1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
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
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
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/>. */
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. */
32 #include "coretypes.h"
38 #include "hard-reg-set.h"
46 #include "dominance.h"
49 #include "basic-block.h"
50 #include "insn-config.h"
55 #include "diagnostic-core.h"
64 #include "tree-ssa-alias.h"
65 #include "internal-fn.h"
66 #include "gimple-expr.h"
69 #include "gimple-iterator.h"
70 #include "gimple-ssa.h"
71 #include "plugin-api.h"
75 #include "tree-phinodes.h"
76 #include "ssa-iterators.h"
77 #include "tree-ssa-loop-niter.h"
78 #include "tree-ssa-loop.h"
79 #include "tree-pass.h"
80 #include "tree-scalar-evolution.h"
83 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
84 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
85 static sreal real_zero
, real_one
, real_almost_one
, real_br_prob_base
,
86 real_inv_br_prob_base
, real_one_half
, real_bb_freq_max
;
88 static void combine_predictions_for_insn (rtx_insn
*, basic_block
);
89 static void dump_prediction (FILE *, enum br_predictor
, int, basic_block
, int);
90 static void predict_paths_leading_to (basic_block
, enum br_predictor
, enum prediction
);
91 static void predict_paths_leading_to_edge (edge
, enum br_predictor
, enum prediction
);
92 static bool can_predict_insn_p (const rtx_insn
*);
94 /* Information we hold about each branch predictor.
95 Filled using information from predict.def. */
99 const char *const name
; /* Name used in the debugging dumps. */
100 const int hitrate
; /* Expected hitrate used by
101 predict_insn_def call. */
105 /* Use given predictor without Dempster-Shaffer theory if it matches
106 using first_match heuristics. */
107 #define PRED_FLAG_FIRST_MATCH 1
109 /* Recompute hitrate in percent to our representation. */
111 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
113 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
114 static const struct predictor_info predictor_info
[]= {
115 #include "predict.def"
117 /* Upper bound on predictors. */
122 /* Return TRUE if frequency FREQ is considered to be hot. */
125 maybe_hot_frequency_p (struct function
*fun
, int freq
)
127 struct cgraph_node
*node
= cgraph_node::get (fun
->decl
);
129 || !opt_for_fn (fun
->decl
, flag_branch_probabilities
))
131 if (node
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
)
133 if (node
->frequency
== NODE_FREQUENCY_HOT
)
136 if (profile_status_for_fn (fun
) == PROFILE_ABSENT
)
138 if (node
->frequency
== NODE_FREQUENCY_EXECUTED_ONCE
139 && freq
< (ENTRY_BLOCK_PTR_FOR_FN (fun
)->frequency
* 2 / 3))
141 if (PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
) == 0)
143 if (freq
< (ENTRY_BLOCK_PTR_FOR_FN (fun
)->frequency
144 / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
)))
149 static gcov_type min_count
= -1;
151 /* Determine the threshold for hot BB counts. */
154 get_hot_bb_threshold ()
156 gcov_working_set_t
*ws
;
159 ws
= find_working_set (PARAM_VALUE (HOT_BB_COUNT_WS_PERMILLE
));
161 min_count
= ws
->min_counter
;
166 /* Set the threshold for hot BB counts. */
169 set_hot_bb_threshold (gcov_type min
)
174 /* Return TRUE if frequency FREQ is considered to be hot. */
177 maybe_hot_count_p (struct function
*fun
, gcov_type count
)
179 if (fun
&& profile_status_for_fn (fun
) != PROFILE_READ
)
181 /* Code executed at most once is not hot. */
182 if (profile_info
->runs
>= count
)
184 return (count
>= get_hot_bb_threshold ());
187 /* Return true in case BB can be CPU intensive and should be optimized
188 for maximal performance. */
191 maybe_hot_bb_p (struct function
*fun
, const_basic_block bb
)
193 gcc_checking_assert (fun
);
194 if (profile_status_for_fn (fun
) == PROFILE_READ
)
195 return maybe_hot_count_p (fun
, bb
->count
);
196 return maybe_hot_frequency_p (fun
, bb
->frequency
);
199 /* Return true in case BB can be CPU intensive and should be optimized
200 for maximal performance. */
203 maybe_hot_edge_p (edge e
)
205 if (profile_status_for_fn (cfun
) == PROFILE_READ
)
206 return maybe_hot_count_p (cfun
, e
->count
);
207 return maybe_hot_frequency_p (cfun
, EDGE_FREQUENCY (e
));
210 /* Return true if profile COUNT and FREQUENCY, or function FUN static
211 node frequency reflects never being executed. */
214 probably_never_executed (struct function
*fun
,
215 gcov_type count
, int frequency
)
217 gcc_checking_assert (fun
);
218 if (profile_status_for_fn (fun
) == PROFILE_READ
)
220 int unlikely_count_fraction
= PARAM_VALUE (UNLIKELY_BB_COUNT_FRACTION
);
221 if (count
* unlikely_count_fraction
>= profile_info
->runs
)
225 if (!ENTRY_BLOCK_PTR_FOR_FN (fun
)->frequency
)
227 if (ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
)
229 gcov_type computed_count
;
230 /* Check for possibility of overflow, in which case entry bb count
231 is large enough to do the division first without losing much
233 if (ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
< REG_BR_PROB_BASE
*
236 gcov_type scaled_count
237 = frequency
* ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
*
238 unlikely_count_fraction
;
239 computed_count
= RDIV (scaled_count
,
240 ENTRY_BLOCK_PTR_FOR_FN (fun
)->frequency
);
244 computed_count
= RDIV (ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
,
245 ENTRY_BLOCK_PTR_FOR_FN (fun
)->frequency
);
246 computed_count
*= frequency
* unlikely_count_fraction
;
248 if (computed_count
>= profile_info
->runs
)
253 if ((!profile_info
|| !(opt_for_fn (fun
->decl
, flag_branch_probabilities
)))
254 && (cgraph_node::get (fun
->decl
)->frequency
255 == NODE_FREQUENCY_UNLIKELY_EXECUTED
))
261 /* Return true in case BB is probably never executed. */
264 probably_never_executed_bb_p (struct function
*fun
, const_basic_block bb
)
266 return probably_never_executed (fun
, bb
->count
, bb
->frequency
);
270 /* Return true in case edge E is probably never executed. */
273 probably_never_executed_edge_p (struct function
*fun
, edge e
)
275 return probably_never_executed (fun
, e
->count
, EDGE_FREQUENCY (e
));
278 /* Return true when current function should always be optimized for size. */
281 optimize_function_for_size_p (struct function
*fun
)
283 if (!fun
|| !fun
->decl
)
284 return optimize_size
;
285 cgraph_node
*n
= cgraph_node::get (fun
->decl
);
286 return n
&& n
->optimize_for_size_p ();
289 /* Return true when current function should always be optimized for speed. */
292 optimize_function_for_speed_p (struct function
*fun
)
294 return !optimize_function_for_size_p (fun
);
297 /* Return TRUE when BB should be optimized for size. */
300 optimize_bb_for_size_p (const_basic_block bb
)
302 return (optimize_function_for_size_p (cfun
)
303 || (bb
&& !maybe_hot_bb_p (cfun
, bb
)));
306 /* Return TRUE when BB should be optimized for speed. */
309 optimize_bb_for_speed_p (const_basic_block bb
)
311 return !optimize_bb_for_size_p (bb
);
314 /* Return TRUE when BB should be optimized for size. */
317 optimize_edge_for_size_p (edge e
