Concretize gimple_cond_set_{lhs|rhs}
[official-gcc.git] / gcc / lra-spills.c
blob8d494af38eb923513248a66638cd3f1112d031a6
1 /* Change pseudos by memory.
2 Copyright (C) 2010-2014 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* This file contains code for a pass to change spilled pseudos into
23 memory.
25 The pass creates necessary stack slots and assigns spilled pseudos
26 to the stack slots in following way:
28 for all spilled pseudos P most frequently used first do
29 for all stack slots S do
30 if P doesn't conflict with pseudos assigned to S then
31 assign S to P and goto to the next pseudo process
32 end
33 end
34 create new stack slot S and assign P to S
35 end
37 The actual algorithm is bit more complicated because of different
38 pseudo sizes.
40 After that the code changes spilled pseudos (except ones created
41 from scratches) by corresponding stack slot memory in RTL.
43 If at least one stack slot was created, we need to run more passes
44 because we have new addresses which should be checked and because
45 the old address displacements might change and address constraints
46 (or insn memory constraints) might not be satisfied any more.
48 For some targets, the pass can spill some pseudos into hard
49 registers of different class (usually into vector registers)
50 instead of spilling them into memory if it is possible and
51 profitable. Spilling GENERAL_REGS pseudo into SSE registers for
52 Intel Corei7 is an example of such optimization. And this is
53 actually recommended by Intel optimization guide.
55 The file also contains code for final change of pseudos on hard
56 regs correspondingly assigned to them. */
58 #include "config.h"
59 #include "system.h"
60 #include "coretypes.h"
61 #include "tm.h"
62 #include "rtl.h"
63 #include "tm_p.h"
64 #include "insn-config.h"
65 #include "recog.h"
66 #include "output.h"
67 #include "regs.h"
68 #include "hard-reg-set.h"
69 #include "flags.h"
70 #include "function.h"
71 #include "expr.h"
72 #include "basic-block.h"
73 #include "except.h"
74 #include "timevar.h"
75 #include "target.h"
76 #include "lra-int.h"
77 #include "ira.h"
78 #include "df.h"
81 /* Max regno at the start of the pass. */
82 static int regs_num;
84 /* Map spilled regno -> hard regno used instead of memory for
85 spilling. */
86 static rtx *spill_hard_reg;
88 /* The structure describes stack slot of a spilled pseudo. */
89 struct pseudo_slot
91 /* Number (0, 1, ...) of the stack slot to which given pseudo
92 belongs. */
93 int slot_num;
94 /* First or next slot with the same slot number. */
95 struct pseudo_slot *next, *first;
96 /* Memory representing the spilled pseudo. */
97 rtx mem;
100 /* The stack slots for each spilled pseudo. Indexed by regnos. */
101 static struct pseudo_slot *pseudo_slots;
103 /* The structure describes a register or a stack slot which can be
104 used for several spilled pseudos. */
105 struct slot
107 /* First pseudo with given stack slot. */
108 int regno;
109 /* Hard reg into which the slot pseudos are spilled. The value is
110 negative for pseudos spilled into memory. */
111 int hard_regno;
112 /* Memory representing the all stack slot. It can be different from
113 memory representing a pseudo belonging to give stack slot because
114 pseudo can be placed in a part of the corresponding stack slot.
115 The value is NULL for pseudos spilled into a hard reg. */
116 rtx mem;
117 /* Combined live ranges of all pseudos belonging to given slot. It
118 is used to figure out that a new spilled pseudo can use given
119 stack slot. */
120 lra_live_range_t live_ranges;
123 /* Array containing info about the stack slots. The array element is
124 indexed by the stack slot number in the range [0..slots_num). */
125 static struct slot *slots;
126 /* The number of the stack slots currently existing. */
127 static int slots_num;
129 /* Set up memory of the spilled pseudo I. The function can allocate
130 the corresponding stack slot if it is not done yet. */
131 static void
132 assign_mem_slot (int i)
134 rtx x = NULL_RTX;
135 enum machine_mode mode = GET_MODE (regno_reg_rtx[i]);
136 unsigned int inherent_size = PSEUDO_REGNO_BYTES (i);
137 unsigned int inherent_align = GET_MODE_ALIGNMENT (mode);
138 unsigned int max_ref_width = GET_MODE_SIZE (lra_reg_info[i].biggest_mode);
139 unsigned int total_size = MAX (inherent_size, max_ref_width);
140 unsigned int min_align = max_ref_width * BITS_PER_UNIT;
141 int adjust = 0;
143 lra_assert (regno_reg_rtx[i] != NULL_RTX && REG_P (regno_reg_rtx[i])
144 && lra_reg_info[i].nrefs != 0 && reg_renumber[i] < 0);
146 x = slots[pseudo_slots[i].slot_num].mem;
148 /* We can use a slot already allocated because it is guaranteed the
149 slot provides both enough inherent space and enough total
150 space. */
151 if (x)
153 /* Each pseudo has an inherent size which comes from its own mode,
154 and a total size which provides room for paradoxical subregs
155 which refer to the pseudo reg in wider modes. We allocate a new
156 slot, making sure that it has enough inherent space and total
157 space. */
158 else
160 rtx stack_slot;
162 /* No known place to spill from => no slot to reuse. */
163 x = assign_stack_local (mode, total_size,
164 min_align > inherent_align
165 || total_size > inherent_size ? -1 : 0);
166 stack_slot = x;
167 /* Cancel the big-endian correction done in assign_stack_local.
