1 | /* Build live ranges for pseudos. |
2 | Copyright (C) 2010-2024 Free Software Foundation, Inc. |
3 | Contributed by Vladimir Makarov <vmakarov@redhat.com>. |
4 | |
5 | This file is part of GCC. |
6 | |
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. |
11 | |
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. |
16 | |
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/>. */ |
20 | |
21 | |
22 | /* This file contains code to build pseudo live-ranges (analogous |
23 | structures used in IRA, so read comments about the live-ranges |
24 | there) and other info necessary for other passes to assign |
25 | hard-registers to pseudos, coalesce the spilled pseudos, and assign |
26 | stack memory slots to spilled pseudos. */ |
27 | |
28 | #include "config.h" |
29 | #include "system.h" |
30 | #include "coretypes.h" |
31 | #include "backend.h" |
32 | #include "rtl.h" |
33 | #include "tree.h" |
34 | #include "predict.h" |
35 | #include "df.h" |
36 | #include "memmodel.h" |
37 | #include "tm_p.h" |
38 | #include "insn-config.h" |
39 | #include "regs.h" |
40 | #include "ira.h" |
41 | #include "recog.h" |
42 | #include "cfganal.h" |
43 | #include "sparseset.h" |
44 | #include "lra-int.h" |
45 | #include "target.h" |
46 | #include "function-abi.h" |
47 | |
48 | /* Program points are enumerated by numbers from range |
49 | 0..LRA_LIVE_MAX_POINT-1. There are approximately two times more |
50 | program points than insns. Program points are places in the |
51 | program where liveness info can be changed. In most general case |
52 | (there are more complicated cases too) some program points |
53 | correspond to places where input operand dies and other ones |
54 | correspond to places where output operands are born. */ |
55 | int lra_live_max_point; |
56 | |
57 | /* Accumulated execution frequency of all references for each hard |
58 | register. */ |
59 | int lra_hard_reg_usage[FIRST_PSEUDO_REGISTER]; |
60 | |
61 | /* A global flag whose true value says to build live ranges for all |
62 | pseudos, otherwise the live ranges only for pseudos got memory is |
63 | build. True value means also building copies and setting up hard |
64 | register preferences. The complete info is necessary only for the |
65 | assignment pass. The complete info is not needed for the |
66 | coalescing and spill passes. */ |
67 | static bool complete_info_p; |
68 | |
69 | /* Pseudos live at current point in the RTL scan. */ |
70 | static sparseset pseudos_live; |
71 | |
72 | /* Pseudos probably living through calls and setjumps. As setjump is |
73 | a call too, if a bit in PSEUDOS_LIVE_THROUGH_SETJUMPS is set up |
74 | then the corresponding bit in PSEUDOS_LIVE_THROUGH_CALLS is set up |
75 | too. These data are necessary for cases when only one subreg of a |
76 | multi-reg pseudo is set up after a call. So we decide it is |
77 | probably live when traversing bb backward. We are sure about |
78 | living when we see its usage or definition of the pseudo. */ |
79 | static sparseset pseudos_live_through_calls; |
80 | static sparseset pseudos_live_through_setjumps; |
81 | |
82 | /* Set of hard regs (except eliminable ones) currently live. */ |
83 | static HARD_REG_SET hard_regs_live; |
84 | |
85 | /* Set of pseudos and hard registers start living/dying in the current |
86 | insn. These sets are used to update REG_DEAD and REG_UNUSED notes |
87 | in the insn. */ |
88 | static sparseset start_living, start_dying; |
89 | |
90 | /* Set of pseudos and hard regs dead and unused in the current |
91 | insn. */ |
92 | static sparseset unused_set, dead_set; |
93 | |
94 | /* Bitmap used for holding intermediate bitmap operation results. */ |
95 | static bitmap_head temp_bitmap; |
96 | |
97 | /* Pool for pseudo live ranges. */ |
98 | static object_allocator<lra_live_range> lra_live_range_pool ("live ranges" ); |
99 | |
100 | /* Free live range list LR. */ |
101 | static void |
102 | free_live_range_list (lra_live_range_t lr) |
103 | { |
104 | lra_live_range_t next; |
105 | |
106 | while (lr != NULL) |
107 | { |
108 | next = lr->next; |
109 | lra_live_range_pool.remove (object: lr); |
110 | lr = next; |
111 | } |
112 | } |
113 | |
114 | /* Create and return pseudo live range with given attributes. */ |
115 | static lra_live_range_t |
116 | create_live_range (int regno, int start, int finish, lra_live_range_t next) |
117 | { |
118 | lra_live_range_t p = lra_live_range_pool.allocate (); |
119 | p->regno = regno; |
120 | p->start = start; |
121 | p->finish = finish; |
122 | p->next = next; |
123 | return p; |
124 | } |
125 | |
126 | /* Copy live range R and return the result. */ |
127 | static lra_live_range_t |
128 | copy_live_range (lra_live_range_t r) |
129 | { |
130 | return new (lra_live_range_pool) lra_live_range (*r); |
131 | } |
132 | |
133 | /* Copy live range list given by its head R and return the result. */ |
134 | lra_live_range_t |
135 | lra_copy_live_range_list (lra_live_range_t r) |
136 | { |
137 | lra_live_range_t p, first, *chain; |
138 | |
139 | first = NULL; |
140 | for (chain = &first; r != NULL; r = r->next) |
141 | { |
142 | p = copy_live_range (r); |
143 | *chain = p; |
144 | chain = &p->next; |
145 | } |
146 | return first; |
147 | } |
148 | |
149 | /* Merge *non-intersected* ranges R1 and R2 and returns the result. |
150 | The function maintains the order of ranges and tries to minimize |
151 | size of the result range list. Ranges R1 and R2 may not be used |
152 | after the call. */ |
153 | lra_live_range_t |
154 | lra_merge_live_ranges (lra_live_range_t r1, lra_live_range_t r2) |
155 | { |
156 | lra_live_range_t first, last; |
157 | |
158 | if (r1 == NULL) |
159 | return r2; |
160 | if (r2 == NULL) |
161 | return r1; |
162 | for (first = last = NULL; r1 != NULL && r2 != NULL;) |
163 | { |
164 | if (r1->start < r2->start) |
165 | std::swap (a&: r1, b&: r2); |
166 | |
167 | if (r1->start == r2->finish + 1) |
168 | { |
169 | /* Joint ranges: merge r1 and r2 into r1. */ |
170 | r1->start = r2->start; |
171 | lra_live_range_t temp = r2; |
172 | r2 = r2->next; |
173 | lra_live_range_pool.remove (object: temp); |
174 | } |
175 | else |
176 | { |
177 | gcc_assert (r2->finish + 1 < r1->start); |
178 | /* Add r1 to the result. */ |
179 | if (first == NULL) |
180 | first = last = r1; |
181 | else |
182 | { |
183 | last->next = r1; |
184 | last = r1; |
185 | } |
186 | r1 = r1->next; |
187 | } |
188 | } |
189 | if (r1 != NULL) |
190 | { |
191 | if (first == NULL) |
192 | first = r1; |
193 | else |
194 | last->next = r1; |
195 | } |
196 | else |
197 | { |
198 | lra_assert (r2 != NULL); |
199 | if (first == NULL) |
200 | first = r2; |
201 | else |
202 | last->next = r2; |
203 | } |
204 | return first; |
205 | } |
206 | |
207 | /* Return TRUE if live ranges R1 and R2 intersect. */ |
208 | bool |
209 | lra_intersected_live_ranges_p (lra_live_range_t r1, lra_live_range_t r2) |
210 | { |
211 | /* Remember the live ranges are always kept ordered. */ |
212 | while (r1 != NULL && r2 != NULL) |
213 | { |
214 | if (r1->start > r2->finish) |
215 | r1 = r1->next; |
216 | else if (r2->start > r1->finish) |
217 | r2 = r2->next; |
218 | else |
219 | return true; |
220 | } |
221 | return false; |
222 | } |
223 | |
