1 | /* Branch prediction routines for the GNU compiler. |
2 | Copyright (C) 2000-2024 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free |
8 | Software Foundation; either version 3, or (at your option) any later |
9 | version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | /* References: |
21 | |
22 | [1] "Branch Prediction for Free" |
23 | Ball and Larus; PLDI '93. |
24 | [2] "Static Branch Frequency and Program Profile Analysis" |
25 | Wu and Larus; MICRO-27. |
26 | [3] "Corpus-based Static Branch Prediction" |
27 | Calder, Grunwald, Lindsay, Martin, Mozer, and Zorn; PLDI '95. */ |
28 | |
29 | |
30 | #include "config.h" |
31 | #include "system.h" |
32 | #include "coretypes.h" |
33 | #include "backend.h" |
34 | #include "rtl.h" |
35 | #include "tree.h" |
36 | #include "gimple.h" |
37 | #include "cfghooks.h" |
38 | #include "tree-pass.h" |
39 | #include "ssa.h" |
40 | #include "memmodel.h" |
41 | #include "emit-rtl.h" |
42 | #include "cgraph.h" |
43 | #include "coverage.h" |
44 | #include "diagnostic-core.h" |
45 | #include "gimple-predict.h" |
46 | #include "fold-const.h" |
47 | #include "calls.h" |
48 | #include "cfganal.h" |
49 | #include "profile.h" |
50 | #include "sreal.h" |
51 | #include "cfgloop.h" |
52 | #include "gimple-iterator.h" |
53 | #include "tree-cfg.h" |
54 | #include "tree-ssa-loop-niter.h" |
55 | #include "tree-ssa-loop.h" |
56 | #include "tree-scalar-evolution.h" |
57 | #include "ipa-utils.h" |
58 | #include "gimple-pretty-print.h" |
59 | #include "selftest.h" |
60 | #include "cfgrtl.h" |
61 | #include "stringpool.h" |
62 | #include "attribs.h" |
63 | |
64 | /* Enum with reasons why a predictor is ignored. */ |
65 | |
66 | enum predictor_reason |
67 | { |
68 | REASON_NONE, |
69 | REASON_IGNORED, |
70 | REASON_SINGLE_EDGE_DUPLICATE, |
71 | REASON_EDGE_PAIR_DUPLICATE |
72 | }; |
73 | |
74 | /* String messages for the aforementioned enum. */ |
75 | |
76 | static const char *reason_messages[] = {"" , " (ignored)" , |
77 | " (single edge duplicate)" , " (edge pair duplicate)" }; |
78 | |
79 | |
80 | static void combine_predictions_for_insn (rtx_insn *, basic_block); |
81 | static void dump_prediction (FILE *, enum br_predictor, int, basic_block, |
82 | enum predictor_reason, edge); |
83 | static void predict_paths_leading_to (basic_block, enum br_predictor, |
84 | enum prediction, |
85 | class loop *in_loop = NULL); |
86 | static void predict_paths_leading_to_edge (edge, enum br_predictor, |
87 | enum prediction, |
88 | class loop *in_loop = NULL); |
89 | static bool can_predict_insn_p (const rtx_insn *); |
90 | static HOST_WIDE_INT get_predictor_value (br_predictor, HOST_WIDE_INT); |
91 | static void determine_unlikely_bbs (); |
92 | static void estimate_bb_frequencies (); |
93 | |
94 | /* Information we hold about each branch predictor. |
95 | Filled using information from predict.def. */ |
96 | |
97 | struct predictor_info |
98 | { |
99 | const char *const name; /* Name used in the debugging dumps. */ |
100 | const int hitrate; /* Expected hitrate used by |
101 | predict_insn_def call. */ |
102 | const int flags; |
103 | }; |
104 | |
105 | /* Use given predictor without Dempster-Shaffer theory if it matches |
106 | using first_match heuristics. */ |
107 | #define PRED_FLAG_FIRST_MATCH 1 |
108 | |
109 | /* Recompute hitrate in percent to our representation. */ |
110 | |
111 | #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100) |
112 | |
113 | #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS}, |
114 | static const struct predictor_info predictor_info[]= { |
115 | #include "predict.def" |
116 | |
117 | /* Upper bound on predictors. */ |
118 | {NULL, .hitrate: 0, .flags: 0} |
119 | }; |
120 | #undef DEF_PREDICTOR |
121 | |
122 | static gcov_type min_count = -1; |
123 | |
124 | /* Determine the threshold for hot BB counts. */ |
125 | |
126 | gcov_type |
127 | get_hot_bb_threshold () |
128 | { |
129 | if (min_count == -1) |
130 | { |
131 | const int hot_frac = param_hot_bb_count_fraction; |
132 | const gcov_type min_hot_count |
133 | = hot_frac |
134 | ? profile_info->sum_max / hot_frac |
135 | : (gcov_type)profile_count::max_count; |
136 | set_hot_bb_threshold (min_hot_count); |
137 | if (dump_file) |
138 | fprintf (stream: dump_file, format: "Setting hotness threshold to %" PRId64 ".\n" , |
139 | min_hot_count); |
140 | } |
141 | return min_count; |
142 | } |
143 | |
144 | /* Set the threshold for hot BB counts. */ |
145 | |
146 | void |
147 | set_hot_bb_threshold (gcov_type min) |
148 | { |
149 | min_count = min; |
150 | } |
151 | |
152 | /* Return TRUE if COUNT is considered to be hot in function FUN. */ |
153 | |
154 | bool |
155 | maybe_hot_count_p (struct function *fun, profile_count count) |
156 | { |
157 | if (!count.initialized_p ()) |
158 | return true; |
159 | if (count.ipa () == profile_count::zero ()) |
160 | return false; |
161 | if (!count.ipa_p ()) |
162 | { |
163 | struct cgraph_node *node = cgraph_node::get (decl: fun->decl); |
164 | if (!profile_info || profile_status_for_fn (fun) != PROFILE_READ) |
165 | { |
166 | if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED) |
167 | return false; |
168 | if (node->frequency == NODE_FREQUENCY_HOT) |
169 | return true; |
170 | } |
171 | if (profile_status_for_fn (fun) == PROFILE_ABSENT) |
172 | return true; |
173 | if (node->frequency == NODE_FREQUENCY_EXECUTED_ONCE |
174 | && count < (ENTRY_BLOCK_PTR_FOR_FN (fun)->count.apply_scale (num: 2, den: 3))) |
175 | return false; |
176 | if (count * param_hot_bb_frequency_fraction |
177 | < ENTRY_BLOCK_PTR_FOR_FN (fun)->count) |
178 | return false; |
179 | return true; |
180 | } |
181 | /* Code executed at most once is not hot. */ |
182 | if (count <= MAX (profile_info ? profile_info->runs : 1, 1)) |
183 | return false; |
184 | return (count >= get_hot_bb_threshold ()); |
185 | } |
186 | |
187 | /* Return true if basic block BB of function FUN can be CPU intensive |
188 | and should thus be optimized for maximum performance. */ |
189 | |
190 | bool |
191 | maybe_hot_bb_p (struct function *fun, const_basic_block bb) |
192 | { |
193 | gcc_checking_assert (fun); |
194 | return maybe_hot_count_p (fun, count: bb->count); |
195 | } |
196 | |
197 | /* Return true if edge E can be CPU intensive and should thus be optimized |
198 | for maximum performance. */ |
199 | |
200 | bool |
201 | maybe_hot_edge_p (edge e) |
202 | { |
203 | return maybe_hot_count_p (cfun, count: e->count ()); |
204 | } |
205 | |
206 | /* Return true if COUNT is considered to be never executed in function FUN |
207 | or if function FUN is considered so in the static profile. */ |
208 | |
209 | static bool |
210 | probably_never_executed (struct function *fun, profile_count count) |
211 | { |
212 | gcc_checking_assert (fun); |
213 | if (count.ipa () == profile_count::zero ()) |
214 | return true; |
215 | /* Do not trust adjusted counts. This will make us to drop int cold section |
216 | code with low execution count as a result of inlining. These low counts |
217 | are not safe even with read profile and may lead us to dropping |
218 | code which actually gets executed into cold section of binary that is not |
219 | desirable. */ |
220 | if (count.precise_p () && profile_status_for_fn (fun) == PROFILE_READ) |
221 | { |
222 | const int unlikely_frac = param_unlikely_bb_count_fraction; |
223 | if (count * unlikely_frac >= profile_info->runs) |
224 | return false; |
225 | return true; |
226 | } |
227 | if ((!profile_info || profile_status_for_fn (fun) != PROFILE_READ) |
228 | && (cgraph_node::get (decl: fun->decl)->frequency |
229 | == NODE_FREQUENCY_UNLIKELY_EXECUTED)) |
230 | return true; |
231 | return false; |
232 | } |
233 | |
234 | /* Return true if basic block BB of function FUN is probably never executed. */ |
235 | |
236 | bool |
237 | probably_never_executed_bb_p (struct function *fun, const_basic_block bb) |
238 | { |
239 | return probably_never_executed (fun, count: bb->count); |
240 | } |
241 | |
242 | /* Return true if edge E is unlikely executed for obvious reasons. */ |
243 | |
244 | static bool |
245 | unlikely_executed_edge_p (edge e) |
246 | { |
247 | return (e->src->count == profile_count::zero () |
248 | || e->probability == profile_probability::never ()) |
249 | || (e->flags & (EDGE_EH | EDGE_FAKE)); |
250 | } |
251 | |
252 | /* Return true if edge E of function FUN is probably never executed. */ |
253 | |
254 | bool |
255 | probably_never_executed_edge_p (struct function *fun, edge e) |
256 | { |
257 | if (unlikely_executed_edge_p (e)) |
258 | return true; |
259 | return probably_never_executed (fun, count: e->count ()); |
260 | } |
261 | |
262 | /* Return true if function FUN should always be optimized for size. */ |
263 | |
264 | optimize_size_level |
265 | optimize_function_for_size_p (struct function *fun) |
266 | { |
267 | if (!fun || !fun->decl) |
268 | return optimize_size ? OPTIMIZE_SIZE_MAX : OPTIMIZE_SIZE_NO; |
269 | cgraph_node *n = cgraph_node::get (decl: fun->decl); |
270 | if (n) |
271 | return n->optimize_for_size_p (); |
272 | return OPTIMIZE_SIZE_NO; |
273 | } |
274 | |
275 | /* Return true if function FUN should always be optimized for speed. */ |
276 | |
277 | bool |
278 | optimize_function_for_speed_p (struct function *fun) |
279 | { |
280 | return !optimize_function_for_size_p (fun); |
281 | } |
282 | |
283 | /* Return the optimization type that should be used for function FUN. */ |
284 | |
285 | optimization_type |
286 | function_optimization_type (struct function *fun) |
287 | { |
288 | return (optimize_function_for_speed_p (fun) |
289 | ? OPTIMIZE_FOR_SPEED |
290 | : OPTIMIZE_FOR_SIZE); |
291 | } |
292 | |
293 | /* Return TRUE if basic block BB should be optimized for size. */ |
294 | |
295 | optimize_size_level |
296 | optimize_bb_for_size_p (const_basic_block bb) |
297 | { |
298 | enum optimize_size_level ret = optimize_function_for_size_p (cfun); |
299 | |
300 | if (bb && ret < OPTIMIZE_SIZE_MAX && bb->count == profile_count::zero ()) |
301 | ret = OPTIMIZE_SIZE_MAX; |
302 | if (bb && ret < OPTIMIZE_SIZE_BALANCED && !maybe_hot_bb_p (cfun, bb)) |
303 | ret = OPTIMIZE_SIZE_BALANCED; |
304 | return ret; |
305 | } |
306 | |
307 | /* Return TRUE if basic block BB should be optimized for speed. */ |
308 | |
309 | bool |
310 | optimize_bb_for_speed_p (const_basic_block bb) |
311 | { |
312 | return !optimize_bb_for_size_p (bb); |
313 | } |
314 | |
315 | /* Return the optimization type that should be used for basic block BB. */ |
316 | |
317 | optimization_type |
318 | bb_optimization_type (const_basic_block bb) |
319 | { |
320 | return (optimize_bb_for_speed_p (bb) |
321 | ? OPTIMIZE_FOR_SPEED |
322 | : OPTIMIZE_FOR_SIZE); |
323 | } |
324 | |
325 | /* Return TRUE if edge E should be optimized for size. */ |
326 | |
327 | optimize_size_level |
328 | optimize_edge_for_size_p (edge e) |
329 | { |
330 | enum optimize_size_level ret = optimize_function_for_size_p (cfun); |
331 | |
332 | if (ret < OPTIMIZE_SIZE_MAX && unlikely_executed_edge_p (e)) |
333 | ret = OPTIMIZE_SIZE_MAX; |
334 | if (ret < OPTIMIZE_SIZE_BALANCED && !maybe_hot_edge_p (e)) |
335 | ret = OPTIMIZE_SIZE_BALANCED; |
336 | return ret; |
337 | } |
338 | |
339 | /* Return TRUE if edge E should be optimized for speed. */ |
340 | |
341 | bool |
342 | optimize_edge_for_speed_p (edge e) |
343 | { |
344 | return !optimize_edge_for_size_p (e); |
345 | } |
346 | |
347 | /* Return TRUE if the current function is optimized for size. */ |
348 | |
349 | optimize_size_level |
350 | optimize_insn_for_size_p (void) |
351 | { |
352 | enum optimize_size_level ret = optimize_function_for_size_p (cfun); |
353 | if (ret < OPTIMIZE_SIZE_BALANCED && !crtl->maybe_hot_insn_p) |
354 | ret = OPTIMIZE_SIZE_BALANCED; |
355 | return ret; |
356 | } |
357 | |
358 | /* Return TRUE if the current function is optimized for speed. */ |
359 | |
360 | bool |
361 | optimize_insn_for_speed_p (void) |
362 | { |
363 | return !optimize_insn_for_size_p (); |
364 | } |
365 | |
366 | /* Return the optimization type that should be used for the current |
367 | instruction. */ |
368 | |
369 | optimization_type |
370 | insn_optimization_type () |
371 | { |
372 | return (optimize_insn_for_speed_p () |
373 | ? OPTIMIZE_FOR_SPEED |
374 | : OPTIMIZE_FOR_SIZE); |
375 | } |
376 | |
377 | /* Return TRUE if LOOP should be optimized for size. */ |
378 | |
379 | optimize_size_level |
380 | optimize_loop_for_size_p (class loop *loop) |
381 | { |
382 | return optimize_bb_for_size_p (bb: loop->header); |
383 | } |
384 | |
385 | /* Return TRUE if LOOP should be optimized for speed. */ |
386 | |
387 | bool |
388 | optimize_loop_for_speed_p (class loop *loop) |
389 | { |
390 | return optimize_bb_for_speed_p (bb: loop->header); |
391 | } |
392 | |
393 | /* Return TRUE if nest rooted at LOOP should be optimized for speed. */ |
394 | |
395 | bool |
396 | optimize_loop_nest_for_speed_p (class loop *loop) |
397 | { |
398 | class loop *l = loop; |
399 | if (optimize_loop_for_speed_p (loop)) |
400 | return true; |
401 | l = loop->inner; |
402 | while (l && l != loop) |
403 | { |
404 | if (optimize_loop_for_speed_p (loop: l)) |
405 | return true; |
406 | if (l->inner) |
407 | l = l->inner; |
408 | else if (l->next) |
409 | l = l->next; |
410 | else |
411 | { |
412 | while (l != loop && !l->next) |
413 | l = loop_outer (loop: l); |
414 | if (l != loop) |
415 | l = l->next; |
416 | } |
417 | } |
418 | return false; |
419 | } |
420 | |
421 | /* Return TRUE if nest rooted at LOOP should be optimized for size. */ |
422 | |
423 | optimize_size_level |
424 | optimize_loop_nest_for_size_p (class loop *loop) |
425 | { |
426 | enum optimize_size_level ret = optimize_loop_for_size_p (loop); |
427 | class loop *l = loop; |
428 | |
429 | l = loop->inner; |
430 | while (l && l != loop) |
431 | { |
432 | if (ret == OPTIMIZE_SIZE_NO) |
433 | break; |
434 | ret = MIN (optimize_loop_for_size_p (l), ret); |
435 | if (l->inner) |
436 | l = l->inner; |
437 | else if (l->next) |
438 | l = l->next; |
439 | else |
440 | { |
441 | while (l != loop && !l->next) |
442 | l = loop_outer (loop: l); |
443 | if (l != loop) |
444 | l = l->next; |
445 | } |
446 | } |
447 | return ret; |
448 | } |
449 | |
450 | /* Return true if edge E is likely to be well predictable by branch |
451 | predictor. */ |
452 | |
453 | bool |
454 | predictable_edge_p (edge e) |
455 | { |
456 | if (!e->probability.initialized_p ()) |
457 | return false; |
458 | if ((e->probability.to_reg_br_prob_base () |
459 | <= param_predictable_branch_outcome * REG_BR_PROB_BASE / 100) |
460 | || (REG_BR_PROB_BASE - e->probability.to_reg_br_prob_base () |
461 | <= param_predictable_branch_outcome * REG_BR_PROB_BASE / 100)) |
462 | return true; |
463 | return false; |
464 | } |
465 | |
466 | |
467 | /* Set RTL expansion for BB profile. */ |
468 | |
469 | void |
470 | rtl_profile_for_bb (basic_block bb) |
471 | { |
472 | crtl->maybe_hot_insn_p = maybe_hot_bb_p (cfun, bb); |
473 | } |
474 | |
475 | /* Set RTL expansion for edge profile. */ |
476 | |
477 | void |
478 | rtl_profile_for_edge (edge e) |
479 | { |
480 | crtl->maybe_hot_insn_p = maybe_hot_edge_p (e); |
481 | } |
482 | |
483 | /* Set RTL expansion to default mode (i.e. when profile info is not known). */ |
484 | void |
485 | default_rtl_profile (void) |
486 | { |
487 | crtl->maybe_hot_insn_p = true; |
488 | } |
489 | |
490 | /* Return true if the one of outgoing edges is already predicted by |
491 | PREDICTOR. */ |
492 | |
493 | bool |
494 | rtl_predicted_by_p (const_basic_block bb, enum br_predictor predictor) |
495 | { |
496 | rtx note; |
497 | if (!INSN_P (BB_END (bb))) |
498 | return false; |
499 | for (note = REG_NOTES (BB_END (bb)); note; note = XEXP (note, 1)) |
500 | if (REG_NOTE_KIND (note) == REG_BR_PRED |
501 | && INTVAL (XEXP (XEXP (note, 0), 0)) == (int)predictor) |
502 | return true; |
503 | return false; |
504 | } |
505 | |
506 | /* Structure representing predictions in tree level. */ |
507 | |
508 | struct edge_prediction { |
509 | struct edge_prediction *ep_next; |
510 | edge ep_edge; |
511 | enum br_predictor ep_predictor; |
512 | int ep_probability; |
513 | }; |
514 | |
515 | /* This map contains for a basic block the list of predictions for the |
516 | outgoing edges. */ |
517 | |
518 | static hash_map<const_basic_block, edge_prediction *> *bb_predictions; |
519 | |
520 | /* Return true if the one of outgoing edges is already predicted by |
521 | PREDICTOR. */ |
522 | |
523 | bool |
524 | gimple_predicted_by_p (const_basic_block bb, enum br_predictor predictor) |
525 | { |
526 | struct edge_prediction *i; |
527 | edge_prediction **preds = bb_predictions->get (k: bb); |
528 | |
529 | if (!preds) |
530 | return false; |
531 | |
532 | for (i = *preds; i; i = i->ep_next) |
533 | if (i->ep_predictor == predictor) |
534 | return true; |
535 | return false; |
536 | } |
537 | |
538 | /* Return true if the one of outgoing edges is already predicted by |
539 | PREDICTOR for edge E predicted as TAKEN. */ |
540 | |
541 | bool |
542 | edge_predicted_by_p (edge e, enum br_predictor predictor, bool taken) |
543 | { |
544 | struct edge_prediction *i; |
545 | basic_block bb = e->src; |
546 | edge_prediction **preds = bb_predictions->get (k: bb); |
547 | if (!preds) |
548 | return false; |
549 | |
550 | int probability = predictor_info[(int) predictor].hitrate; |
551 | |
552 | if (taken != TAKEN) |
553 | probability = REG_BR_PROB_BASE - probability; |
554 | |
555 | for (i = *preds; i; i = i->ep_next) |
556 | if (i->ep_predictor == predictor |
557 | && i->ep_edge == e |
558 | && i->ep_probability == probability) |
559 | return true; |
560 | return false; |
561 | } |
562 | |
563 | /* Same predicate as above, working on edges. */ |
564 | bool |
565 | edge_probability_reliable_p (const_edge e) |
566 | { |
567 | return e->probability.probably_reliable_p (); |
568 | } |
569 | |
570 | /* Same predicate as edge_probability_reliable_p, working on notes. */ |
571 | bool |
572 | br_prob_note_reliable_p (const_rtx note) |
573 | { |
574 | gcc_assert (REG_NOTE_KIND (note) == REG_BR_PROB); |
575 | return profile_probability::from_reg_br_prob_note |
576 | (XINT (note, 0)).probably_reliable_p (); |
577 | } |
578 | |
579 | static void |
580 | predict_insn (rtx_insn *insn, enum br_predictor predictor, int probability) |
581 | { |
582 | gcc_assert (any_condjump_p (insn)); |
583 | if (!flag_guess_branch_prob) |
584 | return; |
585 | |
586 | add_reg_note (insn, REG_BR_PRED, |
587 | gen_rtx_CONCAT (VOIDmode, |
588 | GEN_INT ((int) predictor), |
589 | GEN_INT ((int) probability))); |
590 | } |
591 | |
592 | /* Predict insn by given predictor. */ |
593 | |
594 | void |
595 | predict_insn_def (rtx_insn *insn, enum br_predictor predictor, |
596 | enum prediction taken) |
597 | { |
598 | int probability = predictor_info[(int) predictor].hitrate; |
599 | gcc_assert (probability != PROB_UNINITIALIZED); |
600 | |
601 | if (taken != TAKEN) |
602 | probability = REG_BR_PROB_BASE - probability; |
603 | |
604 | predict_insn (insn, predictor, probability); |
605 | } |
606 | |
607 | /* Predict edge E with given probability if possible. */ |
608 | |
609 | void |
610 | rtl_predict_edge (edge e, enum br_predictor predictor, int probability) |
611 | { |
612 | rtx_insn *last_insn; |
613 | last_insn = BB_END (e->src); |
614 | |
615 | /* We can store the branch prediction information only about |
616 | conditional jumps. */ |
617 | if (!any_condjump_p (last_insn)) |
618 | return; |
619 | |
620 | /* We always store probability of branching. */ |
621 | if (e->flags & EDGE_FALLTHRU) |
622 | probability = REG_BR_PROB_BASE - probability; |
623 | |
624 | predict_insn (insn: last_insn, predictor, probability); |
625 | } |
626 | |
627 | /* Predict edge E with the given PROBABILITY. */ |
628 | void |
629 | gimple_predict_edge (edge e, enum br_predictor predictor, int probability) |
630 | { |
631 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
632 | && EDGE_COUNT (e->src->succs) > 1 |
633 | && flag_guess_branch_prob |
634 | && optimize) |
635 | { |
636 | struct edge_prediction *i = XNEW (struct edge_prediction); |
637 | edge_prediction *&preds = bb_predictions->get_or_insert (k: e->src); |
638 | |
639 | i->ep_next = preds; |
640 | preds = i; |
641 | i->ep_probability = probability; |
642 | i->ep_predictor = predictor; |
643 | i->ep_edge = e; |
644 | } |
645 | } |
646 | |
647 | /* Filter edge predictions PREDS by a function FILTER: if FILTER return false |
648 | the prediction is removed. |
649 | DATA are passed to the filter function. */ |
650 | |
651 | static void |
652 | filter_predictions (edge_prediction **preds, |
653 | bool (*filter) (edge_prediction *, void *), void *data) |
654 | { |
655 | if (!bb_predictions) |
656 | return; |
657 | |
658 | if (preds) |
659 | { |
660 | struct edge_prediction **prediction = preds; |
661 | struct edge_prediction *next; |
662 | |
663 | while (*prediction) |
664 | { |
665 | if ((*filter) (*prediction, data)) |
666 | prediction = &((*prediction)->ep_next); |
667 | else |
668 | { |
669 | next = (*prediction)->ep_next; |
670 | free (ptr: *prediction); |
671 | *prediction = next; |
672 | } |
673 | } |
674 | } |
675 | } |
