2017-02-01 Bill Schmidt <wschmidt@linux.vnet.ibm.com>
[official-gcc.git] / gcc / lra-assigns.c
blob0ccd42546b71778db5d9efdd1b6eb7d9e7dcbd5d
1 /* Assign reload pseudos.
2 Copyright (C) 2010-2017 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's main objective is to assign hard registers to reload
23 pseudos. It also tries to allocate hard registers to other
24 pseudos, but at a lower priority than the reload pseudos. The pass
25 does not transform the RTL.
27 We must allocate a hard register to every reload pseudo. We try to
28 increase the chances of finding a viable allocation by assigning
29 the pseudos in order of fewest available hard registers first. If
30 we still fail to find a hard register, we spill other (non-reload)
31 pseudos in order to make room.
33 find_hard_regno_for finds hard registers for allocation without
34 spilling. spill_for does the same with spilling. Both functions
35 use a cost model to determine the most profitable choice of hard
36 and spill registers.
38 Once we have finished allocating reload pseudos, we also try to
39 assign registers to other (non-reload) pseudos. This is useful if
40 hard registers were freed up by the spilling just described.
42 We try to assign hard registers by collecting pseudos into threads.
43 These threads contain reload and inheritance pseudos that are
44 connected by copies (move insns). Doing this improves the chances
45 of pseudos in the thread getting the same hard register and, as a
46 result, of allowing some move insns to be deleted.
48 When we assign a hard register to a pseudo, we decrease the cost of
49 using the same hard register for pseudos that are connected by
50 copies.
52 If two hard registers have the same frequency-derived cost, we
53 prefer hard registers with higher priorities. The mapping of
54 registers to priorities is controlled by the register_priority
55 target hook. For example, x86-64 has a few register priorities:
56 hard registers with and without REX prefixes have different
57 priorities. This permits us to generate smaller code as insns
58 without REX prefixes are shorter.
60 If a few hard registers are still equally good for the assignment,
61 we choose the least used hard register. It is called leveling and
62 may be profitable for some targets.
64 Only insns with changed allocation pseudos are processed on the
65 next constraint pass.
67 The pseudo live-ranges are used to find conflicting pseudos.
69 For understanding the code, it is important to keep in mind that
70 inheritance, split, and reload pseudos created since last
71 constraint pass have regno >= lra_constraint_new_regno_start.
72 Inheritance and split pseudos created on any pass are in the
73 corresponding bitmaps. Inheritance and split pseudos since the
74 last constraint pass have also the corresponding non-negative
75 restore_regno. */
77 #include "config.h"
78 #include "system.h"
79 #include "coretypes.h"
80 #include "backend.h"
81 #include "target.h"
82 #include "rtl.h"
83 #include "tree.h"
84 #include "predict.h"
85 #include "df.h"
86 #include "memmodel.h"
87 #include "tm_p.h"
88 #include "insn-config.h"
89 #include "regs.h"
90 #include "ira.h"
91 #include "recog.h"
92 #include "rtl-error.h"
93 #include "sparseset.h"
94 #include "params.h"
95 #include "lra.h"
96 #include "lra-int.h"
98 /* Current iteration number of the pass and current iteration number
99 of the pass after the latest spill pass when any former reload
100 pseudo was spilled. */
101 int lra_assignment_iter;
102 int lra_assignment_iter_after_spill;
104 /* Flag of spilling former reload pseudos on this pass. */
105 static bool former_reload_pseudo_spill_p;
107 /* Array containing corresponding values of function
108 lra_get_allocno_class. It is used to speed up the code. */
109 static enum reg_class *regno_allocno_class_array;
111 /* Array containing lengths of pseudo live ranges. It is used to
112 speed up the code. */
113 static int *regno_live_length;
115 /* Information about the thread to which a pseudo belongs. Threads are
116 a set of connected reload and inheritance pseudos with the same set of
117 available hard registers. Lone registers belong to their own threads. */
118 struct regno_assign_info
120 /* First/next pseudo of the same thread. */
121 int first, next;
122 /* Frequency of the thread (execution frequency of only reload
123 pseudos in the thread when the thread contains a reload pseudo).
124 Defined only for the first thread pseudo. */
125 int freq;
128 /* Map regno to the corresponding regno assignment info. */
129 static struct regno_assign_info *regno_assign_info;
131 /* All inherited, subreg or optional pseudos created before last spill
132 sub-pass. Such pseudos are permitted to get memory instead of hard
133 regs. */
134 static bitmap_head non_reload_pseudos;
136 /* Process a pseudo copy with execution frequency COPY_FREQ connecting
137 REGNO1 and REGNO2 to form threads. */
138 static void
139 process_copy_to_form_thread (int regno1, int regno2, int copy_freq)
141 int last, regno1_first, regno2_first;
143 lra_assert (regno1 >= lra_constraint_new_regno_start
144 && regno2 >= lra_constraint_new_regno_start);
145 regno1_first = regno_assign_info[regno1].first;
146 regno2_first = regno_assign_info[regno2].first;
147 if (regno1_first != regno2_first)
149 for (last = regno2_first;
150 regno_assign_info[last].next >= 0;
151 last = regno_assign_info[last].next)
152 regno_assign_info[last].first = regno1_first;
153 regno_assign_info[last].first = regno1_first;
154 regno_assign_info[last].next = regno_assign_info[regno1_first].next;
155 regno_assign_info[regno1_first].next = regno2_first;
156 regno_assign_info[regno1_first].freq
157 += regno_assign_info[regno2_first].freq;
159 regno_assign_info[regno1_first].freq -= 2 * copy_freq;
160 lra_assert (regno_assign_info[regno1_first].freq >= 0);
163 /* Initialize REGNO_ASSIGN_INFO and form threads. */
164 static void
165 init_regno_assign_info (void)
167 int i, regno1, regno2, max_regno = max_reg_num ();
168 lra_copy_t cp;
170 regno_assign_info = XNEWVEC (struct regno_assign_info, max_regno);
171 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
173 regno_assign_info[i].first = i;
174 regno_assign_info[i].next = -1;
175 regno_assign_info[i].freq = lra_reg_info[i].freq;
177 /* Form the threads. */
178 for (i = 0; (cp = lra_get_copy (i)) != NULL; i++)
179 if ((regno1 = cp->regno1) >= lra_constraint_new_regno_start
180 && (regno2 = cp->regno2) >= lra_constraint_new_regno_start
181 && reg_renumber[regno1] < 0 && lra_reg_info[regno1].nrefs != 0
182 && reg_renumber[regno2] < 0 && lra_reg_info[regno2].nrefs != 0
183 && (ira_class_hard_regs_num[regno_allocno_class_array[regno1]]
184 == ira_class_hard_regs_num[regno_allocno_class_array[regno2]]))
185 process_copy_to_form_thread (regno1, regno2, cp->freq);
188 /* Free REGNO_ASSIGN_INFO. */
189 static void
190 finish_regno_assign_info (void)
192 free (regno_assign_info);
195 /* The function is used to sort *reload* and *inheritance* pseudos to
196 try to assign them hard registers. We put pseudos from the same
197 thread always nearby. */
198 static int
199 reload_pseudo_compare_func (const void *v1p, const void *v2p)
201 int r1 = *(const int *) v1p, r2 = *(const int *) v2p;
202 enum reg_class cl1 = regno_allocno_class_array[r1];
203 enum reg_class cl2 = regno_allocno_class_array[r2];
204 int diff;
206 lra_assert (r1 >= lra_constraint_new_regno_start
207 && r2 >= lra_constraint_new_regno_start);
209 /* Prefer to assign reload registers with smaller classes first to
210 guarantee assignment to all reload registers. */
211 if ((diff = (ira_class_hard_regs_num[cl1]
212 - ira_class_hard_regs_num[cl2])) != 0)
213 return diff;
214 if ((diff
215 = (ira_reg_class_max_nregs[cl2][lra_reg_info[r2].biggest_mode]
216 - ira_reg_class_max_nregs[cl1][lra_reg_info[r1].biggest_mode])) != 0
217 /* The code below executes rarely as nregs == 1 in most cases.