224 | enum point_type { |
225 | DEF_POINT, |
226 | USE_POINT |
227 | }; |
228 | |
229 | /* Return TRUE if set A contains a pseudo register, otherwise, return FALSE. */ |
230 | static bool |
231 | sparseset_contains_pseudos_p (sparseset a) |
232 | { |
233 | int regno; |
234 | EXECUTE_IF_SET_IN_SPARSESET (a, regno) |
235 | if (!HARD_REGISTER_NUM_P (regno)) |
236 | return true; |
237 | return false; |
238 | } |
239 | |
240 | /* Mark pseudo REGNO as living or dying at program point POINT, depending on |
241 | whether TYPE is a definition or a use. If this is the first reference to |
242 | REGNO that we've encountered, then create a new live range for it. */ |
243 | |
244 | static void |
245 | update_pseudo_point (int regno, int point, enum point_type type) |
246 | { |
247 | lra_live_range_t p; |
248 | |
249 | /* Don't compute points for hard registers. */ |
250 | if (HARD_REGISTER_NUM_P (regno)) |
251 | return; |
252 | |
253 | if (complete_info_p || lra_get_regno_hard_regno (regno) < 0) |
254 | { |
255 | if (type == DEF_POINT) |
256 | { |
257 | if (sparseset_bit_p (s: pseudos_live, e: regno)) |
258 | { |
259 | p = lra_reg_info[regno].live_ranges; |
260 | lra_assert (p != NULL); |
261 | p->finish = point; |
262 | } |
263 | } |
264 | else /* USE_POINT */ |
265 | { |
266 | if (!sparseset_bit_p (s: pseudos_live, e: regno) |
267 | && ((p = lra_reg_info[regno].live_ranges) == NULL |
268 | || (p->finish != point && p->finish + 1 != point))) |
269 | lra_reg_info[regno].live_ranges |
270 | = create_live_range (regno, start: point, finish: -1, next: p); |
271 | } |
272 | } |
273 | } |
274 | |
275 | /* The corresponding bitmaps of BB currently being processed. */ |
276 | static bitmap bb_killed_pseudos, bb_gen_pseudos; |
277 | |
278 | /* Record hard register REGNO as now being live. It updates |
279 | living hard regs and START_LIVING. */ |
280 | static void |
281 | make_hard_regno_live (int regno) |
282 | { |
283 | lra_assert (HARD_REGISTER_NUM_P (regno)); |
284 | if (TEST_HARD_REG_BIT (set: hard_regs_live, bit: regno) |
285 | || TEST_HARD_REG_BIT (set: eliminable_regset, bit: regno)) |
286 | return; |
287 | SET_HARD_REG_BIT (set&: hard_regs_live, bit: regno); |
288 | sparseset_set_bit (s: start_living, e: regno); |
289 | if (fixed_regs[regno] || TEST_HARD_REG_BIT (set: hard_regs_spilled_into, bit: regno)) |
290 | bitmap_set_bit (bb_gen_pseudos, regno); |
291 | } |
292 | |
293 | /* Process the definition of hard register REGNO. This updates |
294 | hard_regs_live, START_DYING and conflict hard regs for living |
295 | pseudos. */ |
296 | static void |
297 | make_hard_regno_dead (int regno) |
298 | { |
299 | if (TEST_HARD_REG_BIT (set: eliminable_regset, bit: regno)) |
300 | return; |
301 | |
302 | lra_assert (HARD_REGISTER_NUM_P (regno)); |
303 | unsigned int i; |
304 | EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, i) |
305 | SET_HARD_REG_BIT (set&: lra_reg_info[i].conflict_hard_regs, bit: regno); |
306 | |
307 | if (! TEST_HARD_REG_BIT (set: hard_regs_live, bit: regno)) |
308 | return; |
309 | CLEAR_HARD_REG_BIT (set&: hard_regs_live, bit: regno); |
310 | sparseset_set_bit (s: start_dying, e: regno); |
311 | if (fixed_regs[regno] || TEST_HARD_REG_BIT (set: hard_regs_spilled_into, bit: regno)) |
312 | { |
313 | bitmap_clear_bit (bb_gen_pseudos, regno); |
314 | bitmap_set_bit (bb_killed_pseudos, regno); |
315 | } |
316 | } |
317 | |
318 | /* Mark pseudo REGNO as now being live and update START_LIVING. */ |
319 | static void |
320 | mark_pseudo_live (int regno) |
321 | { |
322 | lra_assert (!HARD_REGISTER_NUM_P (regno)); |
323 | if (sparseset_bit_p (s: pseudos_live, e: regno)) |
324 | return; |
325 | |
326 | sparseset_set_bit (s: pseudos_live, e: regno); |
327 | sparseset_set_bit (s: start_living, e: regno); |
328 | } |
329 | |
330 | /* Mark pseudo REGNO as now being dead and update START_DYING. */ |
331 | static void |
332 | mark_pseudo_dead (int regno) |
333 | { |
334 | lra_assert (!HARD_REGISTER_NUM_P (regno)); |
335 | lra_reg_info[regno].conflict_hard_regs |= hard_regs_live; |
336 | if (!sparseset_bit_p (s: pseudos_live, e: regno)) |
337 | return; |
338 | |
339 | sparseset_clear_bit (pseudos_live, regno); |
340 | sparseset_set_bit (s: start_dying, e: regno); |
341 | } |
342 | |
343 | /* Mark register REGNO (pseudo or hard register) in MODE as being live |
344 | and update BB_GEN_PSEUDOS. */ |
345 | static void |
346 | mark_regno_live (int regno, machine_mode mode) |
347 | { |
348 | int last; |
349 | |
350 | if (HARD_REGISTER_NUM_P (regno)) |
351 | { |
352 | for (last = end_hard_regno (mode, regno); regno < last; regno++) |
353 | make_hard_regno_live (regno); |
354 | } |
355 | else |
356 | { |
357 | mark_pseudo_live (regno); |
358 | bitmap_set_bit (bb_gen_pseudos, regno); |
359 | } |
360 | } |
361 | |
362 | |
363 | /* Mark register REGNO (pseudo or hard register) in MODE as being dead |
364 | and update BB_GEN_PSEUDOS and BB_KILLED_PSEUDOS. */ |
365 | static void |
366 | mark_regno_dead (int regno, machine_mode mode) |
367 | { |
368 | int last; |
369 | |
370 | if (HARD_REGISTER_NUM_P (regno)) |
371 | { |
372 | for (last = end_hard_regno (mode, regno); regno < last; regno++) |
373 | make_hard_regno_dead (regno); |
374 | } |
375 | else |
376 | { |
377 | mark_pseudo_dead (regno); |
378 | bitmap_clear_bit (bb_gen_pseudos, regno); |
379 | bitmap_set_bit (bb_killed_pseudos, regno); |
380 | } |
381 | } |
382 | |
383 | |
384 | |
385 | /* This page contains code for making global live analysis of pseudos. |
386 | The code works only when pseudo live info is changed on a BB |
387 | border. That might be a consequence of some global transformations |
388 | in LRA, e.g. PIC pseudo reuse or rematerialization. */ |
389 | |
390 | /* Structure describing local BB data used for pseudo |
391 | live-analysis. */ |
392 | class bb_data_pseudos |
393 | { |
394 | public: |
395 | /* Basic block about which the below data are. */ |
396 | basic_block bb; |
397 | bitmap_head killed_pseudos; /* pseudos killed in the BB. */ |
398 | bitmap_head gen_pseudos; /* pseudos generated in the BB. */ |
399 | }; |
400 | |
401 | /* Array for all BB data. Indexed by the corresponding BB index. */ |
402 | typedef class bb_data_pseudos *bb_data_t; |
403 | |
404 | /* All basic block data are referred through the following array. */ |
405 | static bb_data_t bb_data; |
406 | |
407 | /* Two small functions for access to the bb data. */ |
408 | static inline bb_data_t |
409 | get_bb_data (basic_block bb) |
410 | { |
411 | return &bb_data[(bb)->index]; |
412 | } |
413 | |
414 | static inline bb_data_t |
415 | get_bb_data_by_index (int index) |
416 | { |
417 | return &bb_data[index]; |
418 | } |
419 | |
420 | /* Bitmap with all hard regs. */ |
421 | static bitmap_head all_hard_regs_bitmap; |
422 | |
423 | /* The transfer function used by the DF equation solver to propagate |
424 | live info through block with BB_INDEX according to the following |
425 | equation: |
426 | |
427 | bb.livein = (bb.liveout - bb.kill) OR bb.gen |
428 | */ |
429 | static bool |
430 | live_trans_fun (int bb_index) |
431 | { |
432 | basic_block bb = get_bb_data_by_index (index: bb_index)->bb; |
433 | bitmap bb_liveout = df_get_live_out (bb); |