)
319 return optimize_function_for_size_p (cfun
) || !maybe_hot_edge_p (e
);
322 /* Return TRUE when BB should be optimized for speed. */
325 optimize_edge_for_speed_p (edge e
)
327 return !optimize_edge_for_size_p (e
);
330 /* Return TRUE when BB should be optimized for size. */
333 optimize_insn_for_size_p (void)
335 return optimize_function_for_size_p (cfun
) || !crtl
->maybe_hot_insn_p
;
338 /* Return TRUE when BB should be optimized for speed. */
341 optimize_insn_for_speed_p (void)
343 return !optimize_insn_for_size_p ();
346 /* Return TRUE when LOOP should be optimized for size. */
349 optimize_loop_for_size_p (struct loop
*loop
)
351 return optimize_bb_for_size_p (loop
->header
);
354 /* Return TRUE when LOOP should be optimized for speed. */
357 optimize_loop_for_speed_p (struct loop
*loop
)
359 return optimize_bb_for_speed_p (loop
->header
);
362 /* Return TRUE when LOOP nest should be optimized for speed. */
365 optimize_loop_nest_for_speed_p (struct loop
*loop
)
367 struct loop
*l
= loop
;
368 if (optimize_loop_for_speed_p (loop
))
371 while (l
&& l
!= loop
)
373 if (optimize_loop_for_speed_p (l
))
381 while (l
!= loop
&& !l
->next
)
390 /* Return TRUE when LOOP nest should be optimized for size. */
393 optimize_loop_nest_for_size_p (struct loop
*loop
)
395 return !optimize_loop_nest_for_speed_p (loop
);
398 /* Return true when edge E is likely to be well predictable by branch
402 predictable_edge_p (edge e
)
404 if (profile_status_for_fn (cfun
) == PROFILE_ABSENT
)
407 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME
) * REG_BR_PROB_BASE
/ 100)
408 || (REG_BR_PROB_BASE
- e
->probability
409 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME
) * REG_BR_PROB_BASE
/ 100))
415 /* Set RTL expansion for BB profile. */
418 rtl_profile_for_bb (basic_block bb
)
420 crtl
->maybe_hot_insn_p
= maybe_hot_bb_p (cfun
, bb
);
423 /* Set RTL expansion for edge profile. */
426 rtl_profile_for_edge (edge e
)
428 crtl
->maybe_hot_insn_p
= maybe_hot_edge_p (e
);
431 /* Set RTL expansion to default mode (i.e. when profile info is not known). */
433 default_rtl_profile (void)
435 crtl
->maybe_hot_insn_p
= true;
438 /* Return true if the one of outgoing edges is already predicted by
442 rtl_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
445 if (!INSN_P (BB_END (bb
)))
447 for (note
= REG_NOTES (BB_END (bb
)); note
; note
= XEXP (note
, 1))
448 if (REG_NOTE_KIND (note
) == REG_BR_PRED
449 && INTVAL (XEXP (XEXP (note
, 0), 0)) == (int)predictor
)
454 /* Structure representing predictions in tree level. */
456 struct edge_prediction
{
457 struct edge_prediction
*ep_next
;
459 enum br_predictor ep_predictor
;
463 /* This map contains for a basic block the list of predictions for the
466 static hash_map
<const_basic_block
, edge_prediction
*> *bb_predictions
;
468 /* Return true if the one of outgoing edges is already predicted by
472 gimple_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
474 struct edge_prediction
*i
;
475 edge_prediction
**preds
= bb_predictions
->get (bb
);
480 for (i
= *preds
; i
; i
= i
->ep_next
)
481 if (i
->ep_predictor
== predictor
)
486 /* Return true when the probability of edge is reliable.
488 The profile guessing code is good at predicting branch outcome (ie.
489 taken/not taken), that is predicted right slightly over 75% of time.
490 It is however notoriously poor on predicting the probability itself.
491 In general the profile appear a lot flatter (with probabilities closer
492 to 50%) than the reality so it is bad idea to use it to drive optimization
493 such as those disabling dynamic branch prediction for well predictable
496 There are two exceptions - edges leading to noreturn edges and edges
497 predicted by number of iterations heuristics are predicted well. This macro
498 should be able to distinguish those, but at the moment it simply check for
499 noreturn heuristic that is only one giving probability over 99% or bellow
500 1%. In future we might want to propagate reliability information across the
501 CFG if we find this information useful on multiple places. */
503 probability_reliable_p (int prob
)
505 return (profile_status_for_fn (cfun
) == PROFILE_READ
506 || (profile_status_for_fn (cfun
) == PROFILE_GUESSED
507 && (prob
<= HITRATE (1) || prob
>= HITRATE (99))));
510 /* Same predicate as above, working on edges. */
512 edge_probability_reliable_p (const_edge e
)
514 return probability_reliable_p (e
->probability
);
517 /* Same predicate as edge_probability_reliable_p, working on notes. */
519 br_prob_note_reliable_p (const_rtx note
)
521 gcc_assert (REG_NOTE_KIND (note
) == REG_BR_PROB
);
522 return probability_reliable_p (XINT (note
, 0));
526 predict_insn (rtx_insn
*insn
, enum br_predictor predictor
, int probability
)
528 gcc_assert (any_condjump_p (insn
));
529 if (!flag_guess_branch_prob
)
532 add_reg_note (insn
, REG_BR_PRED
,
533 gen_rtx_CONCAT (VOIDmode
,
534 GEN_INT ((int) predictor
),
535 GEN_INT ((int) probability
)));
538 /* Predict insn by given predictor. */
541 predict_insn_def (rtx_insn
*insn
, enum br_predictor predictor
,
542 enum prediction taken
)
544 int probability
= predictor_info
[(int) predictor
].hitrate
;
547 probability
= REG_BR_PROB_BASE
- probability
;
549 predict_insn (insn
, predictor
, probability
);
552 /* Predict edge E with given probability if possible. */
555 rtl_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
558 last_insn
= BB_END (e
->src
);
560 /* We can store the branch prediction information only about
561 conditional jumps. */
562 if (!any_condjump_p (last_insn
))
565 /* We always store probability of branching. */
566 if (e
->flags
& EDGE_FALLTHRU
)
567 probability
= REG_BR_PROB_BASE
- probability
;
569 predict_insn (last_insn
, predictor
, probability
);
572 /* Predict edge E with the given PROBABILITY. */
574 gimple_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
576 gcc_assert (profile_status_for_fn (cfun
) != PROFILE_GUESSED
);
577 if ((e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
) && EDGE_COUNT (e
->src
->succs
) >
579 && flag_guess_branch_prob
&& optimize
)
581 struct edge_prediction
*i
= XNEW (struct edge_prediction
);
582 edge_prediction
*&preds
= bb_predictions
->get_or_insert (e
->src
);
586 i
->ep_probability
= probability
;
587 i
->ep_predictor
= predictor
;
592 /* Remove all predictions on given basic block that are attached
595 remove_predictions_associated_with_edge (edge e
)
600 edge_prediction
**preds
= bb_predictions
->get (e
->src
);
604 struct edge_prediction
**prediction
= preds
;
605 struct edge_prediction
*next
;
609 if ((*prediction
)->ep_edge
== e
)
611 next
= (*prediction
)->ep_next
;
616 prediction
= &((*prediction
)->ep_next
);
621 /* Clears the list of predictions stored for BB. */
624 clear_bb_predictions (basic_block bb
)
626 edge_prediction
**preds
= bb_predictions
->get (bb
);
627 struct edge_prediction
*pred
, *next
;
632 for (pred
= *preds
; pred
; pred
= next
)
634 next
= pred
->ep_next
;
640 /* Return true when we can store prediction on insn INSN.
