTIAL_REG_STALL
)
19689 return reload_in_progress
|| reload_completed
;
19691 /* We handle both integer and floats in the general purpose registers. */
19692 else if (VALID_INT_MODE_P (mode
))
19694 else if (VALID_FP_MODE_P (mode
))
19696 /* Lots of MMX code casts 8 byte vector modes to DImode. If we then go
19697 on to use that value in smaller contexts, this can easily force a
19698 pseudo to be allocated to GENERAL_REGS. Since this is no worse than
19699 supporting DImode, allow it. */
19700 else if (VALID_MMX_REG_MODE_3DNOW (mode
) || VALID_MMX_REG_MODE (mode
))
19706 /* A subroutine of ix86_modes_tieable_p. Return true if MODE is a
19707 tieable integer mode. */
19710 ix86_tieable_integer_mode_p (enum machine_mode mode
)
19719 return TARGET_64BIT
|| !TARGET_PARTIAL_REG_STALL
;
19722 return TARGET_64BIT
;
19729 /* Return true if MODE1 is accessible in a register that can hold MODE2
19730 without copying. That is, all register classes that can hold MODE2
19731 can also hold MODE1. */
19734 ix86_modes_tieable_p (enum machine_mode mode1
, enum machine_mode mode2
)
19736 if (mode1
== mode2
)
19739 if (ix86_tieable_integer_mode_p (mode1
)
19740 && ix86_tieable_integer_mode_p (mode2
))
19743 /* MODE2 being XFmode implies fp stack or general regs, which means we
19744 can tie any smaller floating point modes to it. Note that we do not
19745 tie this with TFmode. */
19746 if (mode2
== XFmode
)
19747 return mode1
== SFmode
|| mode1
== DFmode
;
19749 /* MODE2 being DFmode implies fp stack, general or sse regs, which means
19750 that we can tie it with SFmode. */
19751 if (mode2
== DFmode
)
19752 return mode1
== SFmode
;
19754 /* If MODE2 is only appropriate for an SSE register, then tie with
19755 any other mode acceptable to SSE registers. */
19756 if (GET_MODE_SIZE (mode2
) == 16
19757 && ix86_hard_regno_mode_ok (FIRST_SSE_REG
, mode2
))
19758 return (GET_MODE_SIZE (mode1
) == 16
19759 && ix86_hard_regno_mode_ok (FIRST_SSE_REG
, mode1
));
19761 /* If MODE2 is appropriate for an MMX register, then tie
19762 with any other mode acceptable to MMX registers. */
19763 if (GET_MODE_SIZE (mode2
) == 8
19764 && ix86_hard_regno_mode_ok (FIRST_MMX_REG
, mode2
))
19765 return (GET_MODE_SIZE (mode1
) == 8
19766 && ix86_hard_regno_mode_ok (FIRST_MMX_REG
, mode1
));
19771 /* Return the cost of moving data of mode M between a
19772 register and memory. A value of 2 is the default; this cost is
19773 relative to those in `REGISTER_MOVE_COST'.
19775 If moving between registers and memory is more expensive than
19776 between two registers, you should define this macro to express the
19779 Model also increased moving costs of QImode registers in non
19783 ix86_memory_move_cost (enum machine_mode mode
, enum reg_class
class, int in
)
19785 if (FLOAT_CLASS_P (class))
19802 return in
? ix86_cost
->fp_load
[index
] : ix86_cost
->fp_store
[index
];
19804 if (SSE_CLASS_P (class))
19807 switch (GET_MODE_SIZE (mode
))
19821 return in
? ix86_cost
->sse_load
[index
] : ix86_cost
->sse_store
[index
];
19823 if (MMX_CLASS_P (class))
19826 switch (GET_MODE_SIZE (mode
))
19837 return in
? ix86_cost
->mmx_load
[index
] : ix86_cost
->mmx_store
[index
];
19839 switch (GET_MODE_SIZE (mode
))
19843 return (Q_CLASS_P (class) ? ix86_cost
->int_load
[0]
19844 : ix86_cost
->movzbl_load
);
19846 return (Q_CLASS_P (class) ? ix86_cost
->int_store
[0]
19847 : ix86_cost
->int_store
[0] + 4);
19850 return in
? ix86_cost
->int_load
[1] : ix86_cost
->int_store
[1];
19852 /* Compute number of 32bit moves needed. TFmode is moved as XFmode. */
19853 if (mode
== TFmode
)
19855 return ((in
? ix86_cost
->int_load
[2] : ix86_cost
->int_store
[2])
19856 * (((int) GET_MODE_SIZE (mode
)
19857 + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
));
19861 /* Compute a (partial) cost for rtx X. Return true if the complete
19862 cost has been computed, and false if subexpressions should be
19863 scanned. In either case, *TOTAL contains the cost result. */
19866 ix86_rtx_costs (rtx x
, int code
, int outer_code
, int *total
)
19868 enum machine_mode mode
= GET_MODE (x
);
19876 if (TARGET_64BIT
&& !x86_64_immediate_operand (x
, VOIDmode
))
19878 else if (TARGET_64BIT
&& !x86_64_zext_immediate_operand (x
, VOIDmode
))
19880 else if (flag_pic
&& SYMBOLIC_CONST (x
)
19882 || (!GET_CODE (x
) != LABEL_REF
19883 && (GET_CODE (x
) != SYMBOL_REF
19884 || !SYMBOL_REF_LOCAL_P (x
)))))
19891 if (mode
== VOIDmode
)
19894 switch (standard_80387_constant_p (x
))
19899 default: /* Other constants */
19904 /* Start with (MEM (SYMBOL_REF)), since that's where
19905 it'll probably end up. Add a penalty for size. */
19906 *total
= (COSTS_N_INSNS (1)
19907 + (flag_pic
!= 0 && !TARGET_64BIT
)
19908 + (mode
== SFmode
? 0 : mode
== DFmode
? 1 : 2));
19914 /* The zero extensions is often completely free on x86_64, so make
19915 it as cheap as possible. */
19916 if (TARGET_64BIT
&& mode
== DImode
19917 && GET_MODE (XEXP (x
, 0)) == SImode
)
19919 else if (TARGET_ZERO_EXTEND_WITH_AND
)
19920 *total
= ix86_cost
->add
;
19922 *total
= ix86_cost
->movzx
;
19926 *total
= ix86_cost
->movsx
;
19930 if (CONST_INT_P (XEXP (x
, 1))
19931 && (GET_MODE (XEXP (x
, 0)) != DImode
|| TARGET_64BIT
))
19933 HOST_WIDE_INT value
= INTVAL (XEXP (x
, 1));
19936 *total
= ix86_cost
->add
;
19939 if ((value
== 2 || value
== 3)
19940 && ix86_cost
->lea
<= ix86_cost
->shift_const
)
19942 *total
= ix86_cost
->lea
;
19952 if (!TARGET_64BIT
&& GET_MODE (XEXP (x
, 0)) == DImode
)
19954 if (CONST_INT_P (XEXP (x
, 1)))
19956 if (INTVAL (XEXP (x
, 1)) > 32)
19957 *total
= ix86_cost
->shift_const
+ COSTS_N_INSNS (2);
19959 *total
= ix86_cost
->shift_const
* 2;
19963 if (GET_CODE (XEXP (x
, 1)) == AND
)
19964 *total
= ix86_cost
->shift_var
* 2;
19966 *total
= ix86_cost
->shift_var
* 6 + COSTS_N_INSNS (2);
19971 if (CONST_INT_P (XEXP (x
, 1)))
19972 *total
= ix86_cost
->shift_const
;
19974 *total
= ix86_cost
->shift_var
;
19979 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
19981 /* ??? SSE scalar cost should be used here. */
19982 *total
= ix86_cost
->fmul
;
19985 else if (X87_FLOAT_MODE_P (mode
))
19987 *total
= ix86_cost
->fmul
;
19990 else if (FLOAT_MODE_P (mode
))
19992 /* ??? SSE vector cost should be used here. */
19993 *total
= ix86_cost
->fmul
;
19998 rtx op0
= XEXP (x
, 0);
19999 rtx op1
= XEXP (x
, 1);
20001 if (CONST_INT_P (XEXP (x
, 1)))
20003 unsigned HOST_WIDE_INT value
= INTVAL (XEXP (x
, 1));
20004 for (nbits
= 0; value
!= 0; value
&= value
- 1)
20008 /* This is arbitrary. */
20011 /* Compute costs correctly for widening multiplication. */
20012 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op1
) == ZERO_EXTEND
)
20013 && GET_MODE_SIZE (GET_MODE (XEXP (op0
, 0))) * 2
20014 == GET_MODE_SIZE (mode
))
20016 int is_mulwiden
= 0;
20017 enum machine_mode inner_mode
= GET_MODE (op0
);
20019 if (GET_CODE (op0
) == GET_CODE (op1
))
20020 is_mulwiden
= 1, op1
= XEXP (op1
, 0);
20021 else if (CONST_INT_P (op1
))
20023 if (GET_CODE (op0
) == SIGN_EXTEND
)
20024 is_mulwiden
= trunc_int_for_mode (INTVAL (op1
), inner_mode
)
20027 is_mulwiden
= !(INTVAL (op1
) & ~GET_MODE_MASK (inner_mode
));
20031 op0
= XEXP (op0
, 0), mode
= GET_MODE (op0
);
20034 *total
= (ix86_cost
->mult_init
[MODE_INDEX (mode
)]
20035 + nbits
* ix86_cost
->mult_bit
20036 + rtx_cost (op0
, outer_code
) + rtx_cost (op1
, outer_code
));
20045 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
20046 /* ??? SSE cost should be used here. */
20047 *total
= ix86_cost
->fdiv
;
20048 else if (X87_FLOAT_MODE_P (mode
))
20049 *total
= ix86_cost
->fdiv
;
20050 else if (FLOAT_MODE_P (mode
))
20051 /* ??? SSE vector cost should be used here. */
20052 *total
= ix86_cost
->fdiv
;
20054 *total
= ix86_cost
->divide
[MODE_INDEX (mode
)];
20058 if (GET_MODE_CLASS (mode
) == MODE_INT
20059 && GET_MODE_BITSIZE (mode
) <= GET_MODE_BITSIZE (Pmode
))
20061 if (GET_CODE (XEXP (x
, 0)) == PLUS
20062 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
20063 && CONST_INT_P (XEXP (XEXP (XEXP (x
, 0), 0), 1))
20064 && CONSTANT_P (XEXP (x
, 1)))
20066 HOST_WIDE_INT val
= INTVAL (XEXP (XEXP (XEXP (x
, 0), 0), 1));
20067 if (val
== 2 || val
== 4 || val
== 8)
20069 *total
= ix86_cost
->lea
;
20070 *total
+= rtx_cost (XEXP (XEXP (x
, 0), 1), outer_code
);
20071 *total
+= rtx_cost (XEXP (XEXP (XEXP (x
, 0), 0), 0),
20073 *total
+= rtx_cost (XEXP (x
, 1), outer_code
);
20077 else if (GET_CODE (XEXP (x
, 0)) == MULT
20078 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))
20080 HOST_WIDE_INT val
= INTVAL (XEXP (XEXP (x
, 0), 1));
20081 if (val
== 2 || val
== 4 || val
== 8)
20083 *total
= ix86_cost
->lea
;
20084 *total
+= rtx_cost (XEXP (XEXP (x
, 0), 0), outer_code
);
20085 *total
+= rtx_cost (XEXP (x
, 1), outer_code
);
20089 else if (GET_CODE (XEXP (x
, 0)) == PLUS
)
20091 *total
= ix86_cost
->lea
;
20092 *total
+= rtx_cost (XEXP (XEXP (x
, 0), 0), outer_code
);
20093 *total
+= rtx_cost (XEXP (XEXP (x
, 0), 1), outer_code
);
20094 *total
+= rtx_cost (XEXP (x
, 1), outer_code
);
20101 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
20103 /* ??? SSE cost should be used here. */
20104 *total
= ix86_cost
->fadd
;
20107 else if (X87_FLOAT_MODE_P (mode
))
20109 *total
= ix86_cost
->fadd
;
20112 else if (FLOAT_MODE_P (mode
))
20114 /* ??? SSE vector cost should be used here. */
20115 *total
= ix86_cost
->fadd
;
20123 if (!TARGET_64BIT
&& mode
== DImode
)
20125 *total
= (ix86_cost
->add
* 2
20126 + (rtx_cost (XEXP (x
, 0), outer_code
)
20127 << (GET_MODE (XEXP (x
, 0)) != DImode
))
20128 + (rtx_cost (XEXP (x
, 1), outer_code
)
20129 << (GET_MODE (XEXP (x
, 1)) != DImode
)));
20135 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
20137 /* ??? SSE cost should be used here. */
20138 *total
= ix86_cost
->fchs
;
20141 else if (X87_FLOAT_MODE_P (mode
))
20143 *total
= ix86_cost
->fchs
;
20146 else if (FLOAT_MODE_P (mode
))
20148 /* ??? SSE vector cost should be used here. */
20149 *total
= ix86_cost
->fchs
;
20155 if (!TARGET_64BIT
&& mode
== DImode
)
20156 *total
= ix86_cost
->add
* 2;
20158 *total
= ix86_cost
->add
;
20162 if (GET_CODE (XEXP (x
, 0)) == ZERO_EXTRACT
20163 && XEXP (XEXP (x
, 0), 1) == const1_rtx
20164 && CONST_INT_P (XEXP (XEXP (x
, 0), 2))
20165 && XEXP (x
, 1) == const0_rtx
)
20167 /* This kind of construct is implemented using test[bwl].