434 | bitmap bb_livein = df_get_live_in (bb); |
435 | bb_data_t bb_info = get_bb_data (bb); |
436 | |
437 | bitmap_and_compl (&temp_bitmap, bb_liveout, &all_hard_regs_bitmap); |
438 | return bitmap_ior_and_compl (DST: bb_livein, A: &bb_info->gen_pseudos, |
439 | B: &temp_bitmap, C: &bb_info->killed_pseudos); |
440 | } |
441 | |
442 | /* The confluence function used by the DF equation solver to set up |
443 | live info for a block BB without predecessor. */ |
444 | static void |
445 | live_con_fun_0 (basic_block bb) |
446 | { |
447 | bitmap_and_into (df_get_live_out (bb), &all_hard_regs_bitmap); |
448 | } |
449 | |
450 | /* The confluence function used by the DF equation solver to propagate |
451 | live info from successor to predecessor on edge E according to the |
452 | following equation: |
453 | |
454 | bb.liveout = 0 for entry block | OR (livein of successors) |
455 | */ |
456 | static bool |
457 | live_con_fun_n (edge e) |
458 | { |
459 | basic_block bb = e->src; |
460 | basic_block dest = e->dest; |
461 | bitmap bb_liveout = df_get_live_out (bb); |
462 | bitmap dest_livein = df_get_live_in (bb: dest); |
463 | |
464 | return bitmap_ior_and_compl_into (A: bb_liveout, |
465 | B: dest_livein, C: &all_hard_regs_bitmap); |
466 | } |
467 | |
468 | /* Indexes of all function blocks. */ |
469 | static bitmap_head all_blocks; |
470 | |
471 | /* Allocate and initialize data needed for global pseudo live |
472 | analysis. */ |
473 | static void |
474 | initiate_live_solver (void) |
475 | { |
476 | bitmap_initialize (head: &all_hard_regs_bitmap, obstack: ®_obstack); |
477 | bitmap_set_range (&all_hard_regs_bitmap, 0, FIRST_PSEUDO_REGISTER); |
478 | bb_data = XNEWVEC (class bb_data_pseudos, last_basic_block_for_fn (cfun)); |
479 | bitmap_initialize (head: &all_blocks, obstack: ®_obstack); |
480 | |
481 | basic_block bb; |
482 | FOR_ALL_BB_FN (bb, cfun) |
483 | { |
484 | bb_data_t bb_info = get_bb_data (bb); |
485 | bb_info->bb = bb; |
486 | bitmap_initialize (head: &bb_info->killed_pseudos, obstack: ®_obstack); |
487 | bitmap_initialize (head: &bb_info->gen_pseudos, obstack: ®_obstack); |
488 | bitmap_set_bit (&all_blocks, bb->index); |
489 | } |
490 | } |
491 | |
492 | /* Free all data needed for global pseudo live analysis. */ |
493 | static void |
494 | finish_live_solver (void) |
495 | { |
496 | basic_block bb; |
497 | |
498 | bitmap_clear (&all_blocks); |
499 | FOR_ALL_BB_FN (bb, cfun) |
500 | { |
501 | bb_data_t bb_info = get_bb_data (bb); |
502 | bitmap_clear (&bb_info->killed_pseudos); |
503 | bitmap_clear (&bb_info->gen_pseudos); |
504 | } |
505 | free (ptr: bb_data); |
506 | bitmap_clear (&all_hard_regs_bitmap); |
507 | } |
508 | |
509 | |
510 | |
511 | /* Insn currently scanned. */ |
512 | static rtx_insn *curr_insn; |
513 | /* The insn data. */ |
514 | static lra_insn_recog_data_t curr_id; |
515 | /* The insn static data. */ |
516 | static struct lra_static_insn_data *curr_static_id; |
517 | |
518 | /* Vec containing execution frequencies of program points. */ |
519 | static vec<int> point_freq_vec; |
520 | |
521 | /* The start of the above vector elements. */ |
522 | int *lra_point_freq; |
523 | |
524 | /* Increment the current program point POINT to the next point which has |
525 | execution frequency FREQ. */ |
526 | static void |
527 | next_program_point (int &point, int freq) |
528 | { |
529 | point_freq_vec.safe_push (obj: freq); |
530 | lra_point_freq = point_freq_vec.address (); |
531 | point++; |
532 | } |
533 | |
534 | /* Update the preference of HARD_REGNO for pseudo REGNO by PROFIT. */ |
535 | void |
536 | lra_setup_reload_pseudo_preferenced_hard_reg (int regno, |
537 | int hard_regno, int profit) |
538 | { |
539 | lra_assert (regno >= lra_constraint_new_regno_start); |
540 | if (lra_reg_info[regno].preferred_hard_regno1 == hard_regno) |
541 | lra_reg_info[regno].preferred_hard_regno_profit1 += profit; |
542 | else if (lra_reg_info[regno].preferred_hard_regno2 == hard_regno) |
543 | lra_reg_info[regno].preferred_hard_regno_profit2 += profit; |
544 | else if (lra_reg_info[regno].preferred_hard_regno1 < 0) |
545 | { |
546 | lra_reg_info[regno].preferred_hard_regno1 = hard_regno; |
547 | lra_reg_info[regno].preferred_hard_regno_profit1 = profit; |
548 | } |
549 | else if (lra_reg_info[regno].preferred_hard_regno2 < 0 |
550 | || profit > lra_reg_info[regno].preferred_hard_regno_profit2) |
551 | { |
552 | lra_reg_info[regno].preferred_hard_regno2 = hard_regno; |
553 | lra_reg_info[regno].preferred_hard_regno_profit2 = profit; |
554 | } |
555 | else |
556 | return; |
557 | /* Keep the 1st hard regno as more profitable. */ |
558 | if (lra_reg_info[regno].preferred_hard_regno1 >= 0 |
559 | && lra_reg_info[regno].preferred_hard_regno2 >= 0 |
560 | && (lra_reg_info[regno].preferred_hard_regno_profit2 |
561 | > lra_reg_info[regno].preferred_hard_regno_profit1)) |
562 | { |
563 | std::swap (a&: lra_reg_info[regno].preferred_hard_regno1, |
564 | b&: lra_reg_info[regno].preferred_hard_regno2); |
565 | std::swap (a&: lra_reg_info[regno].preferred_hard_regno_profit1, |
566 | b&: lra_reg_info[regno].preferred_hard_regno_profit2); |
567 | } |
568 | if (lra_dump_file != NULL) |
569 | { |
570 | if ((hard_regno = lra_reg_info[regno].preferred_hard_regno1) >= 0) |
571 | fprintf (stream: lra_dump_file, |
572 | format: " Hard reg %d is preferable by r%d with profit %d\n" , |
573 | hard_regno, regno, |
574 | lra_reg_info[regno].preferred_hard_regno_profit1); |
575 | if ((hard_regno = lra_reg_info[regno].preferred_hard_regno2) >= 0) |
576 | fprintf (stream: lra_dump_file, |
577 | format: " Hard reg %d is preferable by r%d with profit %d\n" , |
578 | hard_regno, regno, |
579 | lra_reg_info[regno].preferred_hard_regno_profit2); |
580 | } |
581 | } |
582 | |
583 | /* Check whether REGNO lives through calls and setjmps and clear |
584 | the corresponding bits in PSEUDOS_LIVE_THROUGH_CALLS and |
585 | PSEUDOS_LIVE_THROUGH_SETJUMPS. All calls in the region described |
586 | by PSEUDOS_LIVE_THROUGH_CALLS have the given ABI. */ |
587 | |
588 | static inline void |
589 | check_pseudos_live_through_calls (int regno, const function_abi &abi) |
590 | { |
591 | if (! sparseset_bit_p (s: pseudos_live_through_calls, e: regno)) |
592 | return; |
593 | |
594 | machine_mode mode = PSEUDO_REGNO_MODE (regno); |
595 | |
596 | sparseset_clear_bit (pseudos_live_through_calls, regno); |
597 | lra_reg_info[regno].conflict_hard_regs |= abi.mode_clobbers (mode); |
598 | if (! sparseset_bit_p (s: pseudos_live_through_setjumps, e: regno)) |
599 | return; |
600 | sparseset_clear_bit (pseudos_live_through_setjumps, regno); |
601 | /* Don't allocate pseudos that cross setjmps or any call, if this |
602 | function receives a nonlocal goto. */ |
603 | SET_HARD_REG_SET (lra_reg_info[regno].conflict_hard_regs); |
604 | } |
605 | |
606 | /* Return true if insn REG is an early clobber operand in alternative |
607 | NALT. Negative NALT means that we don't know the current insn |
608 | alternative. So assume the worst. */ |
609 | static inline bool |
610 | reg_early_clobber_p (const struct lra_insn_reg *reg, int n_alt) |
611 | { |
612 | return (n_alt == LRA_UNKNOWN_ALT |
613 | ? reg->early_clobber_alts != 0 |