641 At the moment we represent predictions only on conditional
642 jumps, not at computed jump or other complicated cases. */
644 can_predict_insn_p (const rtx_insn
*insn
)
646 return (JUMP_P (insn
)
647 && any_condjump_p (insn
)
648 && EDGE_COUNT (BLOCK_FOR_INSN (insn
)->succs
) >= 2);
651 /* Predict edge E by given predictor if possible. */
654 predict_edge_def (edge e
, enum br_predictor predictor
,
655 enum prediction taken
)
657 int probability
= predictor_info
[(int) predictor
].hitrate
;
660 probability
= REG_BR_PROB_BASE
- probability
;
662 predict_edge (e
, predictor
, probability
);
665 /* Invert all branch predictions or probability notes in the INSN. This needs
666 to be done each time we invert the condition used by the jump. */
669 invert_br_probabilities (rtx insn
)
673 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
674 if (REG_NOTE_KIND (note
) == REG_BR_PROB
)
675 XINT (note
, 0) = REG_BR_PROB_BASE
- XINT (note
, 0);
676 else if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
677 XEXP (XEXP (note
, 0), 1)
678 = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (XEXP (note
, 0), 1)));
681 /* Dump information about the branch prediction to the output file. */
684 dump_prediction (FILE *file
, enum br_predictor predictor
, int probability
,
685 basic_block bb
, int used
)
693 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
694 if (! (e
->flags
& EDGE_FALLTHRU
))
697 fprintf (file
, " %s heuristics%s: %.1f%%",
698 predictor_info
[predictor
].name
,
699 used
? "" : " (ignored)", probability
* 100.0 / REG_BR_PROB_BASE
);
703 fprintf (file
, " exec %"PRId64
, bb
->count
);
706 fprintf (file
, " hit %"PRId64
, e
->count
);
707 fprintf (file
, " (%.1f%%)", e
->count
* 100.0 / bb
->count
);
711 fprintf (file
, "\n");
714 /* We can not predict the probabilities of outgoing edges of bb. Set them
715 evenly and hope for the best. */
717 set_even_probabilities (basic_block bb
)
723 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
724 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
726 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
727 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
728 e
->probability
= (REG_BR_PROB_BASE
+ nedges
/ 2) / nedges
;
733 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
734 note if not already present. Remove now useless REG_BR_PRED notes. */
737 combine_predictions_for_insn (rtx_insn
*insn
, basic_block bb
)
742 int best_probability
= PROB_EVEN
;
743 enum br_predictor best_predictor
= END_PREDICTORS
;
744 int combined_probability
= REG_BR_PROB_BASE
/ 2;
746 bool first_match
= false;
749 if (!can_predict_insn_p (insn
))
751 set_even_probabilities (bb
);
755 prob_note
= find_reg_note (insn
, REG_BR_PROB
, 0);
756 pnote
= ®_NOTES (insn
);
758 fprintf (dump_file
, "Predictions for insn %i bb %i\n", INSN_UID (insn
),
761 /* We implement "first match" heuristics and use probability guessed
762 by predictor with smallest index. */
763 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
764 if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
766 enum br_predictor predictor
= ((enum br_predictor
)
767 INTVAL (XEXP (XEXP (note
, 0), 0)));
768 int probability
= INTVAL (XEXP (XEXP (note
, 0), 1));
771 if (best_predictor
> predictor
)
772 best_probability
= probability
, best_predictor
= predictor
;
774 d
= (combined_probability
* probability
775 + (REG_BR_PROB_BASE
- combined_probability
)
776 * (REG_BR_PROB_BASE
- probability
));
778 /* Use FP math to avoid overflows of 32bit integers. */
780 /* If one probability is 0% and one 100%, avoid division by zero. */
781 combined_probability
= REG_BR_PROB_BASE
/ 2;
783 combined_probability
= (((double) combined_probability
) * probability
784 * REG_BR_PROB_BASE
/ d
+ 0.5);
787 /* Decide which heuristic to use. In case we didn't match anything,
788 use no_prediction heuristic, in case we did match, use either
789 first match or Dempster-Shaffer theory depending on the flags. */
791 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
795 dump_prediction (dump_file
, PRED_NO_PREDICTION
,
796 combined_probability
, bb
, true);
799 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
,
801 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
,
806 combined_probability
= best_probability
;
807 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
811 if (REG_NOTE_KIND (*pnote
) == REG_BR_PRED
)
813 enum br_predictor predictor
= ((enum br_predictor
)
814 INTVAL (XEXP (XEXP (*pnote
, 0), 0)));
815 int probability
= INTVAL (XEXP (XEXP (*pnote
, 0), 1));
817 dump_prediction (dump_file
, predictor
, probability
, bb
,
818 !first_match
|| best_predictor
== predictor
);
819 *pnote
= XEXP (*pnote
, 1);
822 pnote
= &XEXP (*pnote
, 1);
827 add_int_reg_note (insn
, REG_BR_PROB
, combined_probability
);
829 /* Save the prediction into CFG in case we are seeing non-degenerated
831 if (!single_succ_p (bb
))
833 BRANCH_EDGE (bb
)->probability
= combined_probability
;
834 FALLTHRU_EDGE (bb
)->probability
835 = REG_BR_PROB_BASE
- combined_probability
;
838 else if (!single_succ_p (bb
))
840 int prob
= XINT (prob_note
, 0);
842 BRANCH_EDGE (bb
)->probability
= prob
;
843 FALLTHRU_EDGE (bb
)->probability
= REG_BR_PROB_BASE
- prob
;
846 single_succ_edge (bb
)->probability
= REG_BR_PROB_BASE
;
849 /* Combine predictions into single probability and store them into CFG.
850 Remove now useless prediction entries. */
853 combine_predictions_for_bb (basic_block bb
)
855 int best_probability
= PROB_EVEN
;
856 enum br_predictor best_predictor
= END_PREDICTORS
;
857 int combined_probability
= REG_BR_PROB_BASE
/ 2;
859 bool first_match
= false;
861 struct edge_prediction
*pred
;
863 edge e
, first
= NULL
, second
= NULL
;
866 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
867 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
870 if (first
&& !second
)
876 /* When there is no successor or only one choice, prediction is easy.
878 We are lazy for now and predict only basic blocks with two outgoing
879 edges. It is possible to predict generic case too, but we have to
880 ignore first match heuristics and do more involved combining. Implement
885 set_even_probabilities (bb
);
886 clear_bb_predictions (bb
);
888 fprintf (dump_file
, "%i edges in bb %i predicted to even probabilities\n",
894 fprintf (dump_file
, "Predictions for bb %i\n", bb
->index
);
896 edge_prediction
**preds
= bb_predictions
->get (bb
);
899 /* We implement "first match" heuristics and use probability guessed
900 by predictor with smallest index. */
901 for (pred
= *preds
; pred
; pred
= pred
->ep_next
)
903 enum br_predictor predictor
= pred
->ep_predictor
;
904 int probability
= pred
->ep_probability
;
906 if (pred
->ep_edge
!= first
)
907 probability
= REG_BR_PROB_BASE
- probability
;
910 /* First match heuristics would be widly confused if we predicted
912 if (best_predictor
> predictor
)
914 struct edge_prediction
*pred2
;
915 int prob
= probability
;
917 for (pred2
= (struct edge_prediction
*) *preds
;
918 pred2
; pred2
= pred2
->ep_next
)
919 if (pred2
!= pred
&& pred2
->ep_predictor
== pred
->ep_predictor
)
921 int probability2
= pred
->ep_probability
;
923 if (pred2
->ep_edge
!= first
)
924 probability2
= REG_BR_PROB_BASE
- probability2
;
926 if ((probability
< REG_BR_PROB_BASE
/ 2) !=
927 (probability2
< REG_BR_PROB_BASE
/ 2))
930 /* If the same predictor later gave better result, go for it! */
931 if ((probability
>= REG_BR_PROB_BASE
/ 2 && (probability2
> probability
))
932 || (probability
<= REG_BR_PROB_BASE
/ 2 && (probability2
< probability
)))
936 best_probability
= prob
, best_predictor
= predictor
;
939 d
= (combined_probability
* probability
940 + (REG_BR_PROB_BASE
- combined_probability
)
941 * (REG_BR_PROB_BASE
- probability
));
943 /* Use FP math to avoid overflows of 32bit integers. */
945 /* If one probability is 0% and one 100%, avoid division by zero. */
946 combined_probability
= REG_BR_PROB_BASE
/ 2;
948 combined_probability
= (((double) combined_probability
)
950 * REG_BR_PROB_BASE
/ d
+ 0.5);
954 /* Decide which heuristic to use. In case we didn't match anything,
955 use no_prediction heuristic, in case we did match, use either
956 first match or Dempster-Shaffer theory depending on the flags. */
958 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
962 dump_prediction (dump_file
, PRED_NO_PREDICTION
, combined_probability
, bb
, true);
965 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
, bb
,
967 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
, bb
,
972 combined_probability
= best_probability
;
973 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
977 for (pred
= (struct edge_prediction
*) *preds
; pred
; pred
= pred
->ep_next
)
979 enum br_predictor predictor
= pred
->ep_predictor
;
980 int probability
= pred
->ep_probability
;
982 if (pred
->ep_edge
!= EDGE_SUCC (bb
, 0))
983 probability
= REG_BR_PROB_BASE
- probability
;
984 dump_prediction (dump_file
, predictor
, probability
, bb
,
985 !first_match
|| best_predictor
== predictor
);
988 clear_bb_predictions (bb
);
992 first
->probability
= combined_probability
;
993 second
->probability
= REG_BR_PROB_BASE
- combined_probability
;
997 /* Check if T1 and T2 satisfy the IV_COMPARE condition.