676 | |
677 | /* Filter function predicate that returns true for a edge predicate P |
678 | if its edge is equal to DATA. */ |
679 | |
680 | static bool |
681 | not_equal_edge_p (edge_prediction *p, void *data) |
682 | { |
683 | return p->ep_edge != (edge)data; |
684 | } |
685 | |
686 | /* Remove all predictions on given basic block that are attached |
687 | to edge E. */ |
688 | void |
689 | remove_predictions_associated_with_edge (edge e) |
690 | { |
691 | if (!bb_predictions) |
692 | return; |
693 | |
694 | edge_prediction **preds = bb_predictions->get (k: e->src); |
695 | filter_predictions (preds, filter: not_equal_edge_p, data: e); |
696 | } |
697 | |
698 | /* Clears the list of predictions stored for BB. */ |
699 | |
700 | static void |
701 | clear_bb_predictions (basic_block bb) |
702 | { |
703 | edge_prediction **preds = bb_predictions->get (k: bb); |
704 | struct edge_prediction *pred, *next; |
705 | |
706 | if (!preds) |
707 | return; |
708 | |
709 | for (pred = *preds; pred; pred = next) |
710 | { |
711 | next = pred->ep_next; |
712 | free (ptr: pred); |
713 | } |
714 | *preds = NULL; |
715 | } |
716 | |
717 | /* Return true when we can store prediction on insn INSN. |
718 | At the moment we represent predictions only on conditional |
719 | jumps, not at computed jump or other complicated cases. */ |
720 | static bool |
721 | can_predict_insn_p (const rtx_insn *insn) |
722 | { |
723 | return (JUMP_P (insn) |
724 | && any_condjump_p (insn) |
725 | && EDGE_COUNT (BLOCK_FOR_INSN (insn)->succs) >= 2); |
726 | } |
727 | |
728 | /* Predict edge E by given predictor if possible. */ |
729 | |
730 | void |
731 | predict_edge_def (edge e, enum br_predictor predictor, |
732 | enum prediction taken) |
733 | { |
734 | int probability = predictor_info[(int) predictor].hitrate; |
735 | |
736 | if (taken != TAKEN) |
737 | probability = REG_BR_PROB_BASE - probability; |
738 | |
739 | predict_edge (e, predictor, probability); |
740 | } |
741 | |
742 | /* Invert all branch predictions or probability notes in the INSN. This needs |
743 | to be done each time we invert the condition used by the jump. */ |
744 | |
745 | void |
746 | invert_br_probabilities (rtx insn) |
747 | { |
748 | rtx note; |
749 | |
750 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
751 | if (REG_NOTE_KIND (note) == REG_BR_PROB) |
752 | XINT (note, 0) = profile_probability::from_reg_br_prob_note |
753 | (XINT (note, 0)).invert ().to_reg_br_prob_note (); |
754 | else if (REG_NOTE_KIND (note) == REG_BR_PRED) |
755 | XEXP (XEXP (note, 0), 1) |
756 | = GEN_INT (REG_BR_PROB_BASE - INTVAL (XEXP (XEXP (note, 0), 1))); |
757 | } |
758 | |
759 | /* Dump information about the branch prediction to the output file. */ |
760 | |
761 | static void |
762 | dump_prediction (FILE *file, enum br_predictor predictor, int probability, |
763 | basic_block bb, enum predictor_reason reason = REASON_NONE, |
764 | edge ep_edge = NULL) |
765 | { |
766 | edge e = ep_edge; |
767 | edge_iterator ei; |
768 | |
769 | if (!file) |
770 | return; |
771 | |
772 | if (e == NULL) |
773 | FOR_EACH_EDGE (e, ei, bb->succs) |
774 | if (! (e->flags & EDGE_FALLTHRU)) |
775 | break; |
776 | |
777 | char edge_info_str[128]; |
778 | if (ep_edge) |
779 | sprintf (s: edge_info_str, format: " of edge %d->%d" , ep_edge->src->index, |
780 | ep_edge->dest->index); |
781 | else |
782 | edge_info_str[0] = '\0'; |
783 | |
784 | fprintf (stream: file, format: " %s heuristics%s%s: %.2f%%" , |
785 | predictor_info[predictor].name, |
786 | edge_info_str, reason_messages[reason], |
787 | probability * 100.0 / REG_BR_PROB_BASE); |
788 | |
789 | if (bb->count.initialized_p ()) |
790 | { |
791 | fprintf (stream: file, format: " exec " ); |
792 | bb->count.dump (f: file); |
793 | if (e && e->count ().initialized_p () && bb->count.to_gcov_type ()) |
794 | { |
795 | fprintf (stream: file, format: " hit " ); |
796 | e->count ().dump (f: file); |
797 | fprintf (stream: file, format: " (%.1f%%)" , e->count ().to_gcov_type() * 100.0 |
798 | / bb->count.to_gcov_type ()); |
799 | } |
800 | } |
801 | |
802 | fprintf (stream: file, format: "\n" ); |
803 | |
804 | /* Print output that be easily read by analyze_brprob.py script. We are |
805 | interested only in counts that are read from GCDA files. */ |
806 | if (dump_file && (dump_flags & TDF_DETAILS) |
807 | && bb->count.precise_p () |
808 | && reason == REASON_NONE) |
809 | { |
810 | fprintf (stream: file, format: ";;heuristics;%s;%" PRId64 ";%" PRId64 ";%.1f;\n" , |
811 | predictor_info[predictor].name, |
812 | bb->count.to_gcov_type (), e->count ().to_gcov_type (), |
813 | probability * 100.0 / REG_BR_PROB_BASE); |
814 | } |
815 | } |
816 | |
817 | /* Return true if STMT is known to be unlikely executed. */ |
818 | |
819 | static bool |
820 | unlikely_executed_stmt_p (gimple *stmt) |
821 | { |
822 | if (!is_gimple_call (gs: stmt)) |
823 | return false; |
824 | /* NORETURN attribute alone is not strong enough: exit() may be quite |
825 | likely executed once during program run. */ |
826 | if (gimple_call_fntype (gs: stmt) |
827 | && lookup_attribute (attr_name: "cold" , |
828 | TYPE_ATTRIBUTES (gimple_call_fntype (stmt))) |
829 | && !lookup_attribute (attr_name: "cold" , DECL_ATTRIBUTES (current_function_decl))) |
830 | return true; |
831 | tree decl = gimple_call_fndecl (gs: stmt); |
832 | if (!decl) |
833 | return false; |
834 | if (lookup_attribute (attr_name: "cold" , DECL_ATTRIBUTES (decl)) |
835 | && !lookup_attribute (attr_name: "cold" , DECL_ATTRIBUTES (current_function_decl))) |
836 | return true; |
837 | |
838 | cgraph_node *n = cgraph_node::get (decl); |
839 | if (!n) |
840 | return false; |
841 | |
842 | availability avail; |
843 | n = n->ultimate_alias_target (availability: &avail); |
844 | if (avail < AVAIL_AVAILABLE) |
845 | return false; |
846 | if (!n->analyzed |
847 | || n->decl == current_function_decl) |
848 | return false; |
849 | return n->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED; |
850 | } |
851 | |
852 | /* Return true if BB is unlikely executed. */ |
853 | |
854 | static bool |
855 | unlikely_executed_bb_p (basic_block bb) |
856 | { |
857 | if (bb->count == profile_count::zero ()) |
858 | return true; |
859 | if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) || bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
860 | return false; |
861 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); |
862 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
863 | { |
864 | if (unlikely_executed_stmt_p (stmt: gsi_stmt (i: gsi))) |
865 | return true; |
866 | if (stmt_can_terminate_bb_p (gsi_stmt (i: gsi))) |
867 | return false; |
868 | } |
869 | return false; |
870 | } |
871 | |
872 | /* We cannot predict the probabilities of outgoing edges of bb. Set them |
873 | evenly and hope for the best. If UNLIKELY_EDGES is not null, distribute |
874 | even probability for all edges not mentioned in the set. These edges |
875 | are given PROB_VERY_UNLIKELY probability. Similarly for LIKELY_EDGES, |
876 | if we have exactly one likely edge, make the other edges predicted |
877 | as not probable. */ |
878 | |
879 | static void |
880 | set_even_probabilities (basic_block bb, |
881 | hash_set<edge> *unlikely_edges = NULL, |
882 | hash_set<edge_prediction *> *likely_edges = NULL) |
883 | { |
884 | unsigned nedges = 0, unlikely_count = 0; |
885 | edge e = NULL; |
886 | edge_iterator ei; |
887 | profile_probability all = profile_probability::always (); |
888 | |
889 | FOR_EACH_EDGE (e, ei, bb->succs) |
890 | if (e->probability.initialized_p ()) |
891 | all -= e->probability; |
892 | else if (!unlikely_executed_edge_p (e)) |
893 | { |
894 | nedges++; |
895 | if (unlikely_edges != NULL && unlikely_edges->contains (k: e)) |
896 | { |
897 | all -= profile_probability::very_unlikely (); |
898 | unlikely_count++; |
899 | } |
900 | } |
901 | |
902 | /* Make the distribution even if all edges are unlikely. */ |
903 | unsigned likely_count = likely_edges ? likely_edges->elements () : 0; |
904 | if (unlikely_count == nedges) |
905 | { |
906 | unlikely_edges = NULL; |
907 | unlikely_count = 0; |
908 | } |
909 | |
910 | /* If we have one likely edge, then use its probability and distribute |
911 | remaining probabilities as even. */ |
912 | if (likely_count == 1) |
913 | { |
914 | FOR_EACH_EDGE (e, ei, bb->succs) |
915 | if (e->probability.initialized_p ()) |
916 | ; |
917 | else if (!unlikely_executed_edge_p (e)) |
918 | { |
919 | edge_prediction *prediction = *likely_edges->begin (); |
920 | int p = prediction->ep_probability; |
921 | profile_probability prob |
922 | = profile_probability::from_reg_br_prob_base (v: p); |
923 | |
924 | if (prediction->ep_edge == e) |
925 | e->probability = prob; |
926 | else if (unlikely_edges != NULL && unlikely_edges->contains (k: e)) |
927 | e->probability = profile_probability::very_unlikely (); |
928 | else |
929 | { |
930 | profile_probability remainder = prob.invert (); |
931 | remainder -= (profile_probability::very_unlikely () |
932 | * unlikely_count); |
933 | int count = nedges - unlikely_count - 1; |
934 | gcc_assert (count >= 0); |
935 | |
936 | e->probability = remainder / count; |
937 | } |
938 | } |
939 | else |
940 | e->probability = profile_probability::never (); |
941 | } |
942 | else |
943 | { |
944 | /* Make all unlikely edges unlikely and the rest will have even |
945 | probability. */ |
946 | unsigned scale = nedges - unlikely_count; |
947 | FOR_EACH_EDGE (e, ei, bb->succs) |
948 | if (e->probability.initialized_p ()) |
949 | ; |
950 | else if (!unlikely_executed_edge_p (e)) |
951 | { |
952 | if (unlikely_edges != NULL && unlikely_edges->contains (k: e)) |
953 | e->probability = profile_probability::very_unlikely (); |
954 | else |
955 | e->probability = all / scale; |
956 | } |
957 | else |
958 | e->probability = profile_probability::never (); |
959 | } |
960 | } |
961 | |
962 | /* Add REG_BR_PROB note to JUMP with PROB. */ |
963 | |
964 | void |
965 | add_reg_br_prob_note (rtx_insn *jump, profile_probability prob) |
966 | { |
967 | gcc_checking_assert (JUMP_P (jump) && !find_reg_note (jump, REG_BR_PROB, 0)); |
968 | add_int_reg_note (jump, REG_BR_PROB, prob.to_reg_br_prob_note ()); |
969 | } |
970 | |
971 | /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB |
972 | note if not already present. Remove now useless REG_BR_PRED notes. */ |
973 | |
974 | static void |
975 | combine_predictions_for_insn (rtx_insn *insn, basic_block bb) |
976 | { |
977 | rtx prob_note; |
978 | rtx *pnote; |
979 | rtx note; |
980 | int best_probability = PROB_EVEN; |
981 | enum br_predictor best_predictor = END_PREDICTORS; |
982 | int combined_probability = REG_BR_PROB_BASE / 2; |
983 | int d; |
984 | bool first_match = false; |
985 | bool found = false; |
986 | |
987 | if (!can_predict_insn_p (insn)) |
988 | { |
989 | set_even_probabilities (bb); |
990 | return; |
991 | } |
992 | |
993 | prob_note = find_reg_note (insn, REG_BR_PROB, 0); |
994 | pnote = ®_NOTES (insn); |
995 | if (dump_file) |
996 | fprintf (stream: dump_file, format: "Predictions for insn %i bb %i\n" , INSN_UID (insn), |
997 | bb->index); |
998 | |
999 | /* We implement "first match" heuristics and use probability guessed |
1000 | by predictor with smallest index. */ |
1001 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) |
1002 | if (REG_NOTE_KIND (note) == REG_BR_PRED) |
1003 | { |
1004 | enum br_predictor predictor = ((enum br_predictor) |
1005 | INTVAL (XEXP (XEXP (note, 0), 0))); |
1006 | int probability = INTVAL (XEXP (XEXP (note, 0), 1)); |
1007 | |
1008 | found = true; |
1009 | if (best_predictor > predictor |
1010 | && predictor_info[predictor].flags & PRED_FLAG_FIRST_MATCH) |
1011 | best_probability = probability, best_predictor = predictor; |
1012 | |
1013 | d = (combined_probability * probability |
1014 | + (REG_BR_PROB_BASE - combined_probability) |
1015 | * (REG_BR_PROB_BASE - probability)); |
1016 | |
1017 | /* Use FP math to avoid overflows of 32bit integers. */ |
1018 | if (d == 0) |
1019 | /* If one probability is 0% and one 100%, avoid division by zero. */ |
1020 | combined_probability = REG_BR_PROB_BASE / 2; |
1021 | else |
1022 | combined_probability = (((double) combined_probability) * probability |
1023 | * REG_BR_PROB_BASE / d + 0.5); |
1024 | } |
1025 | |
1026 | /* Decide which heuristic to use. In case we didn't match anything, |
1027 | use no_prediction heuristic, in case we did match, use either |
1028 | first match or Dempster-Shaffer theory depending on the flags. */ |
1029 | |
1030 | if (best_predictor != END_PREDICTORS) |
1031 | first_match = true; |
1032 | |
1033 | if (!found) |
1034 | dump_prediction (file: dump_file, predictor: PRED_NO_PREDICTION, |
1035 | probability: combined_probability, bb); |
1036 | else |
1037 | { |
1038 | if (!first_match) |
1039 | dump_prediction (file: dump_file, predictor: PRED_DS_THEORY, probability: combined_probability, |
1040 | bb, reason: !first_match ? REASON_NONE : REASON_IGNORED); |
1041 | else |
1042 | dump_prediction (file: dump_file, predictor: PRED_FIRST_MATCH, probability: best_probability, |
1043 | bb, reason: first_match ? REASON_NONE : REASON_IGNORED); |
1044 | } |
1045 | |
1046 | if (first_match) |
1047 | combined_probability = best_probability; |
1048 | dump_prediction (file: dump_file, predictor: PRED_COMBINED, probability: combined_probability, bb); |
1049 | |
1050 | while (*pnote) |
1051 | { |
1052 | if (REG_NOTE_KIND (*pnote) == REG_BR_PRED) |
1053 | { |
1054 | enum br_predictor predictor = ((enum br_predictor) |
1055 | INTVAL (XEXP (XEXP (*pnote, 0), 0))); |
1056 | int probability = INTVAL (XEXP (XEXP (*pnote, 0), 1)); |
1057 | |
1058 | dump_prediction (file: dump_file, predictor, probability, bb, |
1059 | reason: (!first_match || best_predictor == predictor) |
1060 | ? REASON_NONE : REASON_IGNORED); |
1061 | *pnote = XEXP (*pnote, 1); |
1062 | } |
1063 | else |
1064 | pnote = &XEXP (*pnote, 1); |
1065 | } |
1066 | |
1067 | if (!prob_note) |
1068 | { |
1069 | profile_probability p |
1070 | = profile_probability::from_reg_br_prob_base (v: combined_probability); |
1071 | add_reg_br_prob_note (jump: insn, prob: p); |
1072 | |
1073 | /* Save the prediction into CFG in case we are seeing non-degenerated |
1074 | conditional jump. */ |
1075 | if (!single_succ_p (bb)) |
1076 | { |
1077 | BRANCH_EDGE (bb)->probability = p; |
1078 | FALLTHRU_EDGE (bb)->probability |
1079 | = BRANCH_EDGE (bb)->probability.invert (); |
1080 | } |
1081 | } |
1082 | else if (!single_succ_p (bb)) |
1083 | { |
1084 | profile_probability prob = profile_probability::from_reg_br_prob_note |
1085 | (XINT (prob_note, 0)); |
1086 | |
1087 | BRANCH_EDGE (bb)->probability = prob; |
1088 | FALLTHRU_EDGE (bb)->probability = prob.invert (); |
1089 | } |
1090 | else |
1091 | single_succ_edge (bb)->probability = profile_probability::always (); |
1092 | } |
1093 | |
1094 | /* Edge prediction hash traits. */ |
1095 | |
1096 | struct predictor_hash: pointer_hash <edge_prediction> |
1097 | { |
1098 | |
1099 | static inline hashval_t hash (const edge_prediction *); |
1100 | static inline bool equal (const edge_prediction *, const edge_prediction *); |
1101 | }; |
1102 | |
1103 | /* Calculate hash value of an edge prediction P based on predictor and |
1104 | normalized probability. */ |
1105 | |
1106 | inline hashval_t |
1107 | predictor_hash::hash (const edge_prediction *p) |
1108 | { |
1109 | inchash::hash hstate; |
1110 | hstate.add_int (v: p->ep_predictor); |
1111 | |
1112 | int prob = p->ep_probability; |
1113 | if (prob > REG_BR_PROB_BASE / 2) |
1114 | prob = REG_BR_PROB_BASE - prob; |
1115 | |
1116 | hstate.add_int (v: prob); |
1117 | |
1118 | return hstate.end (); |
1119 | } |
1120 | |
1121 | /* Return true whether edge predictions P1 and P2 use the same predictor and |
1122 | have equal (or opposed probability). */ |
1123 | |
1124 | inline bool |
1125 | predictor_hash::equal (const edge_prediction *p1, const edge_prediction *p2) |
1126 | { |
1127 | return (p1->ep_predictor == p2->ep_predictor |
1128 | && (p1->ep_probability == p2->ep_probability |
1129 | || p1->ep_probability == REG_BR_PROB_BASE - p2->ep_probability)); |
1130 | } |
1131 | |
1132 | struct predictor_hash_traits: predictor_hash, |
1133 | typed_noop_remove <edge_prediction *> {}; |
1134 | |
1135 | /* Return true if edge prediction P is not in DATA hash set. */ |
1136 | |
1137 | static bool |
1138 | not_removed_prediction_p (edge_prediction *p, void *data) |
1139 | { |
1140 | hash_set<edge_prediction *> *remove = (hash_set<edge_prediction *> *) data; |
1141 | return !remove->contains (k: p); |
1142 | } |
1143 | |
1144 | /* Prune predictions for a basic block BB. Currently we do following |
1145 | clean-up steps: |
1146 | |
1147 | 1) remove duplicate prediction that is guessed with the same probability |
1148 | (different than 1/2) to both edge |
1149 | 2) remove duplicates for a prediction that belongs with the same probability |
1150 | to a single edge |
1151 | |
1152 | */ |
1153 | |
1154 | static void |
1155 | prune_predictions_for_bb (basic_block bb) |
1156 | { |
1157 | edge_prediction **preds = bb_predictions->get (k: bb); |
1158 | |
1159 | if (preds) |
1160 | { |
1161 | hash_table <predictor_hash_traits> s (13); |
1162 | hash_set <edge_prediction *> remove; |
1163 | |
1164 | /* Step 1: identify predictors that should be removed. */ |
1165 | for (edge_prediction *pred = *preds; pred; pred = pred->ep_next) |
1166 | { |
1167 | edge_prediction *existing = s.find (value: pred); |
1168 | if (existing) |
1169 | { |
1170 | if (pred->ep_edge == existing->ep_edge |
1171 | && pred->ep_probability == existing->ep_probability) |
1172 | { |
1173 | /* Remove a duplicate predictor. */ |
1174 | dump_prediction (file: dump_file, predictor: pred->ep_predictor, |
1175 | probability: pred->ep_probability, bb, |
1176 | reason: REASON_SINGLE_EDGE_DUPLICATE, ep_edge: pred->ep_edge); |
1177 | |
1178 | remove.add (k: pred); |
1179 | } |
1180 | else if (pred->ep_edge != existing->ep_edge |
1181 | && pred->ep_probability == existing->ep_probability |
1182 | && pred->ep_probability != REG_BR_PROB_BASE / 2) |
1183 | { |
1184 | /* Remove both predictors as they predict the same |
1185 | for both edges. */ |
1186 | dump_prediction (file: dump_file, predictor: existing->ep_predictor, |
1187 | probability: pred->ep_probability, bb, |
1188 | reason: REASON_EDGE_PAIR_DUPLICATE, |
1189 | ep_edge: existing->ep_edge); |
1190 | dump_prediction (file: dump_file, predictor: pred->ep_predictor, |
1191 | probability: pred->ep_probability, bb, |
1192 | reason: REASON_EDGE_PAIR_DUPLICATE, |
1193 | ep_edge: pred->ep_edge); |
1194 | |
1195 | remove.add (k: existing); |
1196 | remove.add (k: pred); |
1197 | } |
1198 | } |
1199 | |
1200 | edge_prediction **slot2 = s.find_slot (value: pred, insert: INSERT); |
1201 | *slot2 = pred; |
1202 | } |
1203 | |
1204 | /* Step 2: Remove predictors. */ |
1205 | filter_predictions (preds, filter: not_removed_prediction_p, data: &remove); |
1206 | } |
1207 | } |
1208 | |
1209 | /* Combine predictions into single probability and store them into CFG. |
1210 | Remove now useless prediction entries. |
1211 | If DRY_RUN is set, only produce dumps and do not modify profile. */ |
1212 | |
1213 | static void |
1214 | combine_predictions_for_bb (basic_block bb, bool dry_run) |
1215 | { |
1216 | int best_probability = PROB_EVEN; |
1217 | enum br_predictor best_predictor = END_PREDICTORS; |
1218 | int combined_probability = REG_BR_PROB_BASE / 2; |
1219 | int d; |
1220 | bool first_match = false; |
1221 | bool found = false; |
1222 | struct edge_prediction *pred; |
1223 | int nedges = 0; |
1224 | edge e, first = NULL, second = NULL; |
1225 | edge_iterator ei; |
1226 | int nzero = 0; |
1227 | int nunknown = 0; |
1228 | |
1229 | FOR_EACH_EDGE (e, ei, bb->succs) |
1230 | { |
1231 | if (!unlikely_executed_edge_p (e)) |
1232 | { |
1233 | nedges ++; |
1234 | if (first && !second) |
1235 | second = e; |
1236 | if (!first) |
1237 | first = e; |
1238 | } |
1239 | else if (!e->probability.initialized_p ()) |
1240 | e->probability = profile_probability::never (); |
1241 | if (!e->probability.initialized_p ()) |
1242 | nunknown++; |
1243 | else if (e->probability == profile_probability::never ()) |
1244 | nzero++; |
1245 | } |
1246 | |
1247 | /* When there is no successor or only one choice, prediction is easy. |
1248 | |
1249 | When we have a basic block with more than 2 successors, the situation |
1250 | is more complicated as DS theory cannot be used literally. |
1251 | More precisely, let's assume we predicted edge e1 with probability p1, |
1252 | thus: m1({b1}) = p1. As we're going to combine more than 2 edges, we |
1253 | need to find probability of e.g. m1({b2}), which we don't know. |
1254 | The only approximation is to equally distribute 1-p1 to all edges |
1255 | different from b1. |
1256 | |
1257 | According to numbers we've got from SPEC2006 benchark, there's only |
1258 | one interesting reliable predictor (noreturn call), which can be |
1259 | handled with a bit easier approach. */ |
1260 | if (nedges != 2) |
1261 | { |
1262 | hash_set<edge> unlikely_edges (4); |