20168 Treat it as if we had an AND. */
20169 *total
= (ix86_cost
->add
20170 + rtx_cost (XEXP (XEXP (x
, 0), 0), outer_code
)
20171 + rtx_cost (const1_rtx
, outer_code
));
20177 if (!(SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
))
20182 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
20183 /* ??? SSE cost should be used here. */
20184 *total
= ix86_cost
->fabs
;
20185 else if (X87_FLOAT_MODE_P (mode
))
20186 *total
= ix86_cost
->fabs
;
20187 else if (FLOAT_MODE_P (mode
))
20188 /* ??? SSE vector cost should be used here. */
20189 *total
= ix86_cost
->fabs
;
20193 if (SSE_FLOAT_MODE_P (mode
) && TARGET_SSE_MATH
)
20194 /* ??? SSE cost should be used here. */
20195 *total
= ix86_cost
->fsqrt
;
20196 else if (X87_FLOAT_MODE_P (mode
))
20197 *total
= ix86_cost
->fsqrt
;
20198 else if (FLOAT_MODE_P (mode
))
20199 /* ??? SSE vector cost should be used here. */
20200 *total
= ix86_cost
->fsqrt
;
20204 if (XINT (x
, 1) == UNSPEC_TP
)
20215 static int current_machopic_label_num
;
20217 /* Given a symbol name and its associated stub, write out the
20218 definition of the stub. */
20221 machopic_output_stub (FILE *file
, const char *symb
, const char *stub
)
20223 unsigned int length
;
20224 char *binder_name
, *symbol_name
, lazy_ptr_name
[32];
20225 int label
= ++current_machopic_label_num
;
20227 /* For 64-bit we shouldn't get here. */
20228 gcc_assert (!TARGET_64BIT
);
20230 /* Lose our funky encoding stuff so it doesn't contaminate the stub. */
20231 symb
= (*targetm
.strip_name_encoding
) (symb
);
20233 length
= strlen (stub
);
20234 binder_name
= alloca (length
+ 32);
20235 GEN_BINDER_NAME_FOR_STUB (binder_name
, stub
, length
);
20237 length
= strlen (symb
);
20238 symbol_name
= alloca (length
+ 32);
20239 GEN_SYMBOL_NAME_FOR_SYMBOL (symbol_name
, symb
, length
);
20241 sprintf (lazy_ptr_name
, "L%d$lz", label
);
20244 switch_to_section (darwin_sections
[machopic_picsymbol_stub_section
]);
20246 switch_to_section (darwin_sections
[machopic_symbol_stub_section
]);
20248 fprintf (file
, "%s:\n", stub
);
20249 fprintf (file
, "\t.indirect_symbol %s\n", symbol_name
);
20253 fprintf (file
, "\tcall\tLPC$%d\nLPC$%d:\tpopl\t%%eax\n", label
, label
);
20254 fprintf (file
, "\tmovl\t%s-LPC$%d(%%eax),%%edx\n", lazy_ptr_name
, label
);
20255 fprintf (file
, "\tjmp\t*%%edx\n");
20258 fprintf (file
, "\tjmp\t*%s\n", lazy_ptr_name
);
20260 fprintf (file
, "%s:\n", binder_name
);
20264 fprintf (file
, "\tlea\t%s-LPC$%d(%%eax),%%eax\n", lazy_ptr_name
, label
);
20265 fprintf (file
, "\tpushl\t%%eax\n");
20268 fprintf (file
, "\tpushl\t$%s\n", lazy_ptr_name
);
20270 fprintf (file
, "\tjmp\tdyld_stub_binding_helper\n");
20272 switch_to_section (darwin_sections
[machopic_lazy_symbol_ptr_section
]);
20273 fprintf (file
, "%s:\n", lazy_ptr_name
);
20274 fprintf (file
, "\t.indirect_symbol %s\n", symbol_name
);
20275 fprintf (file
, "\t.long %s\n", binder_name
);
20279 darwin_x86_file_end (void)
20281 darwin_file_end ();
20284 #endif /* TARGET_MACHO */
20286 /* Order the registers for register allocator. */
20289 x86_order_regs_for_local_alloc (void)
20294 /* First allocate the local general purpose registers. */
20295 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
20296 if (GENERAL_REGNO_P (i
) && call_used_regs
[i
])
20297 reg_alloc_order
[pos
++] = i
;
20299 /* Global general purpose registers. */
20300 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
20301 if (GENERAL_REGNO_P (i
) && !call_used_regs
[i
])
20302 reg_alloc_order
[pos
++] = i
;
20304 /* x87 registers come first in case we are doing FP math
20306 if (!TARGET_SSE_MATH
)
20307 for (i
= FIRST_STACK_REG
; i
<= LAST_STACK_REG
; i
++)
20308 reg_alloc_order
[pos
++] = i
;
20310 /* SSE registers. */
20311 for (i
= FIRST_SSE_REG
; i
<= LAST_SSE_REG
; i
++)
20312 reg_alloc_order
[pos
++] = i
;
20313 for (i
= FIRST_REX_SSE_REG
; i
<= LAST_REX_SSE_REG
; i
++)
20314 reg_alloc_order
[pos
++] = i
;
20316 /* x87 registers. */
20317 if (TARGET_SSE_MATH
)
20318 for (i
= FIRST_STACK_REG
; i
<= LAST_STACK_REG
; i
++)
20319 reg_alloc_order
[pos
++] = i
;
20321 for (i
= FIRST_MMX_REG
; i
<= LAST_MMX_REG
; i
++)
20322 reg_alloc_order
[pos
++] = i
;
20324 /* Initialize the rest of array as we do not allocate some registers
20326 while (pos
< FIRST_PSEUDO_REGISTER
)
20327 reg_alloc_order
[pos
++] = 0;
20330 /* Handle a "ms_struct" or "gcc_struct" attribute; arguments as in
20331 struct attribute_spec.handler. */
20333 ix86_handle_struct_attribute (tree
*node
, tree name
,
20334 tree args ATTRIBUTE_UNUSED
,
20335 int flags ATTRIBUTE_UNUSED
, bool *no_add_attrs
)
20338 if (DECL_P (*node
))
20340 if (TREE_CODE (*node
) == TYPE_DECL
)
20341 type
= &TREE_TYPE (*node
);
20346 if (!(type
&& (TREE_CODE (*type
) == RECORD_TYPE
20347 || TREE_CODE (*type
) == UNION_TYPE
)))
20349 warning (OPT_Wattributes
, "%qs attribute ignored",
20350 IDENTIFIER_POINTER (name
));
20351 *no_add_attrs
= true;
20354 else if ((is_attribute_p ("ms_struct", name
)
20355 && lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (*type
)))
20356 || ((is_attribute_p ("gcc_struct", name
)
20357 && lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (*type
)))))
20359 warning (OPT_Wattributes
, "%qs incompatible attribute ignored",
20360 IDENTIFIER_POINTER (name
));
20361 *no_add_attrs
= true;
20368 ix86_ms_bitfield_layout_p (tree record_type
)
20370 return (TARGET_MS_BITFIELD_LAYOUT
&&
20371 !lookup_attribute ("gcc_struct", TYPE_ATTRIBUTES (record_type
)))
20372 || lookup_attribute ("ms_struct", TYPE_ATTRIBUTES (record_type
));
20375 /* Returns an expression indicating where the this parameter is
20376 located on entry to the FUNCTION. */
20379 x86_this_parameter (tree function
)
20381 tree type
= TREE_TYPE (function
);
20382 bool aggr
= aggregate_value_p (TREE_TYPE (type
), type
) != 0;
20386 const int *parm_regs
;
20388 if (TARGET_64BIT_MS_ABI
)
20389 parm_regs
= x86_64_ms_abi_int_parameter_registers
;
20391 parm_regs
= x86_64_int_parameter_registers
;
20392 return gen_rtx_REG (DImode
, parm_regs
[aggr
]);
20395 if (ix86_function_regparm (type
, function
) > 0
20396 && !type_has_variadic_args_p (type
))
20399 if (lookup_attribute ("fastcall", TYPE_ATTRIBUTES (type
)))
20401 return gen_rtx_REG (SImode
, regno
);
20404 return gen_rtx_MEM (SImode
, plus_constant (stack_pointer_rtx
, aggr
? 8 : 4));
20407 /* Determine whether x86_output_mi_thunk can succeed. */
20410 x86_can_output_mi_thunk (tree thunk ATTRIBUTE_UNUSED
,
20411 HOST_WIDE_INT delta ATTRIBUTE_UNUSED
,
20412 HOST_WIDE_INT vcall_offset
, tree function
)
20414 /* 64-bit can handle anything. */
20418 /* For 32-bit, everything's fine if we have one free register. */
20419 if (ix86_function_regparm (TREE_TYPE (function
), function
) < 3)
20422 /* Need a free register for vcall_offset. */
20426 /* Need a free register for GOT references. */
20427 if (flag_pic
&& !(*targetm
.binds_local_p
) (function
))
20430 /* Otherwise ok. */
20434 /* Output the assembler code for a thunk function. THUNK_DECL is the
20435 declaration for the thunk function itself, FUNCTION is the decl for
20436 the target function. DELTA is an immediate constant offset to be
20437 added to THIS. If VCALL_OFFSET is nonzero, the word at
20438 *(*this + vcall_offset) should be added to THIS. */
20441 x86_output_mi_thunk (FILE *file ATTRIBUTE_UNUSED
,
20442 tree thunk ATTRIBUTE_UNUSED
, HOST_WIDE_INT delta
,
20443 HOST_WIDE_INT vcall_offset
, tree function
)
20446 rtx
this = x86_this_parameter (function
);
20449 /* If VCALL_OFFSET, we'll need THIS in a register. Might as well
20450 pull it in now and let DELTA benefit. */
20453 else if (vcall_offset
)
20455 /* Put the this parameter into %eax. */
20457 xops
[1] = this_reg
= gen_rtx_REG (Pmode
, 0);
20458 output_asm_insn ("mov{l}\t{%0, %1|%1, %0}", xops
);
20461 this_reg
= NULL_RTX
;
20463 /* Adjust the this parameter by a fixed constant. */
20466 xops
[0] = GEN_INT (delta
);
20467 xops
[1] = this_reg
? this_reg
: this;
20470 if (!x86_64_general_operand (xops
[0], DImode
))
20472 tmp
= gen_rtx_REG (DImode
, R10_REG
);
20474 output_asm_insn ("mov{q}\t{%1, %0|%0, %1}", xops
);
20478 output_asm_insn ("add{q}\t{%0, %1|%1, %0}", xops
);
20481 output_asm_insn ("add{l}\t{%0, %1|%1, %0}", xops
);
20484 /* Adjust the this parameter by a value stored in the vtable. */
20488 tmp
= gen_rtx_REG (DImode
, R10_REG
);
20491 int tmp_regno
= 2 /* ECX */;
20492 if (lookup_attribute ("fastcall",
20493 TYPE_ATTRIBUTES (TREE_TYPE (function
))))
20494 tmp_regno
= 0 /* EAX */;
20495 tmp
= gen_rtx_REG (SImode
, tmp_regno
);