168 Get the address of the beginning of the slot. This is so we
169 can do a big-endian correction unconditionally below. */
170 if (BYTES_BIG_ENDIAN)
172 adjust = inherent_size - total_size;
173 if (adjust)
174 stack_slot
175 = adjust_address_nv (x,
176 mode_for_size (total_size * BITS_PER_UNIT,
177 MODE_INT, 1),
178 adjust);
180 slots[pseudo_slots[i].slot_num].mem = stack_slot;
183 /* On a big endian machine, the "address" of the slot is the address
184 of the low part that fits its inherent mode. */
185 if (BYTES_BIG_ENDIAN && inherent_size < total_size)
186 adjust += (total_size - inherent_size);
188 x = adjust_address_nv (x, GET_MODE (regno_reg_rtx[i]), adjust);
190 /* Set all of the memory attributes as appropriate for a spill. */
191 set_mem_attrs_for_spill (x);
192 pseudo_slots[i].mem = x;
195 /* Sort pseudos according their usage frequencies. */
196 static int
197 regno_freq_compare (const void *v1p, const void *v2p)
199 const int regno1 = *(const int *) v1p;
200 const int regno2 = *(const int *) v2p;
201 int diff;
203 if ((diff = lra_reg_info[regno2].freq - lra_reg_info[regno1].freq) != 0)
204 return diff;
205 return regno1 - regno2;
208 /* Redefine STACK_GROWS_DOWNWARD in terms of 0 or 1. */
209 #ifdef STACK_GROWS_DOWNWARD
210 # undef STACK_GROWS_DOWNWARD
211 # define STACK_GROWS_DOWNWARD 1
212 #else
213 # define STACK_GROWS_DOWNWARD 0
214 #endif
216 /* Sort pseudos according to their slots, putting the slots in the order
217 that they should be allocated. Slots with lower numbers have the highest
218 priority and should get the smallest displacement from the stack or
219 frame pointer (whichever is being used).
221 The first allocated slot is always closest to the frame pointer,
222 so prefer lower slot numbers when frame_pointer_needed. If the stack
223 and frame grow in the same direction, then the first allocated slot is
224 always closest to the initial stack pointer and furthest away from the
225 final stack pointer, so allocate higher numbers first when using the
226 stack pointer in that case. The reverse is true if the stack and
227 frame grow in opposite directions. */
228 static int
229 pseudo_reg_slot_compare (const void *v1p, const void *v2p)
231 const int regno1 = *(const int *) v1p;
232 const int regno2 = *(const int *) v2p;
233 int diff, slot_num1, slot_num2;
234 int total_size1, total_size2;
236 slot_num1 = pseudo_slots[regno1].slot_num;
237 slot_num2 = pseudo_slots[regno2].slot_num;
238 if ((diff = slot_num1 - slot_num2) != 0)
239 return (frame_pointer_needed
240 || (!FRAME_GROWS_DOWNWARD) == STACK_GROWS_DOWNWARD ? diff : -diff);
241 total_size1 = GET_MODE_SIZE (lra_reg_info[regno1].biggest_mode);
242 total_size2 = GET_MODE_SIZE (lra_reg_info[regno2].biggest_mode);
243 if ((diff = total_size2 - total_size1) != 0)
244 return diff;
245 return regno1 - regno2;
248 /* Assign spill hard registers to N pseudos in PSEUDO_REGNOS which is
249 sorted in order of highest frequency first. Put the pseudos which
250 did not get a spill hard register at the beginning of array
251 PSEUDO_REGNOS. Return the number of such pseudos. */
252 static int
253 assign_spill_hard_regs (int *pseudo_regnos, int n)
255 int i, k, p, regno, res, spill_class_size, hard_regno, nr;
256 enum reg_class rclass, spill_class;
257 enum machine_mode mode;
258 lra_live_range_t r;
259 rtx_insn *insn;
260 rtx set;
261 basic_block bb;
262 HARD_REG_SET conflict_hard_regs;
263 bitmap_head ok_insn_bitmap;
264 bitmap setjump_crosses = regstat_get_setjmp_crosses ();
265 /* Hard registers which can not be used for any purpose at given