998 Return the SSA_NAME if the condition satisfies, NULL otherwise.
1000 T1 and T2 should be one of the following cases:
1001 1. T1 is SSA_NAME, T2 is NULL
1002 2. T1 is SSA_NAME, T2 is INTEGER_CST between [-4, 4]
1003 3. T2 is SSA_NAME, T1 is INTEGER_CST between [-4, 4] */
1006 strips_small_constant (tree t1
, tree t2
)
1013 else if (TREE_CODE (t1
) == SSA_NAME
)
1015 else if (tree_fits_shwi_p (t1
))
1016 value
= tree_to_shwi (t1
);
1022 else if (tree_fits_shwi_p (t2
))
1023 value
= tree_to_shwi (t2
);
1024 else if (TREE_CODE (t2
) == SSA_NAME
)
1032 if (value
<= 4 && value
>= -4)
1038 /* Return the SSA_NAME in T or T's operands.
1039 Return NULL if SSA_NAME cannot be found. */
1042 get_base_value (tree t
)
1044 if (TREE_CODE (t
) == SSA_NAME
)
1047 if (!BINARY_CLASS_P (t
))
1050 switch (TREE_OPERAND_LENGTH (t
))
1053 return strips_small_constant (TREE_OPERAND (t
, 0), NULL
);
1055 return strips_small_constant (TREE_OPERAND (t
, 0),
1056 TREE_OPERAND (t
, 1));
1062 /* Check the compare STMT in LOOP. If it compares an induction
1063 variable to a loop invariant, return true, and save
1064 LOOP_INVARIANT, COMPARE_CODE and LOOP_STEP.
1065 Otherwise return false and set LOOP_INVAIANT to NULL. */
1068 is_comparison_with_loop_invariant_p (gimple stmt
, struct loop
*loop
,
1069 tree
*loop_invariant
,
1070 enum tree_code
*compare_code
,
1074 tree op0
, op1
, bound
, base
;
1076 enum tree_code code
;
1079 code
= gimple_cond_code (stmt
);
1080 *loop_invariant
= NULL
;
1096 op0
= gimple_cond_lhs (stmt
);
1097 op1
= gimple_cond_rhs (stmt
);
1099 if ((TREE_CODE (op0
) != SSA_NAME
&& TREE_CODE (op0
) != INTEGER_CST
)
1100 || (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op1
) != INTEGER_CST
))
1102 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op0
, &iv0
, true))
1104 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op1
, &iv1
, true))
1106 if (TREE_CODE (iv0
.step
) != INTEGER_CST
1107 || TREE_CODE (iv1
.step
) != INTEGER_CST
)
1109 if ((integer_zerop (iv0
.step
) && integer_zerop (iv1
.step
))
1110 || (!integer_zerop (iv0
.step
) && !integer_zerop (iv1
.step
)))
1113 if (integer_zerop (iv0
.step
))
1115 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
1116 code
= invert_tree_comparison (code
, false);
1119 if (tree_fits_shwi_p (iv1
.step
))
1128 if (tree_fits_shwi_p (iv0
.step
))
1134 if (TREE_CODE (bound
) != INTEGER_CST
)
1135 bound
= get_base_value (bound
);
1138 if (TREE_CODE (base
) != INTEGER_CST
)
1139 base
= get_base_value (base
);
1143 *loop_invariant
= bound
;
1144 *compare_code
= code
;
1146 *loop_iv_base
= base
;
1150 /* Compare two SSA_NAMEs: returns TRUE if T1 and T2 are value coherent. */
1153 expr_coherent_p (tree t1
, tree t2
)
1156 tree ssa_name_1
= NULL
;
1157 tree ssa_name_2
= NULL
;
1159 gcc_assert (TREE_CODE (t1
) == SSA_NAME
|| TREE_CODE (t1
) == INTEGER_CST
);
1160 gcc_assert (TREE_CODE (t2
) == SSA_NAME
|| TREE_CODE (t2
) == INTEGER_CST
);
1165 if (TREE_CODE (t1
) == INTEGER_CST
&& TREE_CODE (t2
) == INTEGER_CST
)
1167 if (TREE_CODE (t1
) == INTEGER_CST
|| TREE_CODE (t2
) == INTEGER_CST
)
1170 /* Check to see if t1 is expressed/defined with t2. */
1171 stmt
= SSA_NAME_DEF_STMT (t1
);
1172 gcc_assert (stmt
!= NULL
);
1173 if (is_gimple_assign (stmt
))
1175 ssa_name_1
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
1176 if (ssa_name_1
&& ssa_name_1
== t2
)
1180 /* Check to see if t2 is expressed/defined with t1. */
1181 stmt
= SSA_NAME_DEF_STMT (t2
);
1182 gcc_assert (stmt
!= NULL
);
1183 if (is_gimple_assign (stmt
))
1185 ssa_name_2
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
1186 if (ssa_name_2
&& ssa_name_2
== t1
)
1190 /* Compare if t1 and t2's def_stmts are identical. */
1191 if (ssa_name_2
!= NULL
&& ssa_name_1
== ssa_name_2
)
1197 /* Predict branch probability of BB when BB contains a branch that compares
1198 an induction variable in LOOP with LOOP_IV_BASE_VAR to LOOP_BOUND_VAR. The
1199 loop exit is compared using LOOP_BOUND_CODE, with step of LOOP_BOUND_STEP.
1202 for (int i = 0; i < bound; i++) {
1209 In this loop, we will predict the branch inside the loop to be taken. */
1212 predict_iv_comparison (struct loop
*loop
, basic_block bb
,
1213 tree loop_bound_var
,
1214 tree loop_iv_base_var
,
1215 enum tree_code loop_bound_code
,
1216 int loop_bound_step
)
1219 tree compare_var
, compare_base
;
1220 enum tree_code compare_code
;
1221 tree compare_step_var
;
1225 if (predicted_by_p (bb
, PRED_LOOP_ITERATIONS_GUESSED
)
1226 || predicted_by_p (bb
, PRED_LOOP_ITERATIONS
)
1227 || predicted_by_p (bb
, PRED_LOOP_EXIT
))
1230 stmt
= last_stmt (bb
);
1231 if (!stmt
|| gimple_code (stmt
) != GIMPLE_COND
)
1233 if (!is_comparison_with_loop_invariant_p (stmt
, loop
, &compare_var
,
1239 /* Find the taken edge. */
1240 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
1241 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1244 /* When comparing an IV to a loop invariant, NE is more likely to be
1245 taken while EQ is more likely to be not-taken. */
1246 if (compare_code
== NE_EXPR
)
1248 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1251 else if (compare_code
== EQ_EXPR
)
1253 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1257 if (!expr_coherent_p (loop_iv_base_var
, compare_base
))
1260 /* If loop bound, base and compare bound are all constants, we can
1261 calculate the probability directly. */
1262 if (tree_fits_shwi_p (loop_bound_var
)
1263 && tree_fits_shwi_p (compare_var
)
1264 && tree_fits_shwi_p (compare_base
))
1267 bool overflow
, overall_overflow
= false;
1268 widest_int compare_count
, tem
;
1270 /* (loop_bound - base) / compare_step */
1271 tem
= wi::sub (wi::to_widest (loop_bound_var
),
1272 wi::to_widest (compare_base
), SIGNED
, &overflow
);
1273 overall_overflow
|= overflow
;
1274 widest_int loop_count
= wi::div_trunc (tem
,
1275 wi::to_widest (compare_step_var
),
1277 overall_overflow
|= overflow
;
1279 if (!wi::neg_p (wi::to_widest (compare_step_var
))
1280 ^ (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1282 /* (loop_bound - compare_bound) / compare_step */
1283 tem
= wi::sub (wi::to_widest (loop_bound_var
),
1284 wi::to_widest (compare_var
), SIGNED
, &overflow
);
1285 overall_overflow
|= overflow
;
1286 compare_count
= wi::div_trunc (tem
, wi::to_widest (compare_step_var
),
1288 overall_overflow
|= overflow
;
1292 /* (compare_bound - base) / compare_step */
1293 tem
= wi::sub (wi::to_widest (compare_var
),
1294 wi::to_widest (compare_base
), SIGNED
, &overflow
);
1295 overall_overflow
|= overflow
;
1296 compare_count
= wi::div_trunc (tem
, wi::to_widest (compare_step_var
),
1298 overall_overflow
|= overflow
;
1300 if (compare_code
== LE_EXPR
|| compare_code
== GE_EXPR
)
1302 if (loop_bound_code
== LE_EXPR
|| loop_bound_code
== GE_EXPR
)
1304 if (wi::neg_p (compare_count
))
1306 if (wi::neg_p (loop_count
))
1308 if (loop_count
== 0)
1310 else if (wi::cmps (compare_count
, loop_count
) == 1)
1311 probability
= REG_BR_PROB_BASE
;
1314 tem
= compare_count
* REG_BR_PROB_BASE
;
1315 tem
= wi::udiv_trunc (tem
, loop_count
);
1316 probability
= tem
.to_uhwi ();
1319 if (!overall_overflow
)
1320 predict_edge (then_edge
, PRED_LOOP_IV_COMPARE
, probability
);
1325 if (expr_coherent_p (loop_bound_var
, compare_var
))
1327 if ((loop_bound_code
== LT_EXPR
|| loop_bound_code
== LE_EXPR
)