20498 xops
[0] = gen_rtx_MEM (Pmode
, this_reg
);
20501 output_asm_insn ("mov{q}\t{%0, %1|%1, %0}", xops
);
20503 output_asm_insn ("mov{l}\t{%0, %1|%1, %0}", xops
);
20505 /* Adjust the this parameter. */
20506 xops
[0] = gen_rtx_MEM (Pmode
, plus_constant (tmp
, vcall_offset
));
20507 if (TARGET_64BIT
&& !memory_operand (xops
[0], Pmode
))
20509 rtx tmp2
= gen_rtx_REG (DImode
, R11_REG
);
20510 xops
[0] = GEN_INT (vcall_offset
);
20512 output_asm_insn ("mov{q}\t{%0, %1|%1, %0}", xops
);
20513 xops
[0] = gen_rtx_MEM (Pmode
, gen_rtx_PLUS (Pmode
, tmp
, tmp2
));
20515 xops
[1] = this_reg
;
20517 output_asm_insn ("add{q}\t{%0, %1|%1, %0}", xops
);
20519 output_asm_insn ("add{l}\t{%0, %1|%1, %0}", xops
);
20522 /* If necessary, drop THIS back to its stack slot. */
20523 if (this_reg
&& this_reg
!= this)
20525 xops
[0] = this_reg
;
20527 output_asm_insn ("mov{l}\t{%0, %1|%1, %0}", xops
);
20530 xops
[0] = XEXP (DECL_RTL (function
), 0);
20533 if (!flag_pic
|| (*targetm
.binds_local_p
) (function
))
20534 output_asm_insn ("jmp\t%P0", xops
);
20535 /* All thunks should be in the same object as their target,
20536 and thus binds_local_p should be true. */
20537 else if (TARGET_64BIT_MS_ABI
)
20538 gcc_unreachable ();
20541 tmp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, xops
[0]), UNSPEC_GOTPCREL
);
20542 tmp
= gen_rtx_CONST (Pmode
, tmp
);
20543 tmp
= gen_rtx_MEM (QImode
, tmp
);
20545 output_asm_insn ("jmp\t%A0", xops
);
20550 if (!flag_pic
|| (*targetm
.binds_local_p
) (function
))
20551 output_asm_insn ("jmp\t%P0", xops
);
20556 rtx sym_ref
= XEXP (DECL_RTL (function
), 0);
20557 tmp
= (gen_rtx_SYMBOL_REF
20559 machopic_indirection_name (sym_ref
, /*stub_p=*/true)));
20560 tmp
= gen_rtx_MEM (QImode
, tmp
);
20562 output_asm_insn ("jmp\t%0", xops
);
20565 #endif /* TARGET_MACHO */
20567 tmp
= gen_rtx_REG (SImode
, 2 /* ECX */);
20568 output_set_got (tmp
, NULL_RTX
);
20571 output_asm_insn ("mov{l}\t{%0@GOT(%1), %1|%1, %0@GOT[%1]}", xops
);
20572 output_asm_insn ("jmp\t{*}%1", xops
);
20578 x86_file_start (void)
20580 default_file_start ();
20582 darwin_file_start ();
20584 if (X86_FILE_START_VERSION_DIRECTIVE
)
20585 fputs ("\t.version\t\"01.01\"\n", asm_out_file
);
20586 if (X86_FILE_START_FLTUSED
)
20587 fputs ("\t.global\t__fltused\n", asm_out_file
);
20588 if (ix86_asm_dialect
== ASM_INTEL
)
20589 fputs ("\t.intel_syntax\n", asm_out_file
);
20593 x86_field_alignment (tree field
, int computed
)
20595 enum machine_mode mode
;
20596 tree type
= TREE_TYPE (field
);
20598 if (TARGET_64BIT
|| TARGET_ALIGN_DOUBLE
)
20600 mode
= TYPE_MODE (TREE_CODE (type
) == ARRAY_TYPE
20601 ? get_inner_array_type (type
) : type
);
20602 if (mode
== DFmode
|| mode
== DCmode
20603 || GET_MODE_CLASS (mode
) == MODE_INT
20604 || GET_MODE_CLASS (mode
) == MODE_COMPLEX_INT
)
20605 return MIN (32, computed
);
20609 /* Output assembler code to FILE to increment profiler label # LABELNO
20610 for profiling a function entry. */
20612 x86_function_profiler (FILE *file
, int labelno ATTRIBUTE_UNUSED
)
20616 #ifndef NO_PROFILE_COUNTERS
20617 fprintf (file
, "\tleaq\t%sP%d@(%%rip),%%r11\n", LPREFIX
, labelno
);
20620 if (!TARGET_64BIT_MS_ABI
&& flag_pic
)
20621 fprintf (file
, "\tcall\t*%s@GOTPCREL(%%rip)\n", MCOUNT_NAME
);
20623 fprintf (file
, "\tcall\t%s\n", MCOUNT_NAME
);
20627 #ifndef NO_PROFILE_COUNTERS
20628 fprintf (file
, "\tleal\t%sP%d@GOTOFF(%%ebx),%%%s\n",
20629 LPREFIX
, labelno
, PROFILE_COUNT_REGISTER
);
20631 fprintf (file
, "\tcall\t*%s@GOT(%%ebx)\n", MCOUNT_NAME
);
20635 #ifndef NO_PROFILE_COUNTERS
20636 fprintf (file
, "\tmovl\t$%sP%d,%%%s\n", LPREFIX
, labelno
,
20637 PROFILE_COUNT_REGISTER
);
20639 fprintf (file
, "\tcall\t%s\n", MCOUNT_NAME
);
20643 /* We don't have exact information about the insn sizes, but we may assume
20644 quite safely that we are informed about all 1 byte insns and memory
20645 address sizes. This is enough to eliminate unnecessary padding in
20649 min_insn_size (rtx insn
)
20653 if (!INSN_P (insn
) || !active_insn_p (insn
))
20656 /* Discard alignments we've emit and jump instructions. */
20657 if (GET_CODE (PATTERN (insn
)) == UNSPEC_VOLATILE
20658 && XINT (PATTERN (insn
), 1) == UNSPECV_ALIGN
)
20661 && (GET_CODE (PATTERN (insn
)) == ADDR_VEC
20662 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
))
20665 /* Important case - calls are always 5 bytes.
20666 It is common to have many calls in the row. */
20668 && symbolic_reference_mentioned_p (PATTERN (insn
))
20669 && !SIBLING_CALL_P (insn
))
20671 if (get_attr_length (insn
) <= 1)
20674 /* For normal instructions we may rely on the sizes of addresses
20675 and the presence of symbol to require 4 bytes of encoding.
20676 This is not the case for jumps where references are PC relative. */
20677 if (!JUMP_P (insn
))
20679 l
= get_attr_length_address (insn
);
20680 if (l
< 4 && symbolic_reference_mentioned_p (PATTERN (insn
)))
20689 /* AMD K8 core mispredicts jumps when there are more than 3 jumps in 16 byte
20693 ix86_avoid_jump_misspredicts (void)
20695 rtx insn
, start
= get_insns ();
20696 int nbytes
= 0, njumps
= 0;
20699 /* Look for all minimal intervals of instructions containing 4 jumps.
20700 The intervals are bounded by START and INSN. NBYTES is the total
20701 size of instructions in the interval including INSN and not including
20702 START. When the NBYTES is smaller than 16 bytes, it is possible
20703 that the end of START and INSN ends up in the same 16byte page.
20705 The smallest offset in the page INSN can start is the case where START
20706 ends on the offset 0. Offset of INSN is then NBYTES - sizeof (INSN).
20707 We add p2align to 16byte window with maxskip 17 - NBYTES + sizeof (INSN).
20709 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
20712 nbytes
+= min_insn_size (insn
);
20714 fprintf(dump_file
, "Insn %i estimated to %i bytes\n",
20715 INSN_UID (insn
), min_insn_size (insn
));
20717 && GET_CODE (PATTERN (insn
)) != ADDR_VEC
20718 && GET_CODE (PATTERN (insn
)) != ADDR_DIFF_VEC
)
20726 start
= NEXT_INSN (start
);
20727 if ((JUMP_P (start
)
20728 && GET_CODE (PATTERN (start
)) != ADDR_VEC
20729 && GET_CODE (PATTERN (start
)) != ADDR_DIFF_VEC
)
20731 njumps
--, isjump
= 1;
20734 nbytes
-= min_insn_size (start
);
20736 gcc_assert (njumps
>= 0);
20738 fprintf (dump_file
, "Interval %i to %i has %i bytes\n",
20739 INSN_UID (start
), INSN_UID (insn
), nbytes
);
20741 if (njumps
== 3 && isjump
&& nbytes
< 16)
20743 int padsize
= 15 - nbytes
+ min_insn_size (insn
);
20746 fprintf (dump_file
, "Padding insn %i by %i bytes!\n",
20747 INSN_UID (insn
), padsize
);
20748 emit_insn_before (gen_align (GEN_INT (padsize
)), insn
);
20753 /* AMD Athlon works faster
20754 when RET is not destination of conditional jump or directly preceded
20755 by other jump instruction. We avoid the penalty by inserting NOP just
20756 before the RET instructions in such cases. */
20758 ix86_pad_returns (void)
20763 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
20765 basic_block bb
= e
->src
;
20766 rtx ret
= BB_END (bb
);
20768 bool replace
= false;
20770 if (!JUMP_P (ret
) || GET_CODE (PATTERN (ret
)) != RETURN
20771 || !maybe_hot_bb_p (bb
))
20773 for (prev
= PREV_INSN (ret
); prev
; prev
= PREV_INSN (prev
))
20774 if (active_insn_p (prev
) || LABEL_P (prev
))
20776 if (prev
&& LABEL_P (prev
))
20781 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
20782 if (EDGE_FREQUENCY (e
) && e
->src
->index
>= 0
20783 && !(e
->flags
& EDGE_FALLTHRU
))
20788 prev
= prev_active_insn (ret
);
20790 && ((JUMP_P (prev
) && any_condjump_p (prev
))
20793 /* Empty functions get branch mispredict even when the jump destination
20794 is not visible to us. */
20795 if (!prev
&& cfun
->function_frequency
> FUNCTION_FREQUENCY_UNLIKELY_EXECUTED
)
20800 emit_insn_before (gen_return_internal_long (), ret
);
20806 /* Implement machine specific optimizations. We implement padding of returns
20807 for K8 CPUs and pass to avoid 4 jumps in the single 16 byte window. */
20811 if (TARGET_PAD_RETURNS
&& optimize
&& !optimize_size
)
20812 ix86_pad_returns ();
20813 if (TARGET_FOUR_JUMP_LIMIT
&& optimize
&& !optimize_size
)
20814 ix86_avoid_jump_misspredicts ();
20817 /* Return nonzero when QImode register that must be represented via REX prefix
20820 x86_extended_QIreg_mentioned_p (rtx insn
)
20823 extract_insn_cached (insn
);
20824 for (i
= 0; i
< recog_data
.n_operands
; i
++)
20825 if (REG_P (recog_data
.operand
[i
])
20826 && REGNO (recog_data
.operand
[i
]) >= 4)
20831 /* Return nonzero when P points to register encoded via REX prefix.
