re PR c++/29234 (Call to operator() of temporary object wrongly parsed)
[gcc.git] / gcc / predict.c
1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000-2013 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 "tm.h"
34 #include "tree.h"
35 #include "rtl.h"
36 #include "tm_p.h"
37 #include "hard-reg-set.h"
38 #include "basic-block.h"
39 #include "insn-config.h"
40 #include "regs.h"
41 #include "flags.h"
42 #include "function.h"
43 #include "except.h"
44 #include "diagnostic-core.h"
45 #include "recog.h"
46 #include "expr.h"
47 #include "predict.h"
48 #include "coverage.h"
49 #include "sreal.h"
50 #include "params.h"
51 #include "target.h"
52 #include "cfgloop.h"
53 #include "gimple.h"
54 #include "gimple-ssa.h"
55 #include "cgraph.h"
56 #include "tree-cfg.h"
57 #include "tree-phinodes.h"
58 #include "ssa-iterators.h"
59 #include "tree-ssa-loop-niter.h"
60 #include "tree-ssa-loop.h"
61 #include "ggc.h"
62 #include "tree-pass.h"
63 #include "tree-scalar-evolution.h"
64 #include "cfgloop.h"
65 #include "pointer-set.h"
66
67 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
68 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
69 static sreal real_zero, real_one, real_almost_one, real_br_prob_base,
70 real_inv_br_prob_base, real_one_half, real_bb_freq_max;
71
72 /* Random guesstimation given names.
73 PROV_VERY_UNLIKELY should be small enough so basic block predicted
74 by it gets below HOT_BB_FREQUENCY_FRACTION. */
75 #define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 2000 - 1)
76 #define PROB_EVEN (REG_BR_PROB_BASE / 2)
77 #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
78 #define PROB_ALWAYS (REG_BR_PROB_BASE)
79
80 static void combine_predictions_for_insn (rtx, basic_block);
81 static void dump_prediction (FILE *, enum br_predictor, int, basic_block, int);
82 static void predict_paths_leading_to (basic_block, enum br_predictor, enum prediction);
83 static void predict_paths_leading_to_edge (edge, enum br_predictor, enum prediction);
84 static bool can_predict_insn_p (const_rtx);
85
86 /* Information we hold about each branch predictor.
87 Filled using information from predict.def. */
88
89 struct predictor_info
90 {
91 const char *const name; /* Name used in the debugging dumps. */
92 const int hitrate; /* Expected hitrate used by
93 predict_insn_def call. */
94 const int flags;
95 };
96
97 /* Use given predictor without Dempster-Shaffer theory if it matches
98 using first_match heuristics. */
99 #define PRED_FLAG_FIRST_MATCH 1
100
101 /* Recompute hitrate in percent to our representation. */
102
103 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
104
105 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
106 static const struct predictor_info predictor_info[]= {
107 #include "predict.def"
108
109 /* Upper bound on predictors. */
110 {NULL, 0, 0}
111 };
112 #undef DEF_PREDICTOR
113
114 /* Return TRUE if frequency FREQ is considered to be hot. */
115
116 static inline bool
117 maybe_hot_frequency_p (struct function *fun, int freq)
118 {
119 struct cgraph_node *node = cgraph_get_node (fun->decl);
120 if (!profile_info || !flag_branch_probabilities)
121 {
122 if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
123 return false;
124 if (node->frequency == NODE_FREQUENCY_HOT)
125 return true;
126 }
127 if (profile_status_for_function (fun) == PROFILE_ABSENT)
128 return true;
129 if (node->frequency == NODE_FREQUENCY_EXECUTED_ONCE
130 && freq < (ENTRY_BLOCK_PTR_FOR_FUNCTION (fun)->frequency * 2 / 3))
131 return false;
132 if (PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION) == 0)
133 return false;
134 if (freq < (ENTRY_BLOCK_PTR_FOR_FUNCTION (fun)->frequency
135 / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)))
136 return false;
137 return true;
138 }
139
140 static gcov_type min_count = -1;
141
142 /* Determine the threshold for hot BB counts. */
143
144 gcov_type
145 get_hot_bb_threshold ()
146 {
147 gcov_working_set_t *ws;
148 if (min_count == -1)
149 {
150 ws = find_working_set (PARAM_VALUE (HOT_BB_COUNT_WS_PERMILLE));
151 gcc_assert (ws);
152 min_count = ws->min_counter;
153 }
154 return min_count;
155 }
156
157 /* Set the threshold for hot BB counts. */
158
159 void
160 set_hot_bb_threshold (gcov_type min)
161 {
162 min_count = min;
163 }
164
165 /* Return TRUE if frequency FREQ is considered to be hot. */
166
167 static inline bool
168 maybe_hot_count_p (struct function *fun, gcov_type count)
169 {
170 if (fun && profile_status_for_function (fun) != PROFILE_READ)
171 return true;
172 /* Code executed at most once is not hot. */
173 if (profile_info->runs >= count)
174 return false;
175 return (count >= get_hot_bb_threshold ());
176 }
177
178 /* Return true in case BB can be CPU intensive and should be optimized
179 for maximal performance. */
180
181 bool
182 maybe_hot_bb_p (struct function *fun, const_basic_block bb)
183 {
184 gcc_checking_assert (fun);
185 if (profile_status_for_function (fun) == PROFILE_READ)
186 return maybe_hot_count_p (fun, bb->count);
187 return maybe_hot_frequency_p (fun, bb->frequency);
188 }
189
190 /* Return true if the call can be hot. */
191
192 bool
193 cgraph_maybe_hot_edge_p (struct cgraph_edge *edge)
194 {
195 if (profile_info && flag_branch_probabilities
196 && !maybe_hot_count_p (NULL,
197 edge->count))
198 return false;
199 if (edge->caller->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED
200 || (edge->callee
201 && edge->callee->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED))
202 return false;
203 if (edge->caller->frequency > NODE_FREQUENCY_UNLIKELY_EXECUTED
204 && (edge->callee
205 && edge->callee->frequency <= NODE_FREQUENCY_EXECUTED_ONCE))
206 return false;
207 if (optimize_size)
208 return false;
209 if (edge->caller->frequency == NODE_FREQUENCY_HOT)
210 return true;
211 if (edge->caller->frequency == NODE_FREQUENCY_EXECUTED_ONCE
212 && edge->frequency < CGRAPH_FREQ_BASE * 3 / 2)
213 return false;
214 if (flag_guess_branch_prob)
215 {
216 if (PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION) == 0
217 || edge->frequency <= (CGRAPH_FREQ_BASE
218 / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)))
219 return false;
220 }
221 return true;
222 }
223
224 /* Return true in case BB can be CPU intensive and should be optimized
225 for maximal performance. */
226
227 bool
228 maybe_hot_edge_p (edge e)
229 {
230 if (profile_status == PROFILE_READ)
231 return maybe_hot_count_p (cfun, e->count);
232 return maybe_hot_frequency_p (cfun, EDGE_FREQUENCY (e));
233 }
234
235
236
237 /* Return true if profile COUNT and FREQUENCY, or function FUN static
238 node frequency reflects never being executed. */
239
240 static bool
241 probably_never_executed (struct function *fun,
242 gcov_type count, int frequency)
243 {
244 gcc_checking_assert (fun);
245 if (profile_status_for_function (fun) == PROFILE_READ)
246 {
247 int unlikely_count_fraction = PARAM_VALUE (UNLIKELY_BB_COUNT_FRACTION);
248 if (count * unlikely_count_fraction >= profile_info->runs)
249 return false;
250 if (!frequency)
251 return true;
252 if (!ENTRY_BLOCK_PTR->frequency)
253 return false;
254 if (ENTRY_BLOCK_PTR->count)
255 {
256 gcov_type computed_count;
257 /* Check for possibility of overflow, in which case entry bb count
258 is large enough to do the division first without losing much
259 precision. */
260 if (ENTRY_BLOCK_PTR->count < REG_BR_PROB_BASE * REG_BR_PROB_BASE)
261 {
262 gcov_type scaled_count
263 = frequency * ENTRY_BLOCK_PTR->count * unlikely_count_fraction;
264 computed_count = RDIV (scaled_count, ENTRY_BLOCK_PTR->frequency);
265 }
266 else
267 {
268 computed_count = RDIV (ENTRY_BLOCK_PTR->count,
269 ENTRY_BLOCK_PTR->frequency);
270 computed_count *= frequency * unlikely_count_fraction;
271 }
272 if (computed_count >= profile_info->runs)
273 return false;
274 }
275 return true;
276 }
277 if ((!profile_info || !flag_branch_probabilities)
278 && (cgraph_get_node (fun->decl)->frequency
279 == NODE_FREQUENCY_UNLIKELY_EXECUTED))
280 return true;
281 return false;
282 }
283
284
285 /* Return true in case BB is probably never executed. */
286
287 bool
288 probably_never_executed_bb_p (struct function *fun, const_basic_block bb)
289 {
290 return probably_never_executed (fun, bb->count, bb->frequency);
291 }
292
293
294 /* Return true in case edge E is probably never executed. */
295
296 bool
297 probably_never_executed_edge_p (struct function *fun, edge e)
298 {
299 return probably_never_executed (fun, e->count, EDGE_FREQUENCY (e));
300 }
301
302 /* Return true if NODE should be optimized for size. */
303
304 bool
305 cgraph_optimize_for_size_p (struct cgraph_node *node)
306 {
307 if (optimize_size)
308 return true;
309 if (node && (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED))
310 return true;
311 else
312 return false;
313 }
314
315 /* Return true when current function should always be optimized for size. */
316
317 bool
318 optimize_function_for_size_p (struct function *fun)
319 {
320 if (optimize_size)
321 return true;
322 if (!fun || !fun->decl)
323 return false;
324 return cgraph_optimize_for_size_p (cgraph_get_node (fun->decl));
325 }
326
327 /* Return true when current function should always be optimized for speed. */
328
329 bool
330 optimize_function_for_speed_p (struct function *fun)
331 {
332 return !optimize_function_for_size_p (fun);
333 }
334
335 /* Return TRUE when BB should be optimized for size. */
336
337 bool
338 optimize_bb_for_size_p (const_basic_block bb)
339 {
340 return optimize_function_for_size_p (cfun) || !maybe_hot_bb_p (cfun, bb);
341 }
342
343 /* Return TRUE when BB should be optimized for speed. */
344
345 bool
346 optimize_bb_for_speed_p (const_basic_block bb)
347 {
348 return !optimize_bb_for_size_p (bb);
349 }
350
351 /* Return TRUE when BB should be optimized for size. */
352
353 bool
354 optimize_edge_for_size_p (edge e)
355 {
356 return optimize_function_for_size_p (cfun) || !maybe_hot_edge_p (e);
357 }
358
359 /* Return TRUE when BB should be optimized for speed. */
360
361 bool
362 optimize_edge_for_speed_p (edge e)
363 {
364 return !optimize_edge_for_size_p (e);
365 }
366
367 /* Return TRUE when BB should be optimized for size. */
368
369 bool
370 optimize_insn_for_size_p (void)
371 {
372 return optimize_function_for_size_p (cfun) || !crtl->maybe_hot_insn_p;
373 }
374
375 /* Return TRUE when BB should be optimized for speed. */
376
377 bool
378 optimize_insn_for_speed_p (void)
379 {
380 return !optimize_insn_for_size_p ();
381 }
382
383 /* Return TRUE when LOOP should be optimized for size. */
384
385 bool
386 optimize_loop_for_size_p (struct loop *loop)
387 {
388 return optimize_bb_for_size_p (loop->header);
389 }
390
391 /* Return TRUE when LOOP should be optimized for speed. */
392
393 bool
394 optimize_loop_for_speed_p (struct loop *loop)
395 {
396 return optimize_bb_for_speed_p (loop->header);
397 }
398
399 /* Return TRUE when LOOP nest should be optimized for speed. */
400
401 bool
402 optimize_loop_nest_for_speed_p (struct loop *loop)
403 {
404 struct loop *l = loop;
405 if (optimize_loop_for_speed_p (loop))
406 return true;
407 l = loop->inner;
408 while (l && l != loop)
409 {
410 if (optimize_loop_for_speed_p (l))
411 return true;
412 if (l->inner)
413 l = l->inner;
414 else if (l->next)
415 l = l->next;
416 else
417 {
418 while (l != loop && !l->next)
419 l = loop_outer (l);
420 if (l != loop)
421 l = l->next;
422 }
423 }
424 return false;
425 }
426
427 /* Return TRUE when LOOP nest should be optimized for size. */
428
429 bool
430 optimize_loop_nest_for_size_p (struct loop *loop)
431 {
432 return !optimize_loop_nest_for_speed_p (loop);
433 }
434
435 /* Return true when edge E is likely to be well predictable by branch
436 predictor. */
437
438 bool
439 predictable_edge_p (edge e)
440 {
441 if (profile_status == PROFILE_ABSENT)
442 return false;
443 if ((e->probability
444 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME) * REG_BR_PROB_BASE / 100)
445 || (REG_BR_PROB_BASE - e->probability
446 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME) * REG_BR_PROB_BASE / 100))
447 return true;
448 return false;
449 }
450
451
452 /* Set RTL expansion for BB profile. */
453
454 void
455 rtl_profile_for_bb (basic_block bb)
456 {
457 crtl->maybe_hot_insn_p = maybe_hot_bb_p (cfun, bb);
458 }
459
460 /* Set RTL expansion for edge profile. */
461
462 void
463 rtl_profile_for_edge (edge e)
464 {
465 crtl->maybe_hot_insn_p = maybe_hot_edge_p (e);
466 }
467
468 /* Set RTL expansion to default mode (i.e. when profile info is not known). */
469 void
470 default_rtl_profile (void)
471 {
472 crtl->maybe_hot_insn_p = true;
473 }
474
475 /* Return true if the one of outgoing edges is already predicted by
476 PREDICTOR. */
477
478 bool
479 rtl_predicted_by_p (const_basic_block bb, enum br_predictor predictor)
480 {
481 rtx note;
482 if (!INSN_P (BB_END (bb)))
483 return false;
484 for (note = REG_NOTES (BB_END (bb)); note; note = XEXP (note, 1))
485 if (REG_NOTE_KIND (note) == REG_BR_PRED
486 && INTVAL (XEXP (XEXP (note, 0), 0)) == (int)predictor)
487 return true;
488 return false;
489 }
490
491 /* This map contains for a basic block the list of predictions for the
492 outgoing edges. */
493
494 static struct pointer_map_t *bb_predictions;
495
496 /* Structure representing predictions in tree level. */
497
498 struct edge_prediction {
499 struct edge_prediction *ep_next;
500 edge ep_edge;
501 enum br_predictor ep_predictor;
502 int ep_probability;
503 };
504
505 /* Return true if the one of outgoing edges is already predicted by
506 PREDICTOR. */
507
508 bool
509 gimple_predicted_by_p (const_basic_block bb, enum br_predictor predictor)
510 {
511 struct edge_prediction *i;
512 void **preds = pointer_map_contains (bb_predictions, bb);
513
514 if (!preds)
515 return false;
516
517 for (i = (struct edge_prediction *) *preds; i; i = i->ep_next)
518 if (i->ep_predictor == predictor)
519 return true;
520 return false;
521 }
522
523 /* Return true when the probability of edge is reliable.