218 So we should not worry about using faster data structures to
219 check reload pseudos. */
220 && ! bitmap_bit_p (&non_reload_pseudos, r1)
221 && ! bitmap_bit_p (&non_reload_pseudos, r2))
222 return diff;
223 if ((diff = (regno_assign_info[regno_assign_info[r2].first].freq
224 - regno_assign_info[regno_assign_info[r1].first].freq)) != 0)
225 return diff;
226 /* Allocate bigger pseudos first to avoid register file
227 fragmentation. */
228 if ((diff
229 = (ira_reg_class_max_nregs[cl2][lra_reg_info[r2].biggest_mode]
230 - ira_reg_class_max_nregs[cl1][lra_reg_info[r1].biggest_mode])) != 0)
231 return diff;
232 /* Put pseudos from the thread nearby. */
233 if ((diff = regno_assign_info[r1].first - regno_assign_info[r2].first) != 0)
234 return diff;
235 /* Prefer pseudos with longer live ranges. It sets up better
236 prefered hard registers for the thread pseudos and decreases
237 register-register moves between the thread pseudos. */
238 if ((diff = regno_live_length[r2] - regno_live_length[r1]) != 0)
239 return diff;
240 /* If regs are equally good, sort by their numbers, so that the
241 results of qsort leave nothing to chance. */
242 return r1 - r2;
245 /* The function is used to sort *non-reload* pseudos to try to assign
246 them hard registers. The order calculation is simpler than in the
247 previous function and based on the pseudo frequency usage. */
248 static int
249 pseudo_compare_func (const void *v1p, const void *v2p)
251 int r1 = *(const int *) v1p, r2 = *(const int *) v2p;
252 int diff;
254 /* Assign hard reg to static chain pointer first pseudo when
255 non-local goto is used. */
256 if (non_spilled_static_chain_regno_p (r1))
257 return -1;
258 else if (non_spilled_static_chain_regno_p (r2))
259 return 1;
261 /* Prefer to assign more frequently used registers first. */
262 if ((diff = lra_reg_info[r2].freq - lra_reg_info[r1].freq) != 0)
263 return diff;
265 /* If regs are equally good, sort by their numbers, so that the
266 results of qsort leave nothing to chance. */
267 return r1 - r2;
270 /* Arrays of size LRA_LIVE_MAX_POINT mapping a program point to the
271 pseudo live ranges with given start point. We insert only live
272 ranges of pseudos interesting for assignment purposes. They are
273 reload pseudos and pseudos assigned to hard registers. */
274 static lra_live_range_t *start_point_ranges;
276 /* Used as a flag that a live range is not inserted in the start point
277 chain. */
278 static struct lra_live_range not_in_chain_mark;
280 /* Create and set up START_POINT_RANGES. */
281 static void
282 create_live_range_start_chains (void)
284 int i, max_regno;
285 lra_live_range_t r;
287 start_point_ranges = XCNEWVEC (lra_live_range_t, lra_live_max_point);
288 max_regno = max_reg_num ();
289 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
290 if (i >= lra_constraint_new_regno_start || reg_renumber[i] >= 0)
292 for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
294 r->start_next = start_point_ranges[r->start];
295 start_point_ranges[r->start] = r;
298 else
300 for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
301 r->start_next = &not_in_chain_mark;
305 /* Insert live ranges of pseudo REGNO into start chains if they are
306 not there yet. */
307 static void
308 insert_in_live_range_start_chain (int regno)
310 lra_live_range_t r = lra_reg_info[regno].live_ranges;
312 if (r->start_next != &not_in_chain_mark)
313 return;
314 for (; r != NULL; r = r->next)
316 r->start_next = start_point_ranges[r->start];
317 start_point_ranges[r->start] = r;
321 /* Free START_POINT_RANGES. */
322 static void
323 finish_live_range_start_chains (void)
325 gcc_assert (start_point_ranges != NULL);
326 free (start_point_ranges);
327 start_point_ranges = NULL;
330 /* Map: program point -> bitmap of all pseudos living at the point and
331 assigned to hard registers. */
332 static bitmap_head *live_hard_reg_pseudos;
333 static bitmap_obstack live_hard_reg_pseudos_bitmap_obstack;
335 /* reg_renumber corresponding to pseudos marked in
336 live_hard_reg_pseudos. reg_renumber might be not matched to
337 live_hard_reg_pseudos but live_pseudos_reg_renumber always reflects
338 live_hard_reg_pseudos. */
339 static int *live_pseudos_reg_renumber;
341 /* Sparseset used to calculate living hard reg pseudos for some program
342 point range. */
343 static sparseset live_range_hard_reg_pseudos;
345 /* Sparseset used to calculate living reload/inheritance pseudos for
346 some program point range. */
347 static sparseset live_range_reload_inheritance_pseudos;
349 /* Allocate and initialize the data about living pseudos at program
350 points. */
351 static void
352 init_lives (void)
354 int i, max_regno = max_reg_num ();
356 live_range_hard_reg_pseudos = sparseset_alloc (max_regno);
357 live_range_reload_inheritance_pseudos = sparseset_alloc (max_regno);
358 live_hard_reg_pseudos = XNEWVEC (bitmap_head, lra_live_max_point);
359 bitmap_obstack_initialize (&live_hard_reg_pseudos_bitmap_obstack);
360 for (i = 0; i < lra_live_max_point; i++)
361 bitmap_initialize (&live_hard_reg_pseudos[i],
362 &live_hard_reg_pseudos_bitmap_obstack);
363 live_pseudos_reg_renumber = XNEWVEC (int, max_regno);
364 for (i = 0; i < max_regno; i++)
365 live_pseudos_reg_renumber[i] = -1;
368 /* Free the data about living pseudos at program points. */
369 static void
370 finish_lives (void)
372 sparseset_free (live_range_hard_reg_pseudos);
373 sparseset_free (live_range_reload_inheritance_pseudos);
374 free (live_hard_reg_pseudos);
375 bitmap_obstack_release (&live_hard_reg_pseudos_bitmap_obstack);
376 free (live_pseudos_reg_renumber);
379 /* Update the LIVE_HARD_REG_PSEUDOS and LIVE_PSEUDOS_REG_RENUMBER
380 entries for pseudo REGNO. Assume that the register has been
381 spilled if FREE_P, otherwise assume that it has been assigned
382 reg_renumber[REGNO] (if >= 0). We also insert the pseudo live
383 ranges in the start chains when it is assumed to be assigned to a
384 hard register because we use the chains of pseudos assigned to hard
385 registers during allocation. */
386 static void
387 update_lives (int regno, bool free_p)
389 int p;
390 lra_live_range_t r;
392 if (reg_renumber[regno] < 0)
393 return;
394 live_pseudos_reg_renumber[regno] = free_p ? -1 : reg_renumber[regno];
395 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
397 for (p = r->start; p <= r->finish; p++)
398 if (free_p)
399 bitmap_clear_bit (&live_hard_reg_pseudos[p], regno);
400 else
402 bitmap_set_bit (&live_hard_reg_pseudos[p], regno);
403 insert_in_live_range_start_chain (regno);
408 /* Sparseset used to calculate reload pseudos conflicting with a given
409 pseudo when we are trying to find a hard register for the given
410 pseudo. */
411 static sparseset conflict_reload_and_inheritance_pseudos;
413 /* Map: program point -> bitmap of all reload and inheritance pseudos
414 living at the point. */
415 static bitmap_head *live_reload_and_inheritance_pseudos;
416 static bitmap_obstack live_reload_and_inheritance_pseudos_bitmap_obstack;
418 /* Allocate and initialize data about living reload pseudos at any
419 given program point. */
420 static void
421 init_live_reload_and_inheritance_pseudos (void)
423 int i, p, max_regno = max_reg_num ();
424 lra_live_range_t r;
426 conflict_reload_and_inheritance_pseudos = sparseset_alloc (max_regno);
427 live_reload_and_inheritance_pseudos = XNEWVEC (bitmap_head, lra_live_max_point);
428 bitmap_obstack_initialize (&live_reload_and_inheritance_pseudos_bitmap_obstack);
429 for (p = 0; p < lra_live_max_point; p++)
430 bitmap_initialize (&live_reload_and_inheritance_pseudos[p],
431 &live_reload_and_inheritance_pseudos_bitmap_obstack);
432 for (i = lra_constraint_new_regno_start; i < max_regno; i++)
434 for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
435 for (p = r->start; p <= r->finish; p++)
436 bitmap_set_bit (&live_reload_and_inheritance_pseudos[p], i);
440 /* Finalize data about living reload pseudos at any given program
441 point. */
442 static void
443 finish_live_reload_and_inheritance_pseudos (void)
445 sparseset_free (conflict_reload_and_inheritance_pseudos);
446 free (live_reload_and_inheritance_pseudos);
447 bitmap_obstack_release (&live_reload_and_inheritance_pseudos_bitmap_obstack);
450 /* The value used to check that cost of given hard reg is really
451 defined currently. */
452 static int curr_hard_regno_costs_check = 0;
453 /* Array used to check that cost of the corresponding hard reg (the
454 array element index) is really defined currently. */
455 static int hard_regno_costs_check[FIRST_PSEUDO_REGISTER];
456 /* The current costs of allocation of hard regs. Defined only if the
457 value of the corresponding element of the previous array is equal to
458 CURR_HARD_REGNO_COSTS_CHECK. */
459 static int hard_regno_costs[FIRST_PSEUDO_REGISTER];
461 /* Adjust cost of HARD_REGNO by INCR. Reset the cost first if it is
462 not defined yet. */
463 static inline void
464 adjust_hard_regno_cost (int hard_regno, int incr)
466 if (hard_regno_costs_check[hard_regno] != curr_hard_regno_costs_check)
467 hard_regno_costs[hard_regno] = 0;
468 hard_regno_costs_check[hard_regno] = curr_hard_regno_costs_check;
469 hard_regno_costs[hard_regno] += incr;
472 /* Try to find a free hard register for pseudo REGNO. Return the
473 hard register on success and set *COST to the cost of using
474 that register. (If several registers have equal cost, the one with
475 the highest priority wins.) Return -1 on failure.