266 program point because they are unallocatable or already allocated
267 for other pseudos. */
268 HARD_REG_SET *reserved_hard_regs;
270 if (! lra_reg_spill_p)
271 return n;
272 /* Set up reserved hard regs for every program point. */
273 reserved_hard_regs = XNEWVEC (HARD_REG_SET, lra_live_max_point);
274 for (p = 0; p < lra_live_max_point; p++)
275 COPY_HARD_REG_SET (reserved_hard_regs[p], lra_no_alloc_regs);
276 for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
277 if (lra_reg_info[i].nrefs != 0
278 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
279 for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
280 for (p = r->start; p <= r->finish; p++)
281 add_to_hard_reg_set (&reserved_hard_regs[p],
282 lra_reg_info[i].biggest_mode, hard_regno);
283 bitmap_initialize (&ok_insn_bitmap, &reg_obstack);
284 FOR_EACH_BB_FN (bb, cfun)
285 FOR_BB_INSNS (bb, insn)
286 if (DEBUG_INSN_P (insn)
287 || ((set = single_set (insn)) != NULL_RTX
288 && REG_P (SET_SRC (set)) && REG_P (SET_DEST (set))))
289 bitmap_set_bit (&ok_insn_bitmap, INSN_UID (insn));
290 for (res = i = 0; i < n; i++)
292 regno = pseudo_regnos[i];
293 rclass = lra_get_allocno_class (regno);
294 if (bitmap_bit_p (setjump_crosses, regno)
295 || (spill_class
296 = ((enum reg_class)
297 targetm.spill_class ((reg_class_t) rclass,
298 PSEUDO_REGNO_MODE (regno)))) == NO_REGS
299 || bitmap_intersect_compl_p (&lra_reg_info[regno].insn_bitmap,
300 &ok_insn_bitmap))
302 pseudo_regnos[res++] = regno;
303 continue;
305 lra_assert (spill_class != NO_REGS);
306 COPY_HARD_REG_SET (conflict_hard_regs,
307 lra_reg_info[regno].conflict_hard_regs);
308 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
309 for (p = r->start; p <= r->finish; p++)
310 IOR_HARD_REG_SET (conflict_hard_regs, reserved_hard_regs[p]);
311 spill_class_size = ira_class_hard_regs_num[spill_class];
312 mode = lra_reg_info[regno].biggest_mode;
313 for (k = 0; k < spill_class_size; k++)
315 hard_regno = ira_class_hard_regs[spill_class][k];
316 if (! overlaps_hard_reg_set_p (conflict_hard_regs, mode, hard_regno))
317 break;
319 if (k >= spill_class_size)
321 /* There is no available regs -- assign memory later. */
322 pseudo_regnos[res++] = regno;
323 continue;
325 if (lra_dump_file != NULL)
326 fprintf (lra_dump_file, " Spill r%d into hr%d\n", regno, hard_regno);
327 /* Update reserved_hard_regs. */
328 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
329 for (p = r->start; p <= r->finish; p++)
330 add_to_hard_reg_set (&reserved_hard_regs[p],
331 lra_reg_info[regno].biggest_mode, hard_regno);
332 spill_hard_reg[regno]
333 = gen_raw_REG (PSEUDO_REGNO_MODE (regno), hard_regno);
334 for (nr = 0;
335 nr < hard_regno_nregs[hard_regno][lra_reg_info[regno].biggest_mode];
336 nr++)
337 /* Just loop. */
338 df_set_regs_ever_live (hard_regno + nr, true);
340 bitmap_clear (&ok_insn_bitmap);
341 free (reserved_hard_regs);
342 return res;
345 /* Add pseudo REGNO to slot SLOT_NUM. */
346 static void
347 add_pseudo_to_slot (int regno, int slot_num)
349 struct pseudo_slot *first;
351 if (slots[slot_num].regno < 0)
353 /* It is the first pseudo in the slot. */
354 slots[slot_num].regno = regno;
355 pseudo_slots[regno].first = &pseudo_slots[regno];
356 pseudo_slots[regno].next = NULL;
358 else
360 first = pseudo_slots[regno].first = &pseudo_slots[slots[slot_num].regno];
361 pseudo_slots[regno].next = first->next;
362 first->next = &pseudo_slots[regno];
364 pseudo_slots[regno].mem = NULL_RTX;
365 pseudo_slots[regno].slot_num = slot_num;
366 slots[slot_num].live_ranges