20832 Called via for_each_rtx. */
20834 extended_reg_mentioned_1 (rtx
*p
, void *data ATTRIBUTE_UNUSED
)
20836 unsigned int regno
;
20839 regno
= REGNO (*p
);
20840 return REX_INT_REGNO_P (regno
) || REX_SSE_REGNO_P (regno
);
20843 /* Return true when INSN mentions register that must be encoded using REX
20846 x86_extended_reg_mentioned_p (rtx insn
)
20848 return for_each_rtx (&PATTERN (insn
), extended_reg_mentioned_1
, NULL
);
20851 /* Generate an unsigned DImode/SImode to FP conversion. This is the same code
20852 optabs would emit if we didn't have TFmode patterns. */
20855 x86_emit_floatuns (rtx operands
[2])
20857 rtx neglab
, donelab
, i0
, i1
, f0
, in
, out
;
20858 enum machine_mode mode
, inmode
;
20860 inmode
= GET_MODE (operands
[1]);
20861 gcc_assert (inmode
== SImode
|| inmode
== DImode
);
20864 in
= force_reg (inmode
, operands
[1]);
20865 mode
= GET_MODE (out
);
20866 neglab
= gen_label_rtx ();
20867 donelab
= gen_label_rtx ();
20868 f0
= gen_reg_rtx (mode
);
20870 emit_cmp_and_jump_insns (in
, const0_rtx
, LT
, const0_rtx
, inmode
, 0, neglab
);
20872 expand_float (out
, in
, 0);
20874 emit_jump_insn (gen_jump (donelab
));
20877 emit_label (neglab
);
20879 i0
= expand_simple_binop (inmode
, LSHIFTRT
, in
, const1_rtx
, NULL
,
20881 i1
= expand_simple_binop (inmode
, AND
, in
, const1_rtx
, NULL
,
20883 i0
= expand_simple_binop (inmode
, IOR
, i0
, i1
, i0
, 1, OPTAB_DIRECT
);
20885 expand_float (f0
, i0
, 0);
20887 emit_insn (gen_rtx_SET (VOIDmode
, out
, gen_rtx_PLUS (mode
, f0
, f0
)));
20889 emit_label (donelab
);
20892 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
20893 with all elements equal to VAR. Return true if successful. */
20896 ix86_expand_vector_init_duplicate (bool mmx_ok
, enum machine_mode mode
,
20897 rtx target
, rtx val
)
20899 enum machine_mode smode
, wsmode
, wvmode
;
20914 val
= force_reg (GET_MODE_INNER (mode
), val
);
20915 x
= gen_rtx_VEC_DUPLICATE (mode
, val
);
20916 emit_insn (gen_rtx_SET (VOIDmode
, target
, x
));
20922 if (TARGET_SSE
|| TARGET_3DNOW_A
)
20924 val
= gen_lowpart (SImode
, val
);
20925 x
= gen_rtx_TRUNCATE (HImode
, val
);
20926 x
= gen_rtx_VEC_DUPLICATE (mode
, x
);
20927 emit_insn (gen_rtx_SET (VOIDmode
, target
, x
));
20949 /* Extend HImode to SImode using a paradoxical SUBREG. */
20950 tmp1
= gen_reg_rtx (SImode
);
20951 emit_move_insn (tmp1
, gen_lowpart (SImode
, val
));
20952 /* Insert the SImode value as low element of V4SImode vector. */
20953 tmp2
= gen_reg_rtx (V4SImode
);
20954 tmp1
= gen_rtx_VEC_MERGE (V4SImode
,
20955 gen_rtx_VEC_DUPLICATE (V4SImode
, tmp1
),
20956 CONST0_RTX (V4SImode
),
20958 emit_insn (gen_rtx_SET (VOIDmode
, tmp2
, tmp1
));
20959 /* Cast the V4SImode vector back to a V8HImode vector. */
20960 tmp1
= gen_reg_rtx (V8HImode
);
20961 emit_move_insn (tmp1
, gen_lowpart (V8HImode
, tmp2
));
20962 /* Duplicate the low short through the whole low SImode word. */
20963 emit_insn (gen_sse2_punpcklwd (tmp1
, tmp1
, tmp1
));
20964 /* Cast the V8HImode vector back to a V4SImode vector. */
20965 tmp2
= gen_reg_rtx (V4SImode
);
20966 emit_move_insn (tmp2
, gen_lowpart (V4SImode
, tmp1
));
20967 /* Replicate the low element of the V4SImode vector. */
20968 emit_insn (gen_sse2_pshufd (tmp2
, tmp2
, const0_rtx
));
20969 /* Cast the V2SImode back to V8HImode, and store in target. */
20970 emit_move_insn (target
, gen_lowpart (V8HImode
, tmp2
));
20981 /* Extend QImode to SImode using a paradoxical SUBREG. */
20982 tmp1
= gen_reg_rtx (SImode
);
20983 emit_move_insn (tmp1
, gen_lowpart (SImode
, val
));
20984 /* Insert the SImode value as low element of V4SImode vector. */
20985 tmp2
= gen_reg_rtx (V4SImode
);
20986 tmp1
= gen_rtx_VEC_MERGE (V4SImode
,
20987 gen_rtx_VEC_DUPLICATE (V4SImode
, tmp1
),
20988 CONST0_RTX (V4SImode
),
20990 emit_insn (gen_rtx_SET (VOIDmode
, tmp2
, tmp1
));
20991 /* Cast the V4SImode vector back to a V16QImode vector. */
20992 tmp1
= gen_reg_rtx (V16QImode
);
20993 emit_move_insn (tmp1
, gen_lowpart (V16QImode
, tmp2
));
20994 /* Duplicate the low byte through the whole low SImode word. */
20995 emit_insn (gen_sse2_punpcklbw (tmp1
, tmp1
, tmp1
));
20996 emit_insn (gen_sse2_punpcklbw (tmp1
, tmp1
, tmp1
));
20997 /* Cast the V16QImode vector back to a V4SImode vector. */
20998 tmp2
= gen_reg_rtx (V4SImode
);
20999 emit_move_insn (tmp2
, gen_lowpart (V4SImode
, tmp1
));
21000 /* Replicate the low element of the V4SImode vector. */
21001 emit_insn (gen_sse2_pshufd (tmp2
, tmp2
, const0_rtx
));
21002 /* Cast the V2SImode back to V16QImode, and store in target. */
21003 emit_move_insn (target
, gen_lowpart (V16QImode
, tmp2
));
21011 /* Replicate the value once into the next wider mode and recurse. */
21012 val
= convert_modes (wsmode
, smode
, val
, true);
21013 x
= expand_simple_binop (wsmode
, ASHIFT
, val
,
21014 GEN_INT (GET_MODE_BITSIZE (smode
)),
21015 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
21016 val
= expand_simple_binop (wsmode
, IOR
, val
, x
, x
, 1, OPTAB_LIB_WIDEN
);
21018 x
= gen_reg_rtx (wvmode
);
21019 if (!ix86_expand_vector_init_duplicate (mmx_ok
, wvmode
, x
, val
))
21020 gcc_unreachable ();
21021 emit_move_insn (target
, gen_lowpart (mode
, x
));
21029 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
21030 whose ONE_VAR element is VAR, and other elements are zero. Return true
21034 ix86_expand_vector_init_one_nonzero (bool mmx_ok
, enum machine_mode mode
,
21035 rtx target
, rtx var
, int one_var
)
21037 enum machine_mode vsimode
;
21053 var
= force_reg (GET_MODE_INNER (mode
), var
);
21054 x
= gen_rtx_VEC_CONCAT (mode
, var
, CONST0_RTX (GET_MODE_INNER (mode
)));
21055 emit_insn (gen_rtx_SET (VOIDmode
, target
, x
));
21060 if (!REG_P (target
) || REGNO (target
) < FIRST_PSEUDO_REGISTER
)
21061 new_target
= gen_reg_rtx (mode
);
21063 new_target
= target
;
21064 var
= force_reg (GET_MODE_INNER (mode
), var
);
21065 x
= gen_rtx_VEC_DUPLICATE (mode
, var
);
21066 x
= gen_rtx_VEC_MERGE (mode
, x
, CONST0_RTX (mode
), const1_rtx
);
21067 emit_insn (gen_rtx_SET (VOIDmode
, new_target
, x
));
21070 /* We need to shuffle the value to the correct position, so
21071 create a new pseudo to store the intermediate result. */
21073 /* With SSE2, we can use the integer shuffle insns. */
21074 if (mode
!= V4SFmode
&& TARGET_SSE2
)
21076 emit_insn (gen_sse2_pshufd_1 (new_target
, new_target
,
21078 GEN_INT (one_var
== 1 ? 0 : 1),
21079 GEN_INT (one_var
== 2 ? 0 : 1),
21080 GEN_INT (one_var
== 3 ? 0 : 1)));
21081 if (target
!= new_target
)
21082 emit_move_insn (target
, new_target
);
21086 /* Otherwise convert the intermediate result to V4SFmode and
21087 use the SSE1 shuffle instructions. */
21088 if (mode
!= V4SFmode
)
21090 tmp
= gen_reg_rtx (V4SFmode
);
21091 emit_move_insn (tmp
, gen_lowpart (V4SFmode
, new_target
));
21096 emit_insn (gen_sse_shufps_1 (tmp
, tmp
, tmp
,
21098 GEN_INT (one_var
== 1 ? 0 : 1),
21099 GEN_INT (one_var
== 2 ? 0+4 : 1+4),
21100 GEN_INT (one_var
== 3 ? 0+4 : 1+4)));
21102 if (mode
!= V4SFmode
)
21103 emit_move_insn (target
, gen_lowpart (V4SImode
, tmp
));
21104 else if (tmp
!= target
)
21105 emit_move_insn (target
, tmp
);
21107 else if (target
!= new_target
)
21108 emit_move_insn (target
, new_target
);
21113 vsimode
= V4SImode
;
21119 vsimode
= V2SImode
;
21125 /* Zero extend the variable element to SImode and recurse. */
21126 var
= convert_modes (SImode
, GET_MODE_INNER (mode
), var
, true);
21128 x
= gen_reg_rtx (vsimode
);
21129 if (!ix86_expand_vector_init_one_nonzero (mmx_ok
, vsimode
, x
,
21131 gcc_unreachable ();
21133 emit_move_insn (target
, gen_lowpart (mode
, x
));
21141 /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector
21142 consisting of the values in VALS. It is known that all elements
21143 except ONE_VAR are constants. Return true if successful. */
21146 ix86_expand_vector_init_one_var (bool mmx_ok
, enum machine_mode mode
,
21147 rtx target
, rtx vals
, int one_var
)
21149 rtx var
= XVECEXP (vals
, 0, one_var
);
21150 enum machine_mode wmode
;
21153 const_vec
= copy_rtx (vals
);
21154 XVECEXP (const_vec
, 0, one_var
) = CONST0_RTX (GET_MODE_INNER (mode
));
21155 const_vec
= gen_rtx_CONST_VECTOR (mode
, XVEC (const_vec
, 0));
21163 /* For the two element vectors, it's just as easy to use
21164 the general case. */
21180 /* There's no way to set one QImode entry easily. Combine
21181 the variable value with its adjacent constant value, and
21182 promote to an HImode set. */
21183 x
= XVECEXP (vals
, 0, one_var
^ 1);
21186 var
= convert_modes (HImode
, QImode
, var
, true);
21187 var
= expand_simple_binop (HImode
, ASHIFT
, var
, GEN_INT (8),
21188 NULL_RTX
, 1, OPTAB_LIB_WIDEN
);
21189 x
= GEN_INT (INTVAL (x
) & 0xff);
21193 var
= convert_modes (HImode
, QImode
, var
, true);
21194 x
= gen_int_mode (INTVAL (x
) << 8, HImode
);
21196 if (x
!= const0_rtx
)
21197 var
= expand_simple_binop (HImode
, IOR
, var
, x
, var
,
21198 1, OPTAB_LIB_WIDEN
);
21200 x
= gen_reg_rtx (wmode
);
21201 emit_move_insn (x
, gen_lowpart (wmode
, const_vec
));
21202 ix86_expand_vector_set (mmx_ok
, x
, var
, one_var
>> 1);
21204 emit_move_insn (target
, gen_lowpart (mode
, x
));
21211 emit_move_insn (target
, const_vec
);
21212 ix86_expand_vector_set (mmx_ok
, target
, var
, one_var
);
21216 /* A subroutine of ix86_expand_vector_init. Handle the most general case:
21217 all values variable, and none identical. */
21220 ix86_expand_vector_init_general (bool mmx_ok
, enum machine_mode mode
,
21221 rtx target
, rtx vals
)
21223 enum machine_mode half_mode
= GET_MODE_INNER (mode
);
21224 rtx op0
= NULL
, op1
= NULL
;
21225 bool use_vec_concat
= false;
21231 if (!mmx_ok
&& !TARGET_SSE
)
21237 /* For the two element vectors, we always implement VEC_CONCAT. */
21238 op0
= XVECEXP (vals
, 0, 0);
21239 op1
= XVECEXP (vals
, 0, 1);
21240 use_vec_concat
= true;
21244 half_mode
= V2SFmode
;
21247 half_mode
= V2SImode
;
21253 /* For V4SF and V4SI, we implement a concat of two V2 vectors.