477 If FIRST_P, return the first available hard reg ignoring other
478 criteria, e.g. allocation cost. This approach results in less hard
479 reg pool fragmentation and permit to allocate hard regs to reload
480 pseudos in complicated situations where pseudo sizes are different.
482 If TRY_ONLY_HARD_REGNO >= 0, consider only that hard register,
483 otherwise consider all hard registers in REGNO's class.
485 If REGNO_SET is not empty, only hard registers from the set are
486 considered. */
487 static int
488 find_hard_regno_for_1 (int regno, int *cost, int try_only_hard_regno,
489 bool first_p, HARD_REG_SET regno_set)
491 HARD_REG_SET conflict_set;
492 int best_cost = INT_MAX, best_priority = INT_MIN, best_usage = INT_MAX;
493 lra_live_range_t r;
494 int p, i, j, rclass_size, best_hard_regno, priority, hard_regno;
495 int hr, conflict_hr, nregs;
496 machine_mode biggest_mode;
497 unsigned int k, conflict_regno;
498 int offset, val, biggest_nregs, nregs_diff;
499 enum reg_class rclass;
500 bitmap_iterator bi;
501 bool *rclass_intersect_p;
502 HARD_REG_SET impossible_start_hard_regs, available_regs;
504 if (hard_reg_set_empty_p (regno_set))
505 COPY_HARD_REG_SET (conflict_set, lra_no_alloc_regs);
506 else
508 COMPL_HARD_REG_SET (conflict_set, regno_set);
509 IOR_HARD_REG_SET (conflict_set, lra_no_alloc_regs);
511 rclass = regno_allocno_class_array[regno];
512 rclass_intersect_p = ira_reg_classes_intersect_p[rclass];
513 curr_hard_regno_costs_check++;
514 sparseset_clear (conflict_reload_and_inheritance_pseudos);
515 sparseset_clear (live_range_hard_reg_pseudos);
516 IOR_HARD_REG_SET (conflict_set, lra_reg_info[regno].conflict_hard_regs);
517 biggest_mode = lra_reg_info[regno].biggest_mode;
518 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
520 EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi)
521 if (rclass_intersect_p[regno_allocno_class_array[k]])
522 sparseset_set_bit (live_range_hard_reg_pseudos, k);
523 EXECUTE_IF_SET_IN_BITMAP (&live_reload_and_inheritance_pseudos[r->start],
524 0, k, bi)
525 if (lra_reg_info[k].preferred_hard_regno1 >= 0
526 && live_pseudos_reg_renumber[k] < 0
527 && rclass_intersect_p[regno_allocno_class_array[k]])
528 sparseset_set_bit (conflict_reload_and_inheritance_pseudos, k);
529 for (p = r->start + 1; p <= r->finish; p++)
531 lra_live_range_t r2;
533 for (r2 = start_point_ranges[p];
534 r2 != NULL;
535 r2 = r2->start_next)
537 if (r2->regno >= lra_constraint_new_regno_start
538 && lra_reg_info[r2->regno].preferred_hard_regno1 >= 0
539 && live_pseudos_reg_renumber[r2->regno] < 0
540 && rclass_intersect_p[regno_allocno_class_array[r2->regno]])
541 sparseset_set_bit (conflict_reload_and_inheritance_pseudos,
542 r2->regno);
543 if (live_pseudos_reg_renumber[r2->regno] >= 0
544 && rclass_intersect_p[regno_allocno_class_array[r2->regno]])
545 sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno);
549 if ((hard_regno = lra_reg_info[regno].preferred_hard_regno1) >= 0)
551 adjust_hard_regno_cost
552 (hard_regno, -lra_reg_info[regno].preferred_hard_regno_profit1);
553 if ((hard_regno = lra_reg_info[regno].preferred_hard_regno2) >= 0)
554 adjust_hard_regno_cost
555 (hard_regno, -lra_reg_info[regno].preferred_hard_regno_profit2);
557 #ifdef STACK_REGS
558 if (lra_reg_info[regno].no_stack_p)
559 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++)
560 SET_HARD_REG_BIT (conflict_set, i);
561 #endif
562 sparseset_clear_bit (conflict_reload_and_inheritance_pseudos, regno);
563 val = lra_reg_info[regno].val;
564 offset = lra_reg_info[regno].offset;
565 CLEAR_HARD_REG_SET (impossible_start_hard_regs);
566 EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno)
568 conflict_hr = live_pseudos_reg_renumber[conflict_regno];
569 if (lra_reg_val_equal_p (conflict_regno, val, offset))
571 conflict_hr = live_pseudos_reg_renumber[conflict_regno];
572 nregs = (hard_regno_nregs[conflict_hr]
573 [lra_reg_info[conflict_regno].biggest_mode]);
574 /* Remember about multi-register pseudos. For example, 2
575 hard register pseudos can start on the same hard register
576 but can not start on HR and HR+1/HR-1. */
577 for (hr = conflict_hr + 1;
578 hr < FIRST_PSEUDO_REGISTER && hr < conflict_hr + nregs;
579 hr++)
580 SET_HARD_REG_BIT (impossible_start_hard_regs, hr);
581 for (hr = conflict_hr - 1;
582 hr >= 0 && hr + hard_regno_nregs[hr][biggest_mode] > conflict_hr;
583 hr--)
584 SET_HARD_REG_BIT (impossible_start_hard_regs, hr);
586 else
588 enum machine_mode biggest_conflict_mode
589 = lra_reg_info[conflict_regno].biggest_mode;
590 int biggest_conflict_nregs
591 = hard_regno_nregs[conflict_hr][biggest_conflict_mode];
593 nregs_diff = (biggest_conflict_nregs
594 - (hard_regno_nregs
595 [conflict_hr]
596 [PSEUDO_REGNO_MODE (conflict_regno)]));
597 add_to_hard_reg_set (&conflict_set,
598 biggest_conflict_mode,
599 conflict_hr
600 - (WORDS_BIG_ENDIAN ? nregs_diff : 0));
601 if (hard_reg_set_subset_p (reg_class_contents[rclass],
602 conflict_set))
603 return -1;
606 EXECUTE_IF_SET_IN_SPARSESET (conflict_reload_and_inheritance_pseudos,
607 conflict_regno)
608 if (!lra_reg_val_equal_p (conflict_regno, val, offset))
610 lra_assert (live_pseudos_reg_renumber[conflict_regno] < 0);
611 if ((hard_regno
612 = lra_reg_info[conflict_regno].preferred_hard_regno1) >= 0)
614 adjust_hard_regno_cost
615 (hard_regno,
616 lra_reg_info[conflict_regno].preferred_hard_regno_profit1);
617 if ((hard_regno
618 = lra_reg_info[conflict_regno].preferred_hard_regno2) >= 0)
619 adjust_hard_regno_cost
620 (hard_regno,
621 lra_reg_info[conflict_regno].preferred_hard_regno_profit2);
624 /* Make sure that all registers in a multi-word pseudo belong to the
625 required class. */
626 IOR_COMPL_HARD_REG_SET (conflict_set, reg_class_contents[rclass]);
627 lra_assert (rclass != NO_REGS);
628 rclass_size = ira_class_hard_regs_num[rclass];
629 best_hard_regno = -1;
630 hard_regno = ira_class_hard_regs[rclass][0];
631 biggest_nregs = hard_regno_nregs[hard_regno][biggest_mode];
632 nregs_diff = (biggest_nregs
633 - hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)]);
634 COPY_HARD_REG_SET (available_regs, reg_class_contents[rclass]);
635 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
636 for (i = 0; i < rclass_size; i++)
638 if (try_only_hard_regno >= 0)
639 hard_regno = try_only_hard_regno;
640 else
641 hard_regno = ira_class_hard_regs[rclass][i];
642 if (! overlaps_hard_reg_set_p (conflict_set,
643 PSEUDO_REGNO_MODE (regno), hard_regno)
644 && HARD_REGNO_MODE_OK (hard_regno, PSEUDO_REGNO_MODE (regno))
645 /* We can not use prohibited_class_mode_regs for all classes
646 because it is not defined for all classes. */
647 && (ira_allocno_class_translate[rclass] != rclass
648 || ! TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs
649 [rclass][PSEUDO_REGNO_MODE (regno)],
650 hard_regno))
651 && ! TEST_HARD_REG_BIT (impossible_start_hard_regs, hard_regno)
652 && (nregs_diff == 0
653 || (WORDS_BIG_ENDIAN
654 ? (hard_regno - nregs_diff >= 0
655 && TEST_HARD_REG_BIT (available_regs,
656 hard_regno - nregs_diff))
657 : TEST_HARD_REG_BIT (available_regs,
658 hard_regno + nregs_diff))))
660 if (hard_regno_costs_check[hard_regno]
661 != curr_hard_regno_costs_check)
663 hard_regno_costs_check[hard_regno] = curr_hard_regno_costs_check;
664 hard_regno_costs[hard_regno] = 0;
666 for (j = 0;
667 j < hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)];
668 j++)
669 if (! TEST_HARD_REG_BIT (call_used_reg_set, hard_regno + j)
670 && ! df_regs_ever_live_p (hard_regno + j))
671 /* It needs save restore. */
672 hard_regno_costs[hard_regno]
673 += (2
674 * REG_FREQ_FROM_BB (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb)
675 + 1);
676 priority = targetm.register_priority (hard_regno);