1263 | hash_set<edge_prediction *> likely_edges (4); |
1264 | |
1265 | /* Identify all edges that have a probability close to very unlikely. |
1266 | Doing the approach for very unlikely doesn't worth for doing as |
1267 | there's no such probability in SPEC2006 benchmark. */ |
1268 | edge_prediction **preds = bb_predictions->get (k: bb); |
1269 | if (preds) |
1270 | for (pred = *preds; pred; pred = pred->ep_next) |
1271 | { |
1272 | if (pred->ep_probability <= PROB_VERY_UNLIKELY |
1273 | || pred->ep_predictor == PRED_COLD_LABEL) |
1274 | unlikely_edges.add (k: pred->ep_edge); |
1275 | else if (pred->ep_probability >= PROB_VERY_LIKELY |
1276 | || pred->ep_predictor == PRED_BUILTIN_EXPECT |
1277 | || pred->ep_predictor == PRED_HOT_LABEL) |
1278 | likely_edges.add (k: pred); |
1279 | } |
1280 | |
1281 | /* It can happen that an edge is both in likely_edges and unlikely_edges. |
1282 | Clear both sets in that situation. */ |
1283 | for (hash_set<edge_prediction *>::iterator it = likely_edges.begin (); |
1284 | it != likely_edges.end (); ++it) |
1285 | if (unlikely_edges.contains (k: (*it)->ep_edge)) |
1286 | { |
1287 | likely_edges.empty (); |
1288 | unlikely_edges.empty (); |
1289 | break; |
1290 | } |
1291 | |
1292 | if (!dry_run) |
1293 | set_even_probabilities (bb, unlikely_edges: &unlikely_edges, likely_edges: &likely_edges); |
1294 | clear_bb_predictions (bb); |
1295 | if (dump_file) |
1296 | { |
1297 | fprintf (stream: dump_file, format: "Predictions for bb %i\n" , bb->index); |
1298 | if (unlikely_edges.is_empty ()) |
1299 | fprintf (stream: dump_file, |
1300 | format: "%i edges in bb %i predicted to even probabilities\n" , |
1301 | nedges, bb->index); |
1302 | else |
1303 | { |
1304 | fprintf (stream: dump_file, |
1305 | format: "%i edges in bb %i predicted with some unlikely edges\n" , |
1306 | nedges, bb->index); |
1307 | FOR_EACH_EDGE (e, ei, bb->succs) |
1308 | if (!unlikely_executed_edge_p (e)) |
1309 | dump_prediction (file: dump_file, predictor: PRED_COMBINED, |
1310 | probability: e->probability.to_reg_br_prob_base (), bb, reason: REASON_NONE, ep_edge: e); |
1311 | } |
1312 | } |
1313 | return; |
1314 | } |
1315 | |
1316 | if (dump_file) |
1317 | fprintf (stream: dump_file, format: "Predictions for bb %i\n" , bb->index); |
1318 | |
1319 | prune_predictions_for_bb (bb); |
1320 | |
1321 | edge_prediction **preds = bb_predictions->get (k: bb); |
1322 | |
1323 | if (preds) |
1324 | { |
1325 | /* We implement "first match" heuristics and use probability guessed |
1326 | by predictor with smallest index. */ |
1327 | for (pred = *preds; pred; pred = pred->ep_next) |
1328 | { |
1329 | enum br_predictor predictor = pred->ep_predictor; |
1330 | int probability = pred->ep_probability; |
1331 | |
1332 | if (pred->ep_edge != first) |
1333 | probability = REG_BR_PROB_BASE - probability; |
1334 | |
1335 | found = true; |
1336 | /* First match heuristics would be widly confused if we predicted |
1337 | both directions. */ |
1338 | if (best_predictor > predictor |
1339 | && predictor_info[predictor].flags & PRED_FLAG_FIRST_MATCH) |
1340 | { |
1341 | struct edge_prediction *pred2; |
1342 | int prob = probability; |
1343 | |
1344 | for (pred2 = (struct edge_prediction *) *preds; |
1345 | pred2; pred2 = pred2->ep_next) |
1346 | if (pred2 != pred && pred2->ep_predictor == pred->ep_predictor) |
1347 | { |
1348 | int probability2 = pred2->ep_probability; |
1349 | |
1350 | if (pred2->ep_edge != first) |
1351 | probability2 = REG_BR_PROB_BASE - probability2; |
1352 | |
1353 | if ((probability < REG_BR_PROB_BASE / 2) != |
1354 | (probability2 < REG_BR_PROB_BASE / 2)) |
1355 | break; |
1356 | |
1357 | /* If the same predictor later gave better result, go for it! */ |
1358 | if ((probability >= REG_BR_PROB_BASE / 2 && (probability2 > probability)) |
1359 | || (probability <= REG_BR_PROB_BASE / 2 && (probability2 < probability))) |
1360 | prob = probability2; |
1361 | } |
1362 | if (!pred2) |
1363 | best_probability = prob, best_predictor = predictor; |
1364 | } |
1365 | |
1366 | d = (combined_probability * probability |
1367 | + (REG_BR_PROB_BASE - combined_probability) |
1368 | * (REG_BR_PROB_BASE - probability)); |
1369 | |
1370 | /* Use FP math to avoid overflows of 32bit integers. */ |
1371 | if (d == 0) |
1372 | /* If one probability is 0% and one 100%, avoid division by zero. */ |
1373 | combined_probability = REG_BR_PROB_BASE / 2; |
1374 | else |
1375 | combined_probability = (((double) combined_probability) |
1376 | * probability |
1377 | * REG_BR_PROB_BASE / d + 0.5); |
1378 | } |
1379 | } |
1380 | |
1381 | /* Decide which heuristic to use. In case we didn't match anything, |
1382 | use no_prediction heuristic, in case we did match, use either |
1383 | first match or Dempster-Shaffer theory depending on the flags. */ |
1384 | |
1385 | if (best_predictor != END_PREDICTORS) |
1386 | first_match = true; |
1387 | |
1388 | if (!found) |
1389 | dump_prediction (file: dump_file, predictor: PRED_NO_PREDICTION, probability: combined_probability, bb); |
1390 | else |
1391 | { |
1392 | if (!first_match) |
1393 | dump_prediction (file: dump_file, predictor: PRED_DS_THEORY, probability: combined_probability, bb, |
1394 | reason: !first_match ? REASON_NONE : REASON_IGNORED); |
1395 | else |
1396 | dump_prediction (file: dump_file, predictor: PRED_FIRST_MATCH, probability: best_probability, bb, |
1397 | reason: first_match ? REASON_NONE : REASON_IGNORED); |
1398 | } |
1399 | |
1400 | if (first_match) |
1401 | combined_probability = best_probability; |
1402 | dump_prediction (file: dump_file, predictor: PRED_COMBINED, probability: combined_probability, bb); |
1403 | |
1404 | if (preds) |
1405 | { |
1406 | for (pred = (struct edge_prediction *) *preds; pred; pred = pred->ep_next) |
1407 | { |
1408 | enum br_predictor predictor = pred->ep_predictor; |
1409 | int probability = pred->ep_probability; |
1410 | |
1411 | dump_prediction (file: dump_file, predictor, probability, bb, |
1412 | reason: (!first_match || best_predictor == predictor) |
1413 | ? REASON_NONE : REASON_IGNORED, ep_edge: pred->ep_edge); |
1414 | } |
1415 | } |
1416 | clear_bb_predictions (bb); |
1417 | |
1418 | |
1419 | /* If we have only one successor which is unknown, we can compute missing |
1420 | probability. */ |
1421 | if (nunknown == 1) |
1422 | { |
1423 | profile_probability prob = profile_probability::always (); |
1424 | edge missing = NULL; |
1425 | |
1426 | FOR_EACH_EDGE (e, ei, bb->succs) |
1427 | if (e->probability.initialized_p ()) |
1428 | prob -= e->probability; |
1429 | else if (missing == NULL) |
1430 | missing = e; |
1431 | else |
1432 | gcc_unreachable (); |
1433 | missing->probability = prob; |
1434 | } |
1435 | /* If nothing is unknown, we have nothing to update. */ |
1436 | else if (!nunknown && nzero != (int)EDGE_COUNT (bb->succs)) |
1437 | ; |
1438 | else if (!dry_run) |
1439 | { |
1440 | first->probability |
1441 | = profile_probability::from_reg_br_prob_base (v: combined_probability); |
1442 | second->probability = first->probability.invert (); |
1443 | } |
1444 | } |
1445 | |
1446 | /* Check if T1 and T2 satisfy the IV_COMPARE condition. |
1447 | Return the SSA_NAME if the condition satisfies, NULL otherwise. |
1448 | |
1449 | T1 and T2 should be one of the following cases: |
1450 | 1. T1 is SSA_NAME, T2 is NULL |
1451 | 2. T1 is SSA_NAME, T2 is INTEGER_CST between [-4, 4] |
1452 | 3. T2 is SSA_NAME, T1 is INTEGER_CST between [-4, 4] */ |
1453 | |
1454 | static tree |
1455 | strips_small_constant (tree t1, tree t2) |
1456 | { |
1457 | tree ret = NULL; |
1458 | int value = 0; |
1459 | |
1460 | if (!t1) |
1461 | return NULL; |
1462 | else if (TREE_CODE (t1) == SSA_NAME) |
1463 | ret = t1; |
1464 | else if (tree_fits_shwi_p (t1)) |
1465 | value = tree_to_shwi (t1); |
1466 | else |
1467 | return NULL; |
1468 | |
1469 | if (!t2) |
1470 | return ret; |
1471 | else if (tree_fits_shwi_p (t2)) |
1472 | value = tree_to_shwi (t2); |
1473 | else if (TREE_CODE (t2) == SSA_NAME) |
1474 | { |
1475 | if (ret) |
1476 | return NULL; |
1477 | else |
1478 | ret = t2; |
1479 | } |
1480 | |
1481 | if (value <= 4 && value >= -4) |
1482 | return ret; |
1483 | else |
1484 | return NULL; |
1485 | } |
1486 | |
1487 | /* Return the SSA_NAME in T or T's operands. |
1488 | Return NULL if SSA_NAME cannot be found. */ |
1489 | |
1490 | static tree |
1491 | get_base_value (tree t) |
1492 | { |
1493 | if (TREE_CODE (t) == SSA_NAME) |
1494 | return t; |
1495 | |
1496 | if (!BINARY_CLASS_P (t)) |
1497 | return NULL; |
1498 | |
1499 | switch (TREE_OPERAND_LENGTH (t)) |
1500 | { |
1501 | case 1: |
1502 | return strips_small_constant (TREE_OPERAND (t, 0), NULL); |
1503 | case 2: |
1504 | return strips_small_constant (TREE_OPERAND (t, 0), |
1505 | TREE_OPERAND (t, 1)); |
1506 | default: |
1507 | return NULL; |
1508 | } |
1509 | } |
1510 | |
1511 | /* Check the compare STMT in LOOP. If it compares an induction |
1512 | variable to a loop invariant, return true, and save |
1513 | LOOP_INVARIANT, COMPARE_CODE and LOOP_STEP. |
1514 | Otherwise return false and set LOOP_INVAIANT to NULL. */ |
1515 | |
1516 | static bool |
1517 | is_comparison_with_loop_invariant_p (gcond *stmt, class loop *loop, |
1518 | tree *loop_invariant, |
1519 | enum tree_code *compare_code, |
1520 | tree *loop_step, |
1521 | tree *loop_iv_base) |
1522 | { |
1523 | tree op0, op1, bound, base; |
1524 | affine_iv iv0, iv1; |
1525 | enum tree_code code; |
1526 | tree step; |
1527 | |
1528 | code = gimple_cond_code (gs: stmt); |
1529 | *loop_invariant = NULL; |
1530 | |
1531 | switch (code) |
1532 | { |
1533 | case GT_EXPR: |
1534 | case GE_EXPR: |
1535 | case NE_EXPR: |
1536 | case LT_EXPR: |
1537 | case LE_EXPR: |
1538 | case EQ_EXPR: |
1539 | break; |
1540 | |
1541 | default: |
1542 | return false; |
1543 | } |
1544 | |
1545 | op0 = gimple_cond_lhs (gs: stmt); |
1546 | op1 = gimple_cond_rhs (gs: stmt); |
1547 | |
1548 | if ((TREE_CODE (op0) != SSA_NAME && TREE_CODE (op0) != INTEGER_CST) |
1549 | || (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op1) != INTEGER_CST)) |
1550 | return false; |
1551 | if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, true)) |
1552 | return false; |
1553 | if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, true)) |
1554 | return false; |
1555 | if (TREE_CODE (iv0.step) != INTEGER_CST |
1556 | || TREE_CODE (iv1.step) != INTEGER_CST) |
1557 | return false; |
1558 | if ((integer_zerop (iv0.step) && integer_zerop (iv1.step)) |
1559 | || (!integer_zerop (iv0.step) && !integer_zerop (iv1.step))) |
1560 | return false; |
1561 | |
1562 | if (integer_zerop (iv0.step)) |
1563 | { |
1564 | if (code != NE_EXPR && code != EQ_EXPR) |
1565 | code = invert_tree_comparison (code, false); |
1566 | bound = iv0.base; |
1567 | base = iv1.base; |
1568 | if (tree_fits_shwi_p (iv1.step)) |
1569 | step = iv1.step; |
1570 | else |
1571 | return false; |
1572 | } |
1573 | else |
1574 | { |
1575 | bound = iv1.base; |
1576 | base = iv0.base; |
1577 | if (tree_fits_shwi_p (iv0.step)) |
1578 | step = iv0.step; |
1579 | else |
1580 | return false; |
1581 | } |
1582 | |
1583 | if (TREE_CODE (bound) != INTEGER_CST) |
1584 | bound = get_base_value (t: bound); |
1585 | if (!bound) |
1586 | return false; |
1587 | if (TREE_CODE (base) != INTEGER_CST) |
1588 | base = get_base_value (t: base); |
1589 | if (!base) |
1590 | return false; |
1591 | |
1592 | *loop_invariant = bound; |
1593 | *compare_code = code; |
1594 | *loop_step = step; |
1595 | *loop_iv_base = base; |
1596 | return true; |
1597 | } |
1598 | |
1599 | /* Compare two SSA_NAMEs: returns TRUE if T1 and T2 are value coherent. */ |
1600 | |
1601 | static bool |
1602 | expr_coherent_p (tree t1, tree t2) |
1603 | { |
1604 | gimple *stmt; |
1605 | tree ssa_name_1 = NULL; |
1606 | tree ssa_name_2 = NULL; |
1607 | |
1608 | gcc_assert (TREE_CODE (t1) == SSA_NAME || TREE_CODE (t1) == INTEGER_CST); |
1609 | gcc_assert (TREE_CODE (t2) == SSA_NAME || TREE_CODE (t2) == INTEGER_CST); |
1610 | |
1611 | if (t1 == t2) |
1612 | return true; |
1613 | |
1614 | if (TREE_CODE (t1) == INTEGER_CST && TREE_CODE (t2) == INTEGER_CST) |
1615 | return true; |
1616 | if (TREE_CODE (t1) == INTEGER_CST || TREE_CODE (t2) == INTEGER_CST) |
1617 | return false; |
1618 | |
1619 | /* Check to see if t1 is expressed/defined with t2. */ |
1620 | stmt = SSA_NAME_DEF_STMT (t1); |
1621 | gcc_assert (stmt != NULL); |
1622 | if (is_gimple_assign (gs: stmt)) |
1623 | { |
1624 | ssa_name_1 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); |
1625 | if (ssa_name_1 && ssa_name_1 == t2) |
1626 | return true; |
1627 | } |
1628 | |
1629 | /* Check to see if t2 is expressed/defined with t1. */ |
1630 | stmt = SSA_NAME_DEF_STMT (t2); |
1631 | gcc_assert (stmt != NULL); |
1632 | if (is_gimple_assign (gs: stmt)) |
1633 | { |
1634 | ssa_name_2 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); |
1635 | if (ssa_name_2 && ssa_name_2 == t1) |
1636 | return true; |
1637 | } |
1638 | |
1639 | /* Compare if t1 and t2's def_stmts are identical. */ |
1640 | if (ssa_name_2 != NULL && ssa_name_1 == ssa_name_2) |
1641 | return true; |
1642 | else |
1643 | return false; |
1644 | } |
1645 | |
1646 | /* Return true if E is predicted by one of loop heuristics. */ |
1647 | |
1648 | static bool |
1649 | predicted_by_loop_heuristics_p (basic_block bb) |
1650 | { |
1651 | struct edge_prediction *i; |
1652 | edge_prediction **preds = bb_predictions->get (k: bb); |
1653 | |
1654 | if (!preds) |
1655 | return false; |
1656 | |
1657 | for (i = *preds; i; i = i->ep_next) |
1658 | if (i->ep_predictor == PRED_LOOP_ITERATIONS_GUESSED |
1659 | || i->ep_predictor == PRED_LOOP_ITERATIONS_MAX |
1660 | || i->ep_predictor == PRED_LOOP_ITERATIONS |
1661 | || i->ep_predictor == PRED_LOOP_EXIT |
1662 | || i->ep_predictor == PRED_LOOP_EXIT_WITH_RECURSION |
1663 | || i->ep_predictor == PRED_LOOP_EXTRA_EXIT) |
1664 | return true; |
1665 | return false; |
1666 | } |
1667 | |
1668 | /* Predict branch probability of BB when BB contains a branch that compares |
1669 | an induction variable in LOOP with LOOP_IV_BASE_VAR to LOOP_BOUND_VAR. The |
1670 | loop exit is compared using LOOP_BOUND_CODE, with step of LOOP_BOUND_STEP. |
1671 | |
1672 | E.g. |
1673 | for (int i = 0; i < bound; i++) { |
1674 | if (i < bound - 2) |
1675 | computation_1(); |
1676 | else |
1677 | computation_2(); |
1678 | } |
1679 | |
1680 | In this loop, we will predict the branch inside the loop to be taken. */ |
1681 | |
1682 | static void |
1683 | predict_iv_comparison (class loop *loop, basic_block bb, |
1684 | tree loop_bound_var, |
1685 | tree loop_iv_base_var, |
1686 | enum tree_code loop_bound_code, |
1687 | int loop_bound_step) |
1688 | { |
1689 | tree compare_var, compare_base; |
1690 | enum tree_code compare_code; |
1691 | tree compare_step_var; |
1692 | edge then_edge; |
1693 | edge_iterator ei; |
1694 | |
1695 | if (predicted_by_loop_heuristics_p (bb)) |
1696 | return; |
1697 | |
1698 | gcond *stmt = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb)); |
1699 | if (!stmt) |
1700 | return; |
1701 | if (!is_comparison_with_loop_invariant_p (stmt, |
1702 | loop, loop_invariant: &compare_var, |
1703 | compare_code: &compare_code, |
1704 | loop_step: &compare_step_var, |
1705 | loop_iv_base: &compare_base)) |
1706 | return; |
1707 | |
1708 | /* Find the taken edge. */ |
1709 | FOR_EACH_EDGE (then_edge, ei, bb->succs) |
1710 | if (then_edge->flags & EDGE_TRUE_VALUE) |
1711 | break; |
1712 | |
1713 | /* When comparing an IV to a loop invariant, NE is more likely to be |
1714 | taken while EQ is more likely to be not-taken. */ |
1715 | if (compare_code == NE_EXPR) |
1716 | { |
1717 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1718 | return; |
1719 | } |
1720 | else if (compare_code == EQ_EXPR) |
1721 | { |
1722 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: NOT_TAKEN); |
1723 | return; |
1724 | } |
1725 | |
1726 | if (!expr_coherent_p (t1: loop_iv_base_var, t2: compare_base)) |
1727 | return; |
1728 | |
1729 | /* If loop bound, base and compare bound are all constants, we can |
1730 | calculate the probability directly. */ |
1731 | if (tree_fits_shwi_p (loop_bound_var) |
1732 | && tree_fits_shwi_p (compare_var) |
1733 | && tree_fits_shwi_p (compare_base)) |
1734 | { |
1735 | int probability; |
1736 | wi::overflow_type overflow; |
1737 | bool overall_overflow = false; |
1738 | widest_int compare_count, tem; |
1739 | |
1740 | /* (loop_bound - base) / compare_step */ |
1741 | tem = wi::sub (x: wi::to_widest (t: loop_bound_var), |
1742 | y: wi::to_widest (t: compare_base), sgn: SIGNED, overflow: &overflow); |
1743 | overall_overflow |= overflow; |
1744 | widest_int loop_count = wi::div_trunc (x: tem, |
1745 | y: wi::to_widest (t: compare_step_var), |
1746 | sgn: SIGNED, overflow: &overflow); |
1747 | overall_overflow |= overflow; |
1748 | |
1749 | if (!wi::neg_p (x: wi::to_widest (t: compare_step_var)) |
1750 | ^ (compare_code == LT_EXPR || compare_code == LE_EXPR)) |
1751 | { |
1752 | /* (loop_bound - compare_bound) / compare_step */ |
1753 | tem = wi::sub (x: wi::to_widest (t: loop_bound_var), |
1754 | y: wi::to_widest (t: compare_var), sgn: SIGNED, overflow: &overflow); |
1755 | overall_overflow |= overflow; |
1756 | compare_count = wi::div_trunc (x: tem, y: wi::to_widest (t: compare_step_var), |
1757 | sgn: SIGNED, overflow: &overflow); |
1758 | overall_overflow |= overflow; |
1759 | } |
1760 | else |
1761 | { |
1762 | /* (compare_bound - base) / compare_step */ |
1763 | tem = wi::sub (x: wi::to_widest (t: compare_var), |
1764 | y: wi::to_widest (t: compare_base), sgn: SIGNED, overflow: &overflow); |
1765 | overall_overflow |= overflow; |
1766 | compare_count = wi::div_trunc (x: tem, y: wi::to_widest (t: compare_step_var), |
1767 | sgn: SIGNED, overflow: &overflow); |
1768 | overall_overflow |= overflow; |
1769 | } |
1770 | if (compare_code == LE_EXPR || compare_code == GE_EXPR) |
1771 | ++compare_count; |
1772 | if (loop_bound_code == LE_EXPR || loop_bound_code == GE_EXPR) |
1773 | ++loop_count; |
1774 | if (wi::neg_p (x: compare_count)) |
1775 | compare_count = 0; |
1776 | if (wi::neg_p (x: loop_count)) |
1777 | loop_count = 0; |
1778 | if (loop_count == 0) |
1779 | probability = 0; |
1780 | else if (wi::cmps (x: compare_count, y: loop_count) == 1) |
1781 | probability = REG_BR_PROB_BASE; |
1782 | else |
1783 | { |
1784 | tem = compare_count * REG_BR_PROB_BASE; |
1785 | tem = wi::udiv_trunc (x: tem, y: loop_count); |
1786 | probability = tem.to_uhwi (); |
1787 | } |
1788 | |
1789 | /* FIXME: The branch prediction seems broken. It has only 20% hitrate. */ |
1790 | if (!overall_overflow) |
1791 | predict_edge (e: then_edge, predictor: PRED_LOOP_IV_COMPARE, probability); |
1792 | |
1793 | return; |
1794 | } |
1795 | |
1796 | if (expr_coherent_p (t1: loop_bound_var, t2: compare_var)) |
1797 | { |
1798 | if ((loop_bound_code == LT_EXPR || loop_bound_code == LE_EXPR) |
1799 | && (compare_code == LT_EXPR || compare_code == LE_EXPR)) |
1800 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1801 | else if ((loop_bound_code == GT_EXPR || loop_bound_code == GE_EXPR) |
1802 | && (compare_code == GT_EXPR || compare_code == GE_EXPR)) |
1803 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1804 | else if (loop_bound_code == NE_EXPR) |
1805 | { |
1806 | /* If the loop backedge condition is "(i != bound)", we do |
1807 | the comparison based on the step of IV: |
1808 | * step < 0 : backedge condition is like (i > bound) |
1809 | * step > 0 : backedge condition is like (i < bound) */ |
1810 | gcc_assert (loop_bound_step != 0); |
1811 | if (loop_bound_step > 0 |
1812 | && (compare_code == LT_EXPR |
1813 | || compare_code == LE_EXPR)) |
1814 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1815 | else if (loop_bound_step < 0 |
1816 | && (compare_code == GT_EXPR |
1817 | || compare_code == GE_EXPR)) |
1818 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1819 | else |
1820 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: NOT_TAKEN); |
1821 | } |
1822 | else |
1823 | /* The branch is predicted not-taken if loop_bound_code is |
1824 | opposite with compare_code. */ |
1825 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: NOT_TAKEN); |
1826 | } |
1827 | else if (expr_coherent_p (t1: loop_iv_base_var, t2: compare_var)) |
1828 | { |
1829 | /* For cases like: |
1830 | for (i = s; i < h; i++) |
1831 | if (i > s + 2) .... |
1832 | The branch should be predicted taken. */ |
1833 | if (loop_bound_step > 0 |
1834 | && (compare_code == GT_EXPR || compare_code == GE_EXPR)) |
1835 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1836 | else if (loop_bound_step < 0 |
1837 | && (compare_code == LT_EXPR || compare_code == LE_EXPR)) |
1838 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: TAKEN); |
1839 | else |
1840 | predict_edge_def (e: then_edge, predictor: PRED_LOOP_IV_COMPARE_GUESS, taken: NOT_TAKEN); |
1841 | } |
1842 | } |
1843 | |
1844 | /* Predict for extra loop exits that will lead to EXIT_EDGE. The extra loop |
1845 | exits are resulted from short-circuit conditions that will generate an |
1846 | if_tmp. E.g.: |
1847 | |
1848 | if (foo() || global > 10) |