1328 && (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1329 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1330 else if ((loop_bound_code
== GT_EXPR
|| loop_bound_code
== GE_EXPR
)
1331 && (compare_code
== GT_EXPR
|| compare_code
== GE_EXPR
))
1332 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1333 else if (loop_bound_code
== NE_EXPR
)
1335 /* If the loop backedge condition is "(i != bound)", we do
1336 the comparison based on the step of IV:
1337 * step < 0 : backedge condition is like (i > bound)
1338 * step > 0 : backedge condition is like (i < bound) */
1339 gcc_assert (loop_bound_step
!= 0);
1340 if (loop_bound_step
> 0
1341 && (compare_code
== LT_EXPR
1342 || compare_code
== LE_EXPR
))
1343 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1344 else if (loop_bound_step
< 0
1345 && (compare_code
== GT_EXPR
1346 || compare_code
== GE_EXPR
))
1347 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1349 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1352 /* The branch is predicted not-taken if loop_bound_code is
1353 opposite with compare_code. */
1354 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1356 else if (expr_coherent_p (loop_iv_base_var
, compare_var
))
1359 for (i = s; i < h; i++)
1361 The branch should be predicted taken. */
1362 if (loop_bound_step
> 0
1363 && (compare_code
== GT_EXPR
|| compare_code
== GE_EXPR
))
1364 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1365 else if (loop_bound_step
< 0
1366 && (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1367 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1369 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1373 /* Predict for extra loop exits that will lead to EXIT_EDGE. The extra loop
1374 exits are resulted from short-circuit conditions that will generate an
1377 if (foo() || global > 10)
1380 This will be translated into:
1385 if foo() goto BB6 else goto BB5
1387 if global > 10 goto BB6 else goto BB7
1391 iftmp = (PHI 0(BB5), 1(BB6))
1392 if iftmp == 1 goto BB8 else goto BB3
1394 outside of the loop...
1396 The edge BB7->BB8 is loop exit because BB8 is outside of the loop.
1397 From the dataflow, we can infer that BB4->BB6 and BB5->BB6 are also loop
1398 exits. This function takes BB7->BB8 as input, and finds out the extra loop
1399 exits to predict them using PRED_LOOP_EXIT. */
1402 predict_extra_loop_exits (edge exit_edge
)
1405 bool check_value_one
;
1407 tree cmp_rhs
, cmp_lhs
;
1408 gimple cmp_stmt
= last_stmt (exit_edge
->src
);
1410 if (!cmp_stmt
|| gimple_code (cmp_stmt
) != GIMPLE_COND
)
1412 cmp_rhs
= gimple_cond_rhs (cmp_stmt
);
1413 cmp_lhs
= gimple_cond_lhs (cmp_stmt
);
1414 if (!TREE_CONSTANT (cmp_rhs
)
1415 || !(integer_zerop (cmp_rhs
) || integer_onep (cmp_rhs
)))
1417 if (TREE_CODE (cmp_lhs
) != SSA_NAME
)
1420 /* If check_value_one is true, only the phi_args with value '1' will lead
1421 to loop exit. Otherwise, only the phi_args with value '0' will lead to
1423 check_value_one
= (((integer_onep (cmp_rhs
))
1424 ^ (gimple_cond_code (cmp_stmt
) == EQ_EXPR
))
1425 ^ ((exit_edge
->flags
& EDGE_TRUE_VALUE
) != 0));
1427 phi_stmt
= SSA_NAME_DEF_STMT (cmp_lhs
);
1428 if (!phi_stmt
|| gimple_code (phi_stmt
) != GIMPLE_PHI
)
1431 for (i
= 0; i
< gimple_phi_num_args (phi_stmt
); i
++)
1435 tree val
= gimple_phi_arg_def (phi_stmt
, i
);
1436 edge e
= gimple_phi_arg_edge (phi_stmt
, i
);
1438 if (!TREE_CONSTANT (val
) || !(integer_zerop (val
) || integer_onep (val
)))
1440 if ((check_value_one
^ integer_onep (val
)) == 1)
1442 if (EDGE_COUNT (e
->src
->succs
) != 1)
1444 predict_paths_leading_to_edge (e
, PRED_LOOP_EXIT
, NOT_TAKEN
);
1448 FOR_EACH_EDGE (e1
, ei
, e
->src
->preds
)
1449 predict_paths_leading_to_edge (e1
, PRED_LOOP_EXIT
, NOT_TAKEN
);
1453 /* Predict edge probabilities by exploiting loop structure. */
1456 predict_loops (void)
1460 /* Try to predict out blocks in a loop that are not part of a
1462 FOR_EACH_LOOP (loop
, 0)
1464 basic_block bb
, *bbs
;
1465 unsigned j
, n_exits
;
1467 struct tree_niter_desc niter_desc
;
1469 struct nb_iter_bound
*nb_iter
;
1470 enum tree_code loop_bound_code
= ERROR_MARK
;
1471 tree loop_bound_step
= NULL
;
1472 tree loop_bound_var
= NULL
;
1473 tree loop_iv_base
= NULL
;
1476 exits
= get_loop_exit_edges (loop
);
1477 n_exits
= exits
.length ();
1484 FOR_EACH_VEC_ELT (exits
, j
, ex
)
1487 HOST_WIDE_INT nitercst
;
1488 int max
= PARAM_VALUE (PARAM_MAX_PREDICTED_ITERATIONS
);
1490 enum br_predictor predictor
;
1492 predict_extra_loop_exits (ex
);
1494 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false, false))
1495 niter
= niter_desc
.niter
;
1496 if (!niter
|| TREE_CODE (niter_desc
.niter
) != INTEGER_CST
)
1497 niter
= loop_niter_by_eval (loop
, ex
);
1499 if (TREE_CODE (niter
) == INTEGER_CST
)
1501 if (tree_fits_uhwi_p (niter
)
1503 && compare_tree_int (niter
, max
- 1) == -1)
1504 nitercst
= tree_to_uhwi (niter
) + 1;
1507 predictor
= PRED_LOOP_ITERATIONS
;
1509 /* If we have just one exit and we can derive some information about
1510 the number of iterations of the loop from the statements inside
1511 the loop, use it to predict this exit. */
1512 else if (n_exits
== 1)
1514 nitercst
= estimated_stmt_executions_int (loop
);
1520 predictor
= PRED_LOOP_ITERATIONS_GUESSED
;
1525 /* If the prediction for number of iterations is zero, do not
1526 predict the exit edges. */
1530 probability
= ((REG_BR_PROB_BASE
+ nitercst
/ 2) / nitercst
);
1531 predict_edge (ex
, predictor
, probability
);
1535 /* Find information about loop bound variables. */
1536 for (nb_iter
= loop
->bounds
; nb_iter
;
1537 nb_iter
= nb_iter
->next
)
1539 && gimple_code (nb_iter
->stmt
) == GIMPLE_COND
)
1541 stmt
= nb_iter
->stmt
;
1544 if (!stmt
&& last_stmt (loop
->header
)
1545 && gimple_code (last_stmt (loop
->header
)) == GIMPLE_COND
)
1546 stmt
= last_stmt (loop
->header
);
1548 is_comparison_with_loop_invariant_p (stmt
, loop
,
1554 bbs
= get_loop_body (loop
);
1556 for (j
= 0; j
< loop
->num_nodes
; j
++)
1558 int header_found
= 0;
1564 /* Bypass loop heuristics on continue statement. These
1565 statements construct loops via "non-loop" constructs
1566 in the source language and are better to be handled
1568 if (predicted_by_p (bb
, PRED_CONTINUE
))
1571 /* Loop branch heuristics - predict an edge back to a
1572 loop's head as taken. */
1573 if (bb
== loop
->latch
)
1575 e
= find_edge (loop
->latch
, loop
->header
);
1579 predict_edge_def (e
, PRED_LOOP_BRANCH
, TAKEN
);
1583 /* Loop exit heuristics - predict an edge exiting the loop if the
1584 conditional has no loop header successors as not taken. */
1586 /* If we already used more reliable loop exit predictors, do not
1587 bother with PRED_LOOP_EXIT. */
1588 && !predicted_by_p (bb
, PRED_LOOP_ITERATIONS_GUESSED
)
1589 && !predicted_by_p (bb
, PRED_LOOP_ITERATIONS
))
1591 /* For loop with many exits we don't want to predict all exits
1592 with the pretty large probability, because if all exits are
1593 considered in row, the loop would be predicted to iterate
1594 almost never. The code to divide probability by number of
1595 exits is very rough. It should compute the number of exits
1596 taken in each patch through function (not the overall number
1597 of exits that might be a lot higher for loops with wide switch
1598 statements in them) and compute n-th square root.