21254 Recurse to load the two halves. */
21256 op0
= gen_reg_rtx (half_mode
);
21257 v
= gen_rtvec (2, XVECEXP (vals
, 0, 0), XVECEXP (vals
, 0, 1));
21258 ix86_expand_vector_init (false, op0
, gen_rtx_PARALLEL (half_mode
, v
));
21260 op1
= gen_reg_rtx (half_mode
);
21261 v
= gen_rtvec (2, XVECEXP (vals
, 0, 2), XVECEXP (vals
, 0, 3));
21262 ix86_expand_vector_init (false, op1
, gen_rtx_PARALLEL (half_mode
, v
));
21264 use_vec_concat
= true;
21275 gcc_unreachable ();
21278 if (use_vec_concat
)
21280 if (!register_operand (op0
, half_mode
))
21281 op0
= force_reg (half_mode
, op0
);
21282 if (!register_operand (op1
, half_mode
))
21283 op1
= force_reg (half_mode
, op1
);
21285 emit_insn (gen_rtx_SET (VOIDmode
, target
,
21286 gen_rtx_VEC_CONCAT (mode
, op0
, op1
)));
21290 int i
, j
, n_elts
, n_words
, n_elt_per_word
;
21291 enum machine_mode inner_mode
;
21292 rtx words
[4], shift
;
21294 inner_mode
= GET_MODE_INNER (mode
);
21295 n_elts
= GET_MODE_NUNITS (mode
);
21296 n_words
= GET_MODE_SIZE (mode
) / UNITS_PER_WORD
;
21297 n_elt_per_word
= n_elts
/ n_words
;
21298 shift
= GEN_INT (GET_MODE_BITSIZE (inner_mode
));
21300 for (i
= 0; i
< n_words
; ++i
)
21302 rtx word
= NULL_RTX
;
21304 for (j
= 0; j
< n_elt_per_word
; ++j
)
21306 rtx elt
= XVECEXP (vals
, 0, (i
+1)*n_elt_per_word
- j
- 1);
21307 elt
= convert_modes (word_mode
, inner_mode
, elt
, true);
21313 word
= expand_simple_binop (word_mode
, ASHIFT
, word
, shift
,
21314 word
, 1, OPTAB_LIB_WIDEN
);
21315 word
= expand_simple_binop (word_mode
, IOR
, word
, elt
,
21316 word
, 1, OPTAB_LIB_WIDEN
);
21324 emit_move_insn (target
, gen_lowpart (mode
, words
[0]));
21325 else if (n_words
== 2)
21327 rtx tmp
= gen_reg_rtx (mode
);
21328 emit_insn (gen_rtx_CLOBBER (VOIDmode
, tmp
));
21329 emit_move_insn (gen_lowpart (word_mode
, tmp
), words
[0]);
21330 emit_move_insn (gen_highpart (word_mode
, tmp
), words
[1]);
21331 emit_move_insn (target
, tmp
);
21333 else if (n_words
== 4)
21335 rtx tmp
= gen_reg_rtx (V4SImode
);
21336 vals
= gen_rtx_PARALLEL (V4SImode
, gen_rtvec_v (4, words
));
21337 ix86_expand_vector_init_general (false, V4SImode
, tmp
, vals
);
21338 emit_move_insn (target
, gen_lowpart (mode
, tmp
));
21341 gcc_unreachable ();
21345 /* Initialize vector TARGET via VALS. Suppress the use of MMX
21346 instructions unless MMX_OK is true. */
21349 ix86_expand_vector_init (bool mmx_ok
, rtx target
, rtx vals
)
21351 enum machine_mode mode
= GET_MODE (target
);
21352 enum machine_mode inner_mode
= GET_MODE_INNER (mode
);
21353 int n_elts
= GET_MODE_NUNITS (mode
);
21354 int n_var
= 0, one_var
= -1;
21355 bool all_same
= true, all_const_zero
= true;
21359 for (i
= 0; i
< n_elts
; ++i
)
21361 x
= XVECEXP (vals
, 0, i
);
21362 if (!CONSTANT_P (x
))
21363 n_var
++, one_var
= i
;
21364 else if (x
!= CONST0_RTX (inner_mode
))
21365 all_const_zero
= false;
21366 if (i
> 0 && !rtx_equal_p (x
, XVECEXP (vals
, 0, 0)))
21370 /* Constants are best loaded from the constant pool. */
21373 emit_move_insn (target
, gen_rtx_CONST_VECTOR (mode
, XVEC (vals
, 0)));
21377 /* If all values are identical, broadcast the value. */
21379 && ix86_expand_vector_init_duplicate (mmx_ok
, mode
, target
,
21380 XVECEXP (vals
, 0, 0)))
21383 /* Values where only one field is non-constant are best loaded from
21384 the pool and overwritten via move later. */
21388 && ix86_expand_vector_init_one_nonzero (mmx_ok
, mode
, target
,
21389 XVECEXP (vals
, 0, one_var
),
21393 if (ix86_expand_vector_init_one_var (mmx_ok
, mode
, target
, vals
, one_var
))
21397 ix86_expand_vector_init_general (mmx_ok
, mode
, target
, vals
);
21401 ix86_expand_vector_set (bool mmx_ok
, rtx target
, rtx val
, int elt
)
21403 enum machine_mode mode
= GET_MODE (target
);
21404 enum machine_mode inner_mode
= GET_MODE_INNER (mode
);
21405 bool use_vec_merge
= false;
21414 tmp
= gen_reg_rtx (GET_MODE_INNER (mode
));
21415 ix86_expand_vector_extract (true, tmp
, target
, 1 - elt
);
21417 tmp
= gen_rtx_VEC_CONCAT (mode
, tmp
, val
);
21419 tmp
= gen_rtx_VEC_CONCAT (mode
, val
, tmp
);
21420 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
21426 use_vec_merge
= TARGET_SSE4_1
;
21434 /* For the two element vectors, we implement a VEC_CONCAT with
21435 the extraction of the other element. */
21437 tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (1, GEN_INT (1 - elt
)));
21438 tmp
= gen_rtx_VEC_SELECT (inner_mode
, target
, tmp
);
21441 op0
= val
, op1
= tmp
;
21443 op0
= tmp
, op1
= val
;
21445 tmp
= gen_rtx_VEC_CONCAT (mode
, op0
, op1
);
21446 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
21451 use_vec_merge
= TARGET_SSE4_1
;
21458 use_vec_merge
= true;
21462 /* tmp = target = A B C D */
21463 tmp
= copy_to_reg (target
);
21464 /* target = A A B B */
21465 emit_insn (gen_sse_unpcklps (target
, target
, target
));
21466 /* target = X A B B */
21467 ix86_expand_vector_set (false, target
, val
, 0);
21468 /* target = A X C D */
21469 emit_insn (gen_sse_shufps_1 (target
, target
, tmp
,
21470 GEN_INT (1), GEN_INT (0),
21471 GEN_INT (2+4), GEN_INT (3+4)));
21475 /* tmp = target = A B C D */
21476 tmp
= copy_to_reg (target
);
21477 /* tmp = X B C D */
21478 ix86_expand_vector_set (false, tmp
, val
, 0);
21479 /* target = A B X D */
21480 emit_insn (gen_sse_shufps_1 (target
, target
, tmp
,
21481 GEN_INT (0), GEN_INT (1),
21482 GEN_INT (0+4), GEN_INT (3+4)));
21486 /* tmp = target = A B C D */
21487 tmp
= copy_to_reg (target
);
21488 /* tmp = X B C D */
21489 ix86_expand_vector_set (false, tmp
, val
, 0);
21490 /* target = A B X D */
21491 emit_insn (gen_sse_shufps_1 (target
, target
, tmp
,
21492 GEN_INT (0), GEN_INT (1),
21493 GEN_INT (2+4), GEN_INT (0+4)));
21497 gcc_unreachable ();
21502 use_vec_merge
= TARGET_SSE4_1
;
21506 /* Element 0 handled by vec_merge below. */
21509 use_vec_merge
= true;
21515 /* With SSE2, use integer shuffles to swap element 0 and ELT,
21516 store into element 0, then shuffle them back. */
21520 order
[0] = GEN_INT (elt
);
21521 order
[1] = const1_rtx
;
21522 order
[2] = const2_rtx
;
21523 order
[3] = GEN_INT (3);
21524 order
[elt
] = const0_rtx
;
21526 emit_insn (gen_sse2_pshufd_1 (target
, target
, order
[0],
21527 order
[1], order
[2], order
[3]));
21529 ix86_expand_vector_set (false, target
, val
, 0);
21531 emit_insn (gen_sse2_pshufd_1 (target
, target
, order
[0],
21532 order
[1], order
[2], order
[3]));
21536 /* For SSE1, we have to reuse the V4SF code. */
21537 ix86_expand_vector_set (false, gen_lowpart (V4SFmode
, target
),
21538 gen_lowpart (SFmode
, val
), elt
);
21543 use_vec_merge
= TARGET_SSE2
;
21546 use_vec_merge
= mmx_ok
&& (TARGET_SSE
|| TARGET_3DNOW_A
);
21550 use_vec_merge
= TARGET_SSE4_1
;
21560 tmp
= gen_rtx_VEC_DUPLICATE (mode
, val
);
21561 tmp
= gen_rtx_VEC_MERGE (mode
, tmp
, target
, GEN_INT (1 << elt
));
21562 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
21566 rtx mem
= assign_stack_temp (mode
, GET_MODE_SIZE (mode
), false);
21568 emit_move_insn (mem
, target
);
21570 tmp
= adjust_address (mem
, inner_mode
, elt
*GET_MODE_SIZE (inner_mode
));
21571 emit_move_insn (tmp
, val
);
21573 emit_move_insn (target
, mem
);
21578 ix86_expand_vector_extract (bool mmx_ok
, rtx target
, rtx vec
, int elt
)
21580 enum machine_mode mode
= GET_MODE (vec
);
21581 enum machine_mode inner_mode
= GET_MODE_INNER (mode
);
21582 bool use_vec_extr
= false;
21595 use_vec_extr
= true;
21599 use_vec_extr
= TARGET_SSE4_1
;
21611 tmp
= gen_reg_rtx (mode
);
21612 emit_insn (gen_sse_shufps_1 (tmp
, vec
, vec
,
21613 GEN_INT (elt
), GEN_INT (elt
),
21614 GEN_INT (elt
+4), GEN_INT (elt
+4)));
21618 tmp
= gen_reg_rtx (mode
);
21619 emit_insn (gen_sse_unpckhps (tmp
, vec
, vec
));
21623 gcc_unreachable ();
21626 use_vec_extr
= true;
21631 use_vec_extr
= TARGET_SSE4_1
;
21645 tmp
= gen_reg_rtx (mode
);
21646 emit_insn (gen_sse2_pshufd_1 (tmp
, vec
,
21647 GEN_INT (elt
), GEN_INT (elt
),
21648 GEN_INT (elt
), GEN_INT (elt
)));
21652 tmp
= gen_reg_rtx (mode
);
21653 emit_insn (gen_sse2_punpckhdq (tmp
, vec
, vec
));
21657 gcc_unreachable ();
21660 use_vec_extr
= true;
21665 /* For SSE1, we have to reuse the V4SF code. */
21666 ix86_expand_vector_extract (false, gen_lowpart (SFmode
, target
),
21667 gen_lowpart (V4SFmode
, vec
), elt
);
21673 use_vec_extr
= TARGET_SSE2
;
21676 use_vec_extr
= mmx_ok
&& (TARGET_SSE
|| TARGET_3DNOW_A
);
21680 use_vec_extr
= TARGET_SSE4_1
;
21684 /* ??? Could extract the appropriate HImode element and shift. */
21691 tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (1, GEN_INT (elt
)));
21692 tmp
= gen_rtx_VEC_SELECT (inner_mode
, vec
, tmp
);
21694 /* Let the rtl optimizers know about the zero extension performed. */
21695 if (inner_mode
== QImode
|| inner_mode
== HImode
)
21697 tmp
= gen_rtx_ZERO_EXTEND (SImode
, tmp
);
21698 target
= gen_lowpart (SImode
, target
);
21701 emit_insn (gen_rtx_SET (VOIDmode
, target
, tmp
));
21705 rtx mem
= assign_stack_temp (mode
, GET_MODE_SIZE (mode
), false);
21707 emit_move_insn (mem
, vec
);
21709 tmp
= adjust_address (mem
, inner_mode
, elt
*GET_MODE_SIZE (inner_mode
));
21710 emit_move_insn (target
, tmp
);
21714 /* Expand a vector reduction on V4SFmode for SSE1. FN is the binary
21715 pattern to reduce; DEST is the destination; IN is the input vector. */
21718 ix86_expand_reduc_v4sf (rtx (*fn
) (rtx
, rtx
, rtx
), rtx dest
, rtx in
)
21720 rtx tmp1
, tmp2
, tmp3
;
21722 tmp1
= gen_reg_rtx (V4SFmode
);
21723 tmp2
= gen_reg_rtx (V4SFmode
);
21724 tmp3
= gen_reg_rtx (V4SFmode
);
21726 emit_insn (gen_sse_movhlps (tmp1
, in
, in
));
21727 emit_insn (fn (tmp2
, tmp1
, in
));
21729 emit_insn (gen_sse_shufps_1 (tmp3
, tmp2
, tmp2
,
21730 GEN_INT (1), GEN_INT (1),
21731 GEN_INT (1+4), GEN_INT (1+4)));
21732 emit_insn (fn (dest
, tmp2
, tmp3
));
21735 /* Target hook for scalar_mode_supported_p. */
21737 ix86_scalar_mode_supported_p (enum machine_mode mode
)
21739 if (DECIMAL_FLOAT_MODE_P (mode
))
21741 else if (mode
== TFmode
)
21742 return TARGET_64BIT
;
21744 return default_scalar_mode_supported_p (mode
);
21747 /* Implements target hook vector_mode_supported_p. */
21749 ix86_vector_mode_supported_p (enum machine_mode mode
)
21751 if (TARGET_SSE
&& VALID_SSE_REG_MODE (mode
))
21753 if (TARGET_SSE2
&& VALID_SSE2_REG_MODE (mode
))
21755 if (TARGET_MMX
&& VALID_MMX_REG_MODE (mode
))
21757 if (TARGET_3DNOW
&& VALID_MMX_REG_MODE_3DNOW (mode
))
21762 /* Worker function for TARGET_MD_ASM_CLOBBERS.
21764 We do this in the new i386 backend to maintain source compatibility
21765 with the old cc0-based compiler. */
21768 ix86_md_asm_clobbers (tree outputs ATTRIBUTE_UNUSED
,
21769 tree inputs ATTRIBUTE_UNUSED
,
21772 clobbers
= tree_cons (NULL_TREE
, build_string (5, "flags"),
21774 clobbers
= tree_cons (NULL_TREE
, build_string (4, "fpsr"),
21779 /* Implements target vector targetm.asm.encode_section_info. This
21780 is not used by netware. */
21782 static void ATTRIBUTE_UNUSED
21783 ix86_encode_section_info (tree decl
, rtx rtl
, int first
)
21785 default_encode_section_info (decl
, rtl
, first
);
21787 if (TREE_CODE (decl
) == VAR_DECL
21788 && (TREE_STATIC (decl
) || DECL_EXTERNAL (decl
))
21789 && ix86_in_large_data_p (decl
))
21790 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_FAR_ADDR
;
21793 /* Worker function for REVERSE_CONDITION. */
21796 ix86_reverse_condition (enum rtx_code code
, enum machine_mode mode
)
21798 return (mode
!= CCFPmode
&& mode
!= CCFPUmode
21799 ? reverse_condition (code
)
21800 : reverse_condition_maybe_unordered (code
));
21803 /* Output code to perform an x87 FP register move, from OPERANDS[1]
21807 output_387_reg_move (rtx insn
, rtx
*operands
)
21809 if (REG_P (operands
[0]))
21811 if (REG_P (operands
[1])
21812 && find_regno_note (insn
, REG_DEAD
, REGNO (operands
[1])))
21814 if (REGNO (operands
[0]) == FIRST_STACK_REG
)
21815 return output_387_ffreep (operands
, 0);
21816 return "fstp\t%y0";
21818 if (STACK_TOP_P (operands
[0]))
21819 return "fld%z1\t%y1";
21822 else if (MEM_P (operands
[0]))
21824 gcc_assert (REG_P (operands
[1]));
21825 if (find_regno_note (insn
, REG_DEAD
, REGNO (operands
[1])))
21826 return "fstp%z0\t%y0";
21829 /* There is no non-popping store to memory for XFmode.