524
525 The profile guessing code is good at predicting branch outcome (ie.
526 taken/not taken), that is predicted right slightly over 75% of time.
527 It is however notoriously poor on predicting the probability itself.
528 In general the profile appear a lot flatter (with probabilities closer
529 to 50%) than the reality so it is bad idea to use it to drive optimization
530 such as those disabling dynamic branch prediction for well predictable
531 branches.
532
533 There are two exceptions - edges leading to noreturn edges and edges
534 predicted by number of iterations heuristics are predicted well. This macro
535 should be able to distinguish those, but at the moment it simply check for
536 noreturn heuristic that is only one giving probability over 99% or bellow
537 1%. In future we might want to propagate reliability information across the
538 CFG if we find this information useful on multiple places. */
539 static bool
540 probability_reliable_p (int prob)
541 {
542 return (profile_status == PROFILE_READ
543 || (profile_status == PROFILE_GUESSED
544 && (prob <= HITRATE (1) || prob >= HITRATE (99))));
545 }
546
547 /* Same predicate as above, working on edges. */
548 bool
549 edge_probability_reliable_p (const_edge e)
550 {
551 return probability_reliable_p (e->probability);
552 }
553
554 /* Same predicate as edge_probability_reliable_p, working on notes. */
555 bool
556 br_prob_note_reliable_p (const_rtx note)
557 {
558 gcc_assert (REG_NOTE_KIND (note) == REG_BR_PROB);
559 return probability_reliable_p (XINT (note, 0));
560 }
561
562 static void
563 predict_insn (rtx insn, enum br_predictor predictor, int probability)
564 {
565 gcc_assert (any_condjump_p (insn));
566 if (!flag_guess_branch_prob)
567 return;
568
569 add_reg_note (insn, REG_BR_PRED,
570 gen_rtx_CONCAT (VOIDmode,
571 GEN_INT ((int) predictor),
572 GEN_INT ((int) probability)));
573 }
574
575 /* Predict insn by given predictor. */
576
577 void
578 predict_insn_def (rtx insn, enum br_predictor predictor,
579 enum prediction taken)
580 {
581 int probability = predictor_info[(int) predictor].hitrate;
582
583 if (taken != TAKEN)
584 probability = REG_BR_PROB_BASE - probability;
585
586 predict_insn (insn, predictor, probability);
587 }
588
589 /* Predict edge E with given probability if possible. */
590
591 void
592 rtl_predict_edge (edge e, enum br_predictor predictor, int probability)
593 {
594 rtx last_insn;
595 last_insn = BB_END (e->src);
596
597 /* We can store the branch prediction information only about
598 conditional jumps. */
599 if (!any_condjump_p (last_insn))
600 return;
601
602 /* We always store probability of branching. */
603 if (e->flags & EDGE_FALLTHRU)
604 probability = REG_BR_PROB_BASE - probability;
605
606 predict_insn (last_insn, predictor, probability);
607 }
608
609 /* Predict edge E with the given PROBABILITY. */
610 void
611 gimple_predict_edge (edge e, enum br_predictor predictor, int probability)
612 {
613 gcc_assert (profile_status != PROFILE_GUESSED);
614 if ((e->src != ENTRY_BLOCK_PTR && EDGE_COUNT (e->src->succs) > 1)
615 && flag_guess_branch_prob && optimize)
616 {
617 struct edge_prediction *i = XNEW (struct edge_prediction);
618 void **preds = pointer_map_insert (bb_predictions, e->src);
619
620 i->ep_next = (struct edge_prediction *) *preds;
621 *preds = i;
622 i->ep_probability = probability;
623 i->ep_predictor = predictor;
624 i->ep_edge = e;
625 }
626 }
627
628 /* Remove all predictions on given basic block that are attached
629 to edge E. */
630 void
631 remove_predictions_associated_with_edge (edge e)
632 {
633 void **preds;
634
635 if (!bb_predictions)
636 return;
637
638 preds = pointer_map_contains (bb_predictions, e->src);
639
640 if (preds)
641 {
642 struct edge_prediction **prediction = (struct edge_prediction **) preds;
643 struct edge_prediction *next;
644
645 while (*prediction)
646 {
647 if ((*prediction)->ep_edge == e)
648 {
649 next = (*prediction)->ep_next;
650 free (*prediction);
651 *prediction = next;
652 }
653 else
654 prediction = &((*prediction)->ep_next);
655 }
656 }
657 }
658
659 /* Clears the list of predictions stored for BB. */
660
661 static void
662 clear_bb_predictions (basic_block bb)
663 {
664 void **preds = pointer_map_contains (bb_predictions, bb);
665 struct edge_prediction *pred, *next;
666
667 if (!preds)
668 return;
669
670 for (pred = (struct edge_prediction *) *preds; pred; pred = next)
671 {
672 next = pred->ep_next;
673 free (pred);
674 }
675 *preds = NULL;
676 }
677
678 /* Return true when we can store prediction on insn INSN.
679 At the moment we represent predictions only on conditional
680 jumps, not at computed jump or other complicated cases. */
681 static bool
682 can_predict_insn_p (const_rtx insn)
683 {
684 return (JUMP_P (insn)
685 && any_condjump_p (insn)
686 && EDGE_COUNT (BLOCK_FOR_INSN (insn)->succs) >= 2);
687 }
688
689 /* Predict edge E by given predictor if possible. */
690
691 void
692 predict_edge_def (edge e, enum br_predictor predictor,
693 enum prediction taken)
694 {
695 int probability = predictor_info[(int) predictor].hitrate;
696
697 if (taken != TAKEN)
698 probability = REG_BR_PROB_BASE - probability;
699
700 predict_edge (e, predictor, probability);
701 }
702
703 /* Invert all branch predictions or probability notes in the INSN. This needs
704 to be done each time we invert the condition used by the jump. */
705
706 void
707 invert_br_probabilities (rtx insn)
708 {
709 rtx note;
710
711 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
712 if (REG_NOTE_KIND (note) == REG_BR_PROB)
713 XINT (note, 0) = REG_BR_PROB_BASE - XINT (note, 0);
714 else if (REG_NOTE_KIND (note) == REG_BR_PRED)
715 XEXP (XEXP (note, 0), 1)
716 = GEN_INT (REG_BR_PROB_BASE - INTVAL (XEXP (XEXP (note, 0), 1)));
717 }
718
719 /* Dump information about the branch prediction to the output file. */
720
721 static void
722 dump_prediction (FILE *file, enum br_predictor predictor, int probability,
723 basic_block bb, int used)
724 {
725 edge e;
726 edge_iterator ei;
727
728 if (!file)
729 return;
730
731 FOR_EACH_EDGE (e, ei, bb->succs)
732 if (! (e->flags & EDGE_FALLTHRU))
733 break;
734
735 fprintf (file, " %s heuristics%s: %.1f%%",
736 predictor_info[predictor].name,
737 used ? "" : " (ignored)", probability * 100.0 / REG_BR_PROB_BASE);
738
739 if (bb->count)
740 {
741 fprintf (file, " exec ");
742 fprintf (file, HOST_WIDEST_INT_PRINT_DEC, bb->count);
743 if (e)
744 {
745 fprintf (file, " hit ");
746 fprintf (file, HOST_WIDEST_INT_PRINT_DEC, e->count);
747 fprintf (file, " (%.1f%%)", e->count * 100.0 / bb->count);
748 }
749 }
750
751 fprintf (file, "\n");
752 }
753
754 /* We can not predict the probabilities of outgoing edges of bb. Set them
755 evenly and hope for the best. */
756 static void
757 set_even_probabilities (basic_block bb)
758 {
759 int nedges = 0;
760 edge e;
761 edge_iterator ei;
762
763 FOR_EACH_EDGE (e, ei, bb->succs)
764 if (!(e->flags & (EDGE_EH | EDGE_FAKE)))
765 nedges ++;
766 FOR_EACH_EDGE (e, ei, bb->succs)
767 if (!(e->flags & (EDGE_EH | EDGE_FAKE)))
768 e->probability = (REG_BR_PROB_BASE + nedges / 2) / nedges;
769 else
770 e->probability = 0;
771 }
772
773 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
774 note if not already present. Remove now useless REG_BR_PRED notes. */
775
776 static void
777 combine_predictions_for_insn (rtx insn, basic_block bb)
778 {
779 rtx prob_note;
780 rtx *pnote;
781 rtx note;
782 int best_probability = PROB_EVEN;
783 enum br_predictor best_predictor = END_PREDICTORS;
784 int combined_probability = REG_BR_PROB_BASE / 2;
785 int d;
786 bool first_match = false;
787 bool found = false;
788
789 if (!can_predict_insn_p (insn))
790 {
791 set_even_probabilities (bb);
792 return;
793 }
794
795 prob_note = find_reg_note (insn, REG_BR_PROB, 0);
796 pnote = &REG_NOTES (insn);
797 if (dump_file)
798 fprintf (dump_file, "Predictions for insn %i bb %i\n", INSN_UID (insn),
799 bb->index);
800
801 /* We implement "first match" heuristics and use probability guessed
802 by predictor with smallest index. */
803 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
804 if (REG_NOTE_KIND (note) == REG_BR_PRED)
805 {
806 enum br_predictor predictor = ((enum br_predictor)
807 INTVAL (XEXP (XEXP (note, 0), 0)));
808 int probability = INTVAL (XEXP (XEXP (note, 0), 1));
809
810 found = true;
811 if (best_predictor > predictor)
812 best_probability = probability, best_predictor = predictor;
813
814 d = (combined_probability * probability
815 + (REG_BR_PROB_BASE - combined_probability)
816 * (REG_BR_PROB_BASE - probability));
817
818 /* Use FP math to avoid overflows of 32bit integers. */
819 if (d == 0)
820 /* If one probability is 0% and one 100%, avoid division by zero. */
821 combined_probability = REG_BR_PROB_BASE / 2;
822 else
823 combined_probability = (((double) combined_probability) * probability
824 * REG_BR_PROB_BASE / d + 0.5);
825 }
826
827 /* Decide which heuristic to use. In case we didn't match anything,
828 use no_prediction heuristic, in case we did match, use either
829 first match or Dempster-Shaffer theory depending on the flags. */
830
831 if (predictor_info [best_predictor].flags & PRED_FLAG_FIRST_MATCH)
832 first_match = true;
833
834 if (!found)
835 dump_prediction (dump_file, PRED_NO_PREDICTION,
836 combined_probability, bb, true);
837 else
838 {
839 dump_prediction (dump_file, PRED_DS_THEORY, combined_probability,
840 bb, !first_match);
841 dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability,
842 bb, first_match);
843 }
844
845 if (first_match)
846 combined_probability = best_probability;
847 dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb, true);
848
849 while (*pnote)
850 {
851 if (REG_NOTE_KIND (*pnote) == REG_BR_PRED)
852 {
853 enum br_predictor predictor = ((enum br_predictor)
854 INTVAL (XEXP (XEXP (*pnote, 0), 0)));
855 int probability = INTVAL (XEXP (XEXP (*pnote, 0), 1));
856
857 dump_prediction (dump_file, predictor, probability, bb,
858 !first_match || best_predictor == predictor);
859 *pnote = XEXP (*pnote, 1);
860 }
861 else
862 pnote = &XEXP (*pnote, 1);
863 }
864
865 if (!prob_note)
866 {
867 add_int_reg_note (insn, REG_BR_PROB, combined_probability);
868
869 /* Save the prediction into CFG in case we are seeing non-degenerated
870 conditional jump. */
871 if (!single_succ_p (bb))
872 {
873 BRANCH_EDGE (bb)->probability = combined_probability;
874 FALLTHRU_EDGE (bb)->probability
875 = REG_BR_PROB_BASE - combined_probability;
876 }
877 }
878 else if (!single_succ_p (bb))
879 {
880 int prob = XINT (prob_note, 0);
881
882 BRANCH_EDGE (bb)->probability = prob;
883 FALLTHRU_EDGE (bb)->probability = REG_BR_PROB_BASE - prob;
884 }
885 else
886 single_succ_edge (bb)->probability = REG_BR_PROB_BASE;
887 }
888
889 /* Combine predictions into single probability and store them into CFG.