1600 We limit the minimal probability by 2% to avoid
1601 EDGE_PROBABILITY_RELIABLE from trusting the branch prediction
1602 as this was causing regression in perl benchmark containing such
1605 int probability
= ((REG_BR_PROB_BASE
1606 - predictor_info
[(int) PRED_LOOP_EXIT
].hitrate
)
1608 if (probability
< HITRATE (2))
1609 probability
= HITRATE (2);
1610 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1611 if (e
->dest
->index
< NUM_FIXED_BLOCKS
1612 || !flow_bb_inside_loop_p (loop
, e
->dest
))
1613 predict_edge (e
, PRED_LOOP_EXIT
, probability
);
1616 predict_iv_comparison (loop
, bb
, loop_bound_var
, loop_iv_base
,
1618 tree_to_shwi (loop_bound_step
));
1621 /* Free basic blocks from get_loop_body. */
1626 /* Attempt to predict probabilities of BB outgoing edges using local
1629 bb_estimate_probability_locally (basic_block bb
)
1631 rtx_insn
*last_insn
= BB_END (bb
);
1634 if (! can_predict_insn_p (last_insn
))
1636 cond
= get_condition (last_insn
, NULL
, false, false);
1640 /* Try "pointer heuristic."
1641 A comparison ptr == 0 is predicted as false.
1642 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1643 if (COMPARISON_P (cond
)
1644 && ((REG_P (XEXP (cond
, 0)) && REG_POINTER (XEXP (cond
, 0)))
1645 || (REG_P (XEXP (cond
, 1)) && REG_POINTER (XEXP (cond
, 1)))))
1647 if (GET_CODE (cond
) == EQ
)
1648 predict_insn_def (last_insn
, PRED_POINTER
, NOT_TAKEN
);
1649 else if (GET_CODE (cond
) == NE
)
1650 predict_insn_def (last_insn
, PRED_POINTER
, TAKEN
);
1654 /* Try "opcode heuristic."
1655 EQ tests are usually false and NE tests are usually true. Also,
1656 most quantities are positive, so we can make the appropriate guesses
1657 about signed comparisons against zero. */
1658 switch (GET_CODE (cond
))
1661 /* Unconditional branch. */
1662 predict_insn_def (last_insn
, PRED_UNCONDITIONAL
,
1663 cond
== const0_rtx
? NOT_TAKEN
: TAKEN
);
1668 /* Floating point comparisons appears to behave in a very
1669 unpredictable way because of special role of = tests in
1671 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
1673 /* Comparisons with 0 are often used for booleans and there is
1674 nothing useful to predict about them. */
1675 else if (XEXP (cond
, 1) == const0_rtx
1676 || XEXP (cond
, 0) == const0_rtx
)
1679 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, NOT_TAKEN
);
1684 /* Floating point comparisons appears to behave in a very
1685 unpredictable way because of special role of = tests in
1687 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
1689 /* Comparisons with 0 are often used for booleans and there is
1690 nothing useful to predict about them. */
1691 else if (XEXP (cond
, 1) == const0_rtx
1692 || XEXP (cond
, 0) == const0_rtx
)
1695 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, TAKEN
);
1699 predict_insn_def (last_insn
, PRED_FPOPCODE
, TAKEN
);
1703 predict_insn_def (last_insn
, PRED_FPOPCODE
, NOT_TAKEN
);
1708 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
1709 || XEXP (cond
, 1) == constm1_rtx
)
1710 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, NOT_TAKEN
);
1715 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
1716 || XEXP (cond
, 1) == constm1_rtx
)
1717 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, TAKEN
);
1725 /* Set edge->probability for each successor edge of BB. */
1727 guess_outgoing_edge_probabilities (basic_block bb
)
1729 bb_estimate_probability_locally (bb
);
1730 combine_predictions_for_insn (BB_END (bb
), bb
);
1733 static tree
expr_expected_value (tree
, bitmap
, enum br_predictor
*predictor
);
1735 /* Helper function for expr_expected_value. */
1738 expr_expected_value_1 (tree type
, tree op0
, enum tree_code code
,
1739 tree op1
, bitmap visited
, enum br_predictor
*predictor
)
1744 *predictor
= PRED_UNCONDITIONAL
;
1746 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1748 if (TREE_CONSTANT (op0
))
1751 if (code
!= SSA_NAME
)
1754 def
= SSA_NAME_DEF_STMT (op0
);
1756 /* If we were already here, break the infinite cycle. */
1757 if (!bitmap_set_bit (visited
, SSA_NAME_VERSION (op0
)))
1760 if (gimple_code (def
) == GIMPLE_PHI
)
1762 /* All the arguments of the PHI node must have the same constant
1764 int i
, n
= gimple_phi_num_args (def
);
1765 tree val
= NULL
, new_val
;
1767 for (i
= 0; i
< n
; i
++)
1769 tree arg
= PHI_ARG_DEF (def
, i
);
1770 enum br_predictor predictor2
;
1772 /* If this PHI has itself as an argument, we cannot
1773 determine the string length of this argument. However,
1774 if we can find an expected constant value for the other
1775 PHI args then we can still be sure that this is
1776 likely a constant. So be optimistic and just
1777 continue with the next argument. */
1778 if (arg
== PHI_RESULT (def
))
1781 new_val
= expr_expected_value (arg
, visited
, &predictor2
);
1783 /* It is difficult to combine value predictors. Simply assume
1784 that later predictor is weaker and take its prediction. */
1785 if (predictor
&& *predictor
< predictor2
)
1786 *predictor
= predictor2
;
1791 else if (!operand_equal_p (val
, new_val
, false))
1796 if (is_gimple_assign (def
))
1798 if (gimple_assign_lhs (def
) != op0
)
1801 return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def
)),
1802 gimple_assign_rhs1 (def
),
1803 gimple_assign_rhs_code (def
),
1804 gimple_assign_rhs2 (def
),
1805 visited
, predictor
);
1808 if (is_gimple_call (def
))
1810 tree decl
= gimple_call_fndecl (def
);
1813 if (gimple_call_internal_p (def
)
1814 && gimple_call_internal_fn (def
) == IFN_BUILTIN_EXPECT
)
1816 gcc_assert (gimple_call_num_args (def
) == 3);
1817 tree val
= gimple_call_arg (def
, 0);
1818 if (TREE_CONSTANT (val
))
1822 tree val2
= gimple_call_arg (def
, 2);
1823 gcc_assert (TREE_CODE (val2
) == INTEGER_CST
1824 && tree_fits_uhwi_p (val2
)
1825 && tree_to_uhwi (val2
) < END_PREDICTORS
);
1826 *predictor
= (enum br_predictor
) tree_to_uhwi (val2
);
1828 return gimple_call_arg (def
, 1);
1832 if (DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_NORMAL
)
1833 switch (DECL_FUNCTION_CODE (decl
))
1835 case BUILT_IN_EXPECT
:
1838 if (gimple_call_num_args (def
) != 2)
1840 val
= gimple_call_arg (def
, 0);
1841 if (TREE_CONSTANT (val
))
1844 *predictor
= PRED_BUILTIN_EXPECT
;
1845 return gimple_call_arg (def
, 1);
1848 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N
:
1849 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1
:
1850 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2
:
1851 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4
:
1852 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8
:
1853 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16
:
1854 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE
:
1855 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N
:
1856 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1
:
1857 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2
:
1858 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4
:
1859 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8
:
1860 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16
:
1861 /* Assume that any given atomic operation has low contention,
1862 and thus the compare-and-swap operation succeeds. */
1864 *predictor
= PRED_COMPARE_AND_SWAP
;
1865 return boolean_true_node
;
1874 if (get_gimple_rhs_class (code
) == GIMPLE_BINARY_RHS
)
1877 enum br_predictor predictor2
;
1878 op0
= expr_expected_value (op0
, visited
, predictor
);
1881 op1
= expr_expected_value (op1
, visited
, &predictor2
);
1882 if (predictor
&& *predictor
< predictor2
)
1883 *predictor
= predictor2
;
1886 res
= fold_build2 (code
, type
, op0
, op1
);
1887 if (TREE_CONSTANT (res
))
1891 if (get_gimple_rhs_class (code
) == GIMPLE_UNARY_RHS
)
1894 op0
= expr_expected_value (op0
, visited
, predictor
);
1897 res
= fold_build1 (code
, type
, op0
);
1898 if (TREE_CONSTANT (res
))
1905 /* Return constant EXPR will likely have at execution time, NULL if unknown.