21830 So if we need one, follow the store with a load. */
21831 if (GET_MODE (operands
[0]) == XFmode
)
21832 return "fstp%z0\t%y0\n\tfld%z0\t%y0";
21834 return "fst%z0\t%y0";
21841 /* Output code to perform a conditional jump to LABEL, if C2 flag in
21842 FP status register is set. */
21845 ix86_emit_fp_unordered_jump (rtx label
)
21847 rtx reg
= gen_reg_rtx (HImode
);
21850 emit_insn (gen_x86_fnstsw_1 (reg
));
21852 if (TARGET_SAHF
&& (TARGET_USE_SAHF
|| optimize_size
))
21854 emit_insn (gen_x86_sahf_1 (reg
));
21856 temp
= gen_rtx_REG (CCmode
, FLAGS_REG
);
21857 temp
= gen_rtx_UNORDERED (VOIDmode
, temp
, const0_rtx
);
21861 emit_insn (gen_testqi_ext_ccno_0 (reg
, GEN_INT (0x04)));
21863 temp
= gen_rtx_REG (CCNOmode
, FLAGS_REG
);
21864 temp
= gen_rtx_NE (VOIDmode
, temp
, const0_rtx
);
21867 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
21868 gen_rtx_LABEL_REF (VOIDmode
, label
),
21870 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
21872 emit_jump_insn (temp
);
21873 predict_jump (REG_BR_PROB_BASE
* 10 / 100);
21876 /* Output code to perform a log1p XFmode calculation. */
21878 void ix86_emit_i387_log1p (rtx op0
, rtx op1
)
21880 rtx label1
= gen_label_rtx ();
21881 rtx label2
= gen_label_rtx ();
21883 rtx tmp
= gen_reg_rtx (XFmode
);
21884 rtx tmp2
= gen_reg_rtx (XFmode
);
21886 emit_insn (gen_absxf2 (tmp
, op1
));
21887 emit_insn (gen_cmpxf (tmp
,
21888 CONST_DOUBLE_FROM_REAL_VALUE (
21889 REAL_VALUE_ATOF ("0.29289321881345247561810596348408353", XFmode
),
21891 emit_jump_insn (gen_bge (label1
));
21893 emit_move_insn (tmp2
, standard_80387_constant_rtx (4)); /* fldln2 */
21894 emit_insn (gen_fyl2xp1xf3_i387 (op0
, op1
, tmp2
));
21895 emit_jump (label2
);
21897 emit_label (label1
);
21898 emit_move_insn (tmp
, CONST1_RTX (XFmode
));
21899 emit_insn (gen_addxf3 (tmp
, op1
, tmp
));
21900 emit_move_insn (tmp2
, standard_80387_constant_rtx (4)); /* fldln2 */
21901 emit_insn (gen_fyl2xxf3_i387 (op0
, tmp
, tmp2
));
21903 emit_label (label2
);
21906 /* Solaris implementation of TARGET_ASM_NAMED_SECTION. */
21908 static void ATTRIBUTE_UNUSED
21909 i386_solaris_elf_named_section (const char *name
, unsigned int flags
,
21912 /* With Binutils 2.15, the "@unwind" marker must be specified on
21913 every occurrence of the ".eh_frame" section, not just the first
21916 && strcmp (name
, ".eh_frame") == 0)
21918 fprintf (asm_out_file
, "\t.section\t%s,\"%s\",@unwind\n", name
,
21919 flags
& SECTION_WRITE
? "aw" : "a");
21922 default_elf_asm_named_section (name
, flags
, decl
);
21925 /* Return the mangling of TYPE if it is an extended fundamental type. */
21927 static const char *
21928 ix86_mangle_fundamental_type (tree type
)
21930 switch (TYPE_MODE (type
))
21933 /* __float128 is "g". */
21936 /* "long double" or __float80 is "e". */
21943 /* For 32-bit code we can save PIC register setup by using
21944 __stack_chk_fail_local hidden function instead of calling
21945 __stack_chk_fail directly. 64-bit code doesn't need to setup any PIC
21946 register, so it is better to call __stack_chk_fail directly. */
21949 ix86_stack_protect_fail (void)
21951 return TARGET_64BIT
21952 ? default_external_stack_protect_fail ()
21953 : default_hidden_stack_protect_fail ();
21956 /* Select a format to encode pointers in exception handling data. CODE
21957 is 0 for data, 1 for code labels, 2 for function pointers. GLOBAL is
21958 true if the symbol may be affected by dynamic relocations.
21960 ??? All x86 object file formats are capable of representing this.
21961 After all, the relocation needed is the same as for the call insn.
21962 Whether or not a particular assembler allows us to enter such, I
21963 guess we'll have to see. */
21965 asm_preferred_eh_data_format (int code
, int global
)
21969 int type
= DW_EH_PE_sdata8
;
21971 || ix86_cmodel
== CM_SMALL_PIC
21972 || (ix86_cmodel
== CM_MEDIUM_PIC
&& (global
|| code
)))
21973 type
= DW_EH_PE_sdata4
;
21974 return (global
? DW_EH_PE_indirect
: 0) | DW_EH_PE_pcrel
| type
;
21976 if (ix86_cmodel
== CM_SMALL
21977 || (ix86_cmodel
== CM_MEDIUM
&& code
))
21978 return DW_EH_PE_udata4
;
21979 return DW_EH_PE_absptr
;
21982 /* Expand copysign from SIGN to the positive value ABS_VALUE
21983 storing in RESULT. If MASK is non-null, it shall be a mask to mask out
21986 ix86_sse_copysign_to_positive (rtx result
, rtx abs_value
, rtx sign
, rtx mask
)
21988 enum machine_mode mode
= GET_MODE (sign
);
21989 rtx sgn
= gen_reg_rtx (mode
);
21990 if (mask
== NULL_RTX
)
21992 mask
= ix86_build_signbit_mask (mode
, VECTOR_MODE_P (mode
), false);
21993 if (!VECTOR_MODE_P (mode
))
21995 /* We need to generate a scalar mode mask in this case. */
21996 rtx tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (1, const0_rtx
));
21997 tmp
= gen_rtx_VEC_SELECT (mode
, mask
, tmp
);
21998 mask
= gen_reg_rtx (mode
);
21999 emit_insn (gen_rtx_SET (VOIDmode
, mask
, tmp
));
22003 mask
= gen_rtx_NOT (mode
, mask
);
22004 emit_insn (gen_rtx_SET (VOIDmode
, sgn
,
22005 gen_rtx_AND (mode
, mask
, sign
)));
22006 emit_insn (gen_rtx_SET (VOIDmode
, result
,
22007 gen_rtx_IOR (mode
, abs_value
, sgn
)));
22010 /* Expand fabs (OP0) and return a new rtx that holds the result. The
22011 mask for masking out the sign-bit is stored in *SMASK, if that is
22014 ix86_expand_sse_fabs (rtx op0
, rtx
*smask
)
22016 enum machine_mode mode
= GET_MODE (op0
);
22019 xa
= gen_reg_rtx (mode
);
22020 mask
= ix86_build_signbit_mask (mode
, VECTOR_MODE_P (mode
), true);
22021 if (!VECTOR_MODE_P (mode
))
22023 /* We need to generate a scalar mode mask in this case. */
22024 rtx tmp
= gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (1, const0_rtx
));
22025 tmp
= gen_rtx_VEC_SELECT (mode
, mask
, tmp
);
22026 mask
= gen_reg_rtx (mode
);
22027 emit_insn (gen_rtx_SET (VOIDmode
, mask
, tmp
));
22029 emit_insn (gen_rtx_SET (VOIDmode
, xa
,
22030 gen_rtx_AND (mode
, op0
, mask
)));
22038 /* Expands a comparison of OP0 with OP1 using comparison code CODE,
22039 swapping the operands if SWAP_OPERANDS is true. The expanded
22040 code is a forward jump to a newly created label in case the
22041 comparison is true. The generated label rtx is returned. */
22043 ix86_expand_sse_compare_and_jump (enum rtx_code code
, rtx op0
, rtx op1
,
22044 bool swap_operands
)
22055 label
= gen_label_rtx ();
22056 tmp
= gen_rtx_REG (CCFPUmode
, FLAGS_REG
);
22057 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
22058 gen_rtx_COMPARE (CCFPUmode
, op0
, op1
)));
22059 tmp
= gen_rtx_fmt_ee (code
, VOIDmode
, tmp
, const0_rtx
);
22060 tmp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, tmp
,
22061 gen_rtx_LABEL_REF (VOIDmode
, label
), pc_rtx
);
22062 tmp
= emit_jump_insn (gen_rtx_SET (VOIDmode
, pc_rtx
, tmp
));
22063 JUMP_LABEL (tmp
) = label
;
22068 /* Expand a mask generating SSE comparison instruction comparing OP0 with OP1
22069 using comparison code CODE. Operands are swapped for the comparison if
22070 SWAP_OPERANDS is true. Returns a rtx for the generated mask. */
22072 ix86_expand_sse_compare_mask (enum rtx_code code
, rtx op0
, rtx op1
,
22073 bool swap_operands
)
22075 enum machine_mode mode
= GET_MODE (op0
);
22076 rtx mask
= gen_reg_rtx (mode
);
22085 if (mode
== DFmode
)
22086 emit_insn (gen_sse2_maskcmpdf3 (mask
, op0
, op1
,
22087 gen_rtx_fmt_ee (code
, mode
, op0
, op1
)));
22089 emit_insn (gen_sse_maskcmpsf3 (mask
, op0
, op1
,
22090 gen_rtx_fmt_ee (code
, mode
, op0
, op1
)));
22095 /* Generate and return a rtx of mode MODE for 2**n where n is the number
22096 of bits of the mantissa of MODE, which must be one of DFmode or SFmode. */
22098 ix86_gen_TWO52 (enum machine_mode mode
)
22100 REAL_VALUE_TYPE TWO52r
;
22103 real_ldexp (&TWO52r
, &dconst1
, mode
== DFmode
? 52 : 23);
22104 TWO52
= const_double_from_real_value (TWO52r
, mode
);
22105 TWO52
= force_reg (mode
, TWO52
);
22110 /* Expand SSE sequence for computing lround from OP1 storing
22113 ix86_expand_lround (rtx op0
, rtx op1
)
22115 /* C code for the stuff we're doing below:
22116 tmp = op1 + copysign (nextafter (0.5, 0.0), op1)
22119 enum machine_mode mode
= GET_MODE (op1
);
22120 const struct real_format
*fmt
;
22121 REAL_VALUE_TYPE pred_half
, half_minus_pred_half
;
22124 /* load nextafter (0.5, 0.0) */
22125 fmt
= REAL_MODE_FORMAT (mode
);
22126 real_2expN (&half_minus_pred_half
, -(fmt
->p
) - 1);
22127 REAL_ARITHMETIC (pred_half
, MINUS_EXPR
, dconsthalf
, half_minus_pred_half
);
22129 /* adj = copysign (0.5, op1) */
22130 adj
= force_reg (mode
, const_double_from_real_value (pred_half
, mode
));
22131 ix86_sse_copysign_to_positive (adj
, adj
, force_reg (mode
, op1
), NULL_RTX
);
22133 /* adj = op1 + adj */
22134 adj
= expand_simple_binop (mode
, PLUS
, adj
, op1
, NULL_RTX
, 0, OPTAB_DIRECT
);
22136 /* op0 = (imode)adj */
22137 expand_fix (op0
, adj
, 0);
22140 /* Expand SSE2 sequence for computing lround from OPERAND1 storing
22143 ix86_expand_lfloorceil (rtx op0
, rtx op1
, bool do_floor
)
22145 /* C code for the stuff we're doing below (for do_floor):
22147 xi -= (double)xi > op1 ? 1 : 0;
22150 enum machine_mode fmode
= GET_MODE (op1
);
22151 enum machine_mode imode
= GET_MODE (op0
);
22152 rtx ireg
, freg
, label
, tmp
;
22154 /* reg = (long)op1 */
22155 ireg
= gen_reg_rtx (imode
);
22156 expand_fix (ireg
, op1
, 0);
22158 /* freg = (double)reg */
22159 freg
= gen_reg_rtx (fmode
);
22160 expand_float (freg
, ireg
, 0);
22162 /* ireg = (freg > op1) ? ireg - 1 : ireg */
22163 label
= ix86_expand_sse_compare_and_jump (UNLE
,
22164 freg
, op1
, !do_floor
);
22165 tmp
= expand_simple_binop (imode
, do_floor
? MINUS
: PLUS
,
22166 ireg
, const1_rtx
, NULL_RTX
, 0, OPTAB_DIRECT
);
22167 emit_move_insn (ireg
, tmp
);
22169 emit_label (label
);
22170 LABEL_NUSES (label
) = 1;
22172 emit_move_insn (op0
, ireg
);
22175 /* Expand rint (IEEE round to nearest) rounding OPERAND1 and storing the
22176 result in OPERAND0. */
22178 ix86_expand_rint (rtx operand0
, rtx operand1
)
22180 /* C code for the stuff we're doing below:
22181 xa = fabs (operand1);
22182 if (!isless (xa, 2**52))
22184 xa = xa + 2**52 - 2**52;
22185 return copysign (xa, operand1);
22187 enum machine_mode mode
= GET_MODE (operand0
);
22188 rtx res
, xa
, label
, TWO52
, mask
;
22190 res
= gen_reg_rtx (mode
);
22191 emit_move_insn (res
, operand1
);
22193 /* xa = abs (operand1) */
22194 xa
= ix86_expand_sse_fabs (res
, &mask
);
22196 /* if (!isless (xa, TWO52)) goto label; */
22197 TWO52
= ix86_gen_TWO52 (mode
);
22198 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22200 xa
= expand_simple_binop (mode
, PLUS
, xa
, TWO52
, NULL_RTX
, 0, OPTAB_DIRECT
);
22201 xa
= expand_simple_binop (mode
, MINUS
, xa
, TWO52
, xa
, 0, OPTAB_DIRECT
);
22203 ix86_sse_copysign_to_positive (res
, xa
, res
, mask
);
22205 emit_label (label
);
22206 LABEL_NUSES (label
) = 1;
22208 emit_move_insn (operand0
, res
);
22211 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
22214 ix86_expand_floorceildf_32 (rtx operand0
, rtx operand1
, bool do_floor
)
22216 /* C code for the stuff we expand below.