677 if (best_hard_regno < 0 || hard_regno_costs[hard_regno] < best_cost
678 || (hard_regno_costs[hard_regno] == best_cost
679 && (priority > best_priority
680 || (targetm.register_usage_leveling_p ()
681 && priority == best_priority
682 && best_usage > lra_hard_reg_usage[hard_regno]))))
684 best_hard_regno = hard_regno;
685 best_cost = hard_regno_costs[hard_regno];
686 best_priority = priority;
687 best_usage = lra_hard_reg_usage[hard_regno];
690 if (try_only_hard_regno >= 0 || (first_p && best_hard_regno >= 0))
691 break;
693 if (best_hard_regno >= 0)
694 *cost = best_cost - lra_reg_info[regno].freq;
695 return best_hard_regno;
698 /* A wrapper for find_hard_regno_for_1 (see comments for that function
699 description). This function tries to find a hard register for
700 preferred class first if it is worth. */
701 static int
702 find_hard_regno_for (int regno, int *cost, int try_only_hard_regno, bool first_p)
704 int hard_regno;
705 HARD_REG_SET regno_set;
707 /* Only original pseudos can have a different preferred class. */
708 if (try_only_hard_regno < 0 && regno < lra_new_regno_start)
710 enum reg_class pref_class = reg_preferred_class (regno);
712 if (regno_allocno_class_array[regno] != pref_class)
714 hard_regno = find_hard_regno_for_1 (regno, cost, -1, first_p,
715 reg_class_contents[pref_class]);
716 if (hard_regno >= 0)
717 return hard_regno;
720 CLEAR_HARD_REG_SET (regno_set);
721 return find_hard_regno_for_1 (regno, cost, try_only_hard_regno, first_p,
722 regno_set);
725 /* Current value used for checking elements in
726 update_hard_regno_preference_check. */
727 static int curr_update_hard_regno_preference_check;
728 /* If an element value is equal to the above variable value, then the
729 corresponding regno has been processed for preference
730 propagation. */
731 static int *update_hard_regno_preference_check;
733 /* Update the preference for using HARD_REGNO for pseudos that are
734 connected directly or indirectly with REGNO. Apply divisor DIV
735 to any preference adjustments.
737 The more indirectly a pseudo is connected, the smaller its effect
738 should be. We therefore increase DIV on each "hop". */
739 static void
740 update_hard_regno_preference (int regno, int hard_regno, int div)
742 int another_regno, cost;
743 lra_copy_t cp, next_cp;
745 /* Search depth 5 seems to be enough. */
746 if (div > (1 << 5))
747 return;
748 for (cp = lra_reg_info[regno].copies; cp != NULL; cp = next_cp)
750 if (cp->regno1 == regno)
752 next_cp = cp->regno1_next;
753 another_regno = cp->regno2;
755 else if (cp->regno2 == regno)
757 next_cp = cp->regno2_next;
758 another_regno = cp->regno1;
760 else
761 gcc_unreachable ();
762 if (reg_renumber[another_regno] < 0
763 && (update_hard_regno_preference_check[another_regno]
764 != curr_update_hard_regno_preference_check))
766 update_hard_regno_preference_check[another_regno]
767 = curr_update_hard_regno_preference_check;
768 cost = cp->freq < div ? 1 : cp->freq / div;
769 lra_setup_reload_pseudo_preferenced_hard_reg
770 (another_regno, hard_regno, cost);
771 update_hard_regno_preference (another_regno, hard_regno, div * 2);
776 /* Return prefix title for pseudo REGNO. */
777 static const char *
778 pseudo_prefix_title (int regno)
780 return
781 (regno < lra_constraint_new_regno_start ? ""
782 : bitmap_bit_p (&lra_inheritance_pseudos, regno) ? "inheritance "
783 : bitmap_bit_p (&lra_split_regs, regno) ? "split "
784 : bitmap_bit_p (&lra_optional_reload_pseudos, regno) ? "optional reload "
785 : bitmap_bit_p (&lra_subreg_reload_pseudos, regno) ? "subreg reload "
786 : "reload ");
789 /* Update REG_RENUMBER and other pseudo preferences by assignment of
790 HARD_REGNO to pseudo REGNO and print about it if PRINT_P. */
791 void
792 lra_setup_reg_renumber (int regno, int hard_regno, bool print_p)
794 int i, hr;
796 /* We can not just reassign hard register. */
797 lra_assert (hard_regno < 0 || reg_renumber[regno] < 0);
798 if ((hr = hard_regno) < 0)
799 hr = reg_renumber[regno];
800 reg_renumber[regno] = hard_regno;
801 lra_assert (hr >= 0);
802 for (i = 0; i < hard_regno_nregs[hr][PSEUDO_REGNO_MODE (regno)]; i++)
803 if (hard_regno < 0)
804 lra_hard_reg_usage[hr + i] -= lra_reg_info[regno].freq;
805 else
806 lra_hard_reg_usage[hr + i] += lra_reg_info[regno].freq;
807 if (print_p && lra_dump_file != NULL)
808 fprintf (lra_dump_file, " Assign %d to %sr%d (freq=%d)\n",
809 reg_renumber[regno], pseudo_prefix_title (regno),
810 regno, lra_reg_info[regno].freq);
811 if (hard_regno >= 0)
813 curr_update_hard_regno_preference_check++;
814 update_hard_regno_preference (regno, hard_regno, 1);
818 /* Pseudos which occur in insns containing a particular pseudo. */
819 static bitmap_head insn_conflict_pseudos;
821 /* Bitmaps used to contain spill pseudos for given pseudo hard regno
822 and best spill pseudos for given pseudo (and best hard regno). */
823 static bitmap_head spill_pseudos_bitmap, best_spill_pseudos_bitmap;
825 /* Current pseudo check for validity of elements in
826 TRY_HARD_REG_PSEUDOS. */
827 static int curr_pseudo_check;
828 /* Array used for validity of elements in TRY_HARD_REG_PSEUDOS. */
829 static int try_hard_reg_pseudos_check[FIRST_PSEUDO_REGISTER];
830 /* Pseudos who hold given hard register at the considered points. */
831 static bitmap_head try_hard_reg_pseudos[FIRST_PSEUDO_REGISTER];
833 /* Set up try_hard_reg_pseudos for given program point P and class
834 RCLASS. Those are pseudos living at P and assigned to a hard
835 register of RCLASS. In other words, those are pseudos which can be
836 spilled to assign a hard register of RCLASS to a pseudo living at
837 P. */
838 static void
839 setup_try_hard_regno_pseudos (int p, enum reg_class rclass)
841 int i, hard_regno;
842 machine_mode mode;
843 unsigned int spill_regno;
844 bitmap_iterator bi;
846 /* Find what pseudos could be spilled. */
847 EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[p], 0, spill_regno, bi)
849 mode = PSEUDO_REGNO_MODE (spill_regno);
850 hard_regno = live_pseudos_reg_renumber[spill_regno];
851 if (overlaps_hard_reg_set_p (reg_class_contents[rclass],
852 mode, hard_regno))
854 for (i = hard_regno_nregs[hard_regno][mode] - 1; i >= 0; i--)
856 if (try_hard_reg_pseudos_check[hard_regno + i]
857 != curr_pseudo_check)
859 try_hard_reg_pseudos_check[hard_regno + i]
860 = curr_pseudo_check;
861 bitmap_clear (&try_hard_reg_pseudos[hard_regno + i]);
863 bitmap_set_bit (&try_hard_reg_pseudos[hard_regno + i],
864 spill_regno);
870 /* Assign temporarily HARD_REGNO to pseudo REGNO. Temporary
871 assignment means that we might undo the data change. */
872 static void
873 assign_temporarily (int regno, int hard_regno)
875 int p;
876 lra_live_range_t r;
878 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
880 for (p = r->start; p <= r->finish; p++)
881 if (hard_regno < 0)
882 bitmap_clear_bit (&live_hard_reg_pseudos[p], regno);
883 else
885 bitmap_set_bit (&live_hard_reg_pseudos[p], regno);
886 insert_in_live_range_start_chain (regno);
889 live_pseudos_reg_renumber[regno] = hard_regno;
892 /* Array used for sorting reload pseudos for subsequent allocation
893 after spilling some pseudo. */
894 static int *sorted_reload_pseudos;
896 /* Spill some pseudos for a reload pseudo REGNO and return hard
897 register which should be used for pseudo after spilling. The
898 function adds spilled pseudos to SPILLED_PSEUDO_BITMAP. When we
899 choose hard register (and pseudos occupying the hard registers and
900 to be spilled), we take into account not only how REGNO will
901 benefit from the spills but also how other reload pseudos not yet
902 assigned to hard registers benefit from the spills too. In very
903 rare cases, the function can fail and return -1.