614 | : (n_alt != LRA_NON_CLOBBERED_ALT |
615 | && TEST_BIT (reg->early_clobber_alts, n_alt))); |
616 | } |
617 | |
618 | /* Clear pseudo REGNO in SET or all hard registers of REGNO in MODE in SET. */ |
619 | static void |
620 | clear_sparseset_regnos (sparseset set, int regno, enum machine_mode mode) |
621 | { |
622 | if (regno >= FIRST_PSEUDO_REGISTER) |
623 | { |
624 | sparseset_clear_bit (dead_set, regno); |
625 | return; |
626 | } |
627 | for (int last = end_hard_regno (mode, regno); regno < last; regno++) |
628 | sparseset_clear_bit (set, regno); |
629 | } |
630 | |
631 | /* Return true if pseudo REGNO is in SET or all hard registers of REGNO in MODE |
632 | are in SET. */ |
633 | static bool |
634 | regnos_in_sparseset_p (sparseset set, int regno, enum machine_mode mode) |
635 | { |
636 | if (regno >= FIRST_PSEUDO_REGISTER) |
637 | return sparseset_bit_p (s: dead_set, e: regno); |
638 | for (int last = end_hard_regno (mode, regno); regno < last; regno++) |
639 | if (!sparseset_bit_p (s: set, e: regno)) |
640 | return false; |
641 | return true; |
642 | } |
643 | |
644 | /* Process insns of the basic block BB to update pseudo live ranges, |
645 | pseudo hard register conflicts, and insn notes. We do it on |
646 | backward scan of BB insns. CURR_POINT is the program point where |
647 | BB ends. The function updates this counter and returns in |
648 | CURR_POINT the program point where BB starts. The function also |
649 | does local live info updates and can delete the dead insns if |
650 | DEAD_INSN_P. It returns true if pseudo live info was |
651 | changed at the BB start. */ |
652 | static bool |
653 | process_bb_lives (basic_block bb, int &curr_point, bool dead_insn_p) |
654 | { |
655 | int i, regno, freq; |
656 | unsigned int j; |
657 | bitmap_iterator bi; |
658 | bitmap reg_live_out; |
659 | unsigned int px; |
660 | rtx_insn *next; |
661 | rtx link, *link_loc; |
662 | bool need_curr_point_incr; |
663 | /* Only has a meaningful value once we've seen a call. */ |
664 | function_abi last_call_abi = default_function_abi; |
665 | |
666 | reg_live_out = df_get_live_out (bb); |
667 | sparseset_clear (s: pseudos_live); |
668 | sparseset_clear (s: pseudos_live_through_calls); |
669 | sparseset_clear (s: pseudos_live_through_setjumps); |
670 | REG_SET_TO_HARD_REG_SET (hard_regs_live, reg_live_out); |
671 | hard_regs_live &= ~eliminable_regset; |
672 | EXECUTE_IF_SET_IN_BITMAP (reg_live_out, FIRST_PSEUDO_REGISTER, j, bi) |
673 | { |
674 | update_pseudo_point (regno: j, point: curr_point, type: USE_POINT); |
675 | mark_pseudo_live (regno: j); |
676 | } |
677 | |
678 | bb_gen_pseudos = &get_bb_data (bb)->gen_pseudos; |
679 | bb_killed_pseudos = &get_bb_data (bb)->killed_pseudos; |
680 | bitmap_clear (bb_gen_pseudos); |
681 | bitmap_clear (bb_killed_pseudos); |
682 | freq = REG_FREQ_FROM_BB (bb); |
683 | |
684 | if (lra_dump_file != NULL) |
685 | fprintf (stream: lra_dump_file, format: " BB %d\n" , bb->index); |
686 | |
687 | /* Scan the code of this basic block, noting which pseudos and hard |
688 | regs are born or die. |
689 | |
690 | Note that this loop treats uninitialized values as live until the |
691 | beginning of the block. For example, if an instruction uses |
692 | (reg:DI foo), and only (subreg:SI (reg:DI foo) 0) is ever set, |
693 | FOO will remain live until the beginning of the block. Likewise |
694 | if FOO is not set at all. This is unnecessarily pessimistic, but |
695 | it probably doesn't matter much in practice. */ |
696 | FOR_BB_INSNS_REVERSE_SAFE (bb, curr_insn, next) |
697 | { |
698 | bool call_p; |
699 | int n_alt, dst_regno, src_regno; |
700 | rtx set; |
701 | struct lra_insn_reg *reg; |
702 | |
703 | if (!NONDEBUG_INSN_P (curr_insn)) |
704 | continue; |
705 | |
706 | curr_id = lra_get_insn_recog_data (insn: curr_insn); |
707 | curr_static_id = curr_id->insn_static_data; |
708 | n_alt = curr_id->used_insn_alternative; |
709 | if (lra_dump_file != NULL) |
710 | fprintf (stream: lra_dump_file, format: " Insn %u: point = %d, n_alt = %d\n" , |
711 | INSN_UID (insn: curr_insn), curr_point, n_alt); |
712 | |
713 | set = single_set (insn: curr_insn); |
714 | |
715 | if (dead_insn_p && set != NULL_RTX |
716 | && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)) |
717 | && find_reg_note (curr_insn, REG_EH_REGION, NULL_RTX) == NULL_RTX |
718 | && ! may_trap_p (PATTERN (insn: curr_insn)) |
719 | /* Don't do premature remove of pic offset pseudo as we can |
720 | start to use it after some reload generation. */ |
721 | && (pic_offset_table_rtx == NULL_RTX |
722 | || pic_offset_table_rtx != SET_DEST (set))) |
723 | { |
724 | bool remove_p = true; |
725 | |
726 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
727 | if (reg->type != OP_IN |
728 | && (reg->regno < FIRST_PSEUDO_REGISTER |
729 | ? TEST_HARD_REG_BIT (set: hard_regs_live, bit: reg->regno) |
730 | : sparseset_bit_p (s: pseudos_live, e: reg->regno))) |
731 | { |
732 | remove_p = false; |
733 | break; |
734 | } |
735 | for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) |
736 | if (reg->type != OP_IN) |
737 | { |
738 | remove_p = false; |
739 | break; |
740 | } |
741 | |
742 | if (remove_p && ! volatile_refs_p (PATTERN (insn: curr_insn))) |
743 | { |
744 | dst_regno = REGNO (SET_DEST (set)); |
745 | if (lra_dump_file != NULL) |
746 | fprintf (stream: lra_dump_file, format: " Deleting dead insn %u\n" , |
747 | INSN_UID (insn: curr_insn)); |
748 | lra_set_insn_deleted (curr_insn); |
749 | if (lra_reg_info[dst_regno].nrefs == 0) |
750 | { |
751 | /* There might be some debug insns with the pseudo. */ |
752 | unsigned int uid; |
753 | rtx_insn *insn; |
754 | |
755 | bitmap_copy (&temp_bitmap, &lra_reg_info[dst_regno].insn_bitmap); |
756 | EXECUTE_IF_SET_IN_BITMAP (&temp_bitmap, 0, uid, bi) |
757 | { |
758 | insn = lra_insn_recog_data[uid]->insn; |
759 | lra_substitute_pseudo_within_insn (insn, dst_regno, |
760 | SET_SRC (set), true); |
761 | lra_update_insn_regno_info (insn); |
762 | } |
763 | } |
764 | continue; |
765 | } |
766 | } |
767 | |
768 | /* Update max ref width and hard reg usage. */ |
769 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
770 | { |
771 | int regno = reg->regno; |
772 | |
773 | lra_update_biggest_mode (regno, mode: reg->biggest_mode); |
774 | if (HARD_REGISTER_NUM_P (regno)) |
775 | lra_hard_reg_usage[regno] += freq; |
776 | } |
777 | |
778 | call_p = CALL_P (curr_insn); |
779 | |
780 | /* If we have a simple register copy and the source reg is live after |
781 | this instruction, then remove the source reg from the live set so |
782 | that it will not conflict with the destination reg. */ |
783 | rtx ignore_reg = non_conflicting_reg_copy_p (curr_insn); |
784 | if (ignore_reg != NULL_RTX) |
785 | { |
786 | int ignore_regno = REGNO (ignore_reg); |
787 | if (HARD_REGISTER_NUM_P (ignore_regno) |
788 | && TEST_HARD_REG_BIT (set: hard_regs_live, bit: ignore_regno)) |
789 | CLEAR_HARD_REG_BIT (set&: hard_regs_live, bit: ignore_regno); |
790 | else if (!HARD_REGISTER_NUM_P (ignore_regno) |
791 | && sparseset_bit_p (s: pseudos_live, e: ignore_regno)) |
792 | sparseset_clear_bit (pseudos_live, ignore_regno); |
793 | else |
794 | /* We don't need any special handling of the source reg if |