890 Remove now useless prediction entries. */
891
892 static void
893 combine_predictions_for_bb (basic_block bb)
894 {
895 int best_probability = PROB_EVEN;
896 enum br_predictor best_predictor = END_PREDICTORS;
897 int combined_probability = REG_BR_PROB_BASE / 2;
898 int d;
899 bool first_match = false;
900 bool found = false;
901 struct edge_prediction *pred;
902 int nedges = 0;
903 edge e, first = NULL, second = NULL;
904 edge_iterator ei;
905 void **preds;
906
907 FOR_EACH_EDGE (e, ei, bb->succs)
908 if (!(e->flags & (EDGE_EH | EDGE_FAKE)))
909 {
910 nedges ++;
911 if (first && !second)
912 second = e;
913 if (!first)
914 first = e;
915 }
916
917 /* When there is no successor or only one choice, prediction is easy.
918
919 We are lazy for now and predict only basic blocks with two outgoing
920 edges. It is possible to predict generic case too, but we have to
921 ignore first match heuristics and do more involved combining. Implement
922 this later. */
923 if (nedges != 2)
924 {
925 if (!bb->count)
926 set_even_probabilities (bb);
927 clear_bb_predictions (bb);
928 if (dump_file)
929 fprintf (dump_file, "%i edges in bb %i predicted to even probabilities\n",
930 nedges, bb->index);
931 return;
932 }
933
934 if (dump_file)
935 fprintf (dump_file, "Predictions for bb %i\n", bb->index);
936
937 preds = pointer_map_contains (bb_predictions, bb);
938 if (preds)
939 {
940 /* We implement "first match" heuristics and use probability guessed
941 by predictor with smallest index. */
942 for (pred = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
943 {
944 enum br_predictor predictor = pred->ep_predictor;
945 int probability = pred->ep_probability;
946
947 if (pred->ep_edge != first)
948 probability = REG_BR_PROB_BASE - probability;
949
950 found = true;
951 /* First match heuristics would be widly confused if we predicted
952 both directions. */
953 if (best_predictor > predictor)
954 {
955 struct edge_prediction *pred2;
956 int prob = probability;
957
958 for (pred2 = (struct edge_prediction *) *preds; pred2; pred2 = pred2->ep_next)
959 if (pred2 != pred && pred2->ep_predictor == pred->ep_predictor)
960 {
961 int probability2 = pred->ep_probability;
962
963 if (pred2->ep_edge != first)
964 probability2 = REG_BR_PROB_BASE - probability2;
965
966 if ((probability < REG_BR_PROB_BASE / 2) !=
967 (probability2 < REG_BR_PROB_BASE / 2))
968 break;
969
970 /* If the same predictor later gave better result, go for it! */
971 if ((probability >= REG_BR_PROB_BASE / 2 && (probability2 > probability))
972 || (probability <= REG_BR_PROB_BASE / 2 && (probability2 < probability)))
973 prob = probability2;
974 }
975 if (!pred2)
976 best_probability = prob, best_predictor = predictor;
977 }
978
979 d = (combined_probability * probability
980 + (REG_BR_PROB_BASE - combined_probability)
981 * (REG_BR_PROB_BASE - probability));
982
983 /* Use FP math to avoid overflows of 32bit integers. */
984 if (d == 0)
985 /* If one probability is 0% and one 100%, avoid division by zero. */
986 combined_probability = REG_BR_PROB_BASE / 2;
987 else
988 combined_probability = (((double) combined_probability)
989 * probability
990 * REG_BR_PROB_BASE / d + 0.5);
991 }
992 }
993
994 /* Decide which heuristic to use. In case we didn't match anything,
995 use no_prediction heuristic, in case we did match, use either
996 first match or Dempster-Shaffer theory depending on the flags. */
997
998 if (predictor_info [best_predictor].flags & PRED_FLAG_FIRST_MATCH)
999 first_match = true;
1000
1001 if (!found)
1002 dump_prediction (dump_file, PRED_NO_PREDICTION, combined_probability, bb, true);
1003 else
1004 {
1005 dump_prediction (dump_file, PRED_DS_THEORY, combined_probability, bb,
1006 !first_match);
1007 dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability, bb,
1008 first_match);
1009 }
1010
1011 if (first_match)
1012 combined_probability = best_probability;
1013 dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb, true);
1014
1015 if (preds)
1016 {
1017 for (pred = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
1018 {
1019 enum br_predictor predictor = pred->ep_predictor;
1020 int probability = pred->ep_probability;
1021
1022 if (pred->ep_edge != EDGE_SUCC (bb, 0))
1023 probability = REG_BR_PROB_BASE - probability;
1024 dump_prediction (dump_file, predictor, probability, bb,
1025 !first_match || best_predictor == predictor);
1026 }
1027 }
1028 clear_bb_predictions (bb);
1029
1030 if (!bb->count)
1031 {
1032 first->probability = combined_probability;
1033 second->probability = REG_BR_PROB_BASE - combined_probability;
1034 }
1035 }
1036
1037 /* Check if T1 and T2 satisfy the IV_COMPARE condition.
1038 Return the SSA_NAME if the condition satisfies, NULL otherwise.
1039
1040 T1 and T2 should be one of the following cases:
1041 1. T1 is SSA_NAME, T2 is NULL
1042 2. T1 is SSA_NAME, T2 is INTEGER_CST between [-4, 4]
1043 3. T2 is SSA_NAME, T1 is INTEGER_CST between [-4, 4] */
1044
1045 static tree
1046 strips_small_constant (tree t1, tree t2)
1047 {
1048 tree ret = NULL;
1049 int value = 0;
1050
1051 if (!t1)
1052 return NULL;
1053 else if (TREE_CODE (t1) == SSA_NAME)
1054 ret = t1;
1055 else if (host_integerp (t1, 0))
1056 value = tree_low_cst (t1, 0);
1057 else
1058 return NULL;
1059
1060 if (!t2)
1061 return ret;
1062 else if (host_integerp (t2, 0))
1063 value = tree_low_cst (t2, 0);
1064 else if (TREE_CODE (t2) == SSA_NAME)
1065 {
1066 if (ret)
1067 return NULL;
1068 else
1069 ret = t2;
1070 }
1071
1072 if (value <= 4 && value >= -4)
1073 return ret;
1074 else
1075 return NULL;
1076 }
1077
1078 /* Return the SSA_NAME in T or T's operands.
1079 Return NULL if SSA_NAME cannot be found. */
1080
1081 static tree
1082 get_base_value (tree t)
1083 {
1084 if (TREE_CODE (t) == SSA_NAME)
1085 return t;
1086
1087 if (!BINARY_CLASS_P (t))
1088 return NULL;
1089
1090 switch (TREE_OPERAND_LENGTH (t))
1091 {
1092 case 1:
1093 return strips_small_constant (TREE_OPERAND (t, 0), NULL);
1094 case 2:
1095 return strips_small_constant (TREE_OPERAND (t, 0),
1096 TREE_OPERAND (t, 1));
1097 default:
1098 return NULL;
1099 }
1100 }
1101
1102 /* Check the compare STMT in LOOP. If it compares an induction
1103 variable to a loop invariant, return true, and save
1104 LOOP_INVARIANT, COMPARE_CODE and LOOP_STEP.
1105 Otherwise return false and set LOOP_INVAIANT to NULL. */
1106
1107 static bool
1108 is_comparison_with_loop_invariant_p (gimple stmt, struct loop *loop,
1109 tree *loop_invariant,
1110 enum tree_code *compare_code,
1111 tree *loop_step,
1112 tree *loop_iv_base)
1113 {
1114 tree op0, op1, bound, base;
1115 affine_iv iv0, iv1;
1116 enum tree_code code;
1117 tree step;
1118
1119 code = gimple_cond_code (stmt);
1120 *loop_invariant = NULL;
1121
1122 switch (code)
1123 {
1124 case GT_EXPR:
1125 case GE_EXPR:
1126 case NE_EXPR:
1127 case LT_EXPR:
1128 case LE_EXPR:
1129 case EQ_EXPR:
1130 break;
1131
1132 default:
1133 return false;
1134 }
1135
1136 op0 = gimple_cond_lhs (stmt);
1137 op1 = gimple_cond_rhs (stmt);
1138
1139 if ((TREE_CODE (op0) != SSA_NAME && TREE_CODE (op0) != INTEGER_CST)
1140 || (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op1) != INTEGER_CST))
1141 return false;
1142 if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, true))
1143 return false;
1144 if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, true))
1145 return false;
1146 if (TREE_CODE (iv0.step) != INTEGER_CST
1147 || TREE_CODE (iv1.step) != INTEGER_CST)
1148 return false;
1149 if ((integer_zerop (iv0.step) && integer_zerop (iv1.step))
1150 || (!integer_zerop (iv0.step) && !integer_zerop (iv1.step)))
1151 return false;
1152
1153 if (integer_zerop (iv0.step))
1154 {
1155 if (code != NE_EXPR && code != EQ_EXPR)
1156 code = invert_tree_comparison (code, false);
1157 bound = iv0.base;
1158 base = iv1.base;
1159 if (host_integerp (iv1.step, 0))
1160 step = iv1.step;
1161 else
1162 return false;
1163 }
1164 else
1165 {
1166 bound = iv1.base;
1167 base = iv0.base;
1168 if (host_integerp (iv0.step, 0))
1169 step = iv0.step;
1170 else
1171 return false;
1172 }
1173
1174 if (TREE_CODE (bound) != INTEGER_CST)
1175 bound = get_base_value (bound);
1176 if (!bound)
1177 return false;
1178 if (TREE_CODE (base) != INTEGER_CST)
1179 base = get_base_value (base);
1180 if (!base)
1181 return false;
1182
1183 *loop_invariant = bound;
1184 *compare_code = code;
1185 *loop_step = step;
1186 *loop_iv_base = base;
1187 return true;
1188 }
1189
1190 /* Compare two SSA_NAMEs: returns TRUE if T1 and T2 are value coherent. */
1191
1192 static bool
1193 expr_coherent_p (tree t1, tree t2)
1194 {
1195 gimple stmt;
1196 tree ssa_name_1 = NULL;
1197 tree ssa_name_2 = NULL;
1198
1199 gcc_assert (TREE_CODE (t1) == SSA_NAME || TREE_CODE (t1) == INTEGER_CST);
1200 gcc_assert (TREE_CODE (t2) == SSA_NAME || TREE_CODE (t2) == INTEGER_CST);
1201
1202 if (t1 == t2)
1203 return true;
1204
1205 if (TREE_CODE (t1) == INTEGER_CST && TREE_CODE (t2) == INTEGER_CST)
1206 return true;
1207 if (TREE_CODE (t1) == INTEGER_CST || TREE_CODE (t2) == INTEGER_CST)
1208 return false;
1209
1210 /* Check to see if t1 is expressed/defined with t2. */
1211 stmt = SSA_NAME_DEF_STMT (t1);
1212 gcc_assert (stmt != NULL);
1213 if (is_gimple_assign (stmt))
1214 {
1215 ssa_name_1 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
1216 if (ssa_name_1 && ssa_name_1 == t2)
1217 return true;
1218 }
1219
1220 /* Check to see if t2 is expressed/defined with t1. */
1221 stmt = SSA_NAME_DEF_STMT (t2);
1222 gcc_assert (stmt != NULL);
1223 if (is_gimple_assign (stmt))
1224 {
1225 ssa_name_2 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
1226 if (ssa_name_2 && ssa_name_2 == t1)
1227 return true;
1228 }
1229
1230 /* Compare if t1 and t2's def_stmts are identical. */
1231 if (ssa_name_2 != NULL && ssa_name_1 == ssa_name_2)
1232 return true;
1233 else
1234 return false;
1235 }
1236
1237 /* Predict branch probability of BB when BB contains a branch that compares
1238 an induction variable in LOOP with LOOP_IV_BASE_VAR to LOOP_BOUND_VAR. The
1239 loop exit is compared using LOOP_BOUND_CODE, with step of LOOP_BOUND_STEP.