367 = lra_merge_live_ranges (slots[slot_num].live_ranges,
368 lra_copy_live_range_list
369 (lra_reg_info[regno].live_ranges));
372 /* Assign stack slot numbers to pseudos in array PSEUDO_REGNOS of
373 length N. Sort pseudos in PSEUDO_REGNOS for subsequent assigning
374 memory stack slots. */
375 static void
376 assign_stack_slot_num_and_sort_pseudos (int *pseudo_regnos, int n)
378 int i, j, regno;
380 slots_num = 0;
381 /* Assign stack slot numbers to spilled pseudos, use smaller numbers
382 for most frequently used pseudos. */
383 for (i = 0; i < n; i++)
385 regno = pseudo_regnos[i];
386 if (! flag_ira_share_spill_slots)
387 j = slots_num;
388 else
390 for (j = 0; j < slots_num; j++)
391 if (slots[j].hard_regno < 0
392 && ! (lra_intersected_live_ranges_p
393 (slots[j].live_ranges,
394 lra_reg_info[regno].live_ranges)))
395 break;
397 if (j >= slots_num)
399 /* New slot. */
400 slots[j].live_ranges = NULL;
401 slots[j].regno = slots[j].hard_regno = -1;
402 slots[j].mem = NULL_RTX;
403 slots_num++;
405 add_pseudo_to_slot (regno, j);
407 /* Sort regnos according to their slot numbers. */
408 qsort (pseudo_regnos, n, sizeof (int), pseudo_reg_slot_compare);
411 /* Recursively process LOC in INSN and change spilled pseudos to the
412 corresponding memory or spilled hard reg. Ignore spilled pseudos
413 created from the scratches. */
414 static void
415 remove_pseudos (rtx *loc, rtx_insn *insn)
417 int i;
418 rtx hard_reg;
419 const char *fmt;
420 enum rtx_code code;
422 if (*loc == NULL_RTX)
423 return;
424 code = GET_CODE (*loc);
425 if (code == REG && (i = REGNO (*loc)) >= FIRST_PSEUDO_REGISTER
426 && lra_get_regno_hard_regno (i) < 0
427 /* We do not want to assign memory for former scratches because
428 it might result in an address reload for some targets. In
429 any case we transform such pseudos not getting hard registers
430 into scratches back. */
431 && ! lra_former_scratch_p (i))
433 if ((hard_reg = spill_hard_reg[i]) != NULL_RTX)
434 *loc = copy_rtx (hard_reg);
435 else
437 rtx x = lra_eliminate_regs_1 (insn, pseudo_slots[i].mem,
438 GET_MODE (pseudo_slots[i].mem),
439 false, false, true);
440 *loc = x != pseudo_slots[i].mem ? x : copy_rtx (x);
442 return;
445 fmt = GET_RTX_FORMAT (code);
446 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
448 if (fmt[i] == 'e')
449 remove_pseudos (&XEXP (*loc, i), insn);
450 else if (fmt[i] == 'E')
452 int j;
454 for (j = XVECLEN (*loc, i) - 1; j >= 0; j--)
455 remove_pseudos (&XVECEXP (*loc, i, j), insn);
460 /* Convert spilled pseudos into their stack slots or spill hard regs,
461 put insns to process on the constraint stack (that is all insns in
462 which pseudos were changed to memory or spill hard regs). */
463 static void
464 spill_pseudos (void)
466 basic_block bb;
467 rtx_insn *insn;
468 int i;
469 bitmap_head spilled_pseudos, changed_insns;
471 bitmap_initialize (&spilled_pseudos, &reg_obstack);
472 bitmap_initialize (&changed_insns, &reg_obstack);
473 for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
475 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
476 && ! lra_former_scratch_p (i))
478 bitmap_set_bit (&spilled_pseudos, i);
479 bitmap_ior_into (&changed_insns, &lra_reg_info[i].insn_bitmap);
482 FOR_EACH_BB_FN (bb, cfun)
484 FOR_BB_INSNS (bb, insn)
485 if (bitmap_bit_p (&changed_insns, INSN_UID (insn)))
487 rtx *link_loc, link;
488 remove_pseudos (&PATTERN (insn), insn);
489 if (CALL_P (insn))
490 remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn);
491 for (link_loc = &REG_NOTES (insn);
492 (link = *link_loc) != NULL_RTX;
493 link_loc = &XEXP (link, 1))