905 If FIRST_P, return the first available hard reg ignoring other
906 criteria, e.g. allocation cost and cost of spilling non-reload
907 pseudos. This approach results in less hard reg pool fragmentation
908 and permit to allocate hard regs to reload pseudos in complicated
909 situations where pseudo sizes are different. */
910 static int
911 spill_for (int regno, bitmap spilled_pseudo_bitmap, bool first_p)
913 int i, j, n, p, hard_regno, best_hard_regno, cost, best_cost, rclass_size;
914 int reload_hard_regno, reload_cost;
915 bool static_p, best_static_p;
916 machine_mode mode;
917 enum reg_class rclass;
918 unsigned int spill_regno, reload_regno, uid;
919 int insn_pseudos_num, best_insn_pseudos_num;
920 int bad_spills_num, smallest_bad_spills_num;
921 lra_live_range_t r;
922 bitmap_iterator bi;
924 rclass = regno_allocno_class_array[regno];
925 lra_assert (reg_renumber[regno] < 0 && rclass != NO_REGS);
926 bitmap_clear (&insn_conflict_pseudos);
927 bitmap_clear (&best_spill_pseudos_bitmap);
928 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
930 struct lra_insn_reg *ir;
932 for (ir = lra_get_insn_regs (uid); ir != NULL; ir = ir->next)
933 if (ir->regno >= FIRST_PSEUDO_REGISTER)
934 bitmap_set_bit (&insn_conflict_pseudos, ir->regno);
936 best_hard_regno = -1;
937 best_cost = INT_MAX;
938 best_static_p = TRUE;
939 best_insn_pseudos_num = INT_MAX;
940 smallest_bad_spills_num = INT_MAX;
941 rclass_size = ira_class_hard_regs_num[rclass];
942 mode = PSEUDO_REGNO_MODE (regno);
943 /* Invalidate try_hard_reg_pseudos elements. */
944 curr_pseudo_check++;
945 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
946 for (p = r->start; p <= r->finish; p++)
947 setup_try_hard_regno_pseudos (p, rclass);
948 for (i = 0; i < rclass_size; i++)
950 hard_regno = ira_class_hard_regs[rclass][i];
951 bitmap_clear (&spill_pseudos_bitmap);
952 for (j = hard_regno_nregs[hard_regno][mode] - 1; j >= 0; j--)
954 if (try_hard_reg_pseudos_check[hard_regno + j] != curr_pseudo_check)
955 continue;
956 lra_assert (!bitmap_empty_p (&try_hard_reg_pseudos[hard_regno + j]));
957 bitmap_ior_into (&spill_pseudos_bitmap,
958 &try_hard_reg_pseudos[hard_regno + j]);
960 /* Spill pseudos. */
961 static_p = false;
962 EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
963 if ((pic_offset_table_rtx != NULL
964 && spill_regno == REGNO (pic_offset_table_rtx))
965 || ((int) spill_regno >= lra_constraint_new_regno_start
966 && ! bitmap_bit_p (&lra_inheritance_pseudos, spill_regno)
967 && ! bitmap_bit_p (&lra_split_regs, spill_regno)
968 && ! bitmap_bit_p (&lra_subreg_reload_pseudos, spill_regno)
969 && ! bitmap_bit_p (&lra_optional_reload_pseudos, spill_regno)))
970 goto fail;
971 else if (non_spilled_static_chain_regno_p (spill_regno))
972 static_p = true;
973 insn_pseudos_num = 0;
974 bad_spills_num = 0;
975 if (lra_dump_file != NULL)
976 fprintf (lra_dump_file, " Trying %d:", hard_regno);
977 sparseset_clear (live_range_reload_inheritance_pseudos);
978 EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
980 if (bitmap_bit_p (&insn_conflict_pseudos, spill_regno))
981 insn_pseudos_num++;
982 if (spill_regno >= (unsigned int) lra_bad_spill_regno_start)
983 bad_spills_num++;
984 for (r = lra_reg_info[spill_regno].live_ranges;
985 r != NULL;
986 r = r->next)
988 for (p = r->start; p <= r->finish; p++)
990 lra_live_range_t r2;
992 for (r2 = start_point_ranges[p];
993 r2 != NULL;
994 r2 = r2->start_next)
995 if (r2->regno >= lra_constraint_new_regno_start)
996 sparseset_set_bit (live_range_reload_inheritance_pseudos,
997 r2->regno);
1001 n = 0;
1002 if (sparseset_cardinality (live_range_reload_inheritance_pseudos)
1003 <= (unsigned)LRA_MAX_CONSIDERED_RELOAD_PSEUDOS)
1004 EXECUTE_IF_SET_IN_SPARSESET (live_range_reload_inheritance_pseudos,
1005 reload_regno)
1006 if ((int) reload_regno != regno
1007 && (ira_reg_classes_intersect_p
1008 [rclass][regno_allocno_class_array[reload_regno]])
1009 && live_pseudos_reg_renumber[reload_regno] < 0
1010 && find_hard_regno_for (reload_regno, &cost, -1, first_p) < 0)
1011 sorted_reload_pseudos[n++] = reload_regno;
1012 EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
1014 update_lives (spill_regno, true);
1015 if (lra_dump_file != NULL)
1016 fprintf (lra_dump_file, " spill %d(freq=%d)",
1017 spill_regno, lra_reg_info[spill_regno].freq);
1019 hard_regno = find_hard_regno_for (regno, &cost, -1, first_p);
1020 if (hard_regno >= 0)
1022 assign_temporarily (regno, hard_regno);
1023 qsort (sorted_reload_pseudos, n, sizeof (int),
1024 reload_pseudo_compare_func);
1025 for (j = 0; j < n; j++)
1027 reload_regno = sorted_reload_pseudos[j];
1028 lra_assert (live_pseudos_reg_renumber[reload_regno] < 0);
1029 if ((reload_hard_regno
1030 = find_hard_regno_for (reload_regno,
1031 &reload_cost, -1, first_p)) >= 0)
1033 if (lra_dump_file != NULL)
1034 fprintf (lra_dump_file, " assign %d(cost=%d)",
1035 reload_regno, reload_cost);
1036 assign_temporarily (reload_regno, reload_hard_regno);
1037 cost += reload_cost;
1040 EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
1042 rtx_insn_list *x;
1044 cost += lra_reg_info[spill_regno].freq;
1045 if (ira_reg_equiv[spill_regno].memory != NULL
1046 || ira_reg_equiv[spill_regno].constant != NULL)
1047 for (x = ira_reg_equiv[spill_regno].init_insns;
1048 x != NULL;
1049 x = x->next ())
1050 cost -= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (x->insn ()));
1052 /* Avoid spilling static chain pointer pseudo when non-local
1053 goto is used. */
1054 if ((! static_p && best_static_p)
1055 || (static_p == best_static_p
1056 && (best_insn_pseudos_num > insn_pseudos_num
1057 || (best_insn_pseudos_num == insn_pseudos_num
1058 && (bad_spills_num < smallest_bad_spills_num
1059 || (bad_spills_num == smallest_bad_spills_num
1060 && best_cost > cost))))))
1062 best_insn_pseudos_num = insn_pseudos_num;
1063 smallest_bad_spills_num = bad_spills_num;
1064 best_static_p = static_p;
1065 best_cost = cost;
1066 best_hard_regno = hard_regno;
1067 bitmap_copy (&best_spill_pseudos_bitmap, &spill_pseudos_bitmap);
1068 if (lra_dump_file != NULL)
1069 fprintf (lra_dump_file,
1070 " Now best %d(cost=%d, bad_spills=%d, insn_pseudos=%d)\n",
1071 hard_regno, cost, bad_spills_num, insn_pseudos_num);
1073 assign_temporarily (regno, -1);
1074 for (j = 0; j < n; j++)
1076 reload_regno = sorted_reload_pseudos[j];
1077 if (live_pseudos_reg_renumber[reload_regno] >= 0)
1078 assign_temporarily (reload_regno, -1);
1081 if (lra_dump_file != NULL)
1082 fprintf (lra_dump_file, "\n");
1083 /* Restore the live hard reg pseudo info for spilled pseudos. */
1084 EXECUTE_IF_SET_IN_BITMAP (&spill_pseudos_bitmap, 0, spill_regno, bi)
1085 update_lives (spill_regno, false);
1086 fail:
1089 /* Spill: */
1090 EXECUTE_IF_SET_IN_BITMAP (&best_spill_pseudos_bitmap, 0, spill_regno, bi)
1092 if ((int) spill_regno >= lra_constraint_new_regno_start)
1093 former_reload_pseudo_spill_p = true;
1094 if (lra_dump_file != NULL)
1095 fprintf (lra_dump_file, " Spill %sr%d(hr=%d, freq=%d) for r%d\n",
1096 pseudo_prefix_title (spill_regno),
1097 spill_regno, reg_renumber[spill_regno],
1098 lra_reg_info[spill_regno].freq, regno);
1099 update_lives (spill_regno, true);
1100 lra_setup_reg_renumber (spill_regno, -1, false);
1102 bitmap_ior_into (spilled_pseudo_bitmap, &best_spill_pseudos_bitmap);
1103 return best_hard_regno;
1106 /* Assign HARD_REGNO to REGNO. */
1107 static void
1108 assign_hard_regno (int hard_regno, int regno)
1110 int i;
1112 lra_assert (hard_regno >= 0);
1113 lra_setup_reg_renumber (regno, hard_regno, true);
1114 update_lives (regno, false);
1115 for (i = 0;
1116 i < hard_regno_nregs[hard_regno][lra_reg_info[regno].biggest_mode];
1117 i++)
1118 df_set_regs_ever_live (hard_regno + i, true);
1121 /* Array used for sorting different pseudos. */
1122 static int *sorted_pseudos;
1124 /* The constraints pass is allowed to create equivalences between
1125 pseudos that make the current allocation "incorrect" (in the sense
1126 that pseudos are assigned to hard registers from their own conflict
1127 sets). The global variable lra_risky_transformations_p says
1128 whether this might have happened.