795 | it is dead after this instruction. */ |
796 | ignore_reg = NULL_RTX; |
797 | } |
798 | |
799 | src_regno = (set != NULL_RTX && REG_P (SET_SRC (set)) |
800 | ? REGNO (SET_SRC (set)) : -1); |
801 | dst_regno = (set != NULL_RTX && REG_P (SET_DEST (set)) |
802 | ? REGNO (SET_DEST (set)) : -1); |
803 | if (complete_info_p |
804 | && src_regno >= 0 && dst_regno >= 0 |
805 | /* Check that source regno does not conflict with |
806 | destination regno to exclude most impossible |
807 | preferences. */ |
808 | && (((!HARD_REGISTER_NUM_P (src_regno) |
809 | && (! sparseset_bit_p (s: pseudos_live, e: src_regno) |
810 | || (!HARD_REGISTER_NUM_P (dst_regno) |
811 | && lra_reg_val_equal_p (regno: src_regno, |
812 | val: lra_reg_info[dst_regno].val, |
813 | offset: lra_reg_info[dst_regno].offset)))) |
814 | || (HARD_REGISTER_NUM_P (src_regno) |
815 | && ! TEST_HARD_REG_BIT (set: hard_regs_live, bit: src_regno))) |
816 | /* It might be 'inheritance pseudo <- reload pseudo'. */ |
817 | || (src_regno >= lra_constraint_new_regno_start |
818 | && dst_regno >= lra_constraint_new_regno_start |
819 | /* Remember to skip special cases where src/dest regnos are |
820 | the same, e.g. insn SET pattern has matching constraints |
821 | like =r,0. */ |
822 | && src_regno != dst_regno))) |
823 | { |
824 | int hard_regno = -1, regno = -1; |
825 | |
826 | if (dst_regno >= lra_constraint_new_regno_start |
827 | && src_regno >= lra_constraint_new_regno_start) |
828 | { |
829 | /* It might be still an original (non-reload) insn with |
830 | one unused output and a constraint requiring to use |
831 | the same reg for input/output operands. In this case |
832 | dst_regno and src_regno have the same value, we don't |
833 | need a misleading copy for this case. */ |
834 | if (dst_regno != src_regno) |
835 | lra_create_copy (dst_regno, src_regno, freq); |
836 | } |
837 | else if (dst_regno >= lra_constraint_new_regno_start) |
838 | { |
839 | if (!HARD_REGISTER_NUM_P (hard_regno = src_regno)) |
840 | hard_regno = reg_renumber[src_regno]; |
841 | regno = dst_regno; |
842 | } |
843 | else if (src_regno >= lra_constraint_new_regno_start) |
844 | { |
845 | if (!HARD_REGISTER_NUM_P (hard_regno = dst_regno)) |
846 | hard_regno = reg_renumber[dst_regno]; |
847 | regno = src_regno; |
848 | } |
849 | if (regno >= 0 && hard_regno >= 0) |
850 | lra_setup_reload_pseudo_preferenced_hard_reg |
851 | (regno, hard_regno, profit: freq); |
852 | } |
853 | |
854 | sparseset_clear (s: start_living); |
855 | |
856 | /* Mark each defined value as live. We need to do this for |
857 | unused values because they still conflict with quantities |
858 | that are live at the time of the definition. */ |
859 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
860 | if (reg->type != OP_IN) |
861 | { |
862 | update_pseudo_point (regno: reg->regno, point: curr_point, type: USE_POINT); |
863 | mark_regno_live (regno: reg->regno, mode: reg->biggest_mode); |
864 | /* ??? Should be a no-op for unused registers. */ |
865 | check_pseudos_live_through_calls (regno: reg->regno, abi: last_call_abi); |
866 | } |
867 | |
868 | for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) |
869 | if (reg->type != OP_IN) |
870 | make_hard_regno_live (regno: reg->regno); |
871 | |
872 | if (curr_id->arg_hard_regs != NULL) |
873 | for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) |
874 | if (!HARD_REGISTER_NUM_P (regno)) |
875 | /* It is a clobber. */ |
876 | make_hard_regno_live (regno: regno - FIRST_PSEUDO_REGISTER); |
877 | |
878 | sparseset_copy (unused_set, start_living); |
879 | |
880 | sparseset_clear (s: start_dying); |
881 | |
882 | /* See which defined values die here. */ |
883 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
884 | if (reg->type != OP_IN |
885 | && ! reg_early_clobber_p (reg, n_alt) && ! reg->subreg_p) |
886 | { |
887 | if (reg->type == OP_OUT) |
888 | update_pseudo_point (regno: reg->regno, point: curr_point, type: DEF_POINT); |
889 | mark_regno_dead (regno: reg->regno, mode: reg->biggest_mode); |
890 | } |
891 | |
892 | for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) |
893 | if (reg->type != OP_IN |
894 | && ! reg_early_clobber_p (reg, n_alt) && ! reg->subreg_p) |
895 | make_hard_regno_dead (regno: reg->regno); |
896 | |
897 | if (curr_id->arg_hard_regs != NULL) |
898 | for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) |
899 | if (!HARD_REGISTER_NUM_P (regno)) |
900 | /* It is a clobber. */ |
901 | make_hard_regno_dead (regno: regno - FIRST_PSEUDO_REGISTER); |
902 | |
903 | if (call_p) |
904 | { |
905 | function_abi call_abi = insn_callee_abi (curr_insn); |
906 | |
907 | if (last_call_abi != call_abi) |
908 | EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, j) |
909 | check_pseudos_live_through_calls (regno: j, abi: last_call_abi); |
910 | |
911 | last_call_abi = call_abi; |
912 | |
913 | sparseset_ior (pseudos_live_through_calls, |
914 | pseudos_live_through_calls, pseudos_live); |
915 | if (cfun->has_nonlocal_label |
916 | || (!targetm.setjmp_preserves_nonvolatile_regs_p () |
917 | && (find_reg_note (curr_insn, REG_SETJMP, NULL_RTX) |
918 | != NULL_RTX))) |
919 | sparseset_ior (pseudos_live_through_setjumps, |
920 | pseudos_live_through_setjumps, pseudos_live); |
921 | } |
922 | |
923 | /* Increment the current program point if we must. */ |
924 | if (sparseset_contains_pseudos_p (a: unused_set) |
925 | || sparseset_contains_pseudos_p (a: start_dying)) |
926 | next_program_point (point&: curr_point, freq); |
927 | |
928 | /* If we removed the source reg from a simple register copy from the |
929 | live set above, then add it back now so we don't accidentally add |
930 | it to the start_living set below. */ |
931 | if (ignore_reg != NULL_RTX) |
932 | { |
933 | int ignore_regno = REGNO (ignore_reg); |
934 | if (HARD_REGISTER_NUM_P (ignore_regno)) |
935 | SET_HARD_REG_BIT (set&: hard_regs_live, bit: ignore_regno); |
936 | else |
937 | sparseset_set_bit (s: pseudos_live, e: ignore_regno); |
938 | } |
939 | |
940 | sparseset_clear (s: start_living); |
941 | |
942 | /* Mark each used value as live. */ |
943 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
944 | if (reg->type != OP_OUT) |
945 | { |
946 | if (reg->type == OP_IN) |
947 | update_pseudo_point (regno: reg->regno, point: curr_point, type: USE_POINT); |
948 | mark_regno_live (regno: reg->regno, mode: reg->biggest_mode); |
949 | check_pseudos_live_through_calls (regno: reg->regno, abi: last_call_abi); |
950 | } |
951 | |
952 | for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) |
953 | if (reg->type != OP_OUT) |
954 | make_hard_regno_live (regno: reg->regno); |
955 | |
956 | if (curr_id->arg_hard_regs != NULL) |
957 | /* Make argument hard registers live. */ |
958 | for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) |
959 | if (HARD_REGISTER_NUM_P (regno)) |
960 | make_hard_regno_live (regno); |
961 | |
962 | sparseset_and_compl (dead_set, start_living, start_dying); |
963 | |
964 | sparseset_clear (s: start_dying); |
965 | |
966 | /* Mark early clobber outputs dead. */ |
967 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) |
968 | if (reg->type != OP_IN |
969 | && reg_early_clobber_p (reg, n_alt) && ! reg->subreg_p) |