1849 | break; |
1850 | |
1851 | This will be translated into: |
1852 | |
1853 | BB3: |
1854 | loop header... |
1855 | BB4: |
1856 | if foo() goto BB6 else goto BB5 |
1857 | BB5: |
1858 | if global > 10 goto BB6 else goto BB7 |
1859 | BB6: |
1860 | goto BB7 |
1861 | BB7: |
1862 | iftmp = (PHI 0(BB5), 1(BB6)) |
1863 | if iftmp == 1 goto BB8 else goto BB3 |
1864 | BB8: |
1865 | outside of the loop... |
1866 | |
1867 | The edge BB7->BB8 is loop exit because BB8 is outside of the loop. |
1868 | From the dataflow, we can infer that BB4->BB6 and BB5->BB6 are also loop |
1869 | exits. This function takes BB7->BB8 as input, and finds out the extra loop |
1870 | exits to predict them using PRED_LOOP_EXTRA_EXIT. */ |
1871 | |
1872 | static void |
1873 | (class loop *loop, edge exit_edge) |
1874 | { |
1875 | unsigned i; |
1876 | bool check_value_one; |
1877 | gimple *lhs_def_stmt; |
1878 | gphi *phi_stmt; |
1879 | tree cmp_rhs, cmp_lhs; |
1880 | |
1881 | gcond *cmp_stmt = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb: exit_edge->src)); |
1882 | if (!cmp_stmt) |
1883 | return; |
1884 | |
1885 | cmp_rhs = gimple_cond_rhs (gs: cmp_stmt); |
1886 | cmp_lhs = gimple_cond_lhs (gs: cmp_stmt); |
1887 | if (!TREE_CONSTANT (cmp_rhs) |
1888 | || !(integer_zerop (cmp_rhs) || integer_onep (cmp_rhs))) |
1889 | return; |
1890 | if (TREE_CODE (cmp_lhs) != SSA_NAME) |
1891 | return; |
1892 | |
1893 | /* If check_value_one is true, only the phi_args with value '1' will lead |
1894 | to loop exit. Otherwise, only the phi_args with value '0' will lead to |
1895 | loop exit. */ |
1896 | check_value_one = (((integer_onep (cmp_rhs)) |
1897 | ^ (gimple_cond_code (gs: cmp_stmt) == EQ_EXPR)) |
1898 | ^ ((exit_edge->flags & EDGE_TRUE_VALUE) != 0)); |
1899 | |
1900 | lhs_def_stmt = SSA_NAME_DEF_STMT (cmp_lhs); |
1901 | if (!lhs_def_stmt) |
1902 | return; |
1903 | |
1904 | phi_stmt = dyn_cast <gphi *> (p: lhs_def_stmt); |
1905 | if (!phi_stmt) |
1906 | return; |
1907 | |
1908 | for (i = 0; i < gimple_phi_num_args (gs: phi_stmt); i++) |
1909 | { |
1910 | edge e1; |
1911 | edge_iterator ei; |
1912 | tree val = gimple_phi_arg_def (gs: phi_stmt, index: i); |
1913 | edge e = gimple_phi_arg_edge (phi: phi_stmt, i); |
1914 | |
1915 | if (!TREE_CONSTANT (val) || !(integer_zerop (val) || integer_onep (val))) |
1916 | continue; |
1917 | if ((check_value_one ^ integer_onep (val)) == 1) |
1918 | continue; |
1919 | if (EDGE_COUNT (e->src->succs) != 1) |
1920 | { |
1921 | predict_paths_leading_to_edge (e, PRED_LOOP_EXTRA_EXIT, NOT_TAKEN, |
1922 | in_loop: loop); |
1923 | continue; |
1924 | } |
1925 | |
1926 | FOR_EACH_EDGE (e1, ei, e->src->preds) |
1927 | predict_paths_leading_to_edge (e1, PRED_LOOP_EXTRA_EXIT, NOT_TAKEN, |
1928 | in_loop: loop); |
1929 | } |
1930 | } |
1931 | |
1932 | |
1933 | /* Predict edge probabilities by exploiting loop structure. */ |
1934 | |
1935 | static void |
1936 | predict_loops (void) |
1937 | { |
1938 | basic_block bb; |
1939 | hash_set <class loop *> with_recursion(10); |
1940 | |
1941 | FOR_EACH_BB_FN (bb, cfun) |
1942 | { |
1943 | gimple_stmt_iterator gsi; |
1944 | tree decl; |
1945 | |
1946 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
1947 | if (is_gimple_call (gs: gsi_stmt (i: gsi)) |
1948 | && (decl = gimple_call_fndecl (gs: gsi_stmt (i: gsi))) != NULL |
1949 | && recursive_call_p (current_function_decl, decl)) |
1950 | { |
1951 | class loop *loop = bb->loop_father; |
1952 | while (loop && !with_recursion.add (k: loop)) |
1953 | loop = loop_outer (loop); |
1954 | } |
1955 | } |
1956 | |
1957 | /* Try to predict out blocks in a loop that are not part of a |
1958 | natural loop. */ |
1959 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
1960 | { |
1961 | basic_block bb, *bbs; |
1962 | unsigned j, n_exits = 0; |
1963 | class tree_niter_desc niter_desc; |
1964 | edge ex; |
1965 | class nb_iter_bound *nb_iter; |
1966 | enum tree_code loop_bound_code = ERROR_MARK; |
1967 | tree loop_bound_step = NULL; |
1968 | tree loop_bound_var = NULL; |
1969 | tree loop_iv_base = NULL; |
1970 | gcond *stmt = NULL; |
1971 | bool recursion = with_recursion.contains (k: loop); |
1972 | |
1973 | auto_vec<edge> exits = get_loop_exit_edges (loop); |
1974 | FOR_EACH_VEC_ELT (exits, j, ex) |
1975 | if (!unlikely_executed_edge_p (e: ex) && !(ex->flags & EDGE_ABNORMAL_CALL)) |
1976 | n_exits ++; |
1977 | if (!n_exits) |
1978 | continue; |
1979 | |
1980 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1981 | fprintf (stream: dump_file, format: "Predicting loop %i%s with %i exits.\n" , |
1982 | loop->num, recursion ? " (with recursion)" :"" , n_exits); |
1983 | if (dump_file && (dump_flags & TDF_DETAILS) |
1984 | && max_loop_iterations_int (loop) >= 0) |
1985 | { |
1986 | fprintf (stream: dump_file, |
1987 | format: "Loop %d iterates at most %i times.\n" , loop->num, |
1988 | (int)max_loop_iterations_int (loop)); |
1989 | } |
1990 | if (dump_file && (dump_flags & TDF_DETAILS) |
1991 | && likely_max_loop_iterations_int (loop) >= 0) |
1992 | { |
1993 | fprintf (stream: dump_file, format: "Loop %d likely iterates at most %i times.\n" , |
1994 | loop->num, (int)likely_max_loop_iterations_int (loop)); |
1995 | } |
1996 | |
1997 | FOR_EACH_VEC_ELT (exits, j, ex) |
1998 | { |
1999 | tree niter = NULL; |
2000 | HOST_WIDE_INT nitercst; |
2001 | int max = param_max_predicted_iterations; |
2002 | int probability; |
2003 | enum br_predictor predictor; |
2004 | widest_int nit; |
2005 | |
2006 | if (unlikely_executed_edge_p (e: ex) |
2007 | || (ex->flags & EDGE_ABNORMAL_CALL)) |
2008 | continue; |
2009 | /* Loop heuristics do not expect exit conditional to be inside |
2010 | inner loop. We predict from innermost to outermost loop. */ |
2011 | if (predicted_by_loop_heuristics_p (bb: ex->src)) |
2012 | { |
2013 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2014 | fprintf (stream: dump_file, format: "Skipping exit %i->%i because " |
2015 | "it is already predicted.\n" , |
2016 | ex->src->index, ex->dest->index); |
2017 | continue; |
2018 | } |
2019 | predict_extra_loop_exits (loop, exit_edge: ex); |
2020 | |
2021 | if (number_of_iterations_exit (loop, ex, niter: &niter_desc, false, every_iteration: false)) |
2022 | niter = niter_desc.niter; |
2023 | if (!niter || TREE_CODE (niter_desc.niter) != INTEGER_CST) |
2024 | niter = loop_niter_by_eval (loop, ex); |
2025 | if (dump_file && (dump_flags & TDF_DETAILS) |
2026 | && TREE_CODE (niter) == INTEGER_CST) |
2027 | { |
2028 | fprintf (stream: dump_file, format: "Exit %i->%i %d iterates " , |
2029 | ex->src->index, ex->dest->index, |
2030 | loop->num); |
2031 | print_generic_expr (dump_file, niter, TDF_SLIM); |
2032 | fprintf (stream: dump_file, format: " times.\n" ); |
2033 | } |
2034 | |
2035 | if (TREE_CODE (niter) == INTEGER_CST) |
2036 | { |
2037 | if (tree_fits_uhwi_p (niter) |
2038 | && max |
2039 | && compare_tree_int (niter, max - 1) == -1) |
2040 | nitercst = tree_to_uhwi (niter) + 1; |
2041 | else |
2042 | nitercst = max; |
2043 | predictor = PRED_LOOP_ITERATIONS; |
2044 | } |
2045 | /* If we have just one exit and we can derive some information about |
2046 | the number of iterations of the loop from the statements inside |
2047 | the loop, use it to predict this exit. */ |
2048 | else if (n_exits == 1 |
2049 | && estimated_stmt_executions (loop, &nit)) |
2050 | { |
2051 | if (wi::gtu_p (x: nit, y: max)) |
2052 | nitercst = max; |
2053 | else |
2054 | nitercst = nit.to_shwi (); |
2055 | predictor = PRED_LOOP_ITERATIONS_GUESSED; |
2056 | } |
2057 | /* If we have likely upper bound, trust it for very small iteration |
2058 | counts. Such loops would otherwise get mispredicted by standard |
2059 | LOOP_EXIT heuristics. */ |
2060 | else if (n_exits == 1 |
2061 | && likely_max_stmt_executions (loop, &nit) |
2062 | && wi::ltu_p (x: nit, |
2063 | RDIV (REG_BR_PROB_BASE, |
2064 | REG_BR_PROB_BASE |
2065 | - predictor_info |
2066 | [recursion |
2067 | ? PRED_LOOP_EXIT_WITH_RECURSION |
2068 | : PRED_LOOP_EXIT].hitrate))) |
2069 | { |
2070 | nitercst = nit.to_shwi (); |
2071 | predictor = PRED_LOOP_ITERATIONS_MAX; |
2072 | } |
2073 | else |
2074 | { |
2075 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2076 | fprintf (stream: dump_file, format: "Nothing known about exit %i->%i.\n" , |
2077 | ex->src->index, ex->dest->index); |
2078 | continue; |
2079 | } |
2080 | |
2081 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2082 | fprintf (stream: dump_file, format: "Recording prediction to %i iterations by %s.\n" , |
2083 | (int)nitercst, predictor_info[predictor].name); |
2084 | /* If the prediction for number of iterations is zero, do not |
2085 | predict the exit edges. */ |
2086 | if (nitercst == 0) |
2087 | continue; |
2088 | |
2089 | probability = RDIV (REG_BR_PROB_BASE, nitercst); |
2090 | predict_edge (e: ex, predictor, probability); |
2091 | } |
2092 | |
2093 | /* Find information about loop bound variables. */ |
2094 | for (nb_iter = loop->bounds; nb_iter; |
2095 | nb_iter = nb_iter->next) |
2096 | if (nb_iter->stmt |
2097 | && gimple_code (g: nb_iter->stmt) == GIMPLE_COND) |
2098 | { |
2099 | stmt = as_a <gcond *> (p: nb_iter->stmt); |
2100 | break; |
2101 | } |
2102 | if (!stmt) |
2103 | stmt = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb: loop->header)); |
2104 | if (stmt) |
2105 | is_comparison_with_loop_invariant_p (stmt, loop, |
2106 | loop_invariant: &loop_bound_var, |
2107 | compare_code: &loop_bound_code, |
2108 | loop_step: &loop_bound_step, |
2109 | loop_iv_base: &loop_iv_base); |
2110 | |
2111 | bbs = get_loop_body (loop); |
2112 | |
2113 | for (j = 0; j < loop->num_nodes; j++) |
2114 | { |
2115 | edge e; |
2116 | edge_iterator ei; |
2117 | |
2118 | bb = bbs[j]; |
2119 | |
2120 | /* Bypass loop heuristics on continue statement. These |
2121 | statements construct loops via "non-loop" constructs |
2122 | in the source language and are better to be handled |
2123 | separately. */ |
2124 | if (predicted_by_p (bb, predictor: PRED_CONTINUE)) |
2125 | { |
2126 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2127 | fprintf (stream: dump_file, format: "BB %i predicted by continue.\n" , |
2128 | bb->index); |
2129 | continue; |
2130 | } |
2131 | |
2132 | /* If we already used more reliable loop exit predictors, do not |
2133 | bother with PRED_LOOP_EXIT. */ |
2134 | if (!predicted_by_loop_heuristics_p (bb)) |
2135 | { |
2136 | /* For loop with many exits we don't want to predict all exits |
2137 | with the pretty large probability, because if all exits are |
2138 | considered in row, the loop would be predicted to iterate |
2139 | almost never. The code to divide probability by number of |
2140 | exits is very rough. It should compute the number of exits |
2141 | taken in each patch through function (not the overall number |
2142 | of exits that might be a lot higher for loops with wide switch |
2143 | statements in them) and compute n-th square root. |
2144 | |
2145 | We limit the minimal probability by 2% to avoid |
2146 | EDGE_PROBABILITY_RELIABLE from trusting the branch prediction |
2147 | as this was causing regression in perl benchmark containing such |
2148 | a wide loop. */ |
2149 | |
2150 | int probability = ((REG_BR_PROB_BASE |
2151 | - predictor_info |
2152 | [recursion |
2153 | ? PRED_LOOP_EXIT_WITH_RECURSION |
2154 | : PRED_LOOP_EXIT].hitrate) |
2155 | / n_exits); |
2156 | if (probability < HITRATE (2)) |
2157 | probability = HITRATE (2); |
2158 | FOR_EACH_EDGE (e, ei, bb->succs) |
2159 | if (e->dest->index < NUM_FIXED_BLOCKS |
2160 | || !flow_bb_inside_loop_p (loop, e->dest)) |
2161 | { |
2162 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2163 | fprintf (stream: dump_file, |
2164 | format: "Predicting exit %i->%i with prob %i.\n" , |
2165 | e->src->index, e->dest->index, probability); |
2166 | predict_edge (e, |
2167 | predictor: recursion ? PRED_LOOP_EXIT_WITH_RECURSION |
2168 | : PRED_LOOP_EXIT, probability); |
2169 | } |
2170 | } |
2171 | if (loop_bound_var) |
2172 | predict_iv_comparison (loop, bb, loop_bound_var, loop_iv_base_var: loop_iv_base, |
2173 | loop_bound_code, |
2174 | loop_bound_step: tree_to_shwi (loop_bound_step)); |
2175 | } |
2176 | |
2177 | /* In the following code |
2178 | for (loop1) |
2179 | if (cond) |
2180 | for (loop2) |
2181 | body; |
2182 | guess that cond is unlikely. */ |
2183 | if (loop_outer (loop)->num) |
2184 | { |
2185 | basic_block bb = NULL; |
2186 | edge = loop_preheader_edge (loop); |
2187 | |
2188 | if (single_pred_p (bb: preheader_edge->src) |
2189 | && single_succ_p (bb: preheader_edge->src)) |
2190 | preheader_edge = single_pred_edge (bb: preheader_edge->src); |
2191 | |
2192 | /* Pattern match fortran loop preheader: |
2193 | _16 = BUILTIN_EXPECT (_15, 1, PRED_FORTRAN_LOOP_PREHEADER); |
2194 | _17 = (logical(kind=4)) _16; |
2195 | if (_17 != 0) |
2196 | goto <bb 11>; |
2197 | else |
2198 | goto <bb 13>; |
2199 | |
2200 | Loop guard branch prediction says nothing about duplicated loop |
2201 | headers produced by fortran frontend and in this case we want |
2202 | to predict paths leading to this preheader. */ |
2203 | |
2204 | gcond *stmt |
2205 | = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb: preheader_edge->src)); |
2206 | if (stmt |
2207 | && gimple_cond_code (gs: stmt) == NE_EXPR |
2208 | && TREE_CODE (gimple_cond_lhs (stmt)) == SSA_NAME |
2209 | && integer_zerop (gimple_cond_rhs (gs: stmt))) |
2210 | { |
2211 | gimple *call_stmt = SSA_NAME_DEF_STMT (gimple_cond_lhs (stmt)); |
2212 | if (gimple_code (g: call_stmt) == GIMPLE_ASSIGN |
2213 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (call_stmt)) |
2214 | && TREE_CODE (gimple_assign_rhs1 (call_stmt)) == SSA_NAME) |
2215 | call_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (call_stmt)); |
2216 | if (gimple_call_internal_p (gs: call_stmt, fn: IFN_BUILTIN_EXPECT) |
2217 | && TREE_CODE (gimple_call_arg (call_stmt, 2)) == INTEGER_CST |
2218 | && tree_fits_uhwi_p (gimple_call_arg (gs: call_stmt, index: 2)) |
2219 | && tree_to_uhwi (gimple_call_arg (gs: call_stmt, index: 2)) |
2220 | == PRED_FORTRAN_LOOP_PREHEADER) |
2221 | bb = preheader_edge->src; |
2222 | } |
2223 | if (!bb) |
2224 | { |
2225 | if (!dominated_by_p (CDI_DOMINATORS, |
2226 | loop_outer (loop)->latch, loop->header)) |
2227 | predict_paths_leading_to_edge (loop_preheader_edge (loop), |
2228 | recursion |
2229 | ? PRED_LOOP_GUARD_WITH_RECURSION |
2230 | : PRED_LOOP_GUARD, |
2231 | NOT_TAKEN, |
2232 | in_loop: loop_outer (loop)); |
2233 | } |
2234 | else |
2235 | { |
2236 | if (!dominated_by_p (CDI_DOMINATORS, |
2237 | loop_outer (loop)->latch, bb)) |
2238 | predict_paths_leading_to (bb, |
2239 | recursion |
2240 | ? PRED_LOOP_GUARD_WITH_RECURSION |
2241 | : PRED_LOOP_GUARD, |
2242 | NOT_TAKEN, |
2243 | in_loop: loop_outer (loop)); |
2244 | } |
2245 | } |
2246 | |
2247 | /* Free basic blocks from get_loop_body. */ |
2248 | free (ptr: bbs); |
2249 | } |
2250 | } |
2251 | |
2252 | /* Attempt to predict probabilities of BB outgoing edges using local |
2253 | properties. */ |
2254 | static void |
2255 | bb_estimate_probability_locally (basic_block bb) |
2256 | { |
2257 | rtx_insn *last_insn = BB_END (bb); |
2258 | rtx cond; |
2259 | |
2260 | if (! can_predict_insn_p (insn: last_insn)) |
2261 | return; |
2262 | cond = get_condition (last_insn, NULL, false, false); |
2263 | if (! cond) |
2264 | return; |
2265 | |
2266 | /* Try "pointer heuristic." |
2267 | A comparison ptr == 0 is predicted as false. |
2268 | Similarly, a comparison ptr1 == ptr2 is predicted as false. */ |
2269 | if (COMPARISON_P (cond) |
2270 | && ((REG_P (XEXP (cond, 0)) && REG_POINTER (XEXP (cond, 0))) |
2271 | || (REG_P (XEXP (cond, 1)) && REG_POINTER (XEXP (cond, 1))))) |
2272 | { |
2273 | if (GET_CODE (cond) == EQ) |
2274 | predict_insn_def (insn: last_insn, predictor: PRED_POINTER, taken: NOT_TAKEN); |
2275 | else if (GET_CODE (cond) == NE) |
2276 | predict_insn_def (insn: last_insn, predictor: PRED_POINTER, taken: TAKEN); |
2277 | } |
2278 | else |
2279 | |
2280 | /* Try "opcode heuristic." |
2281 | EQ tests are usually false and NE tests are usually true. Also, |
2282 | most quantities are positive, so we can make the appropriate guesses |
2283 | about signed comparisons against zero. */ |
2284 | switch (GET_CODE (cond)) |
2285 | { |
2286 | case CONST_INT: |
2287 | /* Unconditional branch. */ |
2288 | predict_insn_def (insn: last_insn, predictor: PRED_UNCONDITIONAL, |
2289 | taken: cond == const0_rtx ? NOT_TAKEN : TAKEN); |
2290 | break; |
2291 | |
2292 | case EQ: |
2293 | case UNEQ: |
2294 | /* Floating point comparisons appears to behave in a very |
2295 | unpredictable way because of special role of = tests in |
2296 | FP code. */ |
2297 | if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) |
2298 | ; |
2299 | /* Comparisons with 0 are often used for booleans and there is |
2300 | nothing useful to predict about them. */ |
2301 | else if (XEXP (cond, 1) == const0_rtx |
2302 | || XEXP (cond, 0) == const0_rtx) |
2303 | ; |
2304 | else |
2305 | predict_insn_def (insn: last_insn, predictor: PRED_OPCODE_NONEQUAL, taken: NOT_TAKEN); |
2306 | break; |
2307 | |
2308 | case NE: |
2309 | case LTGT: |
2310 | /* Floating point comparisons appears to behave in a very |
2311 | unpredictable way because of special role of = tests in |
2312 | FP code. */ |
2313 | if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) |
2314 | ; |
2315 | /* Comparisons with 0 are often used for booleans and there is |
2316 | nothing useful to predict about them. */ |
2317 | else if (XEXP (cond, 1) == const0_rtx |
2318 | || XEXP (cond, 0) == const0_rtx) |
2319 | ; |
2320 | else |
2321 | predict_insn_def (insn: last_insn, predictor: PRED_OPCODE_NONEQUAL, taken: TAKEN); |
2322 | break; |
2323 | |
2324 | case ORDERED: |
2325 | predict_insn_def (insn: last_insn, predictor: PRED_FPOPCODE, taken: TAKEN); |
2326 | break; |
2327 | |
2328 | case UNORDERED: |
2329 | predict_insn_def (insn: last_insn, predictor: PRED_FPOPCODE, taken: NOT_TAKEN); |
2330 | break; |
2331 | |
2332 | case LE: |
2333 | case LT: |
2334 | if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx |
2335 | || XEXP (cond, 1) == constm1_rtx) |
2336 | predict_insn_def (insn: last_insn, predictor: PRED_OPCODE_POSITIVE, taken: NOT_TAKEN); |
2337 | break; |
2338 | |
2339 | case GE: |
2340 | case GT: |
2341 | if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx |
2342 | || XEXP (cond, 1) == constm1_rtx) |
2343 | predict_insn_def (insn: last_insn, predictor: PRED_OPCODE_POSITIVE, taken: TAKEN); |
2344 | break; |
2345 | |
2346 | default: |
2347 | break; |
2348 | } |
2349 | } |
2350 | |
2351 | /* Set edge->probability for each successor edge of BB. */ |
2352 | void |
2353 | guess_outgoing_edge_probabilities (basic_block bb) |
2354 | { |
2355 | bb_estimate_probability_locally (bb); |
2356 | combine_predictions_for_insn (BB_END (bb), bb); |
2357 | } |
2358 | |
2359 | static tree expr_expected_value (tree, bitmap, enum br_predictor *predictor, |
2360 | HOST_WIDE_INT *probability); |
2361 | |
2362 | /* Helper function for expr_expected_value. */ |
2363 | |
2364 | static tree |
2365 | expr_expected_value_1 (tree type, tree op0, enum tree_code code, |
2366 | tree op1, bitmap visited, enum br_predictor *predictor, |
2367 | HOST_WIDE_INT *probability) |
2368 | { |
2369 | gimple *def; |
2370 | |
2371 | /* Reset returned probability value. */ |
2372 | *probability = -1; |
2373 | *predictor = PRED_UNCONDITIONAL; |
2374 | |
2375 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
2376 | { |
2377 | if (TREE_CONSTANT (op0)) |
2378 | return op0; |
2379 | |
2380 | if (code == IMAGPART_EXPR) |
2381 | { |
2382 | if (TREE_CODE (TREE_OPERAND (op0, 0)) == SSA_NAME) |
2383 | { |
2384 | def = SSA_NAME_DEF_STMT (TREE_OPERAND (op0, 0)); |
2385 | if (is_gimple_call (gs: def) |
2386 | && gimple_call_internal_p (gs: def) |
2387 | && (gimple_call_internal_fn (gs: def) |
2388 | == IFN_ATOMIC_COMPARE_EXCHANGE)) |
2389 | { |
2390 | /* Assume that any given atomic operation has low contention, |
2391 | and thus the compare-and-swap operation succeeds. */ |
2392 | *predictor = PRED_COMPARE_AND_SWAP; |
2393 | return build_one_cst (TREE_TYPE (op0)); |
2394 | } |
2395 | } |
2396 | } |
2397 | |
2398 | if (code != SSA_NAME) |
2399 | return NULL_TREE; |
2400 | |
2401 | def = SSA_NAME_DEF_STMT (op0); |
2402 | |
2403 | /* If we were already here, break the infinite cycle. */ |
2404 | if (!bitmap_set_bit (visited, SSA_NAME_VERSION (op0))) |
2405 | return NULL; |
2406 | |
2407 | if (gphi *phi = dyn_cast <gphi *> (p: def)) |
2408 | { |
2409 | /* All the arguments of the PHI node must have the same constant |
2410 | length. */ |
2411 | int i, n = gimple_phi_num_args (gs: phi); |
2412 | tree val = NULL; |
2413 | bool has_nonzero_edge = false; |
2414 | |
2415 | /* If we already proved that given edge is unlikely, we do not need |
2416 | to handle merging of the probabilities. */ |
2417 | for (i = 0; i < n && !has_nonzero_edge; i++) |
2418 | { |
2419 | tree arg = PHI_ARG_DEF (phi, i); |
2420 | if (arg == PHI_RESULT (phi)) |
2421 | continue; |
2422 | profile_count cnt = gimple_phi_arg_edge (phi, i)->count (); |
2423 | if (!cnt.initialized_p () || cnt.nonzero_p ()) |