1906 The function is used by builtin_expect branch predictor so the evidence
1907 must come from this construct and additional possible constant folding.
1909 We may want to implement more involved value guess (such as value range
1910 propagation based prediction), but such tricks shall go to new
1914 expr_expected_value (tree expr
, bitmap visited
,
1915 enum br_predictor
*predictor
)
1917 enum tree_code code
;
1920 if (TREE_CONSTANT (expr
))
1923 *predictor
= PRED_UNCONDITIONAL
;
1927 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
1928 return expr_expected_value_1 (TREE_TYPE (expr
),
1929 op0
, code
, op1
, visited
, predictor
);
1932 /* Predict using opcode of the last statement in basic block. */
1934 tree_predict_by_opcode (basic_block bb
)
1936 gimple stmt
= last_stmt (bb
);
1944 enum br_predictor predictor
;
1946 if (!stmt
|| gimple_code (stmt
) != GIMPLE_COND
)
1948 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
1949 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1951 op0
= gimple_cond_lhs (stmt
);
1952 op1
= gimple_cond_rhs (stmt
);
1953 cmp
= gimple_cond_code (stmt
);
1954 type
= TREE_TYPE (op0
);
1955 visited
= BITMAP_ALLOC (NULL
);
1956 val
= expr_expected_value_1 (boolean_type_node
, op0
, cmp
, op1
, visited
,
1958 BITMAP_FREE (visited
);
1959 if (val
&& TREE_CODE (val
) == INTEGER_CST
)
1961 if (predictor
== PRED_BUILTIN_EXPECT
)
1963 int percent
= PARAM_VALUE (BUILTIN_EXPECT_PROBABILITY
);
1965 gcc_assert (percent
>= 0 && percent
<= 100);
1966 if (integer_zerop (val
))
1967 percent
= 100 - percent
;
1968 predict_edge (then_edge
, PRED_BUILTIN_EXPECT
, HITRATE (percent
));
1971 predict_edge (then_edge
, predictor
,
1972 integer_zerop (val
) ? NOT_TAKEN
: TAKEN
);
1974 /* Try "pointer heuristic."
1975 A comparison ptr == 0 is predicted as false.
1976 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1977 if (POINTER_TYPE_P (type
))
1980 predict_edge_def (then_edge
, PRED_TREE_POINTER
, NOT_TAKEN
);
1981 else if (cmp
== NE_EXPR
)
1982 predict_edge_def (then_edge
, PRED_TREE_POINTER
, TAKEN
);
1986 /* Try "opcode heuristic."
1987 EQ tests are usually false and NE tests are usually true. Also,
1988 most quantities are positive, so we can make the appropriate guesses
1989 about signed comparisons against zero. */
1994 /* Floating point comparisons appears to behave in a very
1995 unpredictable way because of special role of = tests in
1997 if (FLOAT_TYPE_P (type
))
1999 /* Comparisons with 0 are often used for booleans and there is
2000 nothing useful to predict about them. */
2001 else if (integer_zerop (op0
) || integer_zerop (op1
))
2004 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, NOT_TAKEN
);
2009 /* Floating point comparisons appears to behave in a very
2010 unpredictable way because of special role of = tests in
2012 if (FLOAT_TYPE_P (type
))
2014 /* Comparisons with 0 are often used for booleans and there is
2015 nothing useful to predict about them. */
2016 else if (integer_zerop (op0
)
2017 || integer_zerop (op1
))
2020 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, TAKEN
);
2024 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, TAKEN
);
2027 case UNORDERED_EXPR
:
2028 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, NOT_TAKEN
);
2033 if (integer_zerop (op1
)
2034 || integer_onep (op1
)
2035 || integer_all_onesp (op1
)
2038 || real_minus_onep (op1
))
2039 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, NOT_TAKEN
);
2044 if (integer_zerop (op1
)
2045 || integer_onep (op1
)
2046 || integer_all_onesp (op1
)
2049 || real_minus_onep (op1
))
2050 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, TAKEN
);
2058 /* Try to guess whether the value of return means error code. */
2060 static enum br_predictor
2061 return_prediction (tree val
, enum prediction
*prediction
)
2065 return PRED_NO_PREDICTION
;
2066 /* Different heuristics for pointers and scalars. */
2067 if (POINTER_TYPE_P (TREE_TYPE (val
)))
2069 /* NULL is usually not returned. */
2070 if (integer_zerop (val
))
2072 *prediction
= NOT_TAKEN
;
2073 return PRED_NULL_RETURN
;
2076 else if (INTEGRAL_TYPE_P (TREE_TYPE (val
)))
2078 /* Negative return values are often used to indicate
2080 if (TREE_CODE (val
) == INTEGER_CST
2081 && tree_int_cst_sgn (val
) < 0)
2083 *prediction
= NOT_TAKEN
;
2084 return PRED_NEGATIVE_RETURN
;
2086 /* Constant return values seems to be commonly taken.