970 | { |
971 | if (reg->type == OP_OUT) |
972 | update_pseudo_point (regno: reg->regno, point: curr_point, type: DEF_POINT); |
973 | mark_regno_dead (regno: reg->regno, mode: reg->biggest_mode); |
974 | |
975 | /* We're done processing inputs, so make sure early clobber |
976 | operands that are both inputs and outputs are still live. */ |
977 | if (reg->type == OP_INOUT) |
978 | mark_regno_live (regno: reg->regno, mode: reg->biggest_mode); |
979 | } |
980 | |
981 | for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) |
982 | if (reg->type != OP_IN |
983 | && reg_early_clobber_p (reg, n_alt) && ! reg->subreg_p) |
984 | { |
985 | struct lra_insn_reg *reg2; |
986 | |
987 | /* We can have early clobbered non-operand hard reg and |
988 | the same hard reg as an insn input. Don't make hard |
989 | reg dead before the insns. */ |
990 | for (reg2 = curr_static_id->hard_regs; reg2 != NULL; reg2 = reg2->next) |
991 | if (reg2->type != OP_OUT && reg2->regno == reg->regno) |
992 | break; |
993 | if (reg2 == NULL) |
994 | make_hard_regno_dead (regno: reg->regno); |
995 | } |
996 | |
997 | /* Increment the current program point if we must. */ |
998 | if (sparseset_contains_pseudos_p (a: dead_set) |
999 | || sparseset_contains_pseudos_p (a: start_dying)) |
1000 | next_program_point (point&: curr_point, freq); |
1001 | |
1002 | /* Update notes. */ |
1003 | for (link_loc = ®_NOTES (curr_insn); (link = *link_loc) != NULL_RTX;) |
1004 | { |
1005 | if (REG_NOTE_KIND (link) != REG_DEAD |
1006 | && REG_NOTE_KIND (link) != REG_UNUSED) |
1007 | ; |
1008 | else if (REG_P (XEXP (link, 0))) |
1009 | { |
1010 | rtx note_reg = XEXP (link, 0); |
1011 | int note_regno = REGNO (note_reg); |
1012 | |
1013 | if ((REG_NOTE_KIND (link) == REG_DEAD |
1014 | && ! regnos_in_sparseset_p (set: dead_set, regno: note_regno, |
1015 | GET_MODE (note_reg))) |
1016 | || (REG_NOTE_KIND (link) == REG_UNUSED |
1017 | && ! regnos_in_sparseset_p (set: unused_set, regno: note_regno, |
1018 | GET_MODE (note_reg)))) |
1019 | { |
1020 | *link_loc = XEXP (link, 1); |
1021 | continue; |
1022 | } |
1023 | if (REG_NOTE_KIND (link) == REG_DEAD) |
1024 | clear_sparseset_regnos (set: dead_set, regno: note_regno, |
1025 | GET_MODE (note_reg)); |
1026 | else if (REG_NOTE_KIND (link) == REG_UNUSED) |
1027 | clear_sparseset_regnos (set: unused_set, regno: note_regno, |
1028 | GET_MODE (note_reg)); |
1029 | } |
1030 | link_loc = &XEXP (link, 1); |
1031 | } |
1032 | EXECUTE_IF_SET_IN_SPARSESET (dead_set, j) |
1033 | add_reg_note (curr_insn, REG_DEAD, regno_reg_rtx[j]); |
1034 | EXECUTE_IF_SET_IN_SPARSESET (unused_set, j) |
1035 | add_reg_note (curr_insn, REG_UNUSED, regno_reg_rtx[j]); |
1036 | } |
1037 | |
1038 | if (bb_has_eh_pred (bb)) |
1039 | /* Any pseudos that are currently live conflict with the eh_return |
1040 | data registers. For liveness purposes, these registers are set |
1041 | by artificial definitions at the start of the BB, so are not |
1042 | actually live on entry. */ |
1043 | for (j = 0; ; ++j) |
1044 | { |
1045 | unsigned int regno = EH_RETURN_DATA_REGNO (j); |
1046 | |
1047 | if (regno == INVALID_REGNUM) |
1048 | break; |
1049 | |
1050 | make_hard_regno_live (regno); |
1051 | make_hard_regno_dead (regno); |
1052 | } |
1053 | |
1054 | /* Pseudos can't go in stack regs at the start of a basic block that |
1055 | is reached by an abnormal edge. Likewise for registers that are at |
1056 | least partly call clobbered, because caller-save, fixup_abnormal_edges |
1057 | and possibly the table driven EH machinery are not quite ready to |
1058 | handle such pseudos live across such edges. */ |
1059 | if (bb_has_abnormal_pred (bb)) |
1060 | { |
1061 | HARD_REG_SET clobbers; |
1062 | |
1063 | CLEAR_HARD_REG_SET (set&: clobbers); |
1064 | #ifdef STACK_REGS |
1065 | EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, px) |
1066 | lra_reg_info[px].no_stack_p = true; |
1067 | for (px = FIRST_STACK_REG; px <= LAST_STACK_REG; px++) |
1068 | SET_HARD_REG_BIT (set&: clobbers, bit: px); |
1069 | #endif |
1070 | /* No need to record conflicts for call clobbered regs if we |
1071 | have nonlocal labels around, as we don't ever try to |
1072 | allocate such regs in this case. */ |
1073 | if (!cfun->has_nonlocal_label |
1074 | && has_abnormal_call_or_eh_pred_edge_p (bb)) |
1075 | for (px = 0; HARD_REGISTER_NUM_P (px); px++) |
1076 | if (eh_edge_abi.clobbers_at_least_part_of_reg_p (regno: px) |
1077 | #ifdef REAL_PIC_OFFSET_TABLE_REGNUM |
1078 | /* We should create a conflict of PIC pseudo with PIC |
1079 | hard reg as PIC hard reg can have a wrong value after |
1080 | jump described by the abnormal edge. In this case we |
1081 | cannot allocate PIC hard reg to PIC pseudo as PIC |
1082 | pseudo will also have a wrong value. */ |
1083 | || (px == REAL_PIC_OFFSET_TABLE_REGNUM |
1084 | && pic_offset_table_rtx != NULL_RTX |
1085 | && !HARD_REGISTER_P (pic_offset_table_rtx)) |
1086 | #endif |
1087 | ) |
1088 | SET_HARD_REG_BIT (set&: clobbers, bit: px); |
1089 | |
1090 | clobbers &= ~hard_regs_live; |
1091 | for (px = 0; HARD_REGISTER_NUM_P (px); px++) |
1092 | if (TEST_HARD_REG_BIT (set: clobbers, bit: px)) |
1093 | { |
1094 | make_hard_regno_live (regno: px); |
1095 | make_hard_regno_dead (regno: px); |
1096 | } |
1097 | } |
1098 | |
1099 | bool live_change_p = false; |
1100 | /* Check if bb border live info was changed. */ |
1101 | unsigned int live_pseudos_num = 0; |
1102 | EXECUTE_IF_SET_IN_BITMAP (df_get_live_in (bb), |
1103 | FIRST_PSEUDO_REGISTER, j, bi) |
1104 | { |
1105 | live_pseudos_num++; |
1106 | if (! sparseset_bit_p (s: pseudos_live, e: j)) |
1107 | { |
1108 | live_change_p = true; |
1109 | if (lra_dump_file != NULL) |
1110 | fprintf (stream: lra_dump_file, |
1111 | format: " r%d is removed as live at bb%d start\n" , j, bb->index); |
1112 | break; |
1113 | } |
1114 | } |
1115 | if (! live_change_p |
1116 | && sparseset_cardinality (s: pseudos_live) != live_pseudos_num) |
1117 | { |
1118 | live_change_p = true; |
1119 | if (lra_dump_file != NULL) |
1120 | EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, j) |
1121 | if (! bitmap_bit_p (df_get_live_in (bb), j)) |
1122 | fprintf (stream: lra_dump_file, |
1123 | format: " r%d is added to live at bb%d start\n" , j, bb->index); |
1124 | } |
1125 | /* See if we'll need an increment at the end of this basic block. |
1126 | An increment is needed if the PSEUDOS_LIVE set is not empty, |
1127 | to make sure the finish points are set up correctly. */ |
1128 | need_curr_point_incr = (sparseset_cardinality (s: pseudos_live) > 0); |
1129 | |
1130 | EXECUTE_IF_SET_IN_SPARSESET (pseudos_live, i) |
1131 | { |
1132 | update_pseudo_point (regno: i, point: curr_point, type: DEF_POINT); |
1133 | mark_pseudo_dead (regno: i); |
1134 | } |
1135 | |
1136 | EXECUTE_IF_SET_IN_BITMAP (df_get_live_in (bb), FIRST_PSEUDO_REGISTER, j, bi) |
1137 | { |
1138 | if (sparseset_cardinality (s: pseudos_live_through_calls) == 0) |
1139 | break; |
1140 | if (sparseset_bit_p (s: pseudos_live_through_calls, e: j)) |
1141 | check_pseudos_live_through_calls (regno: j, abi: last_call_abi); |
1142 | } |
1143 | |
1144 | for (i = 0; HARD_REGISTER_NUM_P (i); ++i) |
1145 | { |