1130 Process pseudos assigned to hard registers (less frequently used
1131 first), spill if a conflict is found, and mark the spilled pseudos
1132 in SPILLED_PSEUDO_BITMAP. Set up LIVE_HARD_REG_PSEUDOS from
1133 pseudos, assigned to hard registers. */
1134 static void
1135 setup_live_pseudos_and_spill_after_risky_transforms (bitmap
1136 spilled_pseudo_bitmap)
1138 int p, i, j, n, regno, hard_regno;
1139 unsigned int k, conflict_regno;
1140 int val, offset;
1141 HARD_REG_SET conflict_set;
1142 machine_mode mode;
1143 lra_live_range_t r;
1144 bitmap_iterator bi;
1145 int max_regno = max_reg_num ();
1147 if (! lra_risky_transformations_p)
1149 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1150 if (reg_renumber[i] >= 0 && lra_reg_info[i].nrefs > 0)
1151 update_lives (i, false);
1152 return;
1154 for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1155 if ((pic_offset_table_rtx == NULL_RTX
1156 || i != (int) REGNO (pic_offset_table_rtx))
1157 && reg_renumber[i] >= 0 && lra_reg_info[i].nrefs > 0)
1158 sorted_pseudos[n++] = i;
1159 qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func);
1160 if (pic_offset_table_rtx != NULL_RTX
1161 && (regno = REGNO (pic_offset_table_rtx)) >= FIRST_PSEUDO_REGISTER
1162 && reg_renumber[regno] >= 0 && lra_reg_info[regno].nrefs > 0)
1163 sorted_pseudos[n++] = regno;
1164 for (i = n - 1; i >= 0; i--)
1166 regno = sorted_pseudos[i];
1167 hard_regno = reg_renumber[regno];
1168 lra_assert (hard_regno >= 0);
1169 mode = lra_reg_info[regno].biggest_mode;
1170 sparseset_clear (live_range_hard_reg_pseudos);
1171 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
1173 EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi)
1174 sparseset_set_bit (live_range_hard_reg_pseudos, k);
1175 for (p = r->start + 1; p <= r->finish; p++)
1177 lra_live_range_t r2;
1179 for (r2 = start_point_ranges[p];
1180 r2 != NULL;
1181 r2 = r2->start_next)
1182 if (live_pseudos_reg_renumber[r2->regno] >= 0)
1183 sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno);
1186 COPY_HARD_REG_SET (conflict_set, lra_no_alloc_regs);
1187 IOR_HARD_REG_SET (conflict_set, lra_reg_info[regno].conflict_hard_regs);
1188 val = lra_reg_info[regno].val;
1189 offset = lra_reg_info[regno].offset;
1190 EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno)
1191 if (!lra_reg_val_equal_p (conflict_regno, val, offset)
1192 /* If it is multi-register pseudos they should start on
1193 the same hard register. */
1194 || hard_regno != reg_renumber[conflict_regno])
1196 int conflict_hard_regno = reg_renumber[conflict_regno];
1197 machine_mode biggest_mode = lra_reg_info[conflict_regno].biggest_mode;
1198 int biggest_nregs = hard_regno_nregs[conflict_hard_regno][biggest_mode];
1199 int nregs_diff = (biggest_nregs
1200 - (hard_regno_nregs
1201 [conflict_hard_regno]
1202 [PSEUDO_REGNO_MODE (conflict_regno)]));
1203 add_to_hard_reg_set (&conflict_set,
1204 biggest_mode,
1205 conflict_hard_regno
1206 - (WORDS_BIG_ENDIAN ? nregs_diff : 0));
1208 if (! overlaps_hard_reg_set_p (conflict_set, mode, hard_regno))
1210 update_lives (regno, false);
1211 continue;
1213 bitmap_set_bit (spilled_pseudo_bitmap, regno);
1214 for (j = 0;
1215 j < hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (regno)];
1216 j++)
1217 lra_hard_reg_usage[hard_regno + j] -= lra_reg_info[regno].freq;
1218 reg_renumber[regno] = -1;
1219 if (regno >= lra_constraint_new_regno_start)
1220 former_reload_pseudo_spill_p = true;
1221 if (lra_dump_file != NULL)
1222 fprintf (lra_dump_file, " Spill r%d after risky transformations\n",
1223 regno);
1227 /* Improve allocation by assigning the same hard regno of inheritance
1228 pseudos to the connected pseudos. We need this because inheritance
1229 pseudos are allocated after reload pseudos in the thread and when
1230 we assign a hard register to a reload pseudo we don't know yet that
1231 the connected inheritance pseudos can get the same hard register.