22217 double xa = fabs (x), x2;
22218 if (!isless (xa, TWO52))
22220 xa = xa + TWO52 - TWO52;
22221 x2 = copysign (xa, x);
22230 enum machine_mode mode
= GET_MODE (operand0
);
22231 rtx xa
, TWO52
, tmp
, label
, one
, res
, mask
;
22233 TWO52
= ix86_gen_TWO52 (mode
);
22235 /* Temporary for holding the result, initialized to the input
22236 operand to ease control flow. */
22237 res
= gen_reg_rtx (mode
);
22238 emit_move_insn (res
, operand1
);
22240 /* xa = abs (operand1) */
22241 xa
= ix86_expand_sse_fabs (res
, &mask
);
22243 /* if (!isless (xa, TWO52)) goto label; */
22244 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22246 /* xa = xa + TWO52 - TWO52; */
22247 xa
= expand_simple_binop (mode
, PLUS
, xa
, TWO52
, NULL_RTX
, 0, OPTAB_DIRECT
);
22248 xa
= expand_simple_binop (mode
, MINUS
, xa
, TWO52
, xa
, 0, OPTAB_DIRECT
);
22250 /* xa = copysign (xa, operand1) */
22251 ix86_sse_copysign_to_positive (xa
, xa
, res
, mask
);
22253 /* generate 1.0 or -1.0 */
22254 one
= force_reg (mode
,
22255 const_double_from_real_value (do_floor
22256 ? dconst1
: dconstm1
, mode
));
22258 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
22259 tmp
= ix86_expand_sse_compare_mask (UNGT
, xa
, res
, !do_floor
);
22260 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
22261 gen_rtx_AND (mode
, one
, tmp
)));
22262 /* We always need to subtract here to preserve signed zero. */
22263 tmp
= expand_simple_binop (mode
, MINUS
,
22264 xa
, tmp
, NULL_RTX
, 0, OPTAB_DIRECT
);
22265 emit_move_insn (res
, tmp
);
22267 emit_label (label
);
22268 LABEL_NUSES (label
) = 1;
22270 emit_move_insn (operand0
, res
);
22273 /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing
22276 ix86_expand_floorceil (rtx operand0
, rtx operand1
, bool do_floor
)
22278 /* C code for the stuff we expand below.
22279 double xa = fabs (x), x2;
22280 if (!isless (xa, TWO52))
22282 x2 = (double)(long)x;
22289 if (HONOR_SIGNED_ZEROS (mode))
22290 return copysign (x2, x);
22293 enum machine_mode mode
= GET_MODE (operand0
);
22294 rtx xa
, xi
, TWO52
, tmp
, label
, one
, res
, mask
;
22296 TWO52
= ix86_gen_TWO52 (mode
);
22298 /* Temporary for holding the result, initialized to the input
22299 operand to ease control flow. */
22300 res
= gen_reg_rtx (mode
);
22301 emit_move_insn (res
, operand1
);
22303 /* xa = abs (operand1) */
22304 xa
= ix86_expand_sse_fabs (res
, &mask
);
22306 /* if (!isless (xa, TWO52)) goto label; */
22307 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22309 /* xa = (double)(long)x */
22310 xi
= gen_reg_rtx (mode
== DFmode
? DImode
: SImode
);
22311 expand_fix (xi
, res
, 0);
22312 expand_float (xa
, xi
, 0);
22315 one
= force_reg (mode
, const_double_from_real_value (dconst1
, mode
));
22317 /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */
22318 tmp
= ix86_expand_sse_compare_mask (UNGT
, xa
, res
, !do_floor
);
22319 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
22320 gen_rtx_AND (mode
, one
, tmp
)));
22321 tmp
= expand_simple_binop (mode
, do_floor
? MINUS
: PLUS
,
22322 xa
, tmp
, NULL_RTX
, 0, OPTAB_DIRECT
);
22323 emit_move_insn (res
, tmp
);
22325 if (HONOR_SIGNED_ZEROS (mode
))
22326 ix86_sse_copysign_to_positive (res
, res
, force_reg (mode
, operand1
), mask
);
22328 emit_label (label
);
22329 LABEL_NUSES (label
) = 1;
22331 emit_move_insn (operand0
, res
);
22334 /* Expand SSE sequence for computing round from OPERAND1 storing
22335 into OPERAND0. Sequence that works without relying on DImode truncation
22336 via cvttsd2siq that is only available on 64bit targets. */
22338 ix86_expand_rounddf_32 (rtx operand0
, rtx operand1
)
22340 /* C code for the stuff we expand below.
22341 double xa = fabs (x), xa2, x2;
22342 if (!isless (xa, TWO52))
22344 Using the absolute value and copying back sign makes
22345 -0.0 -> -0.0 correct.
22346 xa2 = xa + TWO52 - TWO52;
22351 else if (dxa > 0.5)
22353 x2 = copysign (xa2, x);
22356 enum machine_mode mode
= GET_MODE (operand0
);
22357 rtx xa
, xa2
, dxa
, TWO52
, tmp
, label
, half
, mhalf
, one
, res
, mask
;
22359 TWO52
= ix86_gen_TWO52 (mode
);
22361 /* Temporary for holding the result, initialized to the input
22362 operand to ease control flow. */
22363 res
= gen_reg_rtx (mode
);
22364 emit_move_insn (res
, operand1
);
22366 /* xa = abs (operand1) */
22367 xa
= ix86_expand_sse_fabs (res
, &mask
);
22369 /* if (!isless (xa, TWO52)) goto label; */
22370 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22372 /* xa2 = xa + TWO52 - TWO52; */
22373 xa2
= expand_simple_binop (mode
, PLUS
, xa
, TWO52
, NULL_RTX
, 0, OPTAB_DIRECT
);
22374 xa2
= expand_simple_binop (mode
, MINUS
, xa2
, TWO52
, xa2
, 0, OPTAB_DIRECT
);
22376 /* dxa = xa2 - xa; */
22377 dxa
= expand_simple_binop (mode
, MINUS
, xa2
, xa
, NULL_RTX
, 0, OPTAB_DIRECT
);
22379 /* generate 0.5, 1.0 and -0.5 */
22380 half
= force_reg (mode
, const_double_from_real_value (dconsthalf
, mode
));
22381 one
= expand_simple_binop (mode
, PLUS
, half
, half
, NULL_RTX
, 0, OPTAB_DIRECT
);
22382 mhalf
= expand_simple_binop (mode
, MINUS
, half
, one
, NULL_RTX
,
22386 tmp
= gen_reg_rtx (mode
);
22387 /* xa2 = xa2 - (dxa > 0.5 ? 1 : 0) */
22388 tmp
= ix86_expand_sse_compare_mask (UNGT
, dxa
, half
, false);
22389 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
22390 gen_rtx_AND (mode
, one
, tmp
)));
22391 xa2
= expand_simple_binop (mode
, MINUS
, xa2
, tmp
, NULL_RTX
, 0, OPTAB_DIRECT
);
22392 /* xa2 = xa2 + (dxa <= -0.5 ? 1 : 0) */
22393 tmp
= ix86_expand_sse_compare_mask (UNGE
, mhalf
, dxa
, false);
22394 emit_insn (gen_rtx_SET (VOIDmode
, tmp
,
22395 gen_rtx_AND (mode
, one
, tmp
)));
22396 xa2
= expand_simple_binop (mode
, PLUS
, xa2
, tmp
, NULL_RTX
, 0, OPTAB_DIRECT
);
22398 /* res = copysign (xa2, operand1) */
22399 ix86_sse_copysign_to_positive (res
, xa2
, force_reg (mode
, operand1
), mask
);
22401 emit_label (label
);
22402 LABEL_NUSES (label
) = 1;
22404 emit_move_insn (operand0
, res
);
22407 /* Expand SSE sequence for computing trunc from OPERAND1 storing
22410 ix86_expand_trunc (rtx operand0
, rtx operand1
)
22412 /* C code for SSE variant we expand below.
22413 double xa = fabs (x), x2;
22414 if (!isless (xa, TWO52))
22416 x2 = (double)(long)x;
22417 if (HONOR_SIGNED_ZEROS (mode))
22418 return copysign (x2, x);
22421 enum machine_mode mode
= GET_MODE (operand0
);
22422 rtx xa
, xi
, TWO52
, label
, res
, mask
;
22424 TWO52
= ix86_gen_TWO52 (mode
);
22426 /* Temporary for holding the result, initialized to the input
22427 operand to ease control flow. */
22428 res
= gen_reg_rtx (mode
);
22429 emit_move_insn (res
, operand1
);
22431 /* xa = abs (operand1) */
22432 xa
= ix86_expand_sse_fabs (res
, &mask
);
22434 /* if (!isless (xa, TWO52)) goto label; */
22435 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22437 /* x = (double)(long)x */
22438 xi
= gen_reg_rtx (mode
== DFmode
? DImode
: SImode
);
22439 expand_fix (xi
, res
, 0);
22440 expand_float (res
, xi
, 0);
22442 if (HONOR_SIGNED_ZEROS (mode
))
22443 ix86_sse_copysign_to_positive (res
, res
, force_reg (mode
, operand1
), mask
);
22445 emit_label (label
);
22446 LABEL_NUSES (label
) = 1;
22448 emit_move_insn (operand0
, res
);
22451 /* Expand SSE sequence for computing trunc from OPERAND1 storing
22454 ix86_expand_truncdf_32 (rtx operand0
, rtx operand1
)
22456 enum machine_mode mode
= GET_MODE (operand0
);
22457 rtx xa
, mask
, TWO52
, label
, one
, res
, smask
, tmp
;
22459 /* C code for SSE variant we expand below.