1146 | if (!TEST_HARD_REG_BIT (set: hard_regs_live, bit: i)) |
1147 | continue; |
1148 | |
1149 | if (!TEST_HARD_REG_BIT (set: hard_regs_spilled_into, bit: i)) |
1150 | continue; |
1151 | |
1152 | if (bitmap_bit_p (df_get_live_in (bb), i)) |
1153 | continue; |
1154 | |
1155 | live_change_p = true; |
1156 | if (lra_dump_file) |
1157 | fprintf (stream: lra_dump_file, |
1158 | format: " hard reg r%d is added to live at bb%d start\n" , i, |
1159 | bb->index); |
1160 | bitmap_set_bit (df_get_live_in (bb), i); |
1161 | } |
1162 | |
1163 | if (need_curr_point_incr) |
1164 | next_program_point (point&: curr_point, freq); |
1165 | |
1166 | return live_change_p; |
1167 | } |
1168 | |
1169 | /* Compress pseudo live ranges by removing program points where |
1170 | nothing happens. Complexity of many algorithms in LRA is linear |
1171 | function of program points number. To speed up the code we try to |
1172 | minimize the number of the program points here. */ |
1173 | static void |
1174 | remove_some_program_points_and_update_live_ranges (void) |
1175 | { |
1176 | unsigned i; |
1177 | int n, max_regno; |
1178 | int *map; |
1179 | lra_live_range_t r, prev_r, next_r; |
1180 | sbitmap_iterator sbi; |
1181 | bool born_p, dead_p, prev_born_p, prev_dead_p; |
1182 | |
1183 | auto_sbitmap born (lra_live_max_point); |
1184 | auto_sbitmap dead (lra_live_max_point); |
1185 | bitmap_clear (born); |
1186 | bitmap_clear (dead); |
1187 | max_regno = max_reg_num (); |
1188 | for (i = FIRST_PSEUDO_REGISTER; i < (unsigned) max_regno; i++) |
1189 | { |
1190 | for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next) |
1191 | { |
1192 | lra_assert (r->start <= r->finish); |
1193 | bitmap_set_bit (map: born, bitno: r->start); |
1194 | bitmap_set_bit (map: dead, bitno: r->finish); |
1195 | } |
1196 | } |
1197 | auto_sbitmap born_or_dead (lra_live_max_point); |
1198 | bitmap_ior (born_or_dead, born, dead); |
1199 | map = XCNEWVEC (int, lra_live_max_point); |
1200 | n = -1; |
1201 | prev_born_p = prev_dead_p = false; |
1202 | EXECUTE_IF_SET_IN_BITMAP (born_or_dead, 0, i, sbi) |
1203 | { |
1204 | born_p = bitmap_bit_p (map: born, bitno: i); |
1205 | dead_p = bitmap_bit_p (map: dead, bitno: i); |
1206 | if ((prev_born_p && ! prev_dead_p && born_p && ! dead_p) |
1207 | || (prev_dead_p && ! prev_born_p && dead_p && ! born_p)) |
1208 | { |
1209 | map[i] = n; |
1210 | lra_point_freq[n] = MAX (lra_point_freq[n], lra_point_freq[i]); |
1211 | } |
1212 | else |
1213 | { |
1214 | map[i] = ++n; |
1215 | lra_point_freq[n] = lra_point_freq[i]; |
1216 | } |
1217 | prev_born_p = born_p; |
1218 | prev_dead_p = dead_p; |
1219 | } |
1220 | n++; |
1221 | if (lra_dump_file != NULL) |
1222 | fprintf (stream: lra_dump_file, format: "Compressing live ranges: from %d to %d - %d%%\n" , |
1223 | lra_live_max_point, n, |
1224 | lra_live_max_point ? 100 * n / lra_live_max_point : 100); |
1225 | if (n < lra_live_max_point) |
1226 | { |
1227 | lra_live_max_point = n; |
1228 | for (i = FIRST_PSEUDO_REGISTER; i < (unsigned) max_regno; i++) |
1229 | { |
1230 | for (prev_r = NULL, r = lra_reg_info[i].live_ranges; |
1231 | r != NULL; |
1232 | r = next_r) |
1233 | { |
1234 | next_r = r->next; |
1235 | r->start = map[r->start]; |
1236 | r->finish = map[r->finish]; |
1237 | if (prev_r == NULL || prev_r->start > r->finish + 1) |
1238 | { |
1239 | prev_r = r; |
1240 | continue; |
1241 | } |
1242 | prev_r->start = r->start; |
1243 | prev_r->next = next_r; |
1244 | lra_live_range_pool.remove (object: r); |
1245 | } |
1246 | } |
1247 | } |
1248 | free (ptr: map); |
1249 | } |
1250 | |
1251 | /* Print live ranges R to file F. */ |
1252 | void |
1253 | lra_print_live_range_list (FILE *f, lra_live_range_t r) |
1254 | { |
1255 | for (; r != NULL; r = r->next) |
1256 | fprintf (stream: f, format: " [%d..%d]" , r->start, r->finish); |
1257 | fprintf (stream: f, format: "\n" ); |
1258 | } |
1259 | |
1260 | DEBUG_FUNCTION void |
1261 | debug (lra_live_range &ref) |
1262 | { |
1263 | lra_print_live_range_list (stderr, r: &ref); |
1264 | } |
1265 | |
1266 | DEBUG_FUNCTION void |
1267 | debug (lra_live_range *ptr) |
1268 | { |
1269 | if (ptr) |
1270 | debug (ref&: *ptr); |
1271 | else |
1272 | fprintf (stderr, format: "<nil>\n" ); |
1273 | } |
1274 | |
1275 | /* Print live ranges R to stderr. */ |
1276 | void |
1277 | lra_debug_live_range_list (lra_live_range_t r) |
1278 | { |
1279 | lra_print_live_range_list (stderr, r); |
1280 | } |
1281 | |
1282 | /* Print live ranges of pseudo REGNO to file F. */ |
1283 | static void |
1284 | print_pseudo_live_ranges (FILE *f, int regno) |
1285 | { |
1286 | if (lra_reg_info[regno].live_ranges == NULL) |
1287 | return; |
1288 | fprintf (stream: f, format: " r%d:" , regno); |
1289 | lra_print_live_range_list (f, r: lra_reg_info[regno].live_ranges); |
1290 | } |
1291 | |
1292 | /* Print live ranges of pseudo REGNO to stderr. */ |
1293 | void |
1294 | lra_debug_pseudo_live_ranges (int regno) |
1295 | { |
1296 | print_pseudo_live_ranges (stderr, regno); |
1297 | } |
1298 | |
1299 | /* Print live ranges of all pseudos to file F. */ |
1300 | static void |
1301 | print_live_ranges (FILE *f) |
1302 | { |
1303 | int i, max_regno; |
1304 | |
1305 | max_regno = max_reg_num (); |
1306 | for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++) |
1307 | print_pseudo_live_ranges (f, regno: i); |
1308 | } |
1309 | |
1310 | /* Print live ranges of all pseudos to stderr. */ |
1311 | void |
1312 | lra_debug_live_ranges (void) |
1313 | { |
1314 | print_live_ranges (stderr); |
1315 | } |
1316 | |
1317 | /* Compress pseudo live ranges. */ |
1318 | static void |
1319 | compress_live_ranges (void) |
1320 | { |
1321 | remove_some_program_points_and_update_live_ranges (); |
1322 | if (lra_dump_file != NULL) |
1323 | { |
1324 | fprintf (stream: lra_dump_file, format: "Ranges after the compression:\n" ); |
1325 | print_live_ranges (f: lra_dump_file); |
1326 | } |
1327 | } |
1328 | |
1329 | |
1330 | |
1331 | /* The number of the current live range pass. */ |
1332 | int lra_live_range_iter; |
1333 | |
1334 | /* The function creates live ranges only for memory pseudos (or for |
1335 | all ones if ALL_P), set up CONFLICT_HARD_REGS for the pseudos. It |
1336 | also does dead insn elimination if DEAD_INSN_P and global live |
1337 | analysis only for pseudos and only if the pseudo live info was |
1338 | changed on a BB border. Return TRUE if the live info was |
1339 | changed. */ |
1340 | static bool |
1341 | lra_create_live_ranges_1 (bool all_p, bool dead_insn_p) |
1342 | { |
1343 | basic_block bb; |
1344 | int i, hard_regno, max_regno = max_reg_num (); |
1345 | int curr_point; |
1346 | bool bb_live_change_p, have_referenced_pseudos = false; |
1347 | |
1348 | timevar_push (tv: TV_LRA_CREATE_LIVE_RANGES); |
1349 | |
1350 | complete_info_p = all_p; |
1351 | if (lra_dump_file != NULL) |
1352 | fprintf (stream: lra_dump_file, |
1353 | format: "\n********** Pseudo live ranges #%d: **********\n\n" , |
1354 | ++lra_live_range_iter); |
1355 | memset (s: lra_hard_reg_usage, c: 0, n: sizeof (lra_hard_reg_usage)); |
1356 | for (i = 0; i < max_regno; i++) |
1357 | { |
1358 | lra_reg_info[i].live_ranges = NULL; |
1359 | CLEAR_HARD_REG_SET (set&: lra_reg_info[i].conflict_hard_regs); |