2424 | has_nonzero_edge = true; |
2425 | } |
2426 | |
2427 | for (i = 0; i < n; i++) |
2428 | { |
2429 | tree arg = PHI_ARG_DEF (phi, i); |
2430 | enum br_predictor predictor2; |
2431 | |
2432 | /* Skip self-referring parameters, since they does not change |
2433 | expected value. */ |
2434 | if (arg == PHI_RESULT (phi)) |
2435 | continue; |
2436 | |
2437 | /* Skip edges which we already predicted as executing |
2438 | zero times. */ |
2439 | if (has_nonzero_edge) |
2440 | { |
2441 | profile_count cnt = gimple_phi_arg_edge (phi, i)->count (); |
2442 | if (cnt.initialized_p () && !cnt.nonzero_p ()) |
2443 | continue; |
2444 | } |
2445 | HOST_WIDE_INT probability2; |
2446 | tree new_val = expr_expected_value (arg, visited, predictor: &predictor2, |
2447 | probability: &probability2); |
2448 | /* If we know nothing about value, give up. */ |
2449 | if (!new_val) |
2450 | return NULL; |
2451 | |
2452 | /* If this is a first edge, trust its prediction. */ |
2453 | if (!val) |
2454 | { |
2455 | val = new_val; |
2456 | *predictor = predictor2; |
2457 | *probability = probability2; |
2458 | continue; |
2459 | } |
2460 | /* If there are two different values, give up. */ |
2461 | if (!operand_equal_p (val, new_val, flags: false)) |
2462 | return NULL; |
2463 | |
2464 | int p1 = get_predictor_value (*predictor, *probability); |
2465 | int p2 = get_predictor_value (predictor2, probability2); |
2466 | /* If both predictors agree, it does not matter from which |
2467 | edge we enter the basic block. */ |
2468 | if (*predictor == predictor2 && p1 == p2) |
2469 | continue; |
2470 | /* The general case has no precise solution, since we do not |
2471 | know probabilities of incomming edges, yet. |
2472 | Still if value is predicted over all incomming edges, we |
2473 | can hope it will be indeed the case. Conservatively |
2474 | downgrade prediction quality (so first match merging is not |
2475 | performed) and take least successful prediction. */ |
2476 | |
2477 | *predictor = PRED_COMBINED_VALUE_PREDICTIONS_PHI; |
2478 | *probability = MIN (p1, p2); |
2479 | } |
2480 | return val; |
2481 | } |
2482 | if (is_gimple_assign (gs: def)) |
2483 | { |
2484 | if (gimple_assign_lhs (gs: def) != op0) |
2485 | return NULL; |
2486 | |
2487 | return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def)), |
2488 | op0: gimple_assign_rhs1 (gs: def), |
2489 | code: gimple_assign_rhs_code (gs: def), |
2490 | op1: gimple_assign_rhs2 (gs: def), |
2491 | visited, predictor, probability); |
2492 | } |
2493 | |
2494 | if (is_gimple_call (gs: def)) |
2495 | { |
2496 | tree decl = gimple_call_fndecl (gs: def); |
2497 | if (!decl) |
2498 | { |
2499 | if (gimple_call_internal_p (gs: def) |
2500 | && gimple_call_internal_fn (gs: def) == IFN_BUILTIN_EXPECT) |
2501 | { |
2502 | gcc_assert (gimple_call_num_args (def) == 3); |
2503 | tree val = gimple_call_arg (gs: def, index: 0); |
2504 | if (TREE_CONSTANT (val)) |
2505 | return val; |
2506 | tree val2 = gimple_call_arg (gs: def, index: 2); |
2507 | gcc_assert (TREE_CODE (val2) == INTEGER_CST |
2508 | && tree_fits_uhwi_p (val2) |
2509 | && tree_to_uhwi (val2) < END_PREDICTORS); |
2510 | *predictor = (enum br_predictor) tree_to_uhwi (val2); |
2511 | if (*predictor == PRED_BUILTIN_EXPECT) |
2512 | *probability |
2513 | = HITRATE (param_builtin_expect_probability); |
2514 | return gimple_call_arg (gs: def, index: 1); |
2515 | } |
2516 | return NULL; |
2517 | } |
2518 | |
2519 | if (DECL_IS_MALLOC (decl) || DECL_IS_OPERATOR_NEW_P (decl)) |
2520 | { |
2521 | if (predictor) |
2522 | *predictor = PRED_MALLOC_NONNULL; |
2523 | /* FIXME: This is wrong and we need to convert the logic |
2524 | to value ranges. This makes predictor to assume that |
2525 | malloc always returns (size_t)1 which is not the same |
2526 | as returning non-NULL. */ |
2527 | return fold_convert (type, boolean_true_node); |
2528 | } |
2529 | |
2530 | if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) |
2531 | switch (DECL_FUNCTION_CODE (decl)) |
2532 | { |
2533 | case BUILT_IN_EXPECT: |
2534 | { |
2535 | tree val; |
2536 | if (gimple_call_num_args (gs: def) != 2) |
2537 | return NULL; |
2538 | val = gimple_call_arg (gs: def, index: 0); |
2539 | if (TREE_CONSTANT (val)) |
2540 | return val; |
2541 | *predictor = PRED_BUILTIN_EXPECT; |
2542 | *probability |
2543 | = HITRATE (param_builtin_expect_probability); |
2544 | return gimple_call_arg (gs: def, index: 1); |
2545 | } |
2546 | case BUILT_IN_EXPECT_WITH_PROBABILITY: |
2547 | { |
2548 | tree val; |
2549 | if (gimple_call_num_args (gs: def) != 3) |
2550 | return NULL; |
2551 | val = gimple_call_arg (gs: def, index: 0); |
2552 | if (TREE_CONSTANT (val)) |
2553 | return val; |
2554 | /* Compute final probability as: |
2555 | probability * REG_BR_PROB_BASE. */ |
2556 | tree prob = gimple_call_arg (gs: def, index: 2); |
2557 | tree t = TREE_TYPE (prob); |
2558 | tree base = build_int_cst (integer_type_node, |
2559 | REG_BR_PROB_BASE); |
2560 | base = build_real_from_int_cst (t, base); |
2561 | tree r = fold_build2_initializer_loc (UNKNOWN_LOCATION, |
2562 | MULT_EXPR, t, prob, base); |
2563 | if (TREE_CODE (r) != REAL_CST) |
2564 | { |
2565 | error_at (gimple_location (g: def), |
2566 | "probability %qE must be " |
2567 | "constant floating-point expression" , prob); |
2568 | return NULL; |
2569 | } |
2570 | HOST_WIDE_INT probi |
2571 | = real_to_integer (TREE_REAL_CST_PTR (r)); |
2572 | if (probi >= 0 && probi <= REG_BR_PROB_BASE) |
2573 | { |
2574 | *predictor = PRED_BUILTIN_EXPECT_WITH_PROBABILITY; |
2575 | *probability = probi; |
2576 | } |
2577 | else |
2578 | error_at (gimple_location (g: def), |
2579 | "probability %qE is outside " |
2580 | "the range [0.0, 1.0]" , prob); |
2581 | |
2582 | return gimple_call_arg (gs: def, index: 1); |
2583 | } |
2584 | |
2585 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N: |
2586 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1: |
2587 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2: |
2588 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4: |
2589 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8: |
2590 | case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16: |
2591 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE: |
2592 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N: |
2593 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1: |
2594 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2: |
2595 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4: |
2596 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8: |
2597 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16: |
2598 | /* Assume that any given atomic operation has low contention, |
2599 | and thus the compare-and-swap operation succeeds. */ |
2600 | *predictor = PRED_COMPARE_AND_SWAP; |
2601 | return boolean_true_node; |
2602 | case BUILT_IN_REALLOC: |
2603 | case BUILT_IN_GOMP_REALLOC: |
2604 | if (predictor) |
2605 | *predictor = PRED_MALLOC_NONNULL; |
2606 | /* FIXME: This is wrong and we need to convert the logic |
2607 | to value ranges. */ |
2608 | return fold_convert (type, boolean_true_node); |
2609 | default: |
2610 | break; |
2611 | } |
2612 | } |
2613 | |
2614 | return NULL; |
2615 | } |
2616 | |
2617 | if (get_gimple_rhs_class (code) == GIMPLE_BINARY_RHS) |
2618 | { |
2619 | tree res; |
2620 | tree nop0 = op0; |
2621 | tree nop1 = op1; |
2622 | |
2623 | /* First handle situation where single op is enough to determine final |
2624 | value. In this case we can do better job by avoiding the prediction |
2625 | merging. */ |
2626 | if (TREE_CODE (op0) != INTEGER_CST) |
2627 | { |
2628 | /* See if expected value of op0 is good enough to determine the result. */ |
2629 | nop0 = expr_expected_value (op0, visited, predictor, probability); |
2630 | if (nop0 |
2631 | && (res = fold_build2 (code, type, nop0, op1)) != NULL |
2632 | && TREE_CODE (res) == INTEGER_CST) |
2633 | /* We are now getting conservative probability. Consider for |
2634 | example: |
2635 | op0 * op1 |
2636 | If op0 is 0 with probability p, then we will ignore the |
2637 | posibility that op0 != 0 and op1 == 0. It does not seem to be |
2638 | worthwhile to downgrade prediciton quality for this. */ |
2639 | return res; |
2640 | if (!nop0) |
2641 | nop0 = op0; |
2642 | } |
2643 | enum br_predictor predictor2 = PRED_UNCONDITIONAL; |
2644 | HOST_WIDE_INT probability2 = -1; |
2645 | if (TREE_CODE (op1) != INTEGER_CST) |
2646 | { |
2647 | /* See if expected value of op1 is good enough to determine the result. */ |
2648 | nop1 = expr_expected_value (op1, visited, predictor: &predictor2, probability: &probability2); |
2649 | if (nop1 |
2650 | && (res = fold_build2 (code, type, op0, nop1)) != NULL |
2651 | && TREE_CODE (res) == INTEGER_CST) |
2652 | { |
2653 | /* Similarly as above we now get conservative probability. */ |
2654 | *predictor = predictor2; |
2655 | *probability = probability2; |
2656 | return res; |
2657 | } |
2658 | if (!nop1) |
2659 | nop1 = op1; |
2660 | } |
2661 | /* We already checked if folding one of arguments to constant is good |
2662 | enough. Consequently failing to fold both means that we will not |
2663 | succeed determining the value. */ |
2664 | if (nop0 == op0 || nop1 == op1) |
2665 | return NULL; |
2666 | /* Finally see if we have two known values. */ |
2667 | res = fold_build2 (code, type, nop0, nop1); |
2668 | if (TREE_CODE (res) == INTEGER_CST) |
2669 | { |
2670 | HOST_WIDE_INT p1 = get_predictor_value (*predictor, *probability); |
2671 | HOST_WIDE_INT p2 = get_predictor_value (predictor2, probability2); |
2672 | |
2673 | /* If one of predictions is sure, such as PRED_UNCONDITIONAL, we |
2674 | can ignore it. */ |
2675 | if (p2 == PROB_ALWAYS) |
2676 | return res; |
2677 | if (p1 == PROB_ALWAYS) |
2678 | { |
2679 | *predictor = predictor2; |
2680 | *probability = probability2; |
2681 | return res; |
2682 | } |
2683 | /* Combine binary predictions. |
2684 | Since we do not know about independence of predictors, we |
2685 | can not determine value precisely. */ |
2686 | *probability = RDIV (p1 * p2, REG_BR_PROB_BASE); |
2687 | /* If we no longer track useful information, give up. */ |
2688 | if (!*probability) |
2689 | return NULL; |
2690 | /* Otherwise mark that prediction is a result of combining |
2691 | different heuristics, since we do not want it to participate |
2692 | in first match merging. It is no longer reliable since |
2693 | we do not know if the probabilities are indpenendet. */ |
2694 | *predictor = PRED_COMBINED_VALUE_PREDICTIONS; |
2695 | |
2696 | return res; |
2697 | } |
2698 | return NULL; |
2699 | } |
2700 | if (get_gimple_rhs_class (code) == GIMPLE_UNARY_RHS) |
2701 | { |
2702 | tree res; |
2703 | op0 = expr_expected_value (op0, visited, predictor, probability); |
2704 | if (!op0) |
2705 | return NULL; |
2706 | res = fold_build1 (code, type, op0); |
2707 | if (TREE_CONSTANT (res)) |
2708 | return res; |
2709 | return NULL; |
2710 | } |
2711 | return NULL; |
2712 | } |
2713 | |
2714 | /* Return constant EXPR will likely have at execution time, NULL if unknown. |
2715 | The function is used by builtin_expect branch predictor so the evidence |
2716 | must come from this construct and additional possible constant folding. |
2717 | |
2718 | We may want to implement more involved value guess (such as value range |
2719 | propagation based prediction), but such tricks shall go to new |
2720 | implementation. */ |
2721 | |
2722 | static tree |
2723 | expr_expected_value (tree expr, bitmap visited, |
2724 | enum br_predictor *predictor, |
2725 | HOST_WIDE_INT *probability) |
2726 | { |
2727 | enum tree_code code; |
2728 | tree op0, op1; |
2729 | |
2730 | if (TREE_CONSTANT (expr)) |
2731 | { |
2732 | *predictor = PRED_UNCONDITIONAL; |
2733 | *probability = -1; |
2734 | return expr; |
2735 | } |
2736 | |
2737 | extract_ops_from_tree (expr, code: &code, op0: &op0, op1: &op1); |
2738 | return expr_expected_value_1 (TREE_TYPE (expr), |
2739 | op0, code, op1, visited, predictor, |
2740 | probability); |
2741 | } |
2742 | |
2743 | |
2744 | /* Return probability of a PREDICTOR. If the predictor has variable |
2745 | probability return passed PROBABILITY. */ |
2746 | |
2747 | static HOST_WIDE_INT |
2748 | get_predictor_value (br_predictor predictor, HOST_WIDE_INT probability) |
2749 | { |
2750 | switch (predictor) |
2751 | { |
2752 | case PRED_BUILTIN_EXPECT: |
2753 | case PRED_BUILTIN_EXPECT_WITH_PROBABILITY: |
2754 | case PRED_COMBINED_VALUE_PREDICTIONS_PHI: |
2755 | case PRED_COMBINED_VALUE_PREDICTIONS: |
2756 | gcc_assert (probability != -1); |
2757 | return probability; |
2758 | default: |
2759 | gcc_assert (probability == -1); |
2760 | return predictor_info[(int) predictor].hitrate; |
2761 | } |
2762 | } |
2763 | |
2764 | /* Predict using opcode of the last statement in basic block. */ |
2765 | static void |
2766 | tree_predict_by_opcode (basic_block bb) |
2767 | { |
2768 | edge then_edge; |
2769 | tree op0, op1; |
2770 | tree type; |
2771 | tree val; |
2772 | enum tree_code cmp; |
2773 | edge_iterator ei; |
2774 | enum br_predictor predictor; |
2775 | HOST_WIDE_INT probability; |
2776 | |
2777 | gimple *stmt = *gsi_last_bb (bb); |
2778 | if (!stmt) |
2779 | return; |
2780 | |
2781 | if (gswitch *sw = dyn_cast <gswitch *> (p: stmt)) |
2782 | { |
2783 | tree index = gimple_switch_index (gs: sw); |
2784 | tree val = expr_expected_value (expr: index, visited: auto_bitmap (), |
2785 | predictor: &predictor, probability: &probability); |
2786 | if (val && TREE_CODE (val) == INTEGER_CST) |
2787 | { |
2788 | edge e = find_taken_edge_switch_expr (switch_stmt: sw, val); |
2789 | if (predictor == PRED_BUILTIN_EXPECT) |
2790 | { |
2791 | int percent = param_builtin_expect_probability; |
2792 | gcc_assert (percent >= 0 && percent <= 100); |
2793 | predict_edge (e, predictor: PRED_BUILTIN_EXPECT, |
2794 | HITRATE (percent)); |
2795 | } |
2796 | else |
2797 | predict_edge_def (e, predictor, taken: TAKEN); |
2798 | } |
2799 | } |
2800 | |
2801 | if (gimple_code (g: stmt) != GIMPLE_COND) |
2802 | return; |
2803 | FOR_EACH_EDGE (then_edge, ei, bb->succs) |
2804 | if (then_edge->flags & EDGE_TRUE_VALUE) |
2805 | break; |
2806 | op0 = gimple_cond_lhs (gs: stmt); |
2807 | op1 = gimple_cond_rhs (gs: stmt); |
2808 | cmp = gimple_cond_code (gs: stmt); |
2809 | type = TREE_TYPE (op0); |
2810 | val = expr_expected_value_1 (boolean_type_node, op0, code: cmp, op1, visited: auto_bitmap (), |
2811 | predictor: &predictor, probability: &probability); |
2812 | if (val && TREE_CODE (val) == INTEGER_CST) |
2813 | { |
2814 | HOST_WIDE_INT prob = get_predictor_value (predictor, probability); |
2815 | if (integer_zerop (val)) |
2816 | prob = REG_BR_PROB_BASE - prob; |
2817 | predict_edge (e: then_edge, predictor, probability: prob); |
2818 | } |
2819 | /* Try "pointer heuristic." |
2820 | A comparison ptr == 0 is predicted as false. |
2821 | Similarly, a comparison ptr1 == ptr2 is predicted as false. */ |
2822 | if (POINTER_TYPE_P (type)) |
2823 | { |
2824 | if (cmp == EQ_EXPR) |
2825 | predict_edge_def (e: then_edge, predictor: PRED_TREE_POINTER, taken: NOT_TAKEN); |
2826 | else if (cmp == NE_EXPR) |
2827 | predict_edge_def (e: then_edge, predictor: PRED_TREE_POINTER, taken: TAKEN); |
2828 | } |
2829 | else |
2830 | |
2831 | /* Try "opcode heuristic." |
2832 | EQ tests are usually false and NE tests are usually true. Also, |
2833 | most quantities are positive, so we can make the appropriate guesses |
2834 | about signed comparisons against zero. */ |
2835 | switch (cmp) |
2836 | { |
2837 | case EQ_EXPR: |
2838 | case UNEQ_EXPR: |
2839 | /* Floating point comparisons appears to behave in a very |
2840 | unpredictable way because of special role of = tests in |
2841 | FP code. */ |
2842 | if (FLOAT_TYPE_P (type)) |
2843 | ; |
2844 | /* Comparisons with 0 are often used for booleans and there is |
2845 | nothing useful to predict about them. */ |
2846 | else if (integer_zerop (op0) || integer_zerop (op1)) |
2847 | ; |
2848 | else |
2849 | predict_edge_def (e: then_edge, predictor: PRED_TREE_OPCODE_NONEQUAL, taken: NOT_TAKEN); |
2850 | break; |
2851 | |
2852 | case NE_EXPR: |
2853 | case LTGT_EXPR: |
2854 | /* Floating point comparisons appears to behave in a very |
2855 | unpredictable way because of special role of = tests in |
2856 | FP code. */ |
2857 | if (FLOAT_TYPE_P (type)) |
2858 | ; |
2859 | /* Comparisons with 0 are often used for booleans and there is |
2860 | nothing useful to predict about them. */ |
2861 | else if (integer_zerop (op0) |
2862 | || integer_zerop (op1)) |
2863 | ; |
2864 | else |
2865 | predict_edge_def (e: then_edge, predictor: PRED_TREE_OPCODE_NONEQUAL, taken: TAKEN); |
2866 | break; |
2867 | |
2868 | case ORDERED_EXPR: |
2869 | predict_edge_def (e: then_edge, predictor: PRED_TREE_FPOPCODE, taken: TAKEN); |
2870 | break; |
2871 | |
2872 | case UNORDERED_EXPR: |
2873 | predict_edge_def (e: then_edge, predictor: PRED_TREE_FPOPCODE, taken: NOT_TAKEN); |
2874 | break; |
2875 | |
2876 | case LE_EXPR: |
2877 | case LT_EXPR: |
2878 | if (integer_zerop (op1) |
2879 | || integer_onep (op1) |
2880 | || integer_all_onesp (op1) |
2881 | || real_zerop (op1) |
2882 | || real_onep (op1) |
2883 | || real_minus_onep (op1)) |
2884 | predict_edge_def (e: then_edge, predictor: PRED_TREE_OPCODE_POSITIVE, taken: NOT_TAKEN); |
2885 | break; |
2886 | |
2887 | case GE_EXPR: |
2888 | case GT_EXPR: |
2889 | if (integer_zerop (op1) |
2890 | || integer_onep (op1) |
2891 | || integer_all_onesp (op1) |
2892 | || real_zerop (op1) |
2893 | || real_onep (op1) |
2894 | || real_minus_onep (op1)) |
2895 | predict_edge_def (e: then_edge, predictor: PRED_TREE_OPCODE_POSITIVE, taken: TAKEN); |
2896 | break; |
2897 | |
2898 | default: |
2899 | break; |
2900 | } |
2901 | } |
2902 | |
2903 | /* Returns TRUE if the STMT is exit(0) like statement. */ |
2904 | |
2905 | static bool |
2906 | is_exit_with_zero_arg (const gimple *stmt) |
2907 | { |
2908 | /* This is not exit, _exit or _Exit. */ |
2909 | if (!gimple_call_builtin_p (stmt, BUILT_IN_EXIT) |
2910 | && !gimple_call_builtin_p (stmt, BUILT_IN__EXIT) |
2911 | && !gimple_call_builtin_p (stmt, BUILT_IN__EXIT2)) |
2912 | return false; |
2913 | |
2914 | /* Argument is an interger zero. */ |
2915 | return integer_zerop (gimple_call_arg (gs: stmt, index: 0)); |
2916 | } |
2917 | |
2918 | /* Try to guess whether the value of return means error code. */ |
2919 | |
2920 | static enum br_predictor |
2921 | return_prediction (tree val, enum prediction *prediction) |
2922 | { |
2923 | /* VOID. */ |
2924 | if (!val) |
2925 | return PRED_NO_PREDICTION; |
2926 | /* Different heuristics for pointers and scalars. */ |
2927 | if (POINTER_TYPE_P (TREE_TYPE (val))) |
2928 | { |
2929 | /* NULL is usually not returned. */ |
2930 | if (integer_zerop (val)) |
2931 | { |
2932 | *prediction = NOT_TAKEN; |
2933 | return PRED_NULL_RETURN; |
2934 | } |
2935 | } |
2936 | else if (INTEGRAL_TYPE_P (TREE_TYPE (val))) |
2937 | { |
2938 | /* Negative return values are often used to indicate |
2939 | errors. */ |
2940 | if (TREE_CODE (val) == INTEGER_CST |
2941 | && tree_int_cst_sgn (val) < 0) |
2942 | { |
2943 | *prediction = NOT_TAKEN; |
2944 | return PRED_NEGATIVE_RETURN; |
2945 | } |
2946 | /* Constant return values seems to be commonly taken. |
2947 | Zero/one often represent booleans so exclude them from the |
2948 | heuristics. */ |
2949 | if (TREE_CONSTANT (val) |
2950 | && (!integer_zerop (val) && !integer_onep (val))) |
2951 | { |
2952 | *prediction = NOT_TAKEN; |
2953 | return PRED_CONST_RETURN; |
2954 | } |
2955 | } |
2956 | return PRED_NO_PREDICTION; |
2957 | } |
2958 | |
2959 | /* Return zero if phi result could have values other than -1, 0 or 1, |
2960 | otherwise return a bitmask, with bits 0, 1 and 2 set if -1, 0 and 1 |
2961 | values are used or likely. */ |
2962 | |
2963 | static int |
2964 | zero_one_minusone (gphi *phi, int limit) |
2965 | { |
2966 | int phi_num_args = gimple_phi_num_args (gs: phi); |
2967 | int ret = 0; |
2968 | for (int i = 0; i < phi_num_args; i++) |
2969 | { |
2970 | tree t = PHI_ARG_DEF (phi, i); |
2971 | if (TREE_CODE (t) != INTEGER_CST) |
2972 | continue; |
2973 | wide_int w = wi::to_wide (t); |
2974 | if (w == -1) |
2975 | ret |= 1; |
2976 | else if (w == 0) |
2977 | ret |= 2; |
2978 | else if (w == 1) |
2979 | ret |= 4; |
2980 | else |
2981 | return 0; |
2982 | } |
2983 | for (int i = 0; i < phi_num_args; i++) |
2984 | { |
2985 | tree t = PHI_ARG_DEF (phi, i); |
2986 | if (TREE_CODE (t) == INTEGER_CST) |
2987 | continue; |
2988 | if (TREE_CODE (t) != SSA_NAME) |
2989 | return 0; |
2990 | gimple *g = SSA_NAME_DEF_STMT (t); |
2991 | if (gimple_code (g) == GIMPLE_PHI && limit > 0) |
2992 | if (int r = zero_one_minusone (phi: as_a <gphi *> (p: g), limit: limit - 1)) |
2993 | { |
2994 | ret |= r; |
2995 | continue; |
2996 | } |
2997 | if (!is_gimple_assign (gs: g)) |
2998 | return 0; |
2999 | if (gimple_assign_cast_p (s: g)) |
3000 | { |
3001 | tree rhs1 = gimple_assign_rhs1 (gs: g); |
3002 | if (TREE_CODE (rhs1) != SSA_NAME |
3003 | || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |
3004 | || TYPE_PRECISION (TREE_TYPE (rhs1)) != 1 |