1232 Add pseudos with changed allocation to bitmap CHANGED_PSEUDOS. */
1233 static void
1234 improve_inheritance (bitmap changed_pseudos)
1236 unsigned int k;
1237 int regno, another_regno, hard_regno, another_hard_regno, cost, i, n;
1238 lra_copy_t cp, next_cp;
1239 bitmap_iterator bi;
1241 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
1242 return;
1243 n = 0;
1244 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, k, bi)
1245 if (reg_renumber[k] >= 0 && lra_reg_info[k].nrefs != 0)
1246 sorted_pseudos[n++] = k;
1247 qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func);
1248 for (i = 0; i < n; i++)
1250 regno = sorted_pseudos[i];
1251 hard_regno = reg_renumber[regno];
1252 lra_assert (hard_regno >= 0);
1253 for (cp = lra_reg_info[regno].copies; cp != NULL; cp = next_cp)
1255 if (cp->regno1 == regno)
1257 next_cp = cp->regno1_next;
1258 another_regno = cp->regno2;
1260 else if (cp->regno2 == regno)
1262 next_cp = cp->regno2_next;
1263 another_regno = cp->regno1;
1265 else
1266 gcc_unreachable ();
1267 /* Don't change reload pseudo allocation. It might have
1268 this allocation for a purpose and changing it can result
1269 in LRA cycling. */
1270 if ((another_regno < lra_constraint_new_regno_start
1271 || bitmap_bit_p (&lra_inheritance_pseudos, another_regno))
1272 && (another_hard_regno = reg_renumber[another_regno]) >= 0
1273 && another_hard_regno != hard_regno)
1275 if (lra_dump_file != NULL)
1276 fprintf
1277 (lra_dump_file,
1278 " Improving inheritance for %d(%d) and %d(%d)...\n",
1279 regno, hard_regno, another_regno, another_hard_regno);
1280 update_lives (another_regno, true);
1281 lra_setup_reg_renumber (another_regno, -1, false);
1282 if (hard_regno == find_hard_regno_for (another_regno, &cost,
1283 hard_regno, false))
1284 assign_hard_regno (hard_regno, another_regno);
1285 else
1286 assign_hard_regno (another_hard_regno, another_regno);
1287 bitmap_set_bit (changed_pseudos, another_regno);
1294 /* Bitmap finally containing all pseudos spilled on this assignment
1295 pass. */
1296 static bitmap_head all_spilled_pseudos;
1297 /* All pseudos whose allocation was changed. */
1298 static bitmap_head changed_pseudo_bitmap;
1301 /* Add to LIVE_RANGE_HARD_REG_PSEUDOS all pseudos conflicting with
1302 REGNO and whose hard regs can be assigned to REGNO. */
1303 static void
1304 find_all_spills_for (int regno)
1306 int p;
1307 lra_live_range_t r;
1308 unsigned int k;
1309 bitmap_iterator bi;
1310 enum reg_class rclass;
1311 bool *rclass_intersect_p;
1313 rclass = regno_allocno_class_array[regno];
1314 rclass_intersect_p = ira_reg_classes_intersect_p[rclass];
1315 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
1317 EXECUTE_IF_SET_IN_BITMAP (&live_hard_reg_pseudos[r->start], 0, k, bi)
1318 if (rclass_intersect_p[regno_allocno_class_array[k]])
1319 sparseset_set_bit (live_range_hard_reg_pseudos, k);
1320 for (p = r->start + 1; p <= r->finish; p++)
1322 lra_live_range_t r2;
1324 for (r2 = start_point_ranges[p];
1325 r2 != NULL;
1326 r2 = r2->start_next)
1328 if (live_pseudos_reg_renumber[r2->regno] >= 0
1329 && rclass_intersect_p[regno_allocno_class_array[r2->regno]])
1330 sparseset_set_bit (live_range_hard_reg_pseudos, r2->regno);
1336 /* Assign hard registers to reload pseudos and other pseudos. */
1337 static void
1338 assign_by_spills (void)
1340 int i, n, nfails, iter, regno, hard_regno, cost;
1341 rtx restore_rtx;
1342 rtx_insn *insn;
1343 bitmap_head changed_insns, do_not_assign_nonreload_pseudos;
1344 unsigned int u, conflict_regno;
1345 bitmap_iterator bi;
1346 bool reload_p;
1347 int max_regno = max_reg_num ();
1349 for (n = 0, i = lra_constraint_new_regno_start; i < max_regno; i++)
1350 if (reg_renumber[i] < 0 && lra_reg_info[i].nrefs != 0
1351 && regno_allocno_class_array[i] != NO_REGS)
1352 sorted_pseudos[n++] = i;
1353 bitmap_initialize (&insn_conflict_pseudos, &reg_obstack);
1354 bitmap_initialize (&spill_pseudos_bitmap, &reg_obstack);
1355 bitmap_initialize (&best_spill_pseudos_bitmap, &reg_obstack);
1356 update_hard_regno_preference_check = XCNEWVEC (int, max_regno);
1357 curr_update_hard_regno_preference_check = 0;
1358 memset (try_hard_reg_pseudos_check, 0, sizeof (try_hard_reg_pseudos_check));
1359 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1360 bitmap_initialize (&try_hard_reg_pseudos[i], &reg_obstack);
1361 curr_pseudo_check = 0;
1362 bitmap_initialize (&changed_insns, &reg_obstack);
1363 bitmap_initialize (&non_reload_pseudos, &reg_obstack);
1364 bitmap_ior (&non_reload_pseudos, &lra_inheritance_pseudos, &lra_split_regs);
1365 bitmap_ior_into (&non_reload_pseudos, &lra_subreg_reload_pseudos);
1366 bitmap_ior_into (&non_reload_pseudos, &lra_optional_reload_pseudos);
1367 for (iter = 0; iter <= 1; iter++)
1369 qsort (sorted_pseudos, n, sizeof (int), reload_pseudo_compare_func);
1370 nfails = 0;
1371 for (i = 0; i < n; i++)
1373 regno = sorted_pseudos[i];
1374 if (reg_renumber[regno] >= 0)
1375 continue;
1376 if (lra_dump_file != NULL)
1377 fprintf (lra_dump_file, " Assigning to %d "
1378 "(cl=%s, orig=%d, freq=%d, tfirst=%d, tfreq=%d)...\n",
1379 regno, reg_class_names[regno_allocno_class_array[regno]],
1380 ORIGINAL_REGNO (regno_reg_rtx[regno]),
1381 lra_reg_info[regno].freq, regno_assign_info[regno].first,
1382 regno_assign_info[regno_assign_info[regno].first].freq);
1383 hard_regno = find_hard_regno_for (regno, &cost, -1, iter == 1);
1384 reload_p = ! bitmap_bit_p (&non_reload_pseudos, regno);
1385 if (hard_regno < 0 && reload_p)
1386 hard_regno = spill_for (regno, &all_spilled_pseudos, iter == 1);
1387 if (hard_regno < 0)
1389 if (reload_p)
1390 sorted_pseudos[nfails++] = regno;
1392 else
1394 /* This register might have been spilled by the previous
1395 pass. Indicate that it is no longer spilled. */
1396 bitmap_clear_bit (&all_spilled_pseudos, regno);
1397 assign_hard_regno (hard_regno, regno);
1398 if (! reload_p)
1399 /* As non-reload pseudo assignment is changed we
1400 should reconsider insns referring for the
1401 pseudo. */
1402 bitmap_set_bit (&changed_pseudo_bitmap, regno);
1405 if (nfails == 0)
1406 break;
1407 if (iter > 0)
1409 /* We did not assign hard regs to reload pseudos after two iterations.
1410 Either it's an asm and something is wrong with the constraints, or
1411 we have run out of spill registers; error out in either case. */
1412 bool asm_p = false;
1413 bitmap_head failed_reload_insns;
1415 bitmap_initialize (&failed_reload_insns, &reg_obstack);
1416 for (i = 0; i < nfails; i++)
1418 regno = sorted_pseudos[i];
1419 bitmap_ior_into (&failed_reload_insns,
1420 &lra_reg_info[regno].insn_bitmap);
1421 /* Assign an arbitrary hard register of regno class to
1422 avoid further trouble with this insn. */
1423 bitmap_clear_bit (&all_spilled_pseudos, regno);
1424 assign_hard_regno
1425 (ira_class_hard_regs[regno_allocno_class_array[regno]][0],
1426 regno);
1428 EXECUTE_IF_SET_IN_BITMAP (&failed_reload_insns, 0, u, bi)
1430 insn = lra_insn_recog_data[u]->insn;
1431 if (asm_noperands (PATTERN (insn)) >= 0)
1433 asm_p = true;
1434 error_for_asm (insn,
1435 "%<asm%> operand has impossible constraints");
1436 /* Avoid further trouble with this insn.
1437 For asm goto, instead of fixing up all the edges
1438 just clear the template and clear input operands
1439 (asm goto doesn't have any output operands). */
1440 if (JUMP_P (insn))
1442 rtx asm_op = extract_asm_operands (PATTERN (insn));
1443 ASM_OPERANDS_TEMPLATE (asm_op) = ggc_strdup ("");
1444 ASM_OPERANDS_INPUT_VEC (asm_op) = rtvec_alloc (0);
1445 ASM_OPERANDS_INPUT_CONSTRAINT_VEC (asm_op) = rtvec_alloc (0);
1446 lra_update_insn_regno_info (insn);
1448 else
1450 PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx);
1451 lra_set_insn_deleted (insn);
1454 else if (!asm_p)
1456 error ("unable to find a register to spill");
1457 fatal_insn ("this is the insn:", insn);
1460 break;
1462 /* This is a very rare event. We can not assign a hard register
1463 to reload pseudo because the hard register was assigned to
1464 another reload pseudo on a previous assignment pass. For x86
1465 example, on the 1st pass we assigned CX (although another
1466 hard register could be used for this) to reload pseudo in an
1467 insn, on the 2nd pass we need CX (and only this) hard
1468 register for a new reload pseudo in the same insn. Another
1469 possible situation may occur in assigning to multi-regs
1470 reload pseudos when hard regs pool is too fragmented even
1471 after spilling non-reload pseudos.