1240
1241 E.g.
1242 for (int i = 0; i < bound; i++) {
1243 if (i < bound - 2)
1244 computation_1();
1245 else
1246 computation_2();
1247 }
1248
1249 In this loop, we will predict the branch inside the loop to be taken. */
1250
1251 static void
1252 predict_iv_comparison (struct loop *loop, basic_block bb,
1253 tree loop_bound_var,
1254 tree loop_iv_base_var,
1255 enum tree_code loop_bound_code,
1256 int loop_bound_step)
1257 {
1258 gimple stmt;
1259 tree compare_var, compare_base;
1260 enum tree_code compare_code;
1261 tree compare_step_var;
1262 edge then_edge;
1263 edge_iterator ei;
1264
1265 if (predicted_by_p (bb, PRED_LOOP_ITERATIONS_GUESSED)
1266 || predicted_by_p (bb, PRED_LOOP_ITERATIONS)
1267 || predicted_by_p (bb, PRED_LOOP_EXIT))
1268 return;
1269
1270 stmt = last_stmt (bb);
1271 if (!stmt || gimple_code (stmt) != GIMPLE_COND)
1272 return;
1273 if (!is_comparison_with_loop_invariant_p (stmt, loop, &compare_var,
1274 &compare_code,
1275 &compare_step_var,
1276 &compare_base))
1277 return;
1278
1279 /* Find the taken edge. */
1280 FOR_EACH_EDGE (then_edge, ei, bb->succs)
1281 if (then_edge->flags & EDGE_TRUE_VALUE)
1282 break;
1283
1284 /* When comparing an IV to a loop invariant, NE is more likely to be
1285 taken while EQ is more likely to be not-taken. */
1286 if (compare_code == NE_EXPR)
1287 {
1288 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1289 return;
1290 }
1291 else if (compare_code == EQ_EXPR)
1292 {
1293 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
1294 return;
1295 }
1296
1297 if (!expr_coherent_p (loop_iv_base_var, compare_base))
1298 return;
1299
1300 /* If loop bound, base and compare bound are all constants, we can
1301 calculate the probability directly. */
1302 if (host_integerp (loop_bound_var, 0)
1303 && host_integerp (compare_var, 0)
1304 && host_integerp (compare_base, 0))
1305 {
1306 int probability;
1307 bool of, overflow = false;
1308 double_int mod, compare_count, tem, loop_count;
1309
1310 double_int loop_bound = tree_to_double_int (loop_bound_var);
1311 double_int compare_bound = tree_to_double_int (compare_var);
1312 double_int base = tree_to_double_int (compare_base);
1313 double_int compare_step = tree_to_double_int (compare_step_var);
1314
1315 /* (loop_bound - base) / compare_step */
1316 tem = loop_bound.sub_with_overflow (base, &of);
1317 overflow |= of;
1318 loop_count = tem.divmod_with_overflow (compare_step,
1319 0, TRUNC_DIV_EXPR,
1320 &mod, &of);
1321 overflow |= of;
1322
1323 if ((!compare_step.is_negative ())
1324 ^ (compare_code == LT_EXPR || compare_code == LE_EXPR))
1325 {
1326 /* (loop_bound - compare_bound) / compare_step */
1327 tem = loop_bound.sub_with_overflow (compare_bound, &of);
1328 overflow |= of;
1329 compare_count = tem.divmod_with_overflow (compare_step,
1330 0, TRUNC_DIV_EXPR,
1331 &mod, &of);
1332 overflow |= of;
1333 }
1334 else
1335 {
1336 /* (compare_bound - base) / compare_step */
1337 tem = compare_bound.sub_with_overflow (base, &of);
1338 overflow |= of;
1339 compare_count = tem.divmod_with_overflow (compare_step,
1340 0, TRUNC_DIV_EXPR,
1341 &mod, &of);
1342 overflow |= of;
1343 }
1344 if (compare_code == LE_EXPR || compare_code == GE_EXPR)
1345 ++compare_count;
1346 if (loop_bound_code == LE_EXPR || loop_bound_code == GE_EXPR)
1347 ++loop_count;
1348 if (compare_count.is_negative ())
1349 compare_count = double_int_zero;
1350 if (loop_count.is_negative ())
1351 loop_count = double_int_zero;
1352 if (loop_count.is_zero ())
1353 probability = 0;
1354 else if (compare_count.scmp (loop_count) == 1)
1355 probability = REG_BR_PROB_BASE;
1356 else
1357 {
1358 /* If loop_count is too big, such that REG_BR_PROB_BASE * loop_count
1359 could overflow, shift both loop_count and compare_count right
1360 a bit so that it doesn't overflow. Note both counts are known not
1361 to be negative at this point. */
1362 int clz_bits = clz_hwi (loop_count.high);
1363 gcc_assert (REG_BR_PROB_BASE < 32768);
1364 if (clz_bits < 16)
1365 {
1366 loop_count.arshift (16 - clz_bits, HOST_BITS_PER_DOUBLE_INT);
1367 compare_count.arshift (16 - clz_bits, HOST_BITS_PER_DOUBLE_INT);
1368 }
1369 tem = compare_count.mul_with_sign (double_int::from_shwi
1370 (REG_BR_PROB_BASE), true, &of);
1371 gcc_assert (!of);
1372 tem = tem.divmod (loop_count, true, TRUNC_DIV_EXPR, &mod);
1373 probability = tem.to_uhwi ();
1374 }
1375
1376 if (!overflow)
1377 predict_edge (then_edge, PRED_LOOP_IV_COMPARE, probability);
1378
1379 return;
1380 }
1381
1382 if (expr_coherent_p (loop_bound_var, compare_var))
1383 {
1384 if ((loop_bound_code == LT_EXPR || loop_bound_code == LE_EXPR)
1385 && (compare_code == LT_EXPR || compare_code == LE_EXPR))
1386 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1387 else if ((loop_bound_code == GT_EXPR || loop_bound_code == GE_EXPR)
1388 && (compare_code == GT_EXPR || compare_code == GE_EXPR))
1389 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1390 else if (loop_bound_code == NE_EXPR)
1391 {
1392 /* If the loop backedge condition is "(i != bound)", we do
1393 the comparison based on the step of IV:
1394 * step < 0 : backedge condition is like (i > bound)
1395 * step > 0 : backedge condition is like (i < bound) */
1396 gcc_assert (loop_bound_step != 0);
1397 if (loop_bound_step > 0
1398 && (compare_code == LT_EXPR
1399 || compare_code == LE_EXPR))
1400 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1401 else if (loop_bound_step < 0
1402 && (compare_code == GT_EXPR
1403 || compare_code == GE_EXPR))
1404 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1405 else
1406 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
1407 }
1408 else
1409 /* The branch is predicted not-taken if loop_bound_code is
1410 opposite with compare_code. */
1411 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
1412 }
1413 else if (expr_coherent_p (loop_iv_base_var, compare_var))
1414 {
1415 /* For cases like:
1416 for (i = s; i < h; i++)
1417 if (i > s + 2) ....
1418 The branch should be predicted taken. */
1419 if (loop_bound_step > 0
1420 && (compare_code == GT_EXPR || compare_code == GE_EXPR))
1421 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1422 else if (loop_bound_step < 0
1423 && (compare_code == LT_EXPR || compare_code == LE_EXPR))
1424 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
1425 else
1426 predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
1427 }
1428 }
1429
1430 /* Predict for extra loop exits that will lead to EXIT_EDGE. The extra loop
1431 exits are resulted from short-circuit conditions that will generate an
1432 if_tmp. E.g.:
1433
1434 if (foo() || global > 10)
1435 break;
1436
1437 This will be translated into:
1438
1439 BB3:
1440 loop header...
1441 BB4:
1442 if foo() goto BB6 else goto BB5
1443 BB5:
1444 if global > 10 goto BB6 else goto BB7
1445 BB6:
1446 goto BB7
1447 BB7:
1448 iftmp = (PHI 0(BB5), 1(BB6))
1449 if iftmp == 1 goto BB8 else goto BB3
1450 BB8:
1451 outside of the loop...
1452
1453 The edge BB7->BB8 is loop exit because BB8 is outside of the loop.
1454 From the dataflow, we can infer that BB4->BB6 and BB5->BB6 are also loop
1455 exits. This function takes BB7->BB8 as input, and finds out the extra loop
1456 exits to predict them using PRED_LOOP_EXIT. */
1457
1458 static void
1459 predict_extra_loop_exits (edge exit_edge)
1460 {
1461 unsigned i;
1462 bool check_value_one;
1463 gimple phi_stmt;
1464 tree cmp_rhs, cmp_lhs;
1465 gimple cmp_stmt = last_stmt (exit_edge->src);
1466
1467 if (!cmp_stmt || gimple_code (cmp_stmt) != GIMPLE_COND)
1468 return;
1469 cmp_rhs = gimple_cond_rhs (cmp_stmt);
1470 cmp_lhs = gimple_cond_lhs (cmp_stmt);
1471 if (!TREE_CONSTANT (cmp_rhs)
1472 || !(integer_zerop (cmp_rhs) || integer_onep (cmp_rhs)))
1473 return;
1474 if (TREE_CODE (cmp_lhs) != SSA_NAME)
1475 return;
1476
1477 /* If check_value_one is true, only the phi_args with value '1' will lead
1478 to loop exit. Otherwise, only the phi_args with value '0' will lead to
1479 loop exit. */
1480 check_value_one = (((integer_onep (cmp_rhs))
1481 ^ (gimple_cond_code (cmp_stmt) == EQ_EXPR))
1482 ^ ((exit_edge->flags & EDGE_TRUE_VALUE) != 0));
1483
1484 phi_stmt = SSA_NAME_DEF_STMT (cmp_lhs);
1485 if (!phi_stmt || gimple_code (phi_stmt) != GIMPLE_PHI)
1486 return;
1487
1488 for (i = 0; i < gimple_phi_num_args (phi_stmt); i++)
1489 {
1490 edge e1;
1491 edge_iterator ei;
1492 tree val = gimple_phi_arg_def (phi_stmt, i);
1493 edge e = gimple_phi_arg_edge (phi_stmt, i);
1494
1495 if (!TREE_CONSTANT (val) || !(integer_zerop (val) || integer_onep (val)))
1496 continue;
1497 if ((check_value_one ^ integer_onep (val)) == 1)
1498 continue;
1499 if (EDGE_COUNT (e->src->succs) != 1)
1500 {
1501 predict_paths_leading_to_edge (e, PRED_LOOP_EXIT, NOT_TAKEN);
1502 continue;
1503 }
1504
1505 FOR_EACH_EDGE (e1, ei, e->src->preds)
1506 predict_paths_leading_to_edge (e1, PRED_LOOP_EXIT, NOT_TAKEN);
1507 }
1508 }
1509
1510 /* Predict edge probabilities by exploiting loop structure. */
1511
1512 static void
1513 predict_loops (void)
1514 {
1515 loop_iterator li;
1516 struct loop *loop;
1517
1518 /* Try to predict out blocks in a loop that are not part of a
1519 natural loop. */
1520 FOR_EACH_LOOP (li, loop, 0)
1521 {
1522 basic_block bb, *bbs;
1523 unsigned j, n_exits;
1524 vec<edge> exits;
1525 struct tree_niter_desc niter_desc;
1526 edge ex;
1527 struct nb_iter_bound *nb_iter;
1528 enum tree_code loop_bound_code = ERROR_MARK;
1529 tree loop_bound_step = NULL;
1530 tree loop_bound_var = NULL;
1531 tree loop_iv_base = NULL;
1532 gimple stmt = NULL;
1533
1534 exits = get_loop_exit_edges (loop);
1535 n_exits = exits.length ();
1536 if (!n_exits)
1537 {
1538 exits.release ();
1539 continue;
1540 }
1541
1542 FOR_EACH_VEC_ELT (exits, j, ex)
1543 {
1544 tree niter = NULL;
1545 HOST_WIDE_INT nitercst;
1546 int max = PARAM_VALUE (PARAM_MAX_PREDICTED_ITERATIONS);
1547 int probability;
1548 enum br_predictor predictor;
1549
1550 predict_extra_loop_exits (ex);
1551
1552 if (number_of_iterations_exit (loop, ex, &niter_desc, false, false))
1553 niter = niter_desc.niter;
1554 if (!niter || TREE_CODE (niter_desc.niter) != INTEGER_CST)
1555 niter = loop_niter_by_eval (loop, ex);
1556
1557 if (TREE_CODE (niter) == INTEGER_CST)
1558 {
1559 if (host_integerp (niter, 1)
1560 && max
1561 && compare_tree_int (niter, max - 1) == -1)
1562 nitercst = tree_low_cst (niter, 1) + 1;
1563 else
1564 nitercst = max;
1565 predictor = PRED_LOOP_ITERATIONS;
1566 }
1567 /* If we have just one exit and we can derive some information about
1568 the number of iterations of the loop from the statements inside
1569 the loop, use it to predict this exit. */
1570 else if (n_exits == 1)
1571 {
1572 nitercst = estimated_stmt_executions_int (loop);
1573 if (nitercst < 0)
1574 continue;
1575 if (nitercst > max)
1576 nitercst = max;
1577
1578 predictor = PRED_LOOP_ITERATIONS_GUESSED;
1579 }
1580 else
1581 continue;
1582
1583 /* If the prediction for number of iterations is zero, do not
1584 predict the exit edges. */
1585 if (nitercst == 0)
1586 continue;
1587
1588 probability = ((REG_BR_PROB_BASE + nitercst / 2) / nitercst);
1589 predict_edge (ex, predictor, probability);
1590 }
1591 exits.release ();
1592
1593 /* Find information about loop bound variables. */
1594 for (nb_iter = loop->bounds; nb_iter;
1595 nb_iter = nb_iter->next)
1596 if (nb_iter->stmt
1597 && gimple_code (nb_iter->stmt) == GIMPLE_COND)
1598 {
1599 stmt = nb_iter->stmt;
1600 break;
1601 }
1602 if (!stmt && last_stmt (loop->header)
1603 && gimple_code (last_stmt (loop->header)) == GIMPLE_COND)
1604 stmt = last_stmt (loop->header);
1605 if (stmt)
1606 is_comparison_with_loop_invariant_p (stmt, loop,
1607 &loop_bound_var,
1608 &loop_bound_code,
1609 &loop_bound_step,
1610 &loop_iv_base);
1611
1612 bbs = get_loop_body (loop);
1613
1614 for (j = 0; j < loop->num_nodes; j++)
1615 {
1616 int header_found = 0;
1617 edge e;
1618 edge_iterator ei;
1619
1620 bb = bbs[j];
1621
1622 /* Bypass loop heuristics on continue statement. These
1623 statements construct loops via "non-loop" constructs
1624 in the source language and are better to be handled
1625 separately. */
1626 if (predicted_by_p (bb, PRED_CONTINUE))
1627 continue;
1628
1629 /* Loop branch heuristics - predict an edge back to a
1630 loop's head as taken. */
1631 if (bb == loop->latch)
1632 {
1633 e = find_edge (loop->latch, loop->header);
1634 if (e)
1635 {
1636 header_found = 1;
1637 predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN);
1638 }
1639 }
1640
1641 /* Loop exit heuristics - predict an edge exiting the loop if the
1642 conditional has no loop header successors as not taken. */
1643 if (!header_found
1644 /* If we already used more reliable loop exit predictors, do not
1645 bother with PRED_LOOP_EXIT. */
1646 && !predicted_by_p (bb, PRED_LOOP_ITERATIONS_GUESSED)
1647 && !predicted_by_p (bb, PRED_LOOP_ITERATIONS))
1648 {
1649 /* For loop with many exits we don't want to predict all exits
1650 with the pretty large probability, because if all exits are
1651 considered in row, the loop would be predicted to iterate
1652 almost never. The code to divide probability by number of
1653 exits is very rough. It should compute the number of exits
1654 taken in each patch through function (not the overall number
1655 of exits that might be a lot higher for loops with wide switch
1656 statements in them) and compute n-th square root.