3005 | || !TYPE_UNSIGNED (TREE_TYPE (rhs1))) |
3006 | return 0; |
3007 | ret |= (2 | 4); |
3008 | continue; |
3009 | } |
3010 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (g)) != tcc_comparison) |
3011 | return 0; |
3012 | ret |= (2 | 4); |
3013 | } |
3014 | return ret; |
3015 | } |
3016 | |
3017 | /* Find the basic block with return expression and look up for possible |
3018 | return value trying to apply RETURN_PREDICTION heuristics. */ |
3019 | static void |
3020 | apply_return_prediction (void) |
3021 | { |
3022 | greturn *return_stmt = NULL; |
3023 | tree return_val; |
3024 | edge e; |
3025 | gphi *phi; |
3026 | int phi_num_args, i; |
3027 | enum br_predictor pred; |
3028 | enum prediction direction; |
3029 | edge_iterator ei; |
3030 | |
3031 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
3032 | { |
3033 | if (greturn *last = safe_dyn_cast <greturn *> (p: *gsi_last_bb (bb: e->src))) |
3034 | { |
3035 | return_stmt = last; |
3036 | break; |
3037 | } |
3038 | } |
3039 | if (!e) |
3040 | return; |
3041 | return_val = gimple_return_retval (gs: return_stmt); |
3042 | if (!return_val) |
3043 | return; |
3044 | if (TREE_CODE (return_val) != SSA_NAME |
3045 | || !SSA_NAME_DEF_STMT (return_val) |
3046 | || gimple_code (SSA_NAME_DEF_STMT (return_val)) != GIMPLE_PHI) |
3047 | return; |
3048 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (return_val)); |
3049 | phi_num_args = gimple_phi_num_args (gs: phi); |
3050 | pred = return_prediction (PHI_ARG_DEF (phi, 0), prediction: &direction); |
3051 | |
3052 | /* Avoid the case where the function returns -1, 0 and 1 values and |
3053 | nothing else. Those could be qsort etc. comparison functions |
3054 | where the negative return isn't less probable than positive. |
3055 | For this require that the function returns at least -1 or 1 |
3056 | or -1 and a boolean value or comparison result, so that functions |
3057 | returning just -1 and 0 are treated as if -1 represents error value. */ |
3058 | if (INTEGRAL_TYPE_P (TREE_TYPE (return_val)) |
3059 | && !TYPE_UNSIGNED (TREE_TYPE (return_val)) |
3060 | && TYPE_PRECISION (TREE_TYPE (return_val)) > 1) |
3061 | if (int r = zero_one_minusone (phi, limit: 3)) |
3062 | if ((r & (1 | 4)) == (1 | 4)) |
3063 | return; |
3064 | |
3065 | /* Avoid the degenerate case where all return values form the function |
3066 | belongs to same category (ie they are all positive constants) |
3067 | so we can hardly say something about them. */ |
3068 | for (i = 1; i < phi_num_args; i++) |
3069 | if (pred != return_prediction (PHI_ARG_DEF (phi, i), prediction: &direction)) |
3070 | break; |
3071 | if (i != phi_num_args) |
3072 | for (i = 0; i < phi_num_args; i++) |
3073 | { |
3074 | pred = return_prediction (PHI_ARG_DEF (phi, i), prediction: &direction); |
3075 | if (pred != PRED_NO_PREDICTION) |
3076 | predict_paths_leading_to_edge (gimple_phi_arg_edge (phi, i), pred, |
3077 | direction); |
3078 | } |
3079 | } |
3080 | |
3081 | /* Look for basic block that contains unlikely to happen events |
3082 | (such as noreturn calls) and mark all paths leading to execution |
3083 | of this basic blocks as unlikely. */ |
3084 | |
3085 | static void |
3086 | tree_bb_level_predictions (void) |
3087 | { |
3088 | basic_block bb; |
3089 | bool has_return_edges = false; |
3090 | edge e; |
3091 | edge_iterator ei; |
3092 | |
3093 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
3094 | if (!unlikely_executed_edge_p (e) && !(e->flags & EDGE_ABNORMAL_CALL)) |
3095 | { |
3096 | has_return_edges = true; |
3097 | break; |
3098 | } |
3099 | |
3100 | apply_return_prediction (); |
3101 | |
3102 | FOR_EACH_BB_FN (bb, cfun) |
3103 | { |
3104 | gimple_stmt_iterator gsi; |
3105 | |
3106 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
3107 | { |
3108 | gimple *stmt = gsi_stmt (i: gsi); |
3109 | tree decl; |
3110 | |
3111 | if (is_gimple_call (gs: stmt)) |
3112 | { |
3113 | if (gimple_call_noreturn_p (s: stmt) |
3114 | && has_return_edges |
3115 | && !is_exit_with_zero_arg (stmt)) |
3116 | predict_paths_leading_to (bb, PRED_NORETURN, |
3117 | NOT_TAKEN); |
3118 | decl = gimple_call_fndecl (gs: stmt); |
3119 | if (decl |
3120 | && lookup_attribute (attr_name: "cold" , |
3121 | DECL_ATTRIBUTES (decl))) |
3122 | predict_paths_leading_to (bb, PRED_COLD_FUNCTION, |
3123 | NOT_TAKEN); |
3124 | if (decl && recursive_call_p (current_function_decl, decl)) |
3125 | predict_paths_leading_to (bb, PRED_RECURSIVE_CALL, |
3126 | NOT_TAKEN); |
3127 | } |
3128 | else if (gimple_code (g: stmt) == GIMPLE_PREDICT) |
3129 | { |
3130 | predict_paths_leading_to (bb, gimple_predict_predictor (gs: stmt), |
3131 | gimple_predict_outcome (gs: stmt)); |
3132 | /* Keep GIMPLE_PREDICT around so early inlining will propagate |
3133 | hints to callers. */ |
3134 | } |
3135 | } |
3136 | } |
3137 | } |
3138 | |
3139 | /* Callback for hash_map::traverse, asserts that the pointer map is |
3140 | empty. */ |
3141 | |
3142 | bool |
3143 | assert_is_empty (const_basic_block const &, edge_prediction *const &value, |
3144 | void *) |
3145 | { |
3146 | gcc_assert (!value); |
3147 | return true; |
3148 | } |
3149 | |
3150 | /* Predict branch probabilities and estimate profile for basic block BB. |
3151 | When LOCAL_ONLY is set do not use any global properties of CFG. */ |
3152 | |
3153 | static void |
3154 | tree_estimate_probability_bb (basic_block bb, bool local_only) |
3155 | { |
3156 | edge e; |
3157 | edge_iterator ei; |
3158 | |
3159 | FOR_EACH_EDGE (e, ei, bb->succs) |
3160 | { |
3161 | /* Look for block we are guarding (ie we dominate it, |
3162 | but it doesn't postdominate us). */ |
3163 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) && e->dest != bb |
3164 | && !local_only |
3165 | && dominated_by_p (CDI_DOMINATORS, e->dest, e->src) |
3166 | && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e->dest)) |
3167 | { |
3168 | gimple_stmt_iterator bi; |
3169 | |
3170 | /* The call heuristic claims that a guarded function call |
3171 | is improbable. This is because such calls are often used |
3172 | to signal exceptional situations such as printing error |
3173 | messages. */ |
3174 | for (bi = gsi_start_bb (bb: e->dest); !gsi_end_p (i: bi); |
3175 | gsi_next (i: &bi)) |
3176 | { |
3177 | gimple *stmt = gsi_stmt (i: bi); |
3178 | if (is_gimple_call (gs: stmt) |
3179 | && !gimple_inexpensive_call_p (as_a <gcall *> (p: stmt)) |
3180 | /* Constant and pure calls are hardly used to signalize |
3181 | something exceptional. */ |
3182 | && gimple_has_side_effects (stmt)) |
3183 | { |
3184 | if (gimple_call_fndecl (gs: stmt)) |
3185 | predict_edge_def (e, predictor: PRED_CALL, taken: NOT_TAKEN); |
3186 | else if (virtual_method_call_p (gimple_call_fn (gs: stmt))) |
3187 | predict_edge_def (e, predictor: PRED_POLYMORPHIC_CALL, taken: NOT_TAKEN); |
3188 | else |
3189 | predict_edge_def (e, predictor: PRED_INDIR_CALL, taken: TAKEN); |
3190 | break; |
3191 | } |
3192 | } |
3193 | } |
3194 | } |
3195 | tree_predict_by_opcode (bb); |
3196 | } |
3197 | |
3198 | /* Predict branch probabilities and estimate profile of the tree CFG. |
3199 | This function can be called from the loop optimizers to recompute |
3200 | the profile information. |
3201 | If DRY_RUN is set, do not modify CFG and only produce dump files. */ |
3202 | |
3203 | void |
3204 | tree_estimate_probability (bool dry_run) |
3205 | { |
3206 | basic_block bb; |
3207 | |
3208 | connect_infinite_loops_to_exit (); |
3209 | /* We use loop_niter_by_eval, which requires that the loops have |
3210 | preheaders. */ |
3211 | create_preheaders (CP_SIMPLE_PREHEADERS); |
3212 | calculate_dominance_info (CDI_POST_DOMINATORS); |
3213 | /* Decide which edges are known to be unlikely. This improves later |
3214 | branch prediction. */ |
3215 | determine_unlikely_bbs (); |
3216 | |
3217 | bb_predictions = new hash_map<const_basic_block, edge_prediction *>; |
3218 | tree_bb_level_predictions (); |
3219 | record_loop_exits (); |
3220 | |
3221 | if (number_of_loops (cfun) > 1) |
3222 | predict_loops (); |
3223 | |
3224 | FOR_EACH_BB_FN (bb, cfun) |
3225 | tree_estimate_probability_bb (bb, local_only: false); |
3226 | |
3227 | FOR_EACH_BB_FN (bb, cfun) |
3228 | combine_predictions_for_bb (bb, dry_run); |
3229 | |
3230 | if (flag_checking) |
3231 | bb_predictions->traverse<void *, assert_is_empty> (NULL); |
3232 | |
3233 | delete bb_predictions; |
3234 | bb_predictions = NULL; |
3235 | |
3236 | if (!dry_run |
3237 | && profile_status_for_fn (cfun) != PROFILE_READ) |
3238 | estimate_bb_frequencies (); |
3239 | free_dominance_info (CDI_POST_DOMINATORS); |
3240 | remove_fake_exit_edges (); |
3241 | } |
3242 | |
3243 | /* Set edge->probability for each successor edge of BB. */ |
3244 | void |
3245 | tree_guess_outgoing_edge_probabilities (basic_block bb) |
3246 | { |
3247 | bb_predictions = new hash_map<const_basic_block, edge_prediction *>; |
3248 | tree_estimate_probability_bb (bb, local_only: true); |
3249 | combine_predictions_for_bb (bb, dry_run: false); |
3250 | if (flag_checking) |
3251 | bb_predictions->traverse<void *, assert_is_empty> (NULL); |
3252 | delete bb_predictions; |
3253 | bb_predictions = NULL; |
3254 | } |
3255 | |
3256 | /* Filter function predicate that returns true for a edge predicate P |
3257 | if its edge is equal to DATA. */ |
3258 | |
3259 | static bool |
3260 | not_loop_guard_equal_edge_p (edge_prediction *p, void *data) |
3261 | { |
3262 | return p->ep_edge != (edge)data || p->ep_predictor != PRED_LOOP_GUARD; |
3263 | } |
3264 | |
3265 | /* Predict edge E with PRED unless it is already predicted by some predictor |
3266 | considered equivalent. */ |
3267 | |
3268 | static void |
3269 | maybe_predict_edge (edge e, enum br_predictor pred, enum prediction taken) |
3270 | { |
3271 | if (edge_predicted_by_p (e, predictor: pred, taken)) |
3272 | return; |
3273 | if (pred == PRED_LOOP_GUARD |
3274 | && edge_predicted_by_p (e, predictor: PRED_LOOP_GUARD_WITH_RECURSION, taken)) |
3275 | return; |
3276 | /* Consider PRED_LOOP_GUARD_WITH_RECURSION superrior to LOOP_GUARD. */ |
3277 | if (pred == PRED_LOOP_GUARD_WITH_RECURSION) |
3278 | { |
3279 | edge_prediction **preds = bb_predictions->get (k: e->src); |
3280 | if (preds) |
3281 | filter_predictions (preds, filter: not_loop_guard_equal_edge_p, data: e); |
3282 | } |
3283 | predict_edge_def (e, predictor: pred, taken); |
3284 | } |
3285 | /* Predict edges to successors of CUR whose sources are not postdominated by |
3286 | BB by PRED and recurse to all postdominators. */ |
3287 | |
3288 | static void |
3289 | predict_paths_for_bb (basic_block cur, basic_block bb, |
3290 | enum br_predictor pred, |
3291 | enum prediction taken, |
3292 | bitmap visited, class loop *in_loop = NULL) |
3293 | { |
3294 | edge e; |
3295 | edge_iterator ei; |
3296 | basic_block son; |
3297 | |
3298 | /* If we exited the loop or CUR is unconditional in the loop, there is |
3299 | nothing to do. */ |
3300 | if (in_loop |
3301 | && (!flow_bb_inside_loop_p (in_loop, cur) |
3302 | || dominated_by_p (CDI_DOMINATORS, in_loop->latch, cur))) |
3303 | return; |
3304 | |
3305 | /* We are looking for all edges forming edge cut induced by |
3306 | set of all blocks postdominated by BB. */ |
3307 | FOR_EACH_EDGE (e, ei, cur->preds) |
3308 | if (e->src->index >= NUM_FIXED_BLOCKS |
3309 | && !dominated_by_p (CDI_POST_DOMINATORS, e->src, bb)) |
3310 | { |
3311 | edge e2; |
3312 | edge_iterator ei2; |
3313 | bool found = false; |
3314 | |
3315 | /* Ignore fake edges and eh, we predict them as not taken anyway. */ |
3316 | if (unlikely_executed_edge_p (e)) |
3317 | continue; |
3318 | gcc_assert (bb == cur || dominated_by_p (CDI_POST_DOMINATORS, cur, bb)); |
3319 | |
3320 | /* See if there is an edge from e->src that is not abnormal |
3321 | and does not lead to BB and does not exit the loop. */ |
3322 | FOR_EACH_EDGE (e2, ei2, e->src->succs) |
3323 | if (e2 != e |
3324 | && !unlikely_executed_edge_p (e: e2) |
3325 | && !dominated_by_p (CDI_POST_DOMINATORS, e2->dest, bb) |
3326 | && (!in_loop || !loop_exit_edge_p (in_loop, e2))) |
3327 | { |
3328 | found = true; |
3329 | break; |
3330 | } |
3331 | |
3332 | /* If there is non-abnormal path leaving e->src, predict edge |
3333 | using predictor. Otherwise we need to look for paths |
3334 | leading to e->src. |
3335 | |
3336 | The second may lead to infinite loop in the case we are predicitng |
3337 | regions that are only reachable by abnormal edges. We simply |
3338 | prevent visiting given BB twice. */ |
3339 | if (found) |
3340 | maybe_predict_edge (e, pred, taken); |
3341 | else if (bitmap_set_bit (visited, e->src->index)) |
3342 | predict_paths_for_bb (cur: e->src, bb: e->src, pred, taken, visited, in_loop); |
3343 | } |
3344 | for (son = first_dom_son (CDI_POST_DOMINATORS, cur); |
3345 | son; |
3346 | son = next_dom_son (CDI_POST_DOMINATORS, son)) |
3347 | predict_paths_for_bb (cur: son, bb, pred, taken, visited, in_loop); |
3348 | } |
3349 | |
3350 | /* Sets branch probabilities according to PREDiction and |
3351 | FLAGS. */ |
3352 | |
3353 | static void |
3354 | predict_paths_leading_to (basic_block bb, enum br_predictor pred, |
3355 | enum prediction taken, class loop *in_loop) |
3356 | { |
3357 | predict_paths_for_bb (cur: bb, bb, pred, taken, visited: auto_bitmap (), in_loop); |
3358 | } |
3359 | |
3360 | /* Like predict_paths_leading_to but take edge instead of basic block. */ |
3361 | |
3362 | static void |
3363 | predict_paths_leading_to_edge (edge e, enum br_predictor pred, |
3364 | enum prediction taken, class loop *in_loop) |
3365 | { |
3366 | bool has_nonloop_edge = false; |
3367 | edge_iterator ei; |
3368 | edge e2; |
3369 | |
3370 | basic_block bb = e->src; |
3371 | FOR_EACH_EDGE (e2, ei, bb->succs) |
3372 | if (e2->dest != e->src && e2->dest != e->dest |
3373 | && !unlikely_executed_edge_p (e: e2) |
3374 | && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e2->dest)) |
3375 | { |
3376 | has_nonloop_edge = true; |
3377 | break; |
3378 | } |
3379 | |
3380 | if (!has_nonloop_edge) |
3381 | predict_paths_for_bb (cur: bb, bb, pred, taken, visited: auto_bitmap (), in_loop); |
3382 | else |
3383 | maybe_predict_edge (e, pred, taken); |
3384 | } |
3385 | |
3386 | /* This is used to carry information about basic blocks. It is |
3387 | attached to the AUX field of the standard CFG block. */ |
3388 | |
3389 | class block_info |
3390 | { |
3391 | public: |
3392 | /* Estimated frequency of execution of basic_block. */ |
3393 | sreal frequency; |
3394 | |
3395 | /* To keep queue of basic blocks to process. */ |
3396 | basic_block next; |
3397 | |
3398 | /* Number of predecessors we need to visit first. */ |
3399 | int npredecessors; |
3400 | }; |
3401 | |
3402 | /* Similar information for edges. */ |
3403 | class edge_prob_info |
3404 | { |
3405 | public: |
3406 | /* In case edge is a loopback edge, the probability edge will be reached |
3407 | in case header is. Estimated number of iterations of the loop can be |
3408 | then computed as 1 / (1 - back_edge_prob). */ |
3409 | sreal back_edge_prob; |
3410 | /* True if the edge is a loopback edge in the natural loop. */ |
3411 | unsigned int back_edge:1; |
3412 | }; |
3413 | |
3414 | #define BLOCK_INFO(B) ((block_info *) (B)->aux) |
3415 | #undef EDGE_INFO |
3416 | #define EDGE_INFO(E) ((edge_prob_info *) (E)->aux) |
3417 | |
3418 | /* Helper function for estimate_bb_frequencies. |
3419 | Propagate the frequencies in blocks marked in |
3420 | TOVISIT, starting in HEAD. */ |
3421 | |
3422 | static void |
3423 | propagate_freq (basic_block head, bitmap tovisit, |
3424 | sreal max_cyclic_prob) |
3425 | { |
3426 | basic_block bb; |
3427 | basic_block last; |
3428 | unsigned i; |
3429 | edge e; |
3430 | basic_block nextbb; |
3431 | bitmap_iterator bi; |
3432 | |
3433 | /* For each basic block we need to visit count number of his predecessors |
3434 | we need to visit first. */ |
3435 | EXECUTE_IF_SET_IN_BITMAP (tovisit, 0, i, bi) |
3436 | { |
3437 | edge_iterator ei; |
3438 | int count = 0; |
3439 | |
3440 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
3441 | |
3442 | FOR_EACH_EDGE (e, ei, bb->preds) |
3443 | { |
3444 | bool visit = bitmap_bit_p (tovisit, e->src->index); |
3445 | |
3446 | if (visit && !(e->flags & EDGE_DFS_BACK)) |
3447 | count++; |
3448 | else if (visit && dump_file && !EDGE_INFO (e)->back_edge) |
3449 | fprintf (stream: dump_file, |
3450 | format: "Irreducible region hit, ignoring edge to %i->%i\n" , |
3451 | e->src->index, bb->index); |
3452 | } |
3453 | BLOCK_INFO (bb)->npredecessors = count; |
3454 | /* When function never returns, we will never process exit block. */ |
3455 | if (!count && bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
3456 | bb->count = profile_count::zero (); |
3457 | } |
3458 | |
3459 | BLOCK_INFO (head)->frequency = 1; |
3460 | last = head; |
3461 | for (bb = head; bb; bb = nextbb) |
3462 | { |
3463 | edge_iterator ei; |
3464 | sreal cyclic_probability = 0; |
3465 | sreal frequency = 0; |
3466 | |
3467 | nextbb = BLOCK_INFO (bb)->next; |
3468 | BLOCK_INFO (bb)->next = NULL; |
3469 | |
3470 | /* Compute frequency of basic block. */ |
3471 | if (bb != head) |
3472 | { |
3473 | if (flag_checking) |
3474 | FOR_EACH_EDGE (e, ei, bb->preds) |
3475 | gcc_assert (!bitmap_bit_p (tovisit, e->src->index) |
3476 | || (e->flags & EDGE_DFS_BACK)); |
3477 | |
3478 | FOR_EACH_EDGE (e, ei, bb->preds) |
3479 | if (EDGE_INFO (e)->back_edge) |
3480 | cyclic_probability += EDGE_INFO (e)->back_edge_prob; |
3481 | else if (!(e->flags & EDGE_DFS_BACK)) |
3482 | { |
3483 | /* FIXME: Graphite is producing edges with no profile. Once |
3484 | this is fixed, drop this. */ |
3485 | sreal tmp = e->probability.initialized_p () ? |
3486 | e->probability.to_sreal () : 0; |
3487 | frequency += tmp * BLOCK_INFO (e->src)->frequency; |
3488 | } |
3489 | |
3490 | if (cyclic_probability == 0) |
3491 | { |
3492 | BLOCK_INFO (bb)->frequency = frequency; |
3493 | } |
3494 | else |
3495 | { |
3496 | if (cyclic_probability > max_cyclic_prob) |
3497 | { |
3498 | if (dump_file) |
3499 | fprintf (stream: dump_file, |
3500 | format: "cyclic probability of bb %i is %f (capped to %f)" |
3501 | "; turning freq %f" , |
3502 | bb->index, cyclic_probability.to_double (), |
3503 | max_cyclic_prob.to_double (), |
3504 | frequency.to_double ()); |
3505 | |
3506 | cyclic_probability = max_cyclic_prob; |
3507 | } |
3508 | else if (dump_file) |
3509 | fprintf (stream: dump_file, |
3510 | format: "cyclic probability of bb %i is %f; turning freq %f" , |
3511 | bb->index, cyclic_probability.to_double (), |
3512 | frequency.to_double ()); |
3513 | |
3514 | BLOCK_INFO (bb)->frequency = frequency |
3515 | / (sreal (1) - cyclic_probability); |
3516 | if (dump_file) |
3517 | fprintf (stream: dump_file, format: " to %f\n" , |
3518 | BLOCK_INFO (bb)->frequency.to_double ()); |
3519 | } |
3520 | } |
3521 | |
3522 | bitmap_clear_bit (tovisit, bb->index); |
3523 | |
3524 | e = find_edge (bb, head); |
3525 | if (e) |
3526 | { |
3527 | /* FIXME: Graphite is producing edges with no profile. Once |
3528 | this is fixed, drop this. */ |
3529 | sreal tmp = e->probability.initialized_p () ? |
3530 | e->probability.to_sreal () : 0; |
3531 | EDGE_INFO (e)->back_edge_prob = tmp * BLOCK_INFO (bb)->frequency; |
3532 | } |
3533 | |
3534 | /* Propagate to successor blocks. */ |
3535 | FOR_EACH_EDGE (e, ei, bb->succs) |
3536 | if (!(e->flags & EDGE_DFS_BACK) |
3537 | && BLOCK_INFO (e->dest)->npredecessors) |
3538 | { |
3539 | BLOCK_INFO (e->dest)->npredecessors--; |
3540 | if (!BLOCK_INFO (e->dest)->npredecessors) |
3541 | { |
3542 | if (!nextbb) |
3543 | nextbb = e->dest; |
3544 | else |
3545 | BLOCK_INFO (last)->next = e->dest; |
3546 | |
3547 | last = e->dest; |
3548 | } |
3549 | } |
3550 | } |
3551 | } |
3552 | |
3553 | /* Estimate frequencies in loops at same nest level. */ |
3554 | |
3555 | static void |
3556 | estimate_loops_at_level (class loop *first_loop, sreal max_cyclic_prob) |
3557 | { |
3558 | class loop *loop; |
3559 | |
3560 | for (loop = first_loop; loop; loop = loop->next) |
3561 | { |
3562 | edge e; |
3563 | basic_block *bbs; |
3564 | unsigned i; |
3565 | auto_bitmap tovisit; |
3566 | |
3567 | estimate_loops_at_level (first_loop: loop->inner, max_cyclic_prob); |
3568 | |
3569 | /* Find current loop back edge and mark it. */ |
3570 | e = loop_latch_edge (loop); |
3571 | EDGE_INFO (e)->back_edge = 1; |
3572 | |
3573 | bbs = get_loop_body (loop); |
3574 | for (i = 0; i < loop->num_nodes; i++) |
3575 | bitmap_set_bit (tovisit, bbs[i]->index); |
3576 | free (ptr: bbs); |
3577 | propagate_freq (head: loop->header, tovisit, max_cyclic_prob); |
3578 | } |
3579 | } |
3580 | |
3581 | /* Propagates frequencies through structure of loops. */ |
3582 | |
3583 | static void |
3584 | estimate_loops (void) |
3585 | { |
3586 | auto_bitmap tovisit; |
3587 | basic_block bb; |
3588 | sreal max_cyclic_prob = (sreal)1 |
3589 | - (sreal)1 / (param_max_predicted_iterations + 1); |
3590 | |
3591 | /* Start by estimating the frequencies in the loops. */ |
3592 | if (number_of_loops (cfun) > 1) |
3593 | estimate_loops_at_level (current_loops->tree_root->inner, max_cyclic_prob); |