1473 We should do something radical here to succeed. Here we
1474 spill *all* conflicting pseudos and reassign them. */
1475 if (lra_dump_file != NULL)
1476 fprintf (lra_dump_file, " 2nd iter for reload pseudo assignments:\n");
1477 sparseset_clear (live_range_hard_reg_pseudos);
1478 for (i = 0; i < nfails; i++)
1480 if (lra_dump_file != NULL)
1481 fprintf (lra_dump_file, " Reload r%d assignment failure\n",
1482 sorted_pseudos[i]);
1483 find_all_spills_for (sorted_pseudos[i]);
1485 EXECUTE_IF_SET_IN_SPARSESET (live_range_hard_reg_pseudos, conflict_regno)
1487 if ((int) conflict_regno >= lra_constraint_new_regno_start)
1489 sorted_pseudos[nfails++] = conflict_regno;
1490 former_reload_pseudo_spill_p = true;
1492 if (lra_dump_file != NULL)
1493 fprintf (lra_dump_file, " Spill %s r%d(hr=%d, freq=%d)\n",
1494 pseudo_prefix_title (conflict_regno), conflict_regno,
1495 reg_renumber[conflict_regno],
1496 lra_reg_info[conflict_regno].freq);
1497 update_lives (conflict_regno, true);
1498 lra_setup_reg_renumber (conflict_regno, -1, false);
1500 n = nfails;
1502 improve_inheritance (&changed_pseudo_bitmap);
1503 bitmap_clear (&non_reload_pseudos);
1504 bitmap_clear (&changed_insns);
1505 if (! lra_simple_p)
1507 /* We should not assign to original pseudos of inheritance
1508 pseudos or split pseudos if any its inheritance pseudo did
1509 not get hard register or any its split pseudo was not split
1510 because undo inheritance/split pass will extend live range of
1511 such inheritance or split pseudos. */
1512 bitmap_initialize (&do_not_assign_nonreload_pseudos, &reg_obstack);
1513 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, u, bi)
1514 if ((restore_rtx = lra_reg_info[u].restore_rtx) != NULL_RTX
1515 && REG_P (restore_rtx)
1516 && reg_renumber[u] < 0
1517 && bitmap_bit_p (&lra_inheritance_pseudos, u))
1518 bitmap_set_bit (&do_not_assign_nonreload_pseudos, REGNO (restore_rtx));
1519 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, u, bi)
1520 if ((restore_rtx = lra_reg_info[u].restore_rtx) != NULL_RTX
1521 && reg_renumber[u] >= 0)
1523 lra_assert (REG_P (restore_rtx));
1524 bitmap_set_bit (&do_not_assign_nonreload_pseudos, REGNO (restore_rtx));
1526 for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1527 if (((i < lra_constraint_new_regno_start
1528 && ! bitmap_bit_p (&do_not_assign_nonreload_pseudos, i))
1529 || (bitmap_bit_p (&lra_inheritance_pseudos, i)
1530 && lra_reg_info[i].restore_rtx != NULL_RTX)
1531 || (bitmap_bit_p (&lra_split_regs, i)
1532 && lra_reg_info[i].restore_rtx != NULL_RTX)
1533 || bitmap_bit_p (&lra_subreg_reload_pseudos, i)
1534 || bitmap_bit_p (&lra_optional_reload_pseudos, i))
1535 && reg_renumber[i] < 0 && lra_reg_info[i].nrefs != 0
1536 && regno_allocno_class_array[i] != NO_REGS)
1537 sorted_pseudos[n++] = i;
1538 bitmap_clear (&do_not_assign_nonreload_pseudos);
1539 if (n != 0 && lra_dump_file != NULL)
1540 fprintf (lra_dump_file, " Reassigning non-reload pseudos\n");
1541 qsort (sorted_pseudos, n, sizeof (int), pseudo_compare_func);
1542 for (i = 0; i < n; i++)
1544 regno = sorted_pseudos[i];
1545 hard_regno = find_hard_regno_for (regno, &cost, -1, false);
1546 if (hard_regno >= 0)
1548 assign_hard_regno (hard_regno, regno);
1549 /* We change allocation for non-reload pseudo on this
1550 iteration -- mark the pseudo for invalidation of used
1551 alternatives of insns containing the pseudo. */
1552 bitmap_set_bit (&changed_pseudo_bitmap, regno);
1554 else
1556 enum reg_class rclass = lra_get_allocno_class (regno);
1557 enum reg_class spill_class;
1559 if (targetm.spill_class == NULL
1560 || lra_reg_info[regno].restore_rtx == NULL_RTX
1561 || ! bitmap_bit_p (&lra_inheritance_pseudos, regno)
1562 || (spill_class
1563 = ((enum reg_class)
1564 targetm.spill_class
1565 ((reg_class_t) rclass,
1566 PSEUDO_REGNO_MODE (regno)))) == NO_REGS)
1567 continue;
1568 regno_allocno_class_array[regno] = spill_class;
1569 hard_regno = find_hard_regno_for (regno, &cost, -1, false);
1570 if (hard_regno < 0)
1571 regno_allocno_class_array[regno] = rclass;
1572 else
1574 setup_reg_classes
1575 (regno, spill_class, spill_class, spill_class);
1576 assign_hard_regno (hard_regno, regno);
1577 bitmap_set_bit (&changed_pseudo_bitmap, regno);
1582 free (update_hard_regno_preference_check);
1583 bitmap_clear (&best_spill_pseudos_bitmap);
1584 bitmap_clear (&spill_pseudos_bitmap);
1585 bitmap_clear (&insn_conflict_pseudos);
1589 /* Entry function to assign hard registers to new reload pseudos
1590 starting with LRA_CONSTRAINT_NEW_REGNO_START (by possible spilling
1591 of old pseudos) and possibly to the old pseudos. The function adds
1592 what insns to process for the next constraint pass. Those are all
1593 insns who contains non-reload and non-inheritance pseudos with
1594 changed allocation.
1596 Return true if we did not spill any non-reload and non-inheritance
1597 pseudos. */
1598 bool
1599 lra_assign (void)
1601 int i;
1602 unsigned int u;
1603 bitmap_iterator bi;
1604 bitmap_head insns_to_process;
1605 bool no_spills_p;
1606 int max_regno = max_reg_num ();
1608 timevar_push (TV_LRA_ASSIGN);
1609 lra_assignment_iter++;
1610 if (lra_dump_file != NULL)
1611 fprintf (lra_dump_file, "\n********** Assignment #%d: **********\n\n",
1612 lra_assignment_iter);
1613 init_lives ();
1614 sorted_pseudos = XNEWVEC (int, max_regno);
1615 sorted_reload_pseudos = XNEWVEC (int, max_regno);
1616 regno_allocno_class_array = XNEWVEC (enum reg_class, max_regno);
1617 regno_live_length = XNEWVEC (int, max_regno);
1618 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1620 int l;
1621 lra_live_range_t r;
1623 regno_allocno_class_array[i] = lra_get_allocno_class (i);
1624 for (l = 0, r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
1625 l += r->finish - r->start + 1;
1626 regno_live_length[i] = l;
1628 former_reload_pseudo_spill_p = false;
1629 init_regno_assign_info ();
1630 bitmap_initialize (&all_spilled_pseudos, &reg_obstack);
1631 create_live_range_start_chains ();
1632 setup_live_pseudos_and_spill_after_risky_transforms (&all_spilled_pseudos);
1633 if (flag_checking && !flag_ipa_ra)
1634 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
1635 if (lra_reg_info[i].nrefs != 0 && reg_renumber[i] >= 0
1636 && lra_reg_info[i].call_p
1637 && overlaps_hard_reg_set_p (call_used_reg_set,
1638 PSEUDO_REGNO_MODE (i), reg_renumber[i]))
1639 gcc_unreachable ();
1640 /* Setup insns to process on the next constraint pass. */
1641 bitmap_initialize (&changed_pseudo_bitmap, &reg_obstack);
1642 init_live_reload_and_inheritance_pseudos ();
1643 assign_by_spills ();
1644 finish_live_reload_and_inheritance_pseudos ();
1645 bitmap_ior_into (&changed_pseudo_bitmap, &all_spilled_pseudos);
1646 no_spills_p = true;
1647 EXECUTE_IF_SET_IN_BITMAP (&all_spilled_pseudos, 0, u, bi)
1648 /* We ignore spilled pseudos created on last inheritance pass
1649 because they will be removed. */
1650 if (lra_reg_info[u].restore_rtx == NULL_RTX)
1652 no_spills_p = false;
1653 break;
1655 finish_live_range_start_chains ();
1656 bitmap_clear (&all_spilled_pseudos);
1657 bitmap_initialize (&insns_to_process, &reg_obstack);
1658 EXECUTE_IF_SET_IN_BITMAP (&changed_pseudo_bitmap, 0, u, bi)
1659 bitmap_ior_into (&insns_to_process, &lra_reg_info[u].insn_bitmap);
1660 bitmap_clear (&changed_pseudo_bitmap);
1661 EXECUTE_IF_SET_IN_BITMAP (&insns_to_process, 0, u, bi)
1663 lra_push_insn_by_uid (u);
1664 /* Invalidate alternatives for insn should be processed. */
1665 lra_set_used_insn_alternative_by_uid (u, -1);
1667 bitmap_clear (&insns_to_process);
1668 finish_regno_assign_info ();
1669 free (regno_live_length);
1670 free (regno_allocno_class_array);
1671 free (sorted_pseudos);
1672 free (sorted_reload_pseudos);
1673 finish_lives ();
1674 timevar_pop (TV_LRA_ASSIGN);
1675 if (former_reload_pseudo_spill_p)
1676 lra_assignment_iter_after_spill++;
1677 /* This is conditional on flag_checking because valid code can take
1678 more than this maximum number of iteration, but at the same time
1679 the test can uncover errors in machine descriptions. */
1680 if (flag_checking
1681 && (lra_assignment_iter_after_spill
1682 > LRA_MAX_ASSIGNMENT_ITERATION_NUMBER))
1683 internal_error
1684 ("Maximum number of LRA assignment passes is achieved (%d)\n",
1685 LRA_MAX_ASSIGNMENT_ITERATION_NUMBER);
1686 return no_spills_p;