1657
1658 We limit the minimal probability by 2% to avoid
1659 EDGE_PROBABILITY_RELIABLE from trusting the branch prediction
1660 as this was causing regression in perl benchmark containing such
1661 a wide loop. */
1662
1663 int probability = ((REG_BR_PROB_BASE
1664 - predictor_info [(int) PRED_LOOP_EXIT].hitrate)
1665 / n_exits);
1666 if (probability < HITRATE (2))
1667 probability = HITRATE (2);
1668 FOR_EACH_EDGE (e, ei, bb->succs)
1669 if (e->dest->index < NUM_FIXED_BLOCKS
1670 || !flow_bb_inside_loop_p (loop, e->dest))
1671 predict_edge (e, PRED_LOOP_EXIT, probability);
1672 }
1673 if (loop_bound_var)
1674 predict_iv_comparison (loop, bb, loop_bound_var, loop_iv_base,
1675 loop_bound_code,
1676 tree_low_cst (loop_bound_step, 0));
1677 }
1678
1679 /* Free basic blocks from get_loop_body. */
1680 free (bbs);
1681 }
1682 }
1683
1684 /* Attempt to predict probabilities of BB outgoing edges using local
1685 properties. */
1686 static void
1687 bb_estimate_probability_locally (basic_block bb)
1688 {
1689 rtx last_insn = BB_END (bb);
1690 rtx cond;
1691
1692 if (! can_predict_insn_p (last_insn))
1693 return;
1694 cond = get_condition (last_insn, NULL, false, false);
1695 if (! cond)
1696 return;
1697
1698 /* Try "pointer heuristic."
1699 A comparison ptr == 0 is predicted as false.
1700 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1701 if (COMPARISON_P (cond)
1702 && ((REG_P (XEXP (cond, 0)) && REG_POINTER (XEXP (cond, 0)))
1703 || (REG_P (XEXP (cond, 1)) && REG_POINTER (XEXP (cond, 1)))))
1704 {
1705 if (GET_CODE (cond) == EQ)
1706 predict_insn_def (last_insn, PRED_POINTER, NOT_TAKEN);
1707 else if (GET_CODE (cond) == NE)
1708 predict_insn_def (last_insn, PRED_POINTER, TAKEN);
1709 }
1710 else
1711
1712 /* Try "opcode heuristic."
1713 EQ tests are usually false and NE tests are usually true. Also,
1714 most quantities are positive, so we can make the appropriate guesses
1715 about signed comparisons against zero. */
1716 switch (GET_CODE (cond))
1717 {
1718 case CONST_INT:
1719 /* Unconditional branch. */
1720 predict_insn_def (last_insn, PRED_UNCONDITIONAL,
1721 cond == const0_rtx ? NOT_TAKEN : TAKEN);
1722 break;
1723
1724 case EQ:
1725 case UNEQ:
1726 /* Floating point comparisons appears to behave in a very
1727 unpredictable way because of special role of = tests in
1728 FP code. */
1729 if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0))))
1730 ;
1731 /* Comparisons with 0 are often used for booleans and there is
1732 nothing useful to predict about them. */
1733 else if (XEXP (cond, 1) == const0_rtx
1734 || XEXP (cond, 0) == const0_rtx)
1735 ;
1736 else
1737 predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, NOT_TAKEN);
1738 break;
1739
1740 case NE:
1741 case LTGT:
1742 /* Floating point comparisons appears to behave in a very
1743 unpredictable way because of special role of = tests in
1744 FP code. */
1745 if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0))))
1746 ;
1747 /* Comparisons with 0 are often used for booleans and there is
1748 nothing useful to predict about them. */
1749 else if (XEXP (cond, 1) == const0_rtx
1750 || XEXP (cond, 0) == const0_rtx)
1751 ;
1752 else
1753 predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, TAKEN);
1754 break;
1755
1756 case ORDERED:
1757 predict_insn_def (last_insn, PRED_FPOPCODE, TAKEN);
1758 break;
1759
1760 case UNORDERED:
1761 predict_insn_def (last_insn, PRED_FPOPCODE, NOT_TAKEN);
1762 break;
1763
1764 case LE:
1765 case LT:
1766 if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx
1767 || XEXP (cond, 1) == constm1_rtx)
1768 predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, NOT_TAKEN);
1769 break;
1770
1771 case GE:
1772 case GT:
1773 if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx
1774 || XEXP (cond, 1) == constm1_rtx)
1775 predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, TAKEN);
1776 break;
1777
1778 default:
1779 break;
1780 }
1781 }
1782
1783 /* Set edge->probability for each successor edge of BB. */
1784 void
1785 guess_outgoing_edge_probabilities (basic_block bb)
1786 {
1787 bb_estimate_probability_locally (bb);
1788 combine_predictions_for_insn (BB_END (bb), bb);
1789 }
1790 \f
1791 static tree expr_expected_value (tree, bitmap);
1792
1793 /* Helper function for expr_expected_value. */
1794
1795 static tree
1796 expr_expected_value_1 (tree type, tree op0, enum tree_code code,
1797 tree op1, bitmap visited)
1798 {
1799 gimple def;
1800
1801 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1802 {
1803 if (TREE_CONSTANT (op0))
1804 return op0;
1805
1806 if (code != SSA_NAME)
1807 return NULL_TREE;
1808
1809 def = SSA_NAME_DEF_STMT (op0);
1810
1811 /* If we were already here, break the infinite cycle. */
1812 if (!bitmap_set_bit (visited, SSA_NAME_VERSION (op0)))
1813 return NULL;
1814
1815 if (gimple_code (def) == GIMPLE_PHI)
1816 {
1817 /* All the arguments of the PHI node must have the same constant
1818 length. */
1819 int i, n = gimple_phi_num_args (def);
1820 tree val = NULL, new_val;
1821
1822 for (i = 0; i < n; i++)
1823 {
1824 tree arg = PHI_ARG_DEF (def, i);
1825
1826 /* If this PHI has itself as an argument, we cannot
1827 determine the string length of this argument. However,
1828 if we can find an expected constant value for the other
1829 PHI args then we can still be sure that this is
1830 likely a constant. So be optimistic and just
1831 continue with the next argument. */
1832 if (arg == PHI_RESULT (def))
1833 continue;
1834
1835 new_val = expr_expected_value (arg, visited);
1836 if (!new_val)
1837 return NULL;
1838 if (!val)
1839 val = new_val;
1840 else if (!operand_equal_p (val, new_val, false))
1841 return NULL;
1842 }
1843 return val;
1844 }
1845 if (is_gimple_assign (def))
1846 {
1847 if (gimple_assign_lhs (def) != op0)
1848 return NULL;
1849
1850 return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def)),
1851 gimple_assign_rhs1 (def),
1852 gimple_assign_rhs_code (def),
1853 gimple_assign_rhs2 (def),
1854 visited);
1855 }
1856
1857 if (is_gimple_call (def))
1858 {
1859 tree decl = gimple_call_fndecl (def);
1860 if (!decl)
1861 return NULL;
1862 if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
1863 switch (DECL_FUNCTION_CODE (decl))
1864 {
1865 case BUILT_IN_EXPECT:
1866 {
1867 tree val;
1868 if (gimple_call_num_args (def) != 2)
1869 return NULL;
1870 val = gimple_call_arg (def, 0);
1871 if (TREE_CONSTANT (val))
1872 return val;
1873 return gimple_call_arg (def, 1);
1874 }
1875
1876 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N:
1877 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1:
1878 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2:
1879 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4:
1880 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8:
1881 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16:
1882 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE:
1883 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N:
1884 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1:
1885 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2:
1886 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4:
1887 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8:
1888 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16:
1889 /* Assume that any given atomic operation has low contention,
1890 and thus the compare-and-swap operation succeeds. */
1891 return boolean_true_node;
1892 }
1893 }
1894
1895 return NULL;
1896 }
1897
1898 if (get_gimple_rhs_class (code) == GIMPLE_BINARY_RHS)
1899 {
1900 tree res;
1901 op0 = expr_expected_value (op0, visited);
1902 if (!op0)
1903 return NULL;
1904 op1 = expr_expected_value (op1, visited);
1905 if (!op1)
1906 return NULL;
1907 res = fold_build2 (code, type, op0, op1);
1908 if (TREE_CONSTANT (res))
1909 return res;
1910 return NULL;
1911 }
1912 if (get_gimple_rhs_class (code) == GIMPLE_UNARY_RHS)
1913 {
1914 tree res;
1915 op0 = expr_expected_value (op0, visited);
1916 if (!op0)
1917 return NULL;
1918 res = fold_build1 (code, type, op0);
1919 if (TREE_CONSTANT (res))
1920 return res;
1921 return NULL;
1922 }
1923 return NULL;
1924 }
1925
1926 /* Return constant EXPR will likely have at execution time, NULL if unknown.
1927 The function is used by builtin_expect branch predictor so the evidence
1928 must come from this construct and additional possible constant folding.
1929
1930 We may want to implement more involved value guess (such as value range
1931 propagation based prediction), but such tricks shall go to new
1932 implementation. */
1933
1934 static tree
1935 expr_expected_value (tree expr, bitmap visited)
1936 {
1937 enum tree_code code;
1938 tree op0, op1;
1939
1940 if (TREE_CONSTANT (expr))
1941 return expr;
1942
1943 extract_ops_from_tree (expr, &code, &op0, &op1);
1944 return expr_expected_value_1 (TREE_TYPE (expr),
1945 op0, code, op1, visited);
1946 }
1947
1948 \f
1949 /* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements
1950 we no longer need. */
1951 static unsigned int
1952 strip_predict_hints (void)
1953 {
1954 basic_block bb;
1955 gimple ass_stmt;
1956 tree var;
1957
1958 FOR_EACH_BB (bb)
1959 {
1960 gimple_stmt_iterator bi;
1961 for (bi = gsi_start_bb (bb); !gsi_end_p (bi);)
1962 {
1963 gimple stmt = gsi_stmt (bi);
1964
1965 if (gimple_code (stmt) == GIMPLE_PREDICT)
1966 {
1967 gsi_remove (&bi, true);
1968 continue;
1969 }
1970 else if (gimple_code (stmt) == GIMPLE_CALL)
1971 {
1972 tree fndecl = gimple_call_fndecl (stmt);
1973
1974 if (fndecl
1975 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
1976 && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_EXPECT
1977 && gimple_call_num_args (stmt) == 2)
1978 {
1979 var = gimple_call_lhs (stmt);
1980 if (var)
1981 {
1982 ass_stmt
1983 = gimple_build_assign (var, gimple_call_arg (stmt, 0));
1984 gsi_replace (&bi, ass_stmt, true);
1985 }
1986 else
1987 {
1988 gsi_remove (&bi, true);
1989 continue;
1990 }
1991 }
1992 }
1993 gsi_next (&bi);
1994 }
1995 }
1996 return 0;
1997 }
1998 \f
1999 /* Predict using opcode of the last statement in basic block. */
2000 static void
2001 tree_predict_by_opcode (basic_block bb)
2002 {
2003 gimple stmt = last_stmt (bb);
2004 edge then_edge;
2005 tree op0, op1;
2006 tree type;
2007 tree val;
2008 enum tree_code cmp;
2009 bitmap visited;
2010 edge_iterator ei;
2011
2012 if (!stmt || gimple_code (stmt) != GIMPLE_COND)
2013 return;
2014 FOR_EACH_EDGE (then_edge, ei, bb->succs)
2015 if (then_edge->flags & EDGE_TRUE_VALUE)
2016 break;
2017 op0 = gimple_cond_lhs (stmt);
2018 op1 = gimple_cond_rhs (stmt);
2019 cmp = gimple_cond_code (stmt);
2020 type = TREE_TYPE (op0);
2021 visited = BITMAP_ALLOC (NULL);
2022 val = expr_expected_value_1 (boolean_type_node, op0, cmp, op1, visited);
2023 BITMAP_FREE (visited);
2024 if (val)
2025 {
2026 int percent = PARAM_VALUE (BUILTIN_EXPECT_PROBABILITY);
2027
2028 gcc_assert (percent >= 0 && percent <= 100);
2029 if (integer_zerop (val))
2030 percent = 100 - percent;
2031 predict_edge (then_edge, PRED_BUILTIN_EXPECT, HITRATE (percent));
2032 }
2033 /* Try "pointer heuristic."