3594 | |
3595 | /* Now propagate the frequencies through all the blocks. */ |
3596 | FOR_ALL_BB_FN (bb, cfun) |
3597 | { |
3598 | bitmap_set_bit (tovisit, bb->index); |
3599 | } |
3600 | propagate_freq (ENTRY_BLOCK_PTR_FOR_FN (cfun), tovisit, max_cyclic_prob); |
3601 | } |
3602 | |
3603 | /* Drop the profile for NODE to guessed, and update its frequency based on |
3604 | whether it is expected to be hot given the CALL_COUNT. */ |
3605 | |
3606 | static void |
3607 | drop_profile (struct cgraph_node *node, profile_count call_count) |
3608 | { |
3609 | struct function *fn = DECL_STRUCT_FUNCTION (node->decl); |
3610 | /* In the case where this was called by another function with a |
3611 | dropped profile, call_count will be 0. Since there are no |
3612 | non-zero call counts to this function, we don't know for sure |
3613 | whether it is hot, and therefore it will be marked normal below. */ |
3614 | bool hot = maybe_hot_count_p (NULL, count: call_count); |
3615 | |
3616 | if (dump_file) |
3617 | fprintf (stream: dump_file, |
3618 | format: "Dropping 0 profile for %s. %s based on calls.\n" , |
3619 | node->dump_name (), |
3620 | hot ? "Function is hot" : "Function is normal" ); |
3621 | /* We only expect to miss profiles for functions that are reached |
3622 | via non-zero call edges in cases where the function may have |
3623 | been linked from another module or library (COMDATs and extern |
3624 | templates). See the comments below for handle_missing_profiles. |
3625 | Also, only warn in cases where the missing counts exceed the |
3626 | number of training runs. In certain cases with an execv followed |
3627 | by a no-return call the profile for the no-return call is not |
3628 | dumped and there can be a mismatch. */ |
3629 | if (!DECL_COMDAT (node->decl) && !DECL_EXTERNAL (node->decl) |
3630 | && call_count > profile_info->runs) |
3631 | { |
3632 | if (flag_profile_correction) |
3633 | { |
3634 | if (dump_file) |
3635 | fprintf (stream: dump_file, |
3636 | format: "Missing counts for called function %s\n" , |
3637 | node->dump_name ()); |
3638 | } |
3639 | else |
3640 | warning (0, "Missing counts for called function %s" , |
3641 | node->dump_name ()); |
3642 | } |
3643 | |
3644 | basic_block bb; |
3645 | if (opt_for_fn (node->decl, flag_guess_branch_prob)) |
3646 | { |
3647 | bool clear_zeros |
3648 | = !ENTRY_BLOCK_PTR_FOR_FN (fn)->count.nonzero_p (); |
3649 | FOR_ALL_BB_FN (bb, fn) |
3650 | if (clear_zeros || !(bb->count == profile_count::zero ())) |
3651 | bb->count = bb->count.guessed_local (); |
3652 | fn->cfg->count_max = fn->cfg->count_max.guessed_local (); |
3653 | } |
3654 | else |
3655 | { |
3656 | FOR_ALL_BB_FN (bb, fn) |
3657 | bb->count = profile_count::uninitialized (); |
3658 | fn->cfg->count_max = profile_count::uninitialized (); |
3659 | } |
3660 | |
3661 | struct cgraph_edge *e; |
3662 | for (e = node->callees; e; e = e->next_callee) |
3663 | e->count = gimple_bb (g: e->call_stmt)->count; |
3664 | for (e = node->indirect_calls; e; e = e->next_callee) |
3665 | e->count = gimple_bb (g: e->call_stmt)->count; |
3666 | node->count = ENTRY_BLOCK_PTR_FOR_FN (fn)->count; |
3667 | |
3668 | profile_status_for_fn (fn) |
3669 | = (flag_guess_branch_prob ? PROFILE_GUESSED : PROFILE_ABSENT); |
3670 | node->frequency |
3671 | = hot ? NODE_FREQUENCY_HOT : NODE_FREQUENCY_NORMAL; |
3672 | } |
3673 | |
3674 | /* In the case of COMDAT routines, multiple object files will contain the same |
3675 | function and the linker will select one for the binary. In that case |
3676 | all the other copies from the profile instrument binary will be missing |
3677 | profile counts. Look for cases where this happened, due to non-zero |
3678 | call counts going to 0-count functions, and drop the profile to guessed |
3679 | so that we can use the estimated probabilities and avoid optimizing only |
3680 | for size. |
3681 | |
3682 | The other case where the profile may be missing is when the routine |
3683 | is not going to be emitted to the object file, e.g. for "extern template" |
3684 | class methods. Those will be marked DECL_EXTERNAL. Emit a warning in |
3685 | all other cases of non-zero calls to 0-count functions. */ |
3686 | |
3687 | void |
3688 | handle_missing_profiles (void) |
3689 | { |
3690 | const int unlikely_frac = param_unlikely_bb_count_fraction; |
3691 | struct cgraph_node *node; |
3692 | auto_vec<struct cgraph_node *, 64> worklist; |
3693 | |
3694 | /* See if 0 count function has non-0 count callers. In this case we |
3695 | lost some profile. Drop its function profile to PROFILE_GUESSED. */ |
3696 | FOR_EACH_DEFINED_FUNCTION (node) |
3697 | { |
3698 | struct cgraph_edge *e; |
3699 | profile_count call_count = profile_count::zero (); |
3700 | gcov_type max_tp_first_run = 0; |
3701 | struct function *fn = DECL_STRUCT_FUNCTION (node->decl); |
3702 | |
3703 | if (node->count.ipa ().nonzero_p ()) |
3704 | continue; |
3705 | for (e = node->callers; e; e = e->next_caller) |
3706 | if (e->count.ipa ().initialized_p () && e->count.ipa () > 0) |
3707 | { |
3708 | call_count = call_count + e->count.ipa (); |
3709 | |
3710 | if (e->caller->tp_first_run > max_tp_first_run) |
3711 | max_tp_first_run = e->caller->tp_first_run; |
3712 | } |
3713 | |
3714 | /* If time profile is missing, let assign the maximum that comes from |
3715 | caller functions. */ |
3716 | if (!node->tp_first_run && max_tp_first_run) |
3717 | node->tp_first_run = max_tp_first_run + 1; |
3718 | |
3719 | if (call_count > 0 |
3720 | && fn && fn->cfg |
3721 | && call_count * unlikely_frac >= profile_info->runs) |
3722 | { |
3723 | drop_profile (node, call_count); |
3724 | worklist.safe_push (obj: node); |
3725 | } |
3726 | } |
3727 | |
3728 | /* Propagate the profile dropping to other 0-count COMDATs that are |
3729 | potentially called by COMDATs we already dropped the profile on. */ |
3730 | while (worklist.length () > 0) |
3731 | { |
3732 | struct cgraph_edge *e; |
3733 | |
3734 | node = worklist.pop (); |
3735 | for (e = node->callees; e; e = e->next_caller) |
3736 | { |
3737 | struct cgraph_node *callee = e->callee; |
3738 | struct function *fn = DECL_STRUCT_FUNCTION (callee->decl); |
3739 | |
3740 | if (!(e->count.ipa () == profile_count::zero ()) |
3741 | && callee->count.ipa ().nonzero_p ()) |
3742 | continue; |
3743 | if ((DECL_COMDAT (callee->decl) || DECL_EXTERNAL (callee->decl)) |
3744 | && fn && fn->cfg |
3745 | && profile_status_for_fn (fn) == PROFILE_READ) |
3746 | { |
3747 | drop_profile (node, call_count: profile_count::zero ()); |
3748 | worklist.safe_push (obj: callee); |
3749 | } |
3750 | } |
3751 | } |
3752 | } |
3753 | |
3754 | /* Convert counts measured by profile driven feedback to frequencies. |
3755 | Return nonzero iff there was any nonzero execution count. */ |
3756 | |
3757 | bool |
3758 | update_max_bb_count (void) |
3759 | { |
3760 | profile_count true_count_max = profile_count::uninitialized (); |
3761 | basic_block bb; |
3762 | |
3763 | FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb) |
3764 | true_count_max = true_count_max.max (other: bb->count); |
3765 | |
3766 | cfun->cfg->count_max = true_count_max; |
3767 | |
3768 | return true_count_max.ipa ().nonzero_p (); |
3769 | } |
3770 | |
3771 | /* Return true if function is likely to be expensive, so there is no point to |
3772 | optimize performance of prologue, epilogue or do inlining at the expense |
3773 | of code size growth. THRESHOLD is the limit of number of instructions |
3774 | function can execute at average to be still considered not expensive. */ |
3775 | |
3776 | bool |
3777 | expensive_function_p (int threshold) |
3778 | { |
3779 | basic_block bb; |
3780 | |
3781 | /* If profile was scaled in a way entry block has count 0, then the function |
3782 | is deifnitly taking a lot of time. */ |
3783 | if (!ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.nonzero_p ()) |
3784 | return true; |
3785 | |
3786 | profile_count limit = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count * threshold; |
3787 | profile_count sum = profile_count::zero (); |
3788 | FOR_EACH_BB_FN (bb, cfun) |
3789 | { |
3790 | rtx_insn *insn; |
3791 | |
3792 | if (!bb->count.initialized_p ()) |
3793 | { |
3794 | if (dump_file) |
3795 | fprintf (stream: dump_file, format: "Function is considered expensive because" |
3796 | " count of bb %i is not initialized\n" , bb->index); |
3797 | return true; |
3798 | } |
3799 | |
3800 | FOR_BB_INSNS (bb, insn) |
3801 | if (active_insn_p (insn)) |
3802 | { |
3803 | sum += bb->count; |
3804 | if (sum > limit) |
3805 | return true; |
3806 | } |
3807 | } |
3808 | |
3809 | return false; |
3810 | } |
3811 | |
3812 | /* All basic blocks that are reachable only from unlikely basic blocks are |
3813 | unlikely. */ |
3814 | |
3815 | void |
3816 | propagate_unlikely_bbs_forward (void) |
3817 | { |
3818 | auto_vec<basic_block, 64> worklist; |
3819 | basic_block bb; |
3820 | edge_iterator ei; |
3821 | edge e; |
3822 | |
3823 | if (!(ENTRY_BLOCK_PTR_FOR_FN (cfun)->count == profile_count::zero ())) |
3824 | { |
3825 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux = (void *)(size_t) 1; |
3826 | worklist.safe_push (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
3827 | |
3828 | while (worklist.length () > 0) |
3829 | { |
3830 | bb = worklist.pop (); |
3831 | FOR_EACH_EDGE (e, ei, bb->succs) |
3832 | if (!(e->count () == profile_count::zero ()) |
3833 | && !(e->dest->count == profile_count::zero ()) |
3834 | && !e->dest->aux) |
3835 | { |
3836 | e->dest->aux = (void *)(size_t) 1; |
3837 | worklist.safe_push (obj: e->dest); |
3838 | } |
3839 | } |
3840 | } |
3841 | |
3842 | FOR_ALL_BB_FN (bb, cfun) |
3843 | { |
3844 | if (!bb->aux) |
3845 | { |
3846 | if (!(bb->count == profile_count::zero ()) |
3847 | && (dump_file && (dump_flags & TDF_DETAILS))) |
3848 | fprintf (stream: dump_file, |
3849 | format: "Basic block %i is marked unlikely by forward prop\n" , |
3850 | bb->index); |
3851 | bb->count = profile_count::zero (); |
3852 | } |
3853 | else |
3854 | bb->aux = NULL; |
3855 | } |
3856 | } |
3857 | |
3858 | /* Determine basic blocks/edges that are known to be unlikely executed and set |
3859 | their counters to zero. |
3860 | This is done with first identifying obviously unlikely BBs/edges and then |
3861 | propagating in both directions. */ |
3862 | |
3863 | static void |
3864 | determine_unlikely_bbs () |
3865 | { |
3866 | basic_block bb; |
3867 | auto_vec<basic_block, 64> worklist; |
3868 | edge_iterator ei; |
3869 | edge e; |
3870 | |
3871 | FOR_EACH_BB_FN (bb, cfun) |
3872 | { |
3873 | if (!(bb->count == profile_count::zero ()) |
3874 | && unlikely_executed_bb_p (bb)) |
3875 | { |
3876 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3877 | fprintf (stream: dump_file, format: "Basic block %i is locally unlikely\n" , |
3878 | bb->index); |
3879 | bb->count = profile_count::zero (); |
3880 | } |
3881 | |
3882 | FOR_EACH_EDGE (e, ei, bb->succs) |
3883 | if (!(e->probability == profile_probability::never ()) |
3884 | && unlikely_executed_edge_p (e)) |
3885 | { |
3886 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3887 | fprintf (stream: dump_file, format: "Edge %i->%i is locally unlikely\n" , |
3888 | bb->index, e->dest->index); |
3889 | e->probability = profile_probability::never (); |
3890 | } |
3891 | |
3892 | gcc_checking_assert (!bb->aux); |
3893 | } |
3894 | propagate_unlikely_bbs_forward (); |
3895 | |
3896 | auto_vec<int, 64> nsuccs; |
3897 | nsuccs.safe_grow_cleared (last_basic_block_for_fn (cfun), exact: true); |
3898 | FOR_ALL_BB_FN (bb, cfun) |
3899 | if (!(bb->count == profile_count::zero ()) |
3900 | && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
3901 | { |
3902 | nsuccs[bb->index] = 0; |
3903 | FOR_EACH_EDGE (e, ei, bb->succs) |
3904 | if (!(e->probability == profile_probability::never ()) |
3905 | && !(e->dest->count == profile_count::zero ())) |
3906 | nsuccs[bb->index]++; |
3907 | if (!nsuccs[bb->index]) |
3908 | worklist.safe_push (obj: bb); |
3909 | } |
3910 | while (worklist.length () > 0) |
3911 | { |
3912 | bb = worklist.pop (); |
3913 | if (bb->count == profile_count::zero ()) |
3914 | continue; |
3915 | if (bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
3916 | { |
3917 | bool found = false; |
3918 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); |
3919 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
3920 | if (stmt_can_terminate_bb_p (gsi_stmt (i: gsi)) |
3921 | /* stmt_can_terminate_bb_p special cases noreturns because it |
3922 | assumes that fake edges are created. We want to know that |
3923 | noreturn alone does not imply BB to be unlikely. */ |
3924 | || (is_gimple_call (gs: gsi_stmt (i: gsi)) |
3925 | && (gimple_call_flags (gsi_stmt (i: gsi)) & ECF_NORETURN))) |
3926 | { |
3927 | found = true; |
3928 | break; |
3929 | } |
3930 | if (found) |
3931 | continue; |
3932 | } |
3933 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3934 | fprintf (stream: dump_file, |
3935 | format: "Basic block %i is marked unlikely by backward prop\n" , |
3936 | bb->index); |
3937 | bb->count = profile_count::zero (); |
3938 | FOR_EACH_EDGE (e, ei, bb->preds) |
3939 | if (!(e->probability == profile_probability::never ())) |
3940 | { |
3941 | if (!(e->src->count == profile_count::zero ())) |
3942 | { |
3943 | gcc_checking_assert (nsuccs[e->src->index] > 0); |
3944 | nsuccs[e->src->index]--; |
3945 | if (!nsuccs[e->src->index]) |
3946 | worklist.safe_push (obj: e->src); |
3947 | } |
3948 | } |
3949 | } |
3950 | /* Finally all edges from non-0 regions to 0 are unlikely. */ |
3951 | FOR_ALL_BB_FN (bb, cfun) |
3952 | { |
3953 | if (!(bb->count == profile_count::zero ())) |
3954 | FOR_EACH_EDGE (e, ei, bb->succs) |
3955 | if (!(e->probability == profile_probability::never ()) |
3956 | && e->dest->count == profile_count::zero ()) |
3957 | { |
3958 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3959 | fprintf (stream: dump_file, format: "Edge %i->%i is unlikely because " |
3960 | "it enters unlikely block\n" , |
3961 | bb->index, e->dest->index); |
3962 | e->probability = profile_probability::never (); |
3963 | } |
3964 | |
3965 | edge other = NULL; |
3966 | |
3967 | FOR_EACH_EDGE (e, ei, bb->succs) |
3968 | if (e->probability == profile_probability::never ()) |
3969 | ; |
3970 | else if (other) |
3971 | { |
3972 | other = NULL; |
3973 | break; |
3974 | } |
3975 | else |
3976 | other = e; |
3977 | if (other |
3978 | && !(other->probability == profile_probability::always ())) |
3979 | { |
3980 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3981 | fprintf (stream: dump_file, format: "Edge %i->%i is locally likely\n" , |
3982 | bb->index, other->dest->index); |
3983 | other->probability = profile_probability::always (); |
3984 | } |
3985 | } |
3986 | if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count == profile_count::zero ()) |
3987 | cgraph_node::get (decl: current_function_decl)->count = profile_count::zero (); |
3988 | } |
3989 | |
3990 | /* Estimate and propagate basic block frequencies using the given branch |
3991 | probabilities. */ |
3992 | |
3993 | static void |
3994 | estimate_bb_frequencies () |
3995 | { |
3996 | basic_block bb; |
3997 | sreal freq_max; |
3998 | |
3999 | determine_unlikely_bbs (); |
4000 | |
4001 | mark_dfs_back_edges (); |
4002 | |
4003 | single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->probability = |
4004 | profile_probability::always (); |
4005 | |
4006 | /* Set up block info for each basic block. */ |
4007 | alloc_aux_for_blocks (sizeof (block_info)); |
4008 | alloc_aux_for_edges (sizeof (edge_prob_info)); |
4009 | FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb) |
4010 | { |
4011 | edge e; |
4012 | edge_iterator ei; |
4013 | |
4014 | FOR_EACH_EDGE (e, ei, bb->succs) |
4015 | { |
4016 | /* FIXME: Graphite is producing edges with no profile. Once |
4017 | this is fixed, drop this. */ |
4018 | if (e->probability.initialized_p ()) |
4019 | EDGE_INFO (e)->back_edge_prob |
4020 | = e->probability.to_sreal (); |
4021 | else |
4022 | /* back_edge_prob = 0.5 */ |
4023 | EDGE_INFO (e)->back_edge_prob = sreal (1, -1); |
4024 | } |
4025 | } |
4026 | |
4027 | /* First compute frequencies locally for each loop from innermost |
4028 | to outermost to examine frequencies for back edges. */ |
4029 | estimate_loops (); |
4030 | |
4031 | freq_max = 0; |
4032 | FOR_EACH_BB_FN (bb, cfun) |
4033 | if (freq_max < BLOCK_INFO (bb)->frequency) |
4034 | freq_max = BLOCK_INFO (bb)->frequency; |
4035 | |
4036 | /* Scaling frequencies up to maximal profile count may result in |
4037 | frequent overflows especially when inlining loops. |
4038 | Small scaling results in unnecesary precision loss. Stay in |
4039 | the half of the (exponential) range. */ |
4040 | freq_max = (sreal (1) << (profile_count::n_bits / 2)) / freq_max; |
4041 | if (freq_max < 16) |
4042 | freq_max = 16; |
4043 | profile_count ipa_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.ipa (); |
4044 | cfun->cfg->count_max = profile_count::uninitialized (); |
4045 | FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb) |
4046 | { |
4047 | sreal tmp = BLOCK_INFO (bb)->frequency; |
4048 | if (tmp >= 1) |
4049 | { |
4050 | gimple_stmt_iterator gsi; |
4051 | tree decl; |
4052 | |
4053 | /* Self recursive calls can not have frequency greater than 1 |
4054 | or program will never terminate. This will result in an |
4055 | inconsistent bb profile but it is better than greatly confusing |
4056 | IPA cost metrics. */ |
4057 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4058 | if (is_gimple_call (gs: gsi_stmt (i: gsi)) |
4059 | && (decl = gimple_call_fndecl (gs: gsi_stmt (i: gsi))) != NULL |
4060 | && recursive_call_p (current_function_decl, decl)) |
4061 | { |
4062 | if (dump_file) |
4063 | fprintf (stream: dump_file, format: "Dropping frequency of recursive call" |
4064 | " in bb %i from %f\n" , bb->index, |
4065 | tmp.to_double ()); |
4066 | tmp = (sreal)9 / (sreal)10; |
4067 | break; |
4068 | } |
4069 | } |
4070 | tmp = tmp * freq_max; |
4071 | profile_count count = profile_count::from_gcov_type (v: tmp.to_nearest_int ()); |
4072 | |
4073 | /* If we have profile feedback in which this function was never |
4074 | executed, then preserve this info. */ |
4075 | if (!(bb->count == profile_count::zero ())) |
4076 | bb->count = count.guessed_local ().combine_with_ipa_count (ipa: ipa_count); |
4077 | cfun->cfg->count_max = cfun->cfg->count_max.max (other: bb->count); |
4078 | } |
4079 | |
4080 | free_aux_for_blocks (); |
4081 | free_aux_for_edges (); |
4082 | compute_function_frequency (); |
4083 | } |
4084 | |
4085 | /* Decide whether function is hot, cold or unlikely executed. */ |
4086 | void |
4087 | compute_function_frequency (void) |
4088 | { |
4089 | basic_block bb; |
4090 | struct cgraph_node *node = cgraph_node::get (decl: current_function_decl); |
4091 | |
4092 | if (DECL_STATIC_CONSTRUCTOR (current_function_decl) |
4093 | || MAIN_NAME_P (DECL_NAME (current_function_decl))) |
4094 | node->only_called_at_startup = true; |
4095 | if (DECL_STATIC_DESTRUCTOR (current_function_decl)) |
4096 | node->only_called_at_exit = true; |
4097 | |
4098 | if (!ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.ipa_p ()) |
4099 | { |
4100 | int flags = flags_from_decl_or_type (current_function_decl); |
4101 | if (lookup_attribute (attr_name: "cold" , DECL_ATTRIBUTES (current_function_decl)) |
4102 | != NULL) |
4103 | node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED; |
4104 | else if (lookup_attribute (attr_name: "hot" , DECL_ATTRIBUTES (current_function_decl)) |
4105 | != NULL) |
4106 | node->frequency = NODE_FREQUENCY_HOT; |
4107 | else if (flags & ECF_NORETURN) |
4108 | node->frequency = NODE_FREQUENCY_EXECUTED_ONCE; |
4109 | else if (MAIN_NAME_P (DECL_NAME (current_function_decl))) |
4110 | node->frequency = NODE_FREQUENCY_EXECUTED_ONCE; |
4111 | else if (DECL_STATIC_CONSTRUCTOR (current_function_decl) |
4112 | || DECL_STATIC_DESTRUCTOR (current_function_decl)) |
4113 | node->frequency = NODE_FREQUENCY_EXECUTED_ONCE; |
4114 | return; |
4115 | } |
4116 | |
4117 | node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED; |
4118 | if (lookup_attribute (attr_name: "cold" , DECL_ATTRIBUTES (current_function_decl)) |
4119 | == NULL) |
4120 | warn_function_cold (current_function_decl); |
4121 | if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.ipa() == profile_count::zero ()) |
4122 | return; |
4123 | FOR_EACH_BB_FN (bb, cfun) |
4124 | { |
4125 | if (maybe_hot_bb_p (cfun, bb)) |
4126 | { |
4127 | node->frequency = NODE_FREQUENCY_HOT; |
4128 | return; |
4129 | } |
4130 | if (!probably_never_executed_bb_p (cfun, bb)) |
4131 | node->frequency = NODE_FREQUENCY_NORMAL; |
4132 | } |
4133 | } |
4134 | |
4135 | /* Build PREDICT_EXPR. */ |
4136 | tree |
4137 | build_predict_expr (enum br_predictor predictor, enum prediction taken) |
4138 | { |
4139 | tree t = build1 (PREDICT_EXPR, void_type_node, |
4140 | build_int_cst (integer_type_node, predictor)); |
4141 | SET_PREDICT_EXPR_OUTCOME (t, taken); |
4142 | return t; |
4143 | } |
4144 | |
4145 | const char * |
4146 | predictor_name (enum br_predictor predictor) |