22460 double xa = fabs (x), x2;
22461 if (!isless (xa, TWO52))
22463 xa2 = xa + TWO52 - TWO52;
22467 x2 = copysign (xa2, x);
22471 TWO52
= ix86_gen_TWO52 (mode
);
22473 /* Temporary for holding the result, initialized to the input
22474 operand to ease control flow. */
22475 res
= gen_reg_rtx (mode
);
22476 emit_move_insn (res
, operand1
);
22478 /* xa = abs (operand1) */
22479 xa
= ix86_expand_sse_fabs (res
, &smask
);
22481 /* if (!isless (xa, TWO52)) goto label; */
22482 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22484 /* res = xa + TWO52 - TWO52; */
22485 tmp
= expand_simple_binop (mode
, PLUS
, xa
, TWO52
, NULL_RTX
, 0, OPTAB_DIRECT
);
22486 tmp
= expand_simple_binop (mode
, MINUS
, tmp
, TWO52
, tmp
, 0, OPTAB_DIRECT
);
22487 emit_move_insn (res
, tmp
);
22490 one
= force_reg (mode
, const_double_from_real_value (dconst1
, mode
));
22492 /* Compensate: res = xa2 - (res > xa ? 1 : 0) */
22493 mask
= ix86_expand_sse_compare_mask (UNGT
, res
, xa
, false);
22494 emit_insn (gen_rtx_SET (VOIDmode
, mask
,
22495 gen_rtx_AND (mode
, mask
, one
)));
22496 tmp
= expand_simple_binop (mode
, MINUS
,
22497 res
, mask
, NULL_RTX
, 0, OPTAB_DIRECT
);
22498 emit_move_insn (res
, tmp
);
22500 /* res = copysign (res, operand1) */
22501 ix86_sse_copysign_to_positive (res
, res
, force_reg (mode
, operand1
), smask
);
22503 emit_label (label
);
22504 LABEL_NUSES (label
) = 1;
22506 emit_move_insn (operand0
, res
);
22509 /* Expand SSE sequence for computing round from OPERAND1 storing
22512 ix86_expand_round (rtx operand0
, rtx operand1
)
22514 /* C code for the stuff we're doing below:
22515 double xa = fabs (x);
22516 if (!isless (xa, TWO52))
22518 xa = (double)(long)(xa + nextafter (0.5, 0.0));
22519 return copysign (xa, x);
22521 enum machine_mode mode
= GET_MODE (operand0
);
22522 rtx res
, TWO52
, xa
, label
, xi
, half
, mask
;
22523 const struct real_format
*fmt
;
22524 REAL_VALUE_TYPE pred_half
, half_minus_pred_half
;
22526 /* Temporary for holding the result, initialized to the input
22527 operand to ease control flow. */
22528 res
= gen_reg_rtx (mode
);
22529 emit_move_insn (res
, operand1
);
22531 TWO52
= ix86_gen_TWO52 (mode
);
22532 xa
= ix86_expand_sse_fabs (res
, &mask
);
22533 label
= ix86_expand_sse_compare_and_jump (UNLE
, TWO52
, xa
, false);
22535 /* load nextafter (0.5, 0.0) */
22536 fmt
= REAL_MODE_FORMAT (mode
);
22537 real_2expN (&half_minus_pred_half
, -(fmt
->p
) - 1);
22538 REAL_ARITHMETIC (pred_half
, MINUS_EXPR
, dconsthalf
, half_minus_pred_half
);
22540 /* xa = xa + 0.5 */
22541 half
= force_reg (mode
, const_double_from_real_value (pred_half
, mode
));
22542 xa
= expand_simple_binop (mode
, PLUS
, xa
, half
, NULL_RTX
, 0, OPTAB_DIRECT
);
22544 /* xa = (double)(int64_t)xa */
22545 xi
= gen_reg_rtx (mode
== DFmode
? DImode
: SImode
);
22546 expand_fix (xi
, xa
, 0);
22547 expand_float (xa
, xi
, 0);
22549 /* res = copysign (xa, operand1) */
22550 ix86_sse_copysign_to_positive (res
, xa
, force_reg (mode
, operand1
), mask
);
22552 emit_label (label
);
22553 LABEL_NUSES (label
) = 1;
22555 emit_move_insn (operand0
, res
);
22559 /* Table of valid machine attributes. */
22560 static const struct attribute_spec ix86_attribute_table
[] =
22562 /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
22563 /* Stdcall attribute says callee is responsible for popping arguments
22564 if they are not variable. */
22565 { "stdcall", 0, 0, false, true, true, ix86_handle_cconv_attribute
},
22566 /* Fastcall attribute says callee is responsible for popping arguments
22567 if they are not variable. */
22568 { "fastcall", 0, 0, false, true, true, ix86_handle_cconv_attribute
},
22569 /* Cdecl attribute says the callee is a normal C declaration */
22570 { "cdecl", 0, 0, false, true, true, ix86_handle_cconv_attribute
},
22571 /* Regparm attribute specifies how many integer arguments are to be
22572 passed in registers. */
22573 { "regparm", 1, 1, false, true, true, ix86_handle_cconv_attribute
},
22574 /* Sseregparm attribute says we are using x86_64 calling conventions
22575 for FP arguments. */
22576 { "sseregparm", 0, 0, false, true, true, ix86_handle_cconv_attribute
},
22577 /* force_align_arg_pointer says this function realigns the stack at entry. */
22578 { (const char *)&ix86_force_align_arg_pointer_string
, 0, 0,
22579 false, true, true, ix86_handle_cconv_attribute
},
22580 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
22581 { "dllimport", 0, 0, false, false, false, handle_dll_attribute
},
22582 { "dllexport", 0, 0, false, false, false, handle_dll_attribute
},
22583 { "shared", 0, 0, true, false, false, ix86_handle_shared_attribute
},
22585 { "ms_struct", 0, 0, false, false, false, ix86_handle_struct_attribute
},
22586 { "gcc_struct", 0, 0, false, false, false, ix86_handle_struct_attribute
},
22587 #ifdef SUBTARGET_ATTRIBUTE_TABLE
22588 SUBTARGET_ATTRIBUTE_TABLE
,
22590 { NULL
, 0, 0, false, false, false, NULL
}
22593 /* Initialize the GCC target structure. */
22594 #undef TARGET_ATTRIBUTE_TABLE
22595 #define TARGET_ATTRIBUTE_TABLE ix86_attribute_table
22596 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
22597 # undef TARGET_MERGE_DECL_ATTRIBUTES
22598 # define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes
22601 #undef TARGET_COMP_TYPE_ATTRIBUTES
22602 #define TARGET_COMP_TYPE_ATTRIBUTES ix86_comp_type_attributes
22604 #undef TARGET_INIT_BUILTINS
22605 #define TARGET_INIT_BUILTINS ix86_init_builtins
22606 #undef TARGET_EXPAND_BUILTIN
22607 #define TARGET_EXPAND_BUILTIN ix86_expand_builtin
22609 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
22610 #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION ix86_builtin_vectorized_function
22611 #undef TARGET_VECTORIZE_BUILTIN_CONVERSION
22612 #define TARGET_VECTORIZE_BUILTIN_CONVERSION ix86_builtin_conversion
22614 #undef TARGET_ASM_FUNCTION_EPILOGUE
22615 #define TARGET_ASM_FUNCTION_EPILOGUE ix86_output_function_epilogue
22617 #undef TARGET_ENCODE_SECTION_INFO
22618 #ifndef SUBTARGET_ENCODE_SECTION_INFO
22619 #define TARGET_ENCODE_SECTION_INFO ix86_encode_section_info
22621 #define TARGET_ENCODE_SECTION_INFO SUBTARGET_ENCODE_SECTION_INFO
22624 #undef TARGET_ASM_OPEN_PAREN
22625 #define TARGET_ASM_OPEN_PAREN ""
22626 #undef TARGET_ASM_CLOSE_PAREN
22627 #define TARGET_ASM_CLOSE_PAREN ""
22629 #undef TARGET_ASM_ALIGNED_HI_OP
22630 #define TARGET_ASM_ALIGNED_HI_OP ASM_SHORT
22631 #undef TARGET_ASM_ALIGNED_SI_OP
22632 #define TARGET_ASM_ALIGNED_SI_OP ASM_LONG
22634 #undef TARGET_ASM_ALIGNED_DI_OP
22635 #define TARGET_ASM_ALIGNED_DI_OP ASM_QUAD
22638 #undef TARGET_ASM_UNALIGNED_HI_OP
22639 #define TARGET_ASM_UNALIGNED_HI_OP TARGET_ASM_ALIGNED_HI_OP
22640 #undef TARGET_ASM_UNALIGNED_SI_OP
22641 #define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP
22642 #undef TARGET_ASM_UNALIGNED_DI_OP
22643 #define TARGET_ASM_UNALIGNED_DI_OP TARGET_ASM_ALIGNED_DI_OP
22645 #undef TARGET_SCHED_ADJUST_COST
22646 #define TARGET_SCHED_ADJUST_COST ix86_adjust_cost
22647 #undef TARGET_SCHED_ISSUE_RATE
22648 #define TARGET_SCHED_ISSUE_RATE ix86_issue_rate
22649 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
22650 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
22651 ia32_multipass_dfa_lookahead
22653 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
22654 #define TARGET_FUNCTION_OK_FOR_SIBCALL ix86_function_ok_for_sibcall
22657 #undef TARGET_HAVE_TLS
22658 #define TARGET_HAVE_TLS true
22660 #undef TARGET_CANNOT_FORCE_CONST_MEM
22661 #define TARGET_CANNOT_FORCE_CONST_MEM ix86_cannot_force_const_mem
22662 #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
22663 #define TARGET_USE_BLOCKS_FOR_CONSTANT_P hook_bool_mode_rtx_true
22665 #undef TARGET_DELEGITIMIZE_ADDRESS
22666 #define TARGET_DELEGITIMIZE_ADDRESS ix86_delegitimize_address
22668 #undef TARGET_MS_BITFIELD_LAYOUT_P
22669 #define TARGET_MS_BITFIELD_LAYOUT_P ix86_ms_bitfield_layout_p
22672 #undef TARGET_BINDS_LOCAL_P
22673 #define TARGET_BINDS_LOCAL_P darwin_binds_local_p
22675 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
22676 #undef TARGET_BINDS_LOCAL_P
22677 #define TARGET_BINDS_LOCAL_P i386_pe_binds_local_p
22680 #undef TARGET_ASM_OUTPUT_MI_THUNK
22681 #define TARGET_ASM_OUTPUT_MI_THUNK x86_output_mi_thunk
22682 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
22683 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK x86_can_output_mi_thunk
22685 #undef TARGET_ASM_FILE_START
22686 #define TARGET_ASM_FILE_START x86_file_start
22688 #undef TARGET_DEFAULT_TARGET_FLAGS
22689 #define TARGET_DEFAULT_TARGET_FLAGS \
22691 | TARGET_SUBTARGET_DEFAULT \
22692 | TARGET_TLS_DIRECT_SEG_REFS_DEFAULT)
22694 #undef TARGET_HANDLE_OPTION
22695 #define TARGET_HANDLE_OPTION ix86_handle_option
22697 #undef TARGET_RTX_COSTS
22698 #define TARGET_RTX_COSTS ix86_rtx_costs
22699 #undef TARGET_ADDRESS_COST
22700 #define TARGET_ADDRESS_COST ix86_address_cost
22702 #undef TARGET_FIXED_CONDITION_CODE_REGS
22703 #define TARGET_FIXED_CONDITION_CODE_REGS ix86_fixed_condition_code_regs
22704 #undef TARGET_CC_MODES_COMPATIBLE
22705 #define TARGET_CC_MODES_COMPATIBLE ix86_cc_modes_compatible
22707 #undef TARGET_MACHINE_DEPENDENT_REORG
22708 #define TARGET_MACHINE_DEPENDENT_REORG ix86_reorg
22710 #undef TARGET_BUILD_BUILTIN_VA_LIST
22711 #define TARGET_BUILD_BUILTIN_VA_LIST ix86_build_builtin_va_list
22713 #undef TARGET_MD_ASM_CLOBBERS
22714 #define TARGET_MD_ASM_CLOBBERS ix86_md_asm_clobbers
22716 #undef TARGET_PROMOTE_PROTOTYPES
22717 #define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
22718 #undef TARGET_STRUCT_VALUE_RTX
22719 #define TARGET_STRUCT_VALUE_RTX ix86_struct_value_rtx
22720 #undef TARGET_SETUP_INCOMING_VARARGS
22721 #define TARGET_SETUP_INCOMING_VARARGS ix86_setup_incoming_varargs
22722 #undef TARGET_MUST_PASS_IN_STACK
22723 #define TARGET_MUST_PASS_IN_STACK ix86_must_pass_in_stack
22724 #undef TARGET_PASS_BY_REFERENCE
22725 #define TARGET_PASS_BY_REFERENCE ix86_pass_by_reference
22726 #undef TARGET_INTERNAL_ARG_POINTER
22727 #define TARGET_INTERNAL_ARG_POINTER ix86_internal_arg_pointer
22728 #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC
22729 #define TARGET_DWARF_HANDLE_FRAME_UNSPEC ix86_dwarf_handle_frame_unspec
22730 #undef TARGET_STRICT_ARGUMENT_NAMING
22731 #define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
22733 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
22734 #define TARGET_GIMPLIFY_VA_ARG_EXPR ix86_gimplify_va_arg
22736 #undef TARGET_SCALAR_MODE_SUPPORTED_P
22737 #define TARGET_SCALAR_MODE_SUPPORTED_P ix86_scalar_mode_supported_p
22739 #undef TARGET_VECTOR_MODE_SUPPORTED_P
22740 #define TARGET_VECTOR_MODE_SUPPORTED_P ix86_vector_mode_supported_p
22743 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
22744 #define TARGET_ASM_OUTPUT_DWARF_DTPREL i386_output_dwarf_dtprel
22747 #ifdef SUBTARGET_INSERT_ATTRIBUTES
22748 #undef TARGET_INSERT_ATTRIBUTES
22749 #define TARGET_INSERT_ATTRIBUTES SUBTARGET_INSERT_ATTRIBUTES
22752 #undef TARGET_MANGLE_FUNDAMENTAL_TYPE
22753 #define TARGET_MANGLE_FUNDAMENTAL_TYPE ix86_mangle_fundamental_type
22755 #undef TARGET_STACK_PROTECT_FAIL
22756 #define TARGET_STACK_PROTECT_FAIL ix86_stack_protect_fail
22758 #undef TARGET_FUNCTION_VALUE
22759 #define TARGET_FUNCTION_VALUE ix86_function_value
22761 struct gcc_target targetm
= TARGET_INITIALIZER
;
22763 #include "gt-i386.h"