1360 | lra_reg_info[i].preferred_hard_regno1 = -1; |
1361 | lra_reg_info[i].preferred_hard_regno2 = -1; |
1362 | lra_reg_info[i].preferred_hard_regno_profit1 = 0; |
1363 | lra_reg_info[i].preferred_hard_regno_profit2 = 0; |
1364 | #ifdef STACK_REGS |
1365 | lra_reg_info[i].no_stack_p = false; |
1366 | #endif |
1367 | /* The biggest mode is already set but its value might be to |
1368 | conservative because of recent transformation. Here in this |
1369 | file we recalculate it again as it costs practically |
1370 | nothing. */ |
1371 | if (!HARD_REGISTER_NUM_P (i) && regno_reg_rtx[i] != NULL_RTX) |
1372 | lra_reg_info[i].biggest_mode = GET_MODE (regno_reg_rtx[i]); |
1373 | else |
1374 | lra_reg_info[i].biggest_mode = VOIDmode; |
1375 | if (!HARD_REGISTER_NUM_P (i) |
1376 | && lra_reg_info[i].nrefs != 0) |
1377 | { |
1378 | if ((hard_regno = reg_renumber[i]) >= 0) |
1379 | lra_hard_reg_usage[hard_regno] += lra_reg_info[i].freq; |
1380 | have_referenced_pseudos = true; |
1381 | } |
1382 | } |
1383 | lra_free_copies (); |
1384 | |
1385 | /* Under some circumstances, we can have functions without pseudo |
1386 | registers. For such functions, lra_live_max_point will be 0, |
1387 | see e.g. PR55604, and there's nothing more to do for us here. */ |
1388 | if (! have_referenced_pseudos) |
1389 | { |
1390 | timevar_pop (tv: TV_LRA_CREATE_LIVE_RANGES); |
1391 | return false; |
1392 | } |
1393 | |
1394 | pseudos_live = sparseset_alloc (n_elms: max_regno); |
1395 | pseudos_live_through_calls = sparseset_alloc (n_elms: max_regno); |
1396 | pseudos_live_through_setjumps = sparseset_alloc (n_elms: max_regno); |
1397 | start_living = sparseset_alloc (n_elms: max_regno); |
1398 | start_dying = sparseset_alloc (n_elms: max_regno); |
1399 | dead_set = sparseset_alloc (n_elms: max_regno); |
1400 | unused_set = sparseset_alloc (n_elms: max_regno); |
1401 | curr_point = 0; |
1402 | unsigned new_length = get_max_uid () * 2; |
1403 | point_freq_vec.truncate (size: 0); |
1404 | point_freq_vec.reserve_exact (nelems: new_length); |
1405 | lra_point_freq = point_freq_vec.address (); |
1406 | int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
1407 | int n = inverted_rev_post_order_compute (cfun, rpo); |
1408 | lra_assert (n == n_basic_blocks_for_fn (cfun)); |
1409 | bb_live_change_p = false; |
1410 | for (i = 0; i < n; ++i) |
1411 | { |
1412 | bb = BASIC_BLOCK_FOR_FN (cfun, rpo[i]); |
1413 | if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun) || bb |
1414 | == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
1415 | continue; |
1416 | if (process_bb_lives (bb, curr_point, dead_insn_p)) |
1417 | bb_live_change_p = true; |
1418 | } |
1419 | free (ptr: rpo); |
1420 | if (bb_live_change_p) |
1421 | { |
1422 | /* We need to clear pseudo live info as some pseudos can |
1423 | disappear, e.g. pseudos with used equivalences. */ |
1424 | FOR_EACH_BB_FN (bb, cfun) |
1425 | { |
1426 | bitmap_clear_range (df_get_live_in (bb), FIRST_PSEUDO_REGISTER, |
1427 | max_regno - FIRST_PSEUDO_REGISTER); |
1428 | bitmap_clear_range (df_get_live_out (bb), FIRST_PSEUDO_REGISTER, |
1429 | max_regno - FIRST_PSEUDO_REGISTER); |
1430 | } |
1431 | /* As we did not change CFG since LRA start we can use |
1432 | DF-infrastructure solver to solve live data flow problem. */ |
1433 | for (int i = 0; HARD_REGISTER_NUM_P (i); ++i) |
1434 | { |
1435 | if (TEST_HARD_REG_BIT (set: hard_regs_spilled_into, bit: i)) |
1436 | bitmap_clear_bit (&all_hard_regs_bitmap, i); |
1437 | } |
1438 | df_simple_dataflow |
1439 | (DF_BACKWARD, NULL, live_con_fun_0, live_con_fun_n, |
1440 | live_trans_fun, &all_blocks, |
1441 | df_get_postorder (DF_BACKWARD), df_get_n_blocks (DF_BACKWARD)); |
1442 | if (lra_dump_file != NULL) |
1443 | { |
1444 | fprintf (stream: lra_dump_file, |
1445 | format: "Global pseudo live data have been updated:\n" ); |
1446 | basic_block bb; |
1447 | FOR_EACH_BB_FN (bb, cfun) |
1448 | { |
1449 | bb_data_t bb_info = get_bb_data (bb); |
1450 | bitmap bb_livein = df_get_live_in (bb); |
1451 | bitmap bb_liveout = df_get_live_out (bb); |
1452 | |
1453 | fprintf (stream: lra_dump_file, format: "\nBB %d:\n" , bb->index); |
1454 | lra_dump_bitmap_with_title (" gen:" , |
1455 | &bb_info->gen_pseudos, bb->index); |
1456 | lra_dump_bitmap_with_title (" killed:" , |
1457 | &bb_info->killed_pseudos, bb->index); |
1458 | lra_dump_bitmap_with_title (" livein:" , bb_livein, bb->index); |
1459 | lra_dump_bitmap_with_title (" liveout:" , bb_liveout, bb->index); |
1460 | } |
1461 | } |
1462 | } |
1463 | lra_live_max_point = curr_point; |
1464 | if (lra_dump_file != NULL) |
1465 | print_live_ranges (f: lra_dump_file); |
1466 | /* Clean up. */ |
1467 | sparseset_free (unused_set); |
1468 | sparseset_free (dead_set); |
1469 | sparseset_free (start_dying); |
1470 | sparseset_free (start_living); |
1471 | sparseset_free (pseudos_live_through_calls); |
1472 | sparseset_free (pseudos_live_through_setjumps); |
1473 | sparseset_free (pseudos_live); |
1474 | compress_live_ranges (); |
1475 | timevar_pop (tv: TV_LRA_CREATE_LIVE_RANGES); |
1476 | return bb_live_change_p; |
1477 | } |
1478 | |
1479 | /* The main entry function creates live-ranges and other live info |
1480 | necessary for the assignment sub-pass. It uses |
1481 | lra_creates_live_ranges_1 -- so read comments for the |
1482 | function. */ |
1483 | void |
1484 | lra_create_live_ranges (bool all_p, bool dead_insn_p) |
1485 | { |
1486 | if (! lra_create_live_ranges_1 (all_p, dead_insn_p)) |
1487 | return; |
1488 | if (lra_dump_file != NULL) |
1489 | fprintf (stream: lra_dump_file, format: "Live info was changed -- recalculate it\n" ); |
1490 | /* Live info was changed on a bb border. It means that some info, |
1491 | e.g. about conflict regs, calls crossed, and live ranges may be |
1492 | wrong. We need this info for allocation. So recalculate it |
1493 | again but without removing dead insns which can change live info |
1494 | again. Repetitive live range calculations are expensive therefore |
1495 | we stop here as we already have correct info although some |
1496 | improvement in rare cases could be possible on this sub-pass if |
1497 | we do dead insn elimination again (still the improvement may |
1498 | happen later). */ |
1499 | lra_clear_live_ranges (); |
1500 | bool res = lra_create_live_ranges_1 (all_p, dead_insn_p: false); |
1501 | lra_assert (! res); |
1502 | } |
1503 | |
1504 | /* Finish all live ranges. */ |
1505 | void |
1506 | lra_clear_live_ranges (void) |
1507 | { |
1508 | int i; |
1509 | |
1510 | for (i = 0; i < max_reg_num (); i++) |
1511 | free_live_range_list (lr: lra_reg_info[i].live_ranges); |
1512 | point_freq_vec.release (); |
1513 | } |
1514 | |
1515 | /* Initialize live ranges data once per function. */ |
1516 | void |
1517 | lra_live_ranges_init (void) |
1518 | { |
1519 | bitmap_initialize (head: &temp_bitmap, obstack: ®_obstack); |
1520 | initiate_live_solver (); |
1521 | } |
1522 | |
1523 | /* Finish live ranges data once per function. */ |
1524 | void |
1525 | lra_live_ranges_finish (void) |
1526 | { |
1527 | finish_live_solver (); |
1528 | bitmap_clear (&temp_bitmap); |
1529 | lra_live_range_pool.release (); |
1530 | } |
1531 | |