2034 A comparison ptr == 0 is predicted as false.
2035 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
2036 if (POINTER_TYPE_P (type))
2037 {
2038 if (cmp == EQ_EXPR)
2039 predict_edge_def (then_edge, PRED_TREE_POINTER, NOT_TAKEN);
2040 else if (cmp == NE_EXPR)
2041 predict_edge_def (then_edge, PRED_TREE_POINTER, TAKEN);
2042 }
2043 else
2044
2045 /* Try "opcode heuristic."
2046 EQ tests are usually false and NE tests are usually true. Also,
2047 most quantities are positive, so we can make the appropriate guesses
2048 about signed comparisons against zero. */
2049 switch (cmp)
2050 {
2051 case EQ_EXPR:
2052 case UNEQ_EXPR:
2053 /* Floating point comparisons appears to behave in a very
2054 unpredictable way because of special role of = tests in
2055 FP code. */
2056 if (FLOAT_TYPE_P (type))
2057 ;
2058 /* Comparisons with 0 are often used for booleans and there is
2059 nothing useful to predict about them. */
2060 else if (integer_zerop (op0) || integer_zerop (op1))
2061 ;
2062 else
2063 predict_edge_def (then_edge, PRED_TREE_OPCODE_NONEQUAL, NOT_TAKEN);
2064 break;
2065
2066 case NE_EXPR:
2067 case LTGT_EXPR:
2068 /* Floating point comparisons appears to behave in a very
2069 unpredictable way because of special role of = tests in
2070 FP code. */
2071 if (FLOAT_TYPE_P (type))
2072 ;
2073 /* Comparisons with 0 are often used for booleans and there is
2074 nothing useful to predict about them. */
2075 else if (integer_zerop (op0)
2076 || integer_zerop (op1))
2077 ;
2078 else
2079 predict_edge_def (then_edge, PRED_TREE_OPCODE_NONEQUAL, TAKEN);
2080 break;
2081
2082 case ORDERED_EXPR:
2083 predict_edge_def (then_edge, PRED_TREE_FPOPCODE, TAKEN);
2084 break;
2085
2086 case UNORDERED_EXPR:
2087 predict_edge_def (then_edge, PRED_TREE_FPOPCODE, NOT_TAKEN);
2088 break;
2089
2090 case LE_EXPR:
2091 case LT_EXPR:
2092 if (integer_zerop (op1)
2093 || integer_onep (op1)
2094 || integer_all_onesp (op1)
2095 || real_zerop (op1)
2096 || real_onep (op1)
2097 || real_minus_onep (op1))
2098 predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, NOT_TAKEN);
2099 break;
2100
2101 case GE_EXPR:
2102 case GT_EXPR:
2103 if (integer_zerop (op1)
2104 || integer_onep (op1)
2105 || integer_all_onesp (op1)
2106 || real_zerop (op1)
2107 || real_onep (op1)
2108 || real_minus_onep (op1))
2109 predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, TAKEN);
2110 break;
2111
2112 default:
2113 break;
2114 }
2115 }
2116
2117 /* Try to guess whether the value of return means error code. */
2118
2119 static enum br_predictor
2120 return_prediction (tree val, enum prediction *prediction)
2121 {
2122 /* VOID. */
2123 if (!val)
2124 return PRED_NO_PREDICTION;
2125 /* Different heuristics for pointers and scalars. */
2126 if (POINTER_TYPE_P (TREE_TYPE (val)))
2127 {
2128 /* NULL is usually not returned. */
2129 if (integer_zerop (val))
2130 {
2131 *prediction = NOT_TAKEN;
2132 return PRED_NULL_RETURN;
2133 }
2134 }
2135 else if (INTEGRAL_TYPE_P (TREE_TYPE (val)))
2136 {
2137 /* Negative return values are often used to indicate
2138 errors. */
2139 if (TREE_CODE (val) == INTEGER_CST
2140 && tree_int_cst_sgn (val) < 0)
2141 {
2142 *prediction = NOT_TAKEN;
2143 return PRED_NEGATIVE_RETURN;
2144 }
2145 /* Constant return values seems to be commonly taken.
2146 Zero/one often represent booleans so exclude them from the
2147 heuristics. */
2148 if (TREE_CONSTANT (val)
2149 && (!integer_zerop (val) && !integer_onep (val)))
2150 {
2151 *prediction = TAKEN;
2152 return PRED_CONST_RETURN;
2153 }
2154 }
2155 return PRED_NO_PREDICTION;
2156 }
2157
2158 /* Find the basic block with return expression and look up for possible
2159 return value trying to apply RETURN_PREDICTION heuristics. */
2160 static void
2161 apply_return_prediction (void)
2162 {
2163 gimple return_stmt = NULL;
2164 tree return_val;
2165 edge e;
2166 gimple phi;
2167 int phi_num_args, i;
2168 enum br_predictor pred;
2169 enum prediction direction;
2170 edge_iterator ei;
2171
2172 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
2173 {
2174 return_stmt = last_stmt (e->src);
2175 if (return_stmt
2176 && gimple_code (return_stmt) == GIMPLE_RETURN)
2177 break;
2178 }
2179 if (!e)
2180 return;
2181 return_val = gimple_return_retval (return_stmt);
2182 if (!return_val)
2183 return;
2184 if (TREE_CODE (return_val) != SSA_NAME
2185 || !SSA_NAME_DEF_STMT (return_val)
2186 || gimple_code (SSA_NAME_DEF_STMT (return_val)) != GIMPLE_PHI)
2187 return;
2188 phi = SSA_NAME_DEF_STMT (return_val);
2189 phi_num_args = gimple_phi_num_args (phi);
2190 pred = return_prediction (PHI_ARG_DEF (phi, 0), &direction);
2191
2192 /* Avoid the degenerate case where all return values form the function
2193 belongs to same category (ie they are all positive constants)
2194 so we can hardly say something about them. */
2195 for (i = 1; i < phi_num_args; i++)
2196 if (pred != return_prediction (PHI_ARG_DEF (phi, i), &direction))
2197 break;
2198 if (i != phi_num_args)
2199 for (i = 0; i < phi_num_args; i++)
2200 {
2201 pred = return_prediction (PHI_ARG_DEF (phi, i), &direction);
2202 if (pred != PRED_NO_PREDICTION)
2203 predict_paths_leading_to_edge (gimple_phi_arg_edge (phi, i), pred,
2204 direction);
2205 }
2206 }
2207
2208 /* Look for basic block that contains unlikely to happen events
2209 (such as noreturn calls) and mark all paths leading to execution
2210 of this basic blocks as unlikely. */
2211
2212 static void
2213 tree_bb_level_predictions (void)
2214 {
2215 basic_block bb;
2216 bool has_return_edges = false;
2217 edge e;
2218 edge_iterator ei;
2219
2220 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
2221 if (!(e->flags & (EDGE_ABNORMAL | EDGE_FAKE | EDGE_EH)))
2222 {
2223 has_return_edges = true;
2224 break;
2225 }
2226
2227 apply_return_prediction ();
2228
2229 FOR_EACH_BB (bb)
2230 {
2231 gimple_stmt_iterator gsi;
2232
2233 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2234 {
2235 gimple stmt = gsi_stmt (gsi);
2236 tree decl;
2237
2238 if (is_gimple_call (stmt))
2239 {
2240 if ((gimple_call_flags (stmt) & ECF_NORETURN)
2241 && has_return_edges)
2242 predict_paths_leading_to (bb, PRED_NORETURN,
2243 NOT_TAKEN);
2244 decl = gimple_call_fndecl (stmt);
2245 if (decl
2246 && lookup_attribute ("cold",
2247 DECL_ATTRIBUTES (decl)))
2248 predict_paths_leading_to (bb, PRED_COLD_FUNCTION,
2249 NOT_TAKEN);
2250 }
2251 else if (gimple_code (stmt) == GIMPLE_PREDICT)
2252 {
2253 predict_paths_leading_to (bb, gimple_predict_predictor (stmt),
2254 gimple_predict_outcome (stmt));
2255 /* Keep GIMPLE_PREDICT around so early inlining will propagate
2256 hints to callers. */
2257 }
2258 }
2259 }
2260 }
2261
2262 #ifdef ENABLE_CHECKING
2263
2264 /* Callback for pointer_map_traverse, asserts that the pointer map is
2265 empty. */
2266
2267 static bool
2268 assert_is_empty (const void *key ATTRIBUTE_UNUSED, void **value,
2269 void *data ATTRIBUTE_UNUSED)
2270 {
2271 gcc_assert (!*value);
2272 return false;
2273 }
2274 #endif
2275
2276 /* Predict branch probabilities and estimate profile for basic block BB. */
2277
2278 static void
2279 tree_estimate_probability_bb (basic_block bb)
2280 {
2281 edge e;
2282 edge_iterator ei;
2283 gimple last;
2284
2285 FOR_EACH_EDGE (e, ei, bb->succs)
2286 {
2287 /* Predict edges to user labels with attributes. */
2288 if (e->dest != EXIT_BLOCK_PTR)
2289 {
2290 gimple_stmt_iterator gi;
2291 for (gi = gsi_start_bb (e->dest); !gsi_end_p (gi); gsi_next (&gi))
2292 {
2293 gimple stmt = gsi_stmt (gi);
2294 tree decl;
2295
2296 if (gimple_code (stmt) != GIMPLE_LABEL)
2297 break;
2298 decl = gimple_label_label (stmt);
2299 if (DECL_ARTIFICIAL (decl))
2300 continue;
2301
2302 /* Finally, we have a user-defined label. */
2303 if (lookup_attribute ("cold", DECL_ATTRIBUTES (decl)))
2304 predict_edge_def (e, PRED_COLD_LABEL, NOT_TAKEN);
2305 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (decl)))
2306 predict_edge_def (e, PRED_HOT_LABEL, TAKEN);
2307 }
2308 }
2309
2310 /* Predict early returns to be probable, as we've already taken
2311 care for error returns and other cases are often used for
2312 fast paths through function.
2313
2314 Since we've already removed the return statements, we are
2315 looking for CFG like:
2316
2317 if (conditional)
2318 {
2319 ..
2320 goto return_block
2321 }
2322 some other blocks
2323 return_block:
2324 return_stmt. */
2325 if (e->dest != bb->next_bb
2326 && e->dest != EXIT_BLOCK_PTR
2327 && single_succ_p (e->dest)
2328 && single_succ_edge (e->dest)->dest == EXIT_BLOCK_PTR
2329 && (last = last_stmt (e->dest)) != NULL
2330 && gimple_code (last) == GIMPLE_RETURN)
2331 {
2332 edge e1;
2333 edge_iterator ei1;
2334
2335 if (single_succ_p (bb))
2336 {
2337 FOR_EACH_EDGE (e1, ei1, bb->preds)
2338 if (!predicted_by_p (e1->src, PRED_NULL_RETURN)
2339 && !predicted_by_p (e1->src, PRED_CONST_RETURN)
2340 && !predicted_by_p (e1->src, PRED_NEGATIVE_RETURN))
2341 predict_edge_def (e1, PRED_TREE_EARLY_RETURN, NOT_TAKEN);
2342 }
2343 else
2344 if (!predicted_by_p (e->src, PRED_NULL_RETURN)
2345 && !predicted_by_p (e->src, PRED_CONST_RETURN)
2346 && !predicted_by_p (e->src, PRED_NEGATIVE_RETURN))
2347 predict_edge_def (e, PRED_TREE_EARLY_RETURN, NOT_TAKEN);
2348 }
2349
2350 /* Look for block we are guarding (ie we dominate it,
2351 but it doesn't postdominate us). */
2352 if (e->dest != EXIT_BLOCK_PTR && e->dest != bb
2353 && dominated_by_p (CDI_DOMINATORS, e->dest, e->src)
2354 && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e->dest))
2355 {
2356 gimple_stmt_iterator bi;
2357
2358 /* The call heuristic claims that a guarded function call
2359 is improbable. This is because such calls are often used
2360 to signal exceptional situations such as printing error
2361 messages. */
2362 for (bi = gsi_start_bb (e->dest); !gsi_end_p (bi);
2363 gsi_next (&bi))
2364 {
2365 gimple stmt = gsi_stmt (bi);
2366 if (is_gimple_call (stmt)
2367 /* Constant and pure calls are hardly used to signalize
2368 something exceptional. */
2369 && gimple_has_side_effects (stmt))
2370 {
2371 predict_edge_def (e, PRED_CALL, NOT_TAKEN);
2372 break;
2373 }
2374 }
2375 }
2376 }
2377 tree_predict_by_opcode (bb);
2378 }
2379
2380 /* Predict branch probabilities and estimate profile of the tree CFG.