495 switch (REG_NOTE_KIND (link))
497 case REG_FRAME_RELATED_EXPR:
498 case REG_CFA_DEF_CFA:
499 case REG_CFA_ADJUST_CFA:
500 case REG_CFA_OFFSET:
501 case REG_CFA_REGISTER:
502 case REG_CFA_EXPRESSION:
503 case REG_CFA_RESTORE:
504 case REG_CFA_SET_VDRAP:
505 remove_pseudos (&XEXP (link, 0), insn);
506 break;
507 default:
508 break;
511 if (lra_dump_file != NULL)
512 fprintf (lra_dump_file,
513 "Changing spilled pseudos to memory in insn #%u\n",
514 INSN_UID (insn));
515 lra_push_insn (insn);
516 if (lra_reg_spill_p || targetm.different_addr_displacement_p ())
517 lra_set_used_insn_alternative (insn, -1);
519 else if (CALL_P (insn))
520 /* Presence of any pseudo in CALL_INSN_FUNCTION_USAGE does
521 not affect value of insn_bitmap of the corresponding
522 lra_reg_info. That is because we don't need to reload
523 pseudos in CALL_INSN_FUNCTION_USAGEs. So if we process
524 only insns in the insn_bitmap of given pseudo here, we
525 can miss the pseudo in some
526 CALL_INSN_FUNCTION_USAGEs. */
527 remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn);
528 bitmap_and_compl_into (df_get_live_in (bb), &spilled_pseudos);
529 bitmap_and_compl_into (df_get_live_out (bb), &spilled_pseudos);
531 bitmap_clear (&spilled_pseudos);
532 bitmap_clear (&changed_insns);
535 /* Return true if we need to change some pseudos into memory. */
536 bool
537 lra_need_for_spills_p (void)
539 int i; max_regno = max_reg_num ();
541 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
542 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
543 && ! lra_former_scratch_p (i))
544 return true;
545 return false;
548 /* Change spilled pseudos into memory or spill hard regs. Put changed
549 insns on the constraint stack (these insns will be considered on
550 the next constraint pass). The changed insns are all insns in
551 which pseudos were changed. */
552 void
553 lra_spill (void)
555 int i, n, curr_regno;
556 int *pseudo_regnos;
558 regs_num = max_reg_num ();
559 spill_hard_reg = XNEWVEC (rtx, regs_num);
560 pseudo_regnos = XNEWVEC (int, regs_num);
561 for (n = 0, i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
562 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
563 /* We do not want to assign memory for former scratches. */
564 && ! lra_former_scratch_p (i))
566 spill_hard_reg[i] = NULL_RTX;
567 pseudo_regnos[n++] = i;
569 lra_assert (n > 0);
570 pseudo_slots = XNEWVEC (struct pseudo_slot, regs_num);
571 slots = XNEWVEC (struct slot, regs_num);
572 /* Sort regnos according their usage frequencies. */
573 qsort (pseudo_regnos, n, sizeof (int), regno_freq_compare);
574 n = assign_spill_hard_regs (pseudo_regnos, n);
575 assign_stack_slot_num_and_sort_pseudos (pseudo_regnos, n);
576 for (i = 0; i < n; i++)
577 if (pseudo_slots[pseudo_regnos[i]].mem == NULL_RTX)
578 assign_mem_slot (pseudo_regnos[i]);
579 if (n > 0 && crtl->stack_alignment_needed)
580 /* If we have a stack frame, we must align it now. The stack size
581 may be a part of the offset computation for register
582 elimination. */
583 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
584 if (lra_dump_file != NULL)
586 for (i = 0; i < slots_num; i++)
588 fprintf (lra_dump_file, " Slot %d regnos (width = %d):", i,
589 GET_MODE_SIZE (GET_MODE (slots[i].mem)));
590 for (curr_regno = slots[i].regno;;
591 curr_regno = pseudo_slots[curr_regno].next - pseudo_slots)
593 fprintf (lra_dump_file, " %d", curr_regno);
594 if (pseudo_slots[curr_regno].next == NULL)
595 break;
597 fprintf (lra_dump_file, "\n");
600 spill_pseudos ();
601 free (slots);
602 free (pseudo_slots);