4147 | { |
4148 | return predictor_info[predictor].name; |
4149 | } |
4150 | |
4151 | /* Predict branch probabilities and estimate profile of the tree CFG. */ |
4152 | |
4153 | namespace { |
4154 | |
4155 | const pass_data pass_data_profile = |
4156 | { |
4157 | .type: GIMPLE_PASS, /* type */ |
4158 | .name: "profile_estimate" , /* name */ |
4159 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
4160 | .tv_id: TV_BRANCH_PROB, /* tv_id */ |
4161 | PROP_cfg, /* properties_required */ |
4162 | .properties_provided: 0, /* properties_provided */ |
4163 | .properties_destroyed: 0, /* properties_destroyed */ |
4164 | .todo_flags_start: 0, /* todo_flags_start */ |
4165 | .todo_flags_finish: 0, /* todo_flags_finish */ |
4166 | }; |
4167 | |
4168 | class pass_profile : public gimple_opt_pass |
4169 | { |
4170 | public: |
4171 | pass_profile (gcc::context *ctxt) |
4172 | : gimple_opt_pass (pass_data_profile, ctxt) |
4173 | {} |
4174 | |
4175 | /* opt_pass methods: */ |
4176 | bool gate (function *) final override { return flag_guess_branch_prob; } |
4177 | unsigned int execute (function *) final override; |
4178 | |
4179 | }; // class pass_profile |
4180 | |
4181 | unsigned int |
4182 | pass_profile::execute (function *fun) |
4183 | { |
4184 | unsigned nb_loops; |
4185 | |
4186 | if (profile_status_for_fn (cfun) == PROFILE_GUESSED) |
4187 | return 0; |
4188 | |
4189 | loop_optimizer_init (LOOPS_NORMAL); |
4190 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4191 | flow_loops_dump (dump_file, NULL, 0); |
4192 | |
4193 | nb_loops = number_of_loops (fn: fun); |
4194 | if (nb_loops > 1) |
4195 | scev_initialize (); |
4196 | |
4197 | tree_estimate_probability (dry_run: false); |
4198 | cfun->cfg->full_profile = true; |
4199 | |
4200 | if (nb_loops > 1) |
4201 | scev_finalize (); |
4202 | |
4203 | loop_optimizer_finalize (); |
4204 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4205 | gimple_dump_cfg (dump_file, dump_flags); |
4206 | if (profile_status_for_fn (fun) == PROFILE_ABSENT) |
4207 | profile_status_for_fn (fun) = PROFILE_GUESSED; |
4208 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4209 | { |
4210 | sreal iterations; |
4211 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
4212 | if (expected_loop_iterations_by_profile (loop, ret: &iterations)) |
4213 | fprintf (stream: dump_file, format: "Loop got predicted %d to iterate %f times.\n" , |
4214 | loop->num, iterations.to_double ()); |
4215 | } |
4216 | return 0; |
4217 | } |
4218 | |
4219 | } // anon namespace |
4220 | |
4221 | gimple_opt_pass * |
4222 | make_pass_profile (gcc::context *ctxt) |
4223 | { |
4224 | return new pass_profile (ctxt); |
4225 | } |
4226 | |
4227 | /* Return true when PRED predictor should be removed after early |
4228 | tree passes. Most of the predictors are beneficial to survive |
4229 | as early inlining can also distribute then into caller's bodies. */ |
4230 | |
4231 | static bool |
4232 | strip_predictor_early (enum br_predictor pred) |
4233 | { |
4234 | switch (pred) |
4235 | { |
4236 | case PRED_TREE_EARLY_RETURN: |
4237 | return true; |
4238 | default: |
4239 | return false; |
4240 | } |
4241 | } |
4242 | |
4243 | /* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements |
4244 | we no longer need. EARLY is set to true when called from early |
4245 | optimizations. */ |
4246 | |
4247 | unsigned int |
4248 | strip_predict_hints (function *fun, bool early) |
4249 | { |
4250 | basic_block bb; |
4251 | gimple *ass_stmt; |
4252 | tree var; |
4253 | bool changed = false; |
4254 | |
4255 | FOR_EACH_BB_FN (bb, fun) |
4256 | { |
4257 | gimple_stmt_iterator bi; |
4258 | for (bi = gsi_start_bb (bb); !gsi_end_p (i: bi);) |
4259 | { |
4260 | gimple *stmt = gsi_stmt (i: bi); |
4261 | |
4262 | if (gimple_code (g: stmt) == GIMPLE_PREDICT) |
4263 | { |
4264 | if (!early |
4265 | || strip_predictor_early (pred: gimple_predict_predictor (gs: stmt))) |
4266 | { |
4267 | gsi_remove (&bi, true); |
4268 | changed = true; |
4269 | continue; |
4270 | } |
4271 | } |
4272 | else if (is_gimple_call (gs: stmt)) |
4273 | { |
4274 | tree fndecl = gimple_call_fndecl (gs: stmt); |
4275 | |
4276 | if (!early |
4277 | && ((fndecl != NULL_TREE |
4278 | && fndecl_built_in_p (node: fndecl, name1: BUILT_IN_EXPECT) |
4279 | && gimple_call_num_args (gs: stmt) == 2) |
4280 | || (fndecl != NULL_TREE |
4281 | && fndecl_built_in_p (node: fndecl, |
4282 | name1: BUILT_IN_EXPECT_WITH_PROBABILITY) |
4283 | && gimple_call_num_args (gs: stmt) == 3) |
4284 | || (gimple_call_internal_p (gs: stmt) |
4285 | && gimple_call_internal_fn (gs: stmt) == IFN_BUILTIN_EXPECT))) |
4286 | { |
4287 | var = gimple_call_lhs (gs: stmt); |
4288 | changed = true; |
4289 | if (var) |
4290 | { |
4291 | ass_stmt |
4292 | = gimple_build_assign (var, gimple_call_arg (gs: stmt, index: 0)); |
4293 | gsi_replace (&bi, ass_stmt, true); |
4294 | } |
4295 | else |
4296 | { |
4297 | gsi_remove (&bi, true); |
4298 | continue; |
4299 | } |
4300 | } |
4301 | } |
4302 | gsi_next (i: &bi); |
4303 | } |
4304 | } |
4305 | return changed ? TODO_cleanup_cfg : 0; |
4306 | } |
4307 | |
4308 | namespace { |
4309 | |
4310 | const pass_data pass_data_strip_predict_hints = |
4311 | { |
4312 | .type: GIMPLE_PASS, /* type */ |
4313 | .name: "*strip_predict_hints" , /* name */ |
4314 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
4315 | .tv_id: TV_BRANCH_PROB, /* tv_id */ |
4316 | PROP_cfg, /* properties_required */ |
4317 | .properties_provided: 0, /* properties_provided */ |
4318 | .properties_destroyed: 0, /* properties_destroyed */ |
4319 | .todo_flags_start: 0, /* todo_flags_start */ |
4320 | .todo_flags_finish: 0, /* todo_flags_finish */ |
4321 | }; |
4322 | |
4323 | class pass_strip_predict_hints : public gimple_opt_pass |
4324 | { |
4325 | public: |
4326 | pass_strip_predict_hints (gcc::context *ctxt) |
4327 | : gimple_opt_pass (pass_data_strip_predict_hints, ctxt) |
4328 | {} |
4329 | |
4330 | /* opt_pass methods: */ |
4331 | opt_pass * clone () final override |
4332 | { |
4333 | return new pass_strip_predict_hints (m_ctxt); |
4334 | } |
4335 | void set_pass_param (unsigned int n, bool param) final override |
4336 | { |
4337 | gcc_assert (n == 0); |
4338 | early_p = param; |
4339 | } |
4340 | |
4341 | unsigned int execute (function *) final override; |
4342 | |
4343 | private: |
4344 | bool early_p; |
4345 | |
4346 | }; // class pass_strip_predict_hints |
4347 | |
4348 | unsigned int |
4349 | pass_strip_predict_hints::execute (function *fun) |
4350 | { |
4351 | return strip_predict_hints (fun, early: early_p); |
4352 | } |
4353 | |
4354 | } // anon namespace |
4355 | |
4356 | gimple_opt_pass * |
4357 | make_pass_strip_predict_hints (gcc::context *ctxt) |
4358 | { |
4359 | return new pass_strip_predict_hints (ctxt); |
4360 | } |
4361 | |
4362 | /* Rebuild function frequencies. Passes are in general expected to |
4363 | maintain profile by hand, however in some cases this is not possible: |
4364 | for example when inlining several functions with loops freuqencies might run |
4365 | out of scale and thus needs to be recomputed. */ |
4366 | |
4367 | void |
4368 | rebuild_frequencies (void) |
4369 | { |
4370 | /* If we have no profile, do nothing. Note that after inlining |
4371 | profile_status_for_fn may not represent the actual presence/absence of |
4372 | profile. */ |
4373 | if (profile_status_for_fn (cfun) == PROFILE_ABSENT |
4374 | && !ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.initialized_p ()) |
4375 | return; |
4376 | |
4377 | |
4378 | /* See if everything is OK and update count_max. */ |
4379 | basic_block bb; |
4380 | bool inconsistency_found = false; |
4381 | bool uninitialized_probablity_found = false; |
4382 | bool uninitialized_count_found = false; |
4383 | |
4384 | cfun->cfg->count_max = profile_count::uninitialized (); |
4385 | FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun), NULL, next_bb) |
4386 | { |
4387 | cfun->cfg->count_max = cfun->cfg->count_max.max (other: bb->count); |
4388 | /* Uninitialized count may be result of inlining or an omision in an |
4389 | optimization pass. */ |
4390 | if (!bb->count.initialized_p ()) |
4391 | { |
4392 | uninitialized_count_found = true; |
4393 | if (dump_file) |
4394 | fprintf (stream: dump_file, format: "BB %i has uninitialized count\n" , |
4395 | bb->index); |
4396 | } |
4397 | if (bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
4398 | && (!uninitialized_probablity_found || !inconsistency_found)) |
4399 | { |
4400 | profile_count sum = profile_count::zero (); |
4401 | edge e; |
4402 | edge_iterator ei; |
4403 | |
4404 | FOR_EACH_EDGE (e, ei, bb->preds) |
4405 | { |
4406 | sum += e->count (); |
4407 | /* Uninitialized probability may be result of inlining or an |
4408 | omision in an optimization pass. */ |
4409 | if (!e->probability.initialized_p ()) |
4410 | { |
4411 | if (dump_file) |
4412 | fprintf (stream: dump_file, |
4413 | format: "Edge %i->%i has uninitialized probability\n" , |
4414 | e->src->index, e->dest->index); |
4415 | } |
4416 | } |
4417 | if (sum.differs_from_p (other: bb->count)) |
4418 | { |
4419 | if (dump_file) |
4420 | fprintf (stream: dump_file, |
4421 | format: "BB %i has invalid sum of incomming counts\n" , |
4422 | bb->index); |
4423 | inconsistency_found = true; |
4424 | } |
4425 | } |
4426 | } |
4427 | |
4428 | /* If everything is OK, do not re-propagate frequencies. */ |
4429 | if (!inconsistency_found |
4430 | && (!uninitialized_count_found || uninitialized_probablity_found) |
4431 | && !cfun->cfg->count_max.very_large_p ()) |
4432 | { |
4433 | if (dump_file) |
4434 | fprintf (stream: dump_file, format: "Profile is consistent\n" ); |
4435 | return; |
4436 | } |
4437 | /* Do not re-propagate if we have profile feedback. Even if the profile is |
4438 | inconsistent from previous transofrmations, it is probably more realistic |
4439 | for hot part of the program than result of repropagating. |
4440 | |
4441 | Consider example where we previously has |
4442 | |
4443 | if (test) |
4444 | then [large probability for true] |
4445 | |
4446 | and we later proved that test is always 0. In this case, if profile was |
4447 | read correctly, we must have duplicated the conditional (for example by |
4448 | inlining) in to a context where test is false. From profile feedback |
4449 | we know that most executions if the conditionals were true, so the |
4450 | important copy is not the one we look on. |
4451 | |
4452 | Propagating from probabilities would make profile look consistent, but |
4453 | because probablities after code duplication may not be representative |
4454 | for a given run, we would only propagate the error further. */ |
4455 | if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.ipa ().nonzero_p () |
4456 | && !uninitialized_count_found) |
4457 | { |
4458 | if (dump_file) |
4459 | fprintf (stream: dump_file, |
4460 | format: "Profile is inconsistent but read from profile feedback;" |
4461 | " not rebuilding\n" ); |
4462 | return; |
4463 | } |
4464 | |
4465 | loop_optimizer_init (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); |
4466 | connect_infinite_loops_to_exit (); |
4467 | estimate_bb_frequencies (); |
4468 | remove_fake_exit_edges (); |
4469 | loop_optimizer_finalize (); |
4470 | if (dump_file) |
4471 | fprintf (stream: dump_file, format: "Rebuilt basic block counts\n" ); |
4472 | |
4473 | return; |
4474 | } |
4475 | |
4476 | namespace { |
4477 | |
4478 | const pass_data pass_data_rebuild_frequencies = |
4479 | { |
4480 | .type: GIMPLE_PASS, /* type */ |
4481 | .name: "rebuild_frequencies" , /* name */ |
4482 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
4483 | .tv_id: TV_REBUILD_FREQUENCIES, /* tv_id */ |
4484 | PROP_cfg, /* properties_required */ |
4485 | .properties_provided: 0, /* properties_provided */ |
4486 | .properties_destroyed: 0, /* properties_destroyed */ |
4487 | .todo_flags_start: 0, /* todo_flags_start */ |
4488 | .todo_flags_finish: 0, /* todo_flags_finish */ |
4489 | }; |
4490 | |
4491 | class pass_rebuild_frequencies : public gimple_opt_pass |
4492 | { |
4493 | public: |
4494 | pass_rebuild_frequencies (gcc::context *ctxt) |
4495 | : gimple_opt_pass (pass_data_rebuild_frequencies, ctxt) |
4496 | {} |
4497 | |
4498 | /* opt_pass methods: */ |
4499 | opt_pass * clone () final override |
4500 | { |
4501 | return new pass_rebuild_frequencies (m_ctxt); |
4502 | } |
4503 | void set_pass_param (unsigned int n, bool param) final override |
4504 | { |
4505 | gcc_assert (n == 0); |
4506 | early_p = param; |
4507 | } |
4508 | |
4509 | unsigned int execute (function *) final override |
4510 | { |
4511 | rebuild_frequencies (); |
4512 | return 0; |
4513 | } |
4514 | |
4515 | private: |
4516 | bool early_p; |
4517 | |
4518 | }; // class pass_rebuild_frequencies |
4519 | |
4520 | } // anon namespace |
4521 | |
4522 | gimple_opt_pass * |
4523 | make_pass_rebuild_frequencies (gcc::context *ctxt) |
4524 | { |
4525 | return new pass_rebuild_frequencies (ctxt); |
4526 | } |
4527 | |
4528 | /* Perform a dry run of the branch prediction pass and report comparsion of |
4529 | the predicted and real profile into the dump file. */ |
4530 | |
4531 | void |
4532 | report_predictor_hitrates (void) |
4533 | { |
4534 | unsigned nb_loops; |
4535 | |
4536 | loop_optimizer_init (LOOPS_NORMAL); |
4537 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4538 | flow_loops_dump (dump_file, NULL, 0); |
4539 | |
4540 | nb_loops = number_of_loops (cfun); |
4541 | if (nb_loops > 1) |
4542 | scev_initialize (); |
4543 | |
4544 | tree_estimate_probability (dry_run: true); |
4545 | |
4546 | if (nb_loops > 1) |
4547 | scev_finalize (); |
4548 | |
4549 | loop_optimizer_finalize (); |
4550 | } |
4551 | |
4552 | /* Force edge E to be cold. |
4553 | If IMPOSSIBLE is true, for edge to have count and probability 0 otherwise |
4554 | keep low probability to represent possible error in a guess. This is used |
4555 | i.e. in case we predict loop to likely iterate given number of times but |
4556 | we are not 100% sure. |
4557 | |
4558 | This function locally updates profile without attempt to keep global |
4559 | consistency which cannot be reached in full generality without full profile |
4560 | rebuild from probabilities alone. Doing so is not necessarily a good idea |
4561 | because frequencies and counts may be more realistic then probabilities. |
4562 | |
4563 | In some cases (such as for elimination of early exits during full loop |
4564 | unrolling) the caller can ensure that profile will get consistent |
4565 | afterwards. */ |
4566 | |
4567 | void |
4568 | force_edge_cold (edge e, bool impossible) |
4569 | { |
4570 | profile_count count_sum = profile_count::zero (); |
4571 | profile_probability prob_sum = profile_probability::never (); |
4572 | edge_iterator ei; |
4573 | edge e2; |
4574 | bool uninitialized_exit = false; |
4575 | |
4576 | /* When branch probability guesses are not known, then do nothing. */ |
4577 | if (!impossible && !e->count ().initialized_p ()) |
4578 | return; |
4579 | |
4580 | profile_probability goal = (impossible ? profile_probability::never () |
4581 | : profile_probability::very_unlikely ()); |
4582 | |
4583 | /* If edge is already improbably or cold, just return. */ |
4584 | if (e->probability <= goal |
4585 | && (!impossible || e->count () == profile_count::zero ())) |
4586 | return; |
4587 | FOR_EACH_EDGE (e2, ei, e->src->succs) |
4588 | if (e2 != e) |
4589 | { |
4590 | if (e->flags & EDGE_FAKE) |
4591 | continue; |
4592 | if (e2->count ().initialized_p ()) |
4593 | count_sum += e2->count (); |
4594 | if (e2->probability.initialized_p ()) |
4595 | prob_sum += e2->probability; |
4596 | else |
4597 | uninitialized_exit = true; |
4598 | } |
4599 | |
4600 | /* If we are not guessing profiles but have some other edges out, |
4601 | just assume the control flow goes elsewhere. */ |
4602 | if (uninitialized_exit) |
4603 | e->probability = goal; |
4604 | /* If there are other edges out of e->src, redistribute probabilitity |
4605 | there. */ |
4606 | else if (prob_sum > profile_probability::never ()) |
4607 | { |
4608 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4609 | { |
4610 | fprintf (stream: dump_file, format: "Making edge %i->%i %s by redistributing " |
4611 | "probability to other edges. Original probability: " , |
4612 | e->src->index, e->dest->index, |
4613 | impossible ? "impossible" : "cold" ); |
4614 | e->probability.dump (f: dump_file); |
4615 | fprintf (stream: dump_file, format: "\n" ); |
4616 | } |
4617 | set_edge_probability_and_rescale_others (e, goal); |
4618 | if (current_ir_type () != IR_GIMPLE |
4619 | && e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
4620 | update_br_prob_note (e->src); |
4621 | } |
4622 | /* If all edges out of e->src are unlikely, the basic block itself |
4623 | is unlikely. */ |
4624 | else |
4625 | { |
4626 | if (prob_sum == profile_probability::never ()) |
4627 | e->probability = profile_probability::always (); |
4628 | else |
4629 | { |
4630 | if (impossible) |
4631 | e->probability = profile_probability::never (); |
4632 | /* If BB has some edges out that are not impossible, we cannot |
4633 | assume that BB itself is. */ |
4634 | impossible = false; |
4635 | } |
4636 | if (current_ir_type () != IR_GIMPLE |
4637 | && e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
4638 | update_br_prob_note (e->src); |
4639 | if (e->src->count == profile_count::zero ()) |
4640 | return; |
4641 | if (count_sum == profile_count::zero () && impossible) |
4642 | { |
4643 | bool found = false; |
4644 | if (e->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
4645 | ; |
4646 | else if (current_ir_type () == IR_GIMPLE) |
4647 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb: e->src); |
4648 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4649 | { |
4650 | if (stmt_can_terminate_bb_p (gsi_stmt (i: gsi))) |
4651 | { |
4652 | found = true; |
4653 | break; |
4654 | } |
4655 | } |
4656 | /* FIXME: Implement RTL path. */ |
4657 | else |
4658 | found = true; |
4659 | if (!found) |
4660 | { |
4661 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4662 | fprintf (stream: dump_file, |
4663 | format: "Making bb %i impossible and dropping count to 0.\n" , |
4664 | e->src->index); |
4665 | e->src->count = profile_count::zero (); |
4666 | FOR_EACH_EDGE (e2, ei, e->src->preds) |
4667 | force_edge_cold (e: e2, impossible); |
4668 | return; |
4669 | } |
4670 | } |
4671 | |
4672 | /* If we did not adjusting, the source basic block has no likely edeges |
4673 | leaving other direction. In that case force that bb cold, too. |
4674 | This in general is difficult task to do, but handle special case when |
4675 | BB has only one predecestor. This is common case when we are updating |
4676 | after loop transforms. */ |
4677 | if (!(prob_sum > profile_probability::never ()) |
4678 | && count_sum == profile_count::zero () |
4679 | && single_pred_p (bb: e->src) && e->src->count.to_frequency (cfun) |
4680 | > (impossible ? 0 : 1)) |
4681 | { |
4682 | int old_frequency = e->src->count.to_frequency (cfun); |
4683 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4684 | fprintf (stream: dump_file, format: "Making bb %i %s.\n" , e->src->index, |
4685 | impossible ? "impossible" : "cold" ); |
4686 | int new_frequency = MIN (e->src->count.to_frequency (cfun), |
4687 | impossible ? 0 : 1); |
4688 | if (impossible) |
4689 | e->src->count = profile_count::zero (); |
4690 | else |
4691 | e->src->count = e->count ().apply_scale (num: new_frequency, |
4692 | den: old_frequency); |
4693 | force_edge_cold (e: single_pred_edge (bb: e->src), impossible); |
4694 | } |
4695 | else if (dump_file && (dump_flags & TDF_DETAILS) |
4696 | && maybe_hot_bb_p (cfun, bb: e->src)) |
4697 | fprintf (stream: dump_file, format: "Giving up on making bb %i %s.\n" , e->src->index, |
4698 | impossible ? "impossible" : "cold" ); |
4699 | } |
4700 | } |
4701 | |
4702 | #if CHECKING_P |
4703 | |
4704 | namespace selftest { |
4705 | |
4706 | /* Test that value range of predictor values defined in predict.def is |
4707 | within range (50, 100]. */ |
4708 | |
4709 | struct branch_predictor |
4710 | { |
4711 | const char *name; |
4712 | int probability; |
4713 | }; |
4714 | |
4715 | #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) { NAME, HITRATE }, |
4716 | |
4717 | static void |
4718 | test_prediction_value_range () |
4719 | { |
4720 | branch_predictor predictors[] = { |
4721 | #include "predict.def" |
4722 | { NULL, PROB_UNINITIALIZED } |
4723 | }; |
4724 | |
4725 | for (unsigned i = 0; predictors[i].name != NULL; i++) |
4726 | { |
4727 | if (predictors[i].probability == PROB_UNINITIALIZED) |
4728 | continue; |
4729 | |
4730 | unsigned p = 100 * predictors[i].probability / REG_BR_PROB_BASE; |
4731 | ASSERT_TRUE (p >= 50 && p <= 100); |
4732 | } |
4733 | } |
4734 | |
4735 | #undef DEF_PREDICTOR |
4736 | |
4737 | /* Run all of the selfests within this file. */ |
4738 | |
4739 | void |
4740 | predict_cc_tests () |
4741 | { |
4742 | test_prediction_value_range (); |
4743 | } |
4744 | |
4745 | } // namespace selftest |
4746 | #endif /* CHECKING_P. */ |
4747 | |