2381 This function can be called from the loop optimizers to recompute
2382 the profile information. */
2383
2384 void
2385 tree_estimate_probability (void)
2386 {
2387 basic_block bb;
2388
2389 add_noreturn_fake_exit_edges ();
2390 connect_infinite_loops_to_exit ();
2391 /* We use loop_niter_by_eval, which requires that the loops have
2392 preheaders. */
2393 create_preheaders (CP_SIMPLE_PREHEADERS);
2394 calculate_dominance_info (CDI_POST_DOMINATORS);
2395
2396 bb_predictions = pointer_map_create ();
2397 tree_bb_level_predictions ();
2398 record_loop_exits ();
2399
2400 if (number_of_loops (cfun) > 1)
2401 predict_loops ();
2402
2403 FOR_EACH_BB (bb)
2404 tree_estimate_probability_bb (bb);
2405
2406 FOR_EACH_BB (bb)
2407 combine_predictions_for_bb (bb);
2408
2409 #ifdef ENABLE_CHECKING
2410 pointer_map_traverse (bb_predictions, assert_is_empty, NULL);
2411 #endif
2412 pointer_map_destroy (bb_predictions);
2413 bb_predictions = NULL;
2414
2415 estimate_bb_frequencies (false);
2416 free_dominance_info (CDI_POST_DOMINATORS);
2417 remove_fake_exit_edges ();
2418 }
2419
2420 /* Predict branch probabilities and estimate profile of the tree CFG.
2421 This is the driver function for PASS_PROFILE. */
2422
2423 static unsigned int
2424 tree_estimate_probability_driver (void)
2425 {
2426 unsigned nb_loops;
2427
2428 loop_optimizer_init (LOOPS_NORMAL);
2429 if (dump_file && (dump_flags & TDF_DETAILS))
2430 flow_loops_dump (dump_file, NULL, 0);
2431
2432 mark_irreducible_loops ();
2433
2434 nb_loops = number_of_loops (cfun);
2435 if (nb_loops > 1)
2436 scev_initialize ();
2437
2438 tree_estimate_probability ();
2439
2440 if (nb_loops > 1)
2441 scev_finalize ();
2442
2443 loop_optimizer_finalize ();
2444 if (dump_file && (dump_flags & TDF_DETAILS))
2445 gimple_dump_cfg (dump_file, dump_flags);
2446 if (profile_status == PROFILE_ABSENT)
2447 profile_status = PROFILE_GUESSED;
2448 return 0;
2449 }
2450 \f
2451 /* Predict edges to successors of CUR whose sources are not postdominated by
2452 BB by PRED and recurse to all postdominators. */
2453
2454 static void
2455 predict_paths_for_bb (basic_block cur, basic_block bb,
2456 enum br_predictor pred,
2457 enum prediction taken,
2458 bitmap visited)
2459 {
2460 edge e;
2461 edge_iterator ei;
2462 basic_block son;
2463
2464 /* We are looking for all edges forming edge cut induced by
2465 set of all blocks postdominated by BB. */
2466 FOR_EACH_EDGE (e, ei, cur->preds)
2467 if (e->src->index >= NUM_FIXED_BLOCKS
2468 && !dominated_by_p (CDI_POST_DOMINATORS, e->src, bb))
2469 {
2470 edge e2;
2471 edge_iterator ei2;
2472 bool found = false;
2473
2474 /* Ignore fake edges and eh, we predict them as not taken anyway. */
2475 if (e->flags & (EDGE_EH | EDGE_FAKE))
2476 continue;
2477 gcc_assert (bb == cur || dominated_by_p (CDI_POST_DOMINATORS, cur, bb));
2478
2479 /* See if there is an edge from e->src that is not abnormal
2480 and does not lead to BB. */
2481 FOR_EACH_EDGE (e2, ei2, e->src->succs)
2482 if (e2 != e
2483 && !(e2->flags & (EDGE_EH | EDGE_FAKE))
2484 && !dominated_by_p (CDI_POST_DOMINATORS, e2->dest, bb))
2485 {
2486 found = true;
2487 break;
2488 }
2489
2490 /* If there is non-abnormal path leaving e->src, predict edge
2491 using predictor. Otherwise we need to look for paths
2492 leading to e->src.
2493
2494 The second may lead to infinite loop in the case we are predicitng
2495 regions that are only reachable by abnormal edges. We simply
2496 prevent visiting given BB twice. */
2497 if (found)
2498 predict_edge_def (e, pred, taken);
2499 else if (bitmap_set_bit (visited, e->src->index))
2500 predict_paths_for_bb (e->src, e->src, pred, taken, visited);
2501 }
2502 for (son = first_dom_son (CDI_POST_DOMINATORS, cur);
2503 son;
2504 son = next_dom_son (CDI_POST_DOMINATORS, son))
2505 predict_paths_for_bb (son, bb, pred, taken, visited);
2506 }
2507
2508 /* Sets branch probabilities according to PREDiction and
2509 FLAGS. */
2510
2511 static void
2512 predict_paths_leading_to (basic_block bb, enum br_predictor pred,
2513 enum prediction taken)
2514 {
2515 bitmap visited = BITMAP_ALLOC (NULL);
2516 predict_paths_for_bb (bb, bb, pred, taken, visited);
2517 BITMAP_FREE (visited);
2518 }
2519
2520 /* Like predict_paths_leading_to but take edge instead of basic block. */
2521
2522 static void
2523 predict_paths_leading_to_edge (edge e, enum br_predictor pred,
2524 enum prediction taken)
2525 {
2526 bool has_nonloop_edge = false;
2527 edge_iterator ei;
2528 edge e2;
2529
2530 basic_block bb = e->src;
2531 FOR_EACH_EDGE (e2, ei, bb->succs)
2532 if (e2->dest != e->src && e2->dest != e->dest
2533 && !(e->flags & (EDGE_EH | EDGE_FAKE))
2534 && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e2->dest))
2535 {
2536 has_nonloop_edge = true;
2537 break;
2538 }
2539 if (!has_nonloop_edge)
2540 {
2541 bitmap visited = BITMAP_ALLOC (NULL);
2542 predict_paths_for_bb (bb, bb, pred, taken, visited);
2543 BITMAP_FREE (visited);
2544 }
2545 else
2546 predict_edge_def (e, pred, taken);
2547 }
2548 \f
2549 /* This is used to carry information about basic blocks. It is
2550 attached to the AUX field of the standard CFG block. */
2551
2552 typedef struct block_info_def
2553 {
2554 /* Estimated frequency of execution of basic_block. */
2555 sreal frequency;
2556
2557 /* To keep queue of basic blocks to process. */
2558 basic_block next;
2559
2560 /* Number of predecessors we need to visit first. */
2561 int npredecessors;
2562 } *block_info;
2563
2564 /* Similar information for edges. */
2565 typedef struct edge_info_def
2566 {
2567 /* In case edge is a loopback edge, the probability edge will be reached
2568 in case header is. Estimated number of iterations of the loop can be
2569 then computed as 1 / (1 - back_edge_prob). */
2570 sreal back_edge_prob;
2571 /* True if the edge is a loopback edge in the natural loop. */
2572 unsigned int back_edge:1;
2573 } *edge_info;
2574
2575 #define BLOCK_INFO(B) ((block_info) (B)->aux)
2576 #define EDGE_INFO(E) ((edge_info) (E)->aux)
2577
2578 /* Helper function for estimate_bb_frequencies.
2579 Propagate the frequencies in blocks marked in
2580 TOVISIT, starting in HEAD. */
2581
2582 static void
2583 propagate_freq (basic_block head, bitmap tovisit)
2584 {
2585 basic_block bb;
2586 basic_block last;
2587 unsigned i;
2588 edge e;
2589 basic_block nextbb;
2590 bitmap_iterator bi;
2591
2592 /* For each basic block we need to visit count number of his predecessors
2593 we need to visit first. */
2594 EXECUTE_IF_SET_IN_BITMAP (tovisit, 0, i, bi)
2595 {
2596 edge_iterator ei;
2597 int count = 0;
2598
2599 bb = BASIC_BLOCK (i);
2600
2601 FOR_EACH_EDGE (e, ei, bb->preds)
2602 {
2603 bool visit = bitmap_bit_p (tovisit, e->src->index);
2604
2605 if (visit && !(e->flags & EDGE_DFS_BACK))
2606 count++;
2607 else if (visit && dump_file && !EDGE_INFO (e)->back_edge)
2608 fprintf (dump_file,
2609 "Irreducible region hit, ignoring edge to %i->%i\n",
2610 e->src->index, bb->index);
2611 }
2612 BLOCK_INFO (bb)->npredecessors = count;
2613 /* When function never returns, we will never process exit block. */
2614 if (!count && bb == EXIT_BLOCK_PTR)
2615 bb->count = bb->frequency = 0;
2616 }
2617
2618 memcpy (&BLOCK_INFO (head)->frequency, &real_one, sizeof (real_one));
2619 last = head;
2620 for (bb = head; bb; bb = nextbb)
2621 {
2622 edge_iterator ei;
2623 sreal cyclic_probability, frequency;
2624
2625 memcpy (&cyclic_probability, &real_zero, sizeof (real_zero));
2626 memcpy (&frequency, &real_zero, sizeof (real_zero));
2627
2628 nextbb = BLOCK_INFO (bb)->next;
2629 BLOCK_INFO (bb)->next = NULL;
2630
2631 /* Compute frequency of basic block. */
2632 if (bb != head)
2633 {
2634 #ifdef ENABLE_CHECKING
2635 FOR_EACH_EDGE (e, ei, bb->preds)
2636 gcc_assert (!bitmap_bit_p (tovisit, e->src->index)
2637 || (e->flags & EDGE_DFS_BACK));
2638 #endif
2639
2640 FOR_EACH_EDGE (e, ei, bb->preds)
2641 if (EDGE_INFO (e)->back_edge)
2642 {
2643 sreal_add (&cyclic_probability, &cyclic_probability,
2644 &EDGE_INFO (e)->back_edge_prob);
2645 }
2646 else if (!(e->flags & EDGE_DFS_BACK))
2647 {
2648 sreal tmp;
2649
2650 /* frequency += (e->probability
2651 * BLOCK_INFO (e->src)->frequency /
2652 REG_BR_PROB_BASE); */
2653
2654 sreal_init (&tmp, e->probability, 0);
2655 sreal_mul (&tmp, &tmp, &BLOCK_INFO (e->src)->frequency);
2656 sreal_mul (&tmp, &tmp, &real_inv_br_prob_base);
2657 sreal_add (&frequency, &frequency, &tmp);
2658 }
2659
2660 if (sreal_compare (&cyclic_probability, &real_zero) == 0)
2661 {
2662 memcpy (&BLOCK_INFO (bb)->frequency, &frequency,
2663 sizeof (frequency));
2664 }
2665 else
2666 {
2667 if (sreal_compare (&cyclic_probability, &real_almost_one) > 0)
2668 {
2669 memcpy (&cyclic_probability, &real_almost_one,
2670 sizeof (real_almost_one));
2671 }
2672
2673 /* BLOCK_INFO (bb)->frequency = frequency
2674 / (1 - cyclic_probability) */
2675
2676 sreal_sub (&cyclic_probability, &real_one, &cyclic_probability);
2677 sreal_div (&BLOCK_INFO (bb)->frequency,
2678 &frequency, &cyclic_probability);
2679 }
2680 }
2681
2682 bitmap_clear_bit (tovisit, bb->index);
2683
2684 e = find_edge (bb, head);
2685 if (e)
2686 {
2687 sreal tmp;
2688
2689 /* EDGE_INFO (e)->back_edge_prob
2690 = ((e->probability * BLOCK_INFO (bb)->frequency)
2691 / REG_BR_PROB_BASE); */
2692
2693 sreal_init (&tmp, e->probability, 0);
2694 sreal_mul (&tmp, &tmp, &BLOCK_INFO (bb)->frequency);
2695 sreal_mul (&EDGE_INFO (e)->back_edge_prob,
2696 &tmp, &real_inv_br_prob_base);
2697 }
2698
2699 /* Propagate to successor blocks. */
2700 FOR_EACH_EDGE (e, ei, bb->succs)
2701 if (!(e->flags & EDGE_DFS_BACK)
2702 && BLOCK_INFO (e->dest)->npredecessors)
2703 {
2704 BLOCK_INFO (e->dest)->npredecessors--;
2705 if (!BLOCK_INFO (e->dest)->npredecessors)
2706 {
2707 if (!nextbb)
2708 nextbb = e->dest;
2709 else
2710 BLOCK_INFO (last)->next = e->dest;
2711
2712 last = e->dest;
2713 }
2714 }
2715 }
2716 }
2717
2718 /* Estimate frequencies in loops at same nest level. */
2719
2720 static void
2721 estimate_loops_at_level (struct loop *first_loop)
2722 {
2723 struct loop *loop;
2724
2725 for (loop = first_loop; loop; loop = loop->next)
2726 {
2727 edge e;
2728 basic_block *bbs;
2729 unsigned i;
2730 bitmap tovisit = BITMAP_ALLOC (NULL);
2731
2732 estimate_loops_at_level (loop->inner);
2733
2734 /* Find current loop back edge and mark it. */
2735 e = loop_latch_edge (loop);
2736 EDGE_INFO (e)->back_edge = 1;
2737
2738 bbs = get_loop_body (loop);
2739 for (i = 0; i < loop->num_nodes; i++)
2740 bitmap_set_bit (tovisit, bbs[i]->index);
2741 free (bbs);
2742 propagate_freq (loop->header, tovisit);
2743 BITMAP_FREE (tovisit);
2744 }
2745 }
2746
2747 /* Propagates frequencies through structure of loops. */
2748
2749 static void
2750 estimate_loops (void)
2751 {
2752 bitmap tovisit = BITMAP_ALLOC (NULL);
2753 basic_block bb;
2754
2755 /* Start by estimating the frequencies in the loops. */
2756 if (number_of_loops (cfun) > 1)
2757 estimate_loops_at_level (current_loops->tree_root->inner);
2758
2759 /* Now propagate the frequencies through all the blocks. */
2760 FOR_ALL_BB (bb)
2761 {
2762 bitmap_set_bit (tovisit, bb->index);
2763 }
2764 propagate_freq (ENTRY_BLOCK_PTR, tovisit);
2765 BITMAP_FREE (tovisit);
2766 }
2767
2768 /* Convert counts measured by profile driven feedback to frequencies.