603 free (pseudo_regnos);
604 free (spill_hard_reg);
607 /* Apply alter_subreg for subregs of regs in *LOC. Use FINAL_P for
608 alter_subreg calls. Return true if any subreg of reg is
609 processed. */
610 static bool
611 alter_subregs (rtx *loc, bool final_p)
613 int i;
614 rtx x = *loc;
615 bool res;
616 const char *fmt;
617 enum rtx_code code;
619 if (x == NULL_RTX)
620 return false;
621 code = GET_CODE (x);
622 if (code == SUBREG && REG_P (SUBREG_REG (x)))
624 lra_assert (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER);
625 alter_subreg (loc, final_p);
626 return true;
628 fmt = GET_RTX_FORMAT (code);
629 res = false;
630 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
632 if (fmt[i] == 'e')
634 if (alter_subregs (&XEXP (x, i), final_p))
635 res = true;
637 else if (fmt[i] == 'E')
639 int j;
641 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
642 if (alter_subregs (&XVECEXP (x, i, j), final_p))
643 res = true;
646 return res;
649 /* Return true if REGNO is used for return in the current
650 function. */
651 static bool
652 return_regno_p (unsigned int regno)
654 rtx outgoing = crtl->return_rtx;
656 if (! outgoing)
657 return false;
659 if (REG_P (outgoing))
660 return REGNO (outgoing) == regno;
661 else if (GET_CODE (outgoing) == PARALLEL)
663 int i;
665 for (i = 0; i < XVECLEN (outgoing, 0); i++)
667 rtx x = XEXP (XVECEXP (outgoing, 0, i), 0);
669 if (REG_P (x) && REGNO (x) == regno)
670 return true;
673 return false;
676 /* Final change of pseudos got hard registers into the corresponding
677 hard registers and removing temporary clobbers. */
678 void
679 lra_final_code_change (void)
681 int i, hard_regno;
682 basic_block bb;
683 rtx_insn *insn, *curr;
684 int max_regno = max_reg_num ();
686 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
687 if (lra_reg_info[i].nrefs != 0
688 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
689 SET_REGNO (regno_reg_rtx[i], hard_regno);
690 FOR_EACH_BB_FN (bb, cfun)
691 FOR_BB_INSNS_SAFE (bb, insn, curr)
692 if (INSN_P (insn))
694 rtx pat = PATTERN (insn);
696 if (GET_CODE (pat) == CLOBBER && LRA_TEMP_CLOBBER_P (pat))
698 /* Remove clobbers temporarily created in LRA. We don't
699 need them anymore and don't want to waste compiler
700 time processing them in a few subsequent passes. */
701 lra_invalidate_insn_data (insn);
702 delete_insn (insn);
703 continue;
706 /* IRA can generate move insns involving pseudos. It is
707 better remove them earlier to speed up compiler a bit.
708 It is also better to do it here as they might not pass
709 final RTL check in LRA, (e.g. insn moving a control
710 register into itself). So remove an useless move insn
711 unless next insn is USE marking the return reg (we should
712 save this as some subsequent optimizations assume that
713 such original insns are saved). */
714 if (NONJUMP_INSN_P (insn) && GET_CODE (pat) == SET
715 && REG_P (SET_SRC (pat)) && REG_P (SET_DEST (pat))
716 && REGNO (SET_SRC (pat)) == REGNO (SET_DEST (pat))
717 && ! return_regno_p (REGNO (SET_SRC (pat))))
719 lra_invalidate_insn_data (insn);
720 delete_insn (insn);
721 continue;
724 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
725 struct lra_static_insn_data *static_id = id->insn_static_data;
726 bool insn_change_p = false;
728 for (i = id->insn_static_data->n_operands - 1; i >= 0; i--)
729 if ((DEBUG_INSN_P (insn) || ! static_id->operand[i].is_operator)
730 && alter_subregs (id->operand_loc[i], ! DEBUG_INSN_P (insn)))
732 lra_update_dup (id, i);
733 insn_change_p = true;
735 if (insn_change_p)
736 lra_update_operator_dups (id);