2769 Return nonzero iff there was any nonzero execution count. */
2770
2771 int
2772 counts_to_freqs (void)
2773 {
2774 gcov_type count_max, true_count_max = 0;
2775 basic_block bb;
2776
2777 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
2778 true_count_max = MAX (bb->count, true_count_max);
2779
2780 count_max = MAX (true_count_max, 1);
2781 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
2782 bb->frequency = (bb->count * BB_FREQ_MAX + count_max / 2) / count_max;
2783
2784 return true_count_max;
2785 }
2786
2787 /* Return true if function is likely to be expensive, so there is no point to
2788 optimize performance of prologue, epilogue or do inlining at the expense
2789 of code size growth. THRESHOLD is the limit of number of instructions
2790 function can execute at average to be still considered not expensive. */
2791
2792 bool
2793 expensive_function_p (int threshold)
2794 {
2795 unsigned int sum = 0;
2796 basic_block bb;
2797 unsigned int limit;
2798
2799 /* We can not compute accurately for large thresholds due to scaled
2800 frequencies. */
2801 gcc_assert (threshold <= BB_FREQ_MAX);
2802
2803 /* Frequencies are out of range. This either means that function contains
2804 internal loop executing more than BB_FREQ_MAX times or profile feedback
2805 is available and function has not been executed at all. */
2806 if (ENTRY_BLOCK_PTR->frequency == 0)
2807 return true;
2808
2809 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
2810 limit = ENTRY_BLOCK_PTR->frequency * threshold;
2811 FOR_EACH_BB (bb)
2812 {
2813 rtx insn;
2814
2815 FOR_BB_INSNS (bb, insn)
2816 if (active_insn_p (insn))
2817 {
2818 sum += bb->frequency;
2819 if (sum > limit)
2820 return true;
2821 }
2822 }
2823
2824 return false;
2825 }
2826
2827 /* Estimate and propagate basic block frequencies using the given branch
2828 probabilities. If FORCE is true, the frequencies are used to estimate
2829 the counts even when there are already non-zero profile counts. */
2830
2831 void
2832 estimate_bb_frequencies (bool force)
2833 {
2834 basic_block bb;
2835 sreal freq_max;
2836
2837 if (force || profile_status != PROFILE_READ || !counts_to_freqs ())
2838 {
2839 static int real_values_initialized = 0;
2840
2841 if (!real_values_initialized)
2842 {
2843 real_values_initialized = 1;
2844 sreal_init (&real_zero, 0, 0);
2845 sreal_init (&real_one, 1, 0);
2846 sreal_init (&real_br_prob_base, REG_BR_PROB_BASE, 0);
2847 sreal_init (&real_bb_freq_max, BB_FREQ_MAX, 0);
2848 sreal_init (&real_one_half, 1, -1);
2849 sreal_div (&real_inv_br_prob_base, &real_one, &real_br_prob_base);
2850 sreal_sub (&real_almost_one, &real_one, &real_inv_br_prob_base);
2851 }
2852
2853 mark_dfs_back_edges ();
2854
2855 single_succ_edge (ENTRY_BLOCK_PTR)->probability = REG_BR_PROB_BASE;
2856
2857 /* Set up block info for each basic block. */
2858 alloc_aux_for_blocks (sizeof (struct block_info_def));
2859 alloc_aux_for_edges (sizeof (struct edge_info_def));
2860 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
2861 {
2862 edge e;
2863 edge_iterator ei;
2864
2865 FOR_EACH_EDGE (e, ei, bb->succs)
2866 {
2867 sreal_init (&EDGE_INFO (e)->back_edge_prob, e->probability, 0);
2868 sreal_mul (&EDGE_INFO (e)->back_edge_prob,
2869 &EDGE_INFO (e)->back_edge_prob,
2870 &real_inv_br_prob_base);
2871 }
2872 }
2873
2874 /* First compute frequencies locally for each loop from innermost
2875 to outermost to examine frequencies for back edges. */
2876 estimate_loops ();
2877
2878 memcpy (&freq_max, &real_zero, sizeof (real_zero));
2879 FOR_EACH_BB (bb)
2880 if (sreal_compare (&freq_max, &BLOCK_INFO (bb)->frequency) < 0)
2881 memcpy (&freq_max, &BLOCK_INFO (bb)->frequency, sizeof (freq_max));
2882
2883 sreal_div (&freq_max, &real_bb_freq_max, &freq_max);
2884 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
2885 {
2886 sreal tmp;
2887
2888 sreal_mul (&tmp, &BLOCK_INFO (bb)->frequency, &freq_max);
2889 sreal_add (&tmp, &tmp, &real_one_half);
2890 bb->frequency = sreal_to_int (&tmp);
2891 }
2892
2893 free_aux_for_blocks ();
2894 free_aux_for_edges ();
2895 }
2896 compute_function_frequency ();
2897 }
2898
2899 /* Decide whether function is hot, cold or unlikely executed. */
2900 void
2901 compute_function_frequency (void)
2902 {
2903 basic_block bb;
2904 struct cgraph_node *node = cgraph_get_node (current_function_decl);
2905
2906 if (DECL_STATIC_CONSTRUCTOR (current_function_decl)
2907 || MAIN_NAME_P (DECL_NAME (current_function_decl)))
2908 node->only_called_at_startup = true;
2909 if (DECL_STATIC_DESTRUCTOR (current_function_decl))
2910 node->only_called_at_exit = true;
2911
2912 if (profile_status != PROFILE_READ)
2913 {
2914 int flags = flags_from_decl_or_type (current_function_decl);
2915 if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl))
2916 != NULL)
2917 node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
2918 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl))
2919 != NULL)
2920 node->frequency = NODE_FREQUENCY_HOT;
2921 else if (flags & ECF_NORETURN)
2922 node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
2923 else if (MAIN_NAME_P (DECL_NAME (current_function_decl)))
2924 node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
2925 else if (DECL_STATIC_CONSTRUCTOR (current_function_decl)
2926 || DECL_STATIC_DESTRUCTOR (current_function_decl))
2927 node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
2928 return;
2929 }
2930
2931 /* Only first time try to drop function into unlikely executed.
2932 After inlining the roundoff errors may confuse us.
2933 Ipa-profile pass will drop functions only called from unlikely
2934 functions to unlikely and that is most of what we care about. */
2935 if (!cfun->after_inlining)
2936 node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
2937 FOR_EACH_BB (bb)
2938 {
2939 if (maybe_hot_bb_p (cfun, bb))
2940 {
2941 node->frequency = NODE_FREQUENCY_HOT;
2942 return;
2943 }
2944 if (!probably_never_executed_bb_p (cfun, bb))
2945 node->frequency = NODE_FREQUENCY_NORMAL;
2946 }
2947 }
2948
2949 static bool
2950 gate_estimate_probability (void)
2951 {
2952 return flag_guess_branch_prob;
2953 }
2954
2955 /* Build PREDICT_EXPR. */
2956 tree
2957 build_predict_expr (enum br_predictor predictor, enum prediction taken)
2958 {
2959 tree t = build1 (PREDICT_EXPR, void_type_node,
2960 build_int_cst (integer_type_node, predictor));
2961 SET_PREDICT_EXPR_OUTCOME (t, taken);
2962 return t;
2963 }
2964
2965 const char *
2966 predictor_name (enum br_predictor predictor)
2967 {
2968 return predictor_info[predictor].name;
2969 }
2970
2971 namespace {
2972
2973 const pass_data pass_data_profile =
2974 {
2975 GIMPLE_PASS, /* type */
2976 "profile_estimate", /* name */
2977 OPTGROUP_NONE, /* optinfo_flags */
2978 true, /* has_gate */
2979 true, /* has_execute */
2980 TV_BRANCH_PROB, /* tv_id */
2981 PROP_cfg, /* properties_required */
2982 0, /* properties_provided */
2983 0, /* properties_destroyed */
2984 0, /* todo_flags_start */
2985 TODO_verify_ssa, /* todo_flags_finish */
2986 };
2987
2988 class pass_profile : public gimple_opt_pass
2989 {
2990 public:
2991 pass_profile (gcc::context *ctxt)
2992 : gimple_opt_pass (pass_data_profile, ctxt)
2993 {}
2994
2995 /* opt_pass methods: */
2996 bool gate () { return gate_estimate_probability (); }
2997 unsigned int execute () { return tree_estimate_probability_driver (); }
2998
2999 }; // class pass_profile
3000
3001 } // anon namespace
3002
3003 gimple_opt_pass *
3004 make_pass_profile (gcc::context *ctxt)
3005 {
3006 return new pass_profile (ctxt);
3007 }
3008
3009 namespace {
3010
3011 const pass_data pass_data_strip_predict_hints =
3012 {
3013 GIMPLE_PASS, /* type */
3014 "*strip_predict_hints", /* name */
3015 OPTGROUP_NONE, /* optinfo_flags */
3016 false, /* has_gate */
3017 true, /* has_execute */
3018 TV_BRANCH_PROB, /* tv_id */
3019 PROP_cfg, /* properties_required */
3020 0, /* properties_provided */
3021 0, /* properties_destroyed */
3022 0, /* todo_flags_start */
3023 TODO_verify_ssa, /* todo_flags_finish */
3024 };
3025
3026 class pass_strip_predict_hints : public gimple_opt_pass
3027 {
3028 public:
3029 pass_strip_predict_hints (gcc::context *ctxt)
3030 : gimple_opt_pass (pass_data_strip_predict_hints, ctxt)
3031 {}
3032
3033 /* opt_pass methods: */
3034 opt_pass * clone () { return new pass_strip_predict_hints (m_ctxt); }
3035 unsigned int execute () { return strip_predict_hints (); }
3036
3037 }; // class pass_strip_predict_hints
3038
3039 } // anon namespace
3040
3041 gimple_opt_pass *
3042 make_pass_strip_predict_hints (gcc::context *ctxt)
3043 {
3044 return new pass_strip_predict_hints (ctxt);
3045 }
3046
3047 /* Rebuild function frequencies. Passes are in general expected to
3048 maintain profile by hand, however in some cases this is not possible:
3049 for example when inlining several functions with loops freuqencies might run
3050 out of scale and thus needs to be recomputed. */
3051
3052 void
3053 rebuild_frequencies (void)
3054 {
3055 timevar_push (TV_REBUILD_FREQUENCIES);
3056
3057 /* When the max bb count in the function is small, there is a higher
3058 chance that there were truncation errors in the integer scaling
3059 of counts by inlining and other optimizations. This could lead
3060 to incorrect classification of code as being cold when it isn't.
3061 In that case, force the estimation of bb counts/frequencies from the
3062 branch probabilities, rather than computing frequencies from counts,
3063 which may also lead to frequencies incorrectly reduced to 0. There
3064 is less precision in the probabilities, so we only do this for small
3065 max counts. */
3066 gcov_type count_max = 0;
3067 basic_block bb;
3068 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
3069 count_max = MAX (bb->count, count_max);
3070
3071 if (profile_status == PROFILE_GUESSED
3072 || (profile_status == PROFILE_READ && count_max < REG_BR_PROB_BASE/10))
3073 {
3074 loop_optimizer_init (0);
3075 add_noreturn_fake_exit_edges ();
3076 mark_irreducible_loops ();
3077 connect_infinite_loops_to_exit ();
3078 estimate_bb_frequencies (true);
3079 remove_fake_exit_edges ();
3080 loop_optimizer_finalize ();
3081 }
3082 else if (profile_status == PROFILE_READ)
3083 counts_to_freqs ();
3084 else
3085 gcc_unreachable ();
3086 timevar_pop (TV_REBUILD_FREQUENCIES);
3087 }