2087 Zero/one often represent booleans so exclude them from the
2089 if (TREE_CONSTANT (val
)
2090 && (!integer_zerop (val
) && !integer_onep (val
)))
2092 *prediction
= TAKEN
;
2093 return PRED_CONST_RETURN
;
2096 return PRED_NO_PREDICTION
;
2099 /* Find the basic block with return expression and look up for possible
2100 return value trying to apply RETURN_PREDICTION heuristics. */
2102 apply_return_prediction (void)
2104 gimple return_stmt
= NULL
;
2108 int phi_num_args
, i
;
2109 enum br_predictor pred
;
2110 enum prediction direction
;
2113 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2115 return_stmt
= last_stmt (e
->src
);
2117 && gimple_code (return_stmt
) == GIMPLE_RETURN
)
2122 return_val
= gimple_return_retval (return_stmt
);
2125 if (TREE_CODE (return_val
) != SSA_NAME
2126 || !SSA_NAME_DEF_STMT (return_val
)
2127 || gimple_code (SSA_NAME_DEF_STMT (return_val
)) != GIMPLE_PHI
)
2129 phi
= SSA_NAME_DEF_STMT (return_val
);
2130 phi_num_args
= gimple_phi_num_args (phi
);
2131 pred
= return_prediction (PHI_ARG_DEF (phi
, 0), &direction
);
2133 /* Avoid the degenerate case where all return values form the function
2134 belongs to same category (ie they are all positive constants)
2135 so we can hardly say something about them. */
2136 for (i
= 1; i
< phi_num_args
; i
++)
2137 if (pred
!= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
))
2139 if (i
!= phi_num_args
)
2140 for (i
= 0; i
< phi_num_args
; i
++)
2142 pred
= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
);
2143 if (pred
!= PRED_NO_PREDICTION
)
2144 predict_paths_leading_to_edge (gimple_phi_arg_edge (phi
, i
), pred
,
2149 /* Look for basic block that contains unlikely to happen events
2150 (such as noreturn calls) and mark all paths leading to execution
2151 of this basic blocks as unlikely. */
2154 tree_bb_level_predictions (void)
2157 bool has_return_edges
= false;
2161 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2162 if (!(e
->flags
& (EDGE_ABNORMAL
| EDGE_FAKE
| EDGE_EH
)))
2164 has_return_edges
= true;
2168 apply_return_prediction ();
2170 FOR_EACH_BB_FN (bb
, cfun
)
2172 gimple_stmt_iterator gsi
;
2174 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2176 gimple stmt
= gsi_stmt (gsi
);
2179 if (is_gimple_call (stmt
))
2181 if ((gimple_call_flags (stmt
) & ECF_NORETURN
)
2182 && has_return_edges
)
2183 predict_paths_leading_to (bb
, PRED_NORETURN
,
2185 decl
= gimple_call_fndecl (stmt
);
2187 && lookup_attribute ("cold",
2188 DECL_ATTRIBUTES (decl
)))
2189 predict_paths_leading_to (bb
, PRED_COLD_FUNCTION
,
2192 else if (gimple_code (stmt
) == GIMPLE_PREDICT
)
2194 predict_paths_leading_to (bb
, gimple_predict_predictor (stmt
),
2195 gimple_predict_outcome (stmt
));
2196 /* Keep GIMPLE_PREDICT around so early inlining will propagate
2197 hints to callers. */
2203 #ifdef ENABLE_CHECKING
2205 /* Callback for hash_map::traverse, asserts that the pointer map is
2209 assert_is_empty (const_basic_block
const &, edge_prediction
*const &value
,
2212 gcc_assert (!value
);
2217 /* Predict branch probabilities and estimate profile for basic block BB. */
2220 tree_estimate_probability_bb (basic_block bb
)
2226 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2228 /* Predict edges to user labels with attributes. */
2229 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
2231 gimple_stmt_iterator gi
;
2232 for (gi
= gsi_start_bb (e
->dest
); !gsi_end_p (gi
); gsi_next (&gi
))
2234 gimple stmt
= gsi_stmt (gi
);
2237 if (gimple_code (stmt
) != GIMPLE_LABEL
)
2239 decl
= gimple_label_label (stmt
);
2240 if (DECL_ARTIFICIAL (decl
))
2243 /* Finally, we have a user-defined label. */
2244 if (lookup_attribute ("cold", DECL_ATTRIBUTES (decl
)))
2245 predict_edge_def (e
, PRED_COLD_LABEL
, NOT_TAKEN
);
2246 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (decl
)))
2247 predict_edge_def (e
, PRED_HOT_LABEL
, TAKEN
);
2251 /* Predict early returns to be probable, as we've already taken
2252 care for error returns and other cases are often used for
2253 fast paths through function.
2255 Since we've already removed the return statements, we are
2256 looking for CFG like:
2266 if (e
->dest
!= bb
->next_bb
2267 && e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
2268 && single_succ_p (e
->dest
)
2269 && single_succ_edge (e
->dest
)->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
)
2270 && (last
= last_stmt (e
->dest
)) != NULL
2271 && gimple_code (last
) == GIMPLE_RETURN
)
2276 if (single_succ_p (bb
))
2278 FOR_EACH_EDGE (e1
, ei1
, bb
->preds
)
2279 if (!predicted_by_p (e1
->src
, PRED_NULL_RETURN
)
2280 && !predicted_by_p (e1
->src
, PRED_CONST_RETURN
)
2281 && !predicted_by_p (e1
->src
, PRED_NEGATIVE_RETURN
))
2282 predict_edge_def (e1
, PRED_TREE_EARLY_RETURN
, NOT_TAKEN
);
2285 if (!predicted_by_p (e
->src
, PRED_NULL_RETURN
)
2286 && !predicted_by_p (e
->src
, PRED_CONST_RETURN
)
2287 && !predicted_by_p (e
->src
, PRED_NEGATIVE_RETURN
))
2288 predict_edge_def (e
, PRED_TREE_EARLY_RETURN
, NOT_TAKEN
);
2291 /* Look for block we are guarding (ie we dominate it,
2292 but it doesn't postdominate us). */
2293 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
) && e
->dest
!= bb
2294 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
2295 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
2297 gimple_stmt_iterator bi
;
2299 /* The call heuristic claims that a guarded function call
2300 is improbable. This is because such calls are often used
2301 to signal exceptional situations such as printing error
2303 for (bi
= gsi_start_bb (e
->dest
); !gsi_end_p (bi
);
2306 gimple stmt
= gsi_stmt (bi
);
2307 if (is_gimple_call (stmt
)
2308 /* Constant and pure calls are hardly used to signalize
2309 something exceptional. */
2310 && gimple_has_side_effects (stmt
))
2312 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
2318 tree_predict_by_opcode (bb
);
2321 /* Predict branch probabilities and estimate profile of the tree CFG.
2322 This function can be called from the loop optimizers to recompute
2323 the profile information. */
2326 tree_estimate_probability (void)
2330 add_noreturn_fake_exit_edges ();
2331 connect_infinite_loops_to_exit ();
2332 /* We use loop_niter_by_eval, which requires that the loops have
2334 create_preheaders (CP_SIMPLE_PREHEADERS
);
2335 calculate_dominance_info (CDI_POST_DOMINATORS
);
2337 bb_predictions
= new hash_map
<const_basic_block
, edge_prediction
*>;
2338 tree_bb_level_predictions ();
2339 record_loop_exits ();
2341 if (number_of_loops (cfun
) > 1)
2344 FOR_EACH_BB_FN (bb
, cfun
)
2345 tree_estimate_probability_bb (bb
);
2347 FOR_EACH_BB_FN (bb
, cfun
)
2348 combine_predictions_for_bb (bb
);
2350 #ifdef ENABLE_CHECKING
2351 bb_predictions
->traverse
<void *, assert_is_empty
> (NULL
);
2353 delete bb_predictions
;
2354 bb_predictions
= NULL
;
2356 estimate_bb_frequencies (false);
2357 free_dominance_info (CDI_POST_DOMINATORS
);
2358 remove_fake_exit_edges ();
2361 /* Predict edges to successors of CUR whose sources are not postdominated by
2362 BB by PRED and recurse to all postdominators. */
2365 predict_paths_for_bb (basic_block cur
, basic_block bb
,
2366 enum br_predictor pred
,
2367 enum prediction taken
,
2374 /* We are looking for all edges forming edge cut induced by
2375 set of all blocks postdominated by BB. */
2376 FOR_EACH_EDGE (e
, ei
, cur
->preds
)
2377 if (e
->src
->index
>= NUM_FIXED_BLOCKS
2378 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, bb
))
2384 /* Ignore fake edges and eh, we predict them as not taken anyway. */
2385 if (e
->flags
& (EDGE_EH
| EDGE_FAKE
))
2387 gcc_assert (bb
== cur
|| dominated_by_p (CDI_POST_DOMINATORS
, cur
, bb
));
2389 /* See if there is an edge from e->src that is not abnormal
2390 and does not lead to BB. */
2391 FOR_EACH_EDGE (e2
, ei2
, e
->src
->succs
)
2393 && !(e2
->flags
& (EDGE_EH
| EDGE_FAKE
))
2394 && !dominated_by_p (CDI_POST_DOMINATORS
, e2
->dest
, bb
))
2400 /* If there is non-abnormal path leaving e->src, predict edge
2401 using predictor. Otherwise we need to look for paths
2404 The second may lead to infinite loop in the case we are predicitng
2405 regions that are only reachable by abnormal edges. We simply
2406 prevent visiting given BB twice. */
2408 predict_edge_def (e
, pred
, taken
);
2409 else if (bitmap_set_bit (visited
, e
->src
->index
))
2410 predict_paths_for_bb (e
->src
, e
->src
, pred
, taken
, visited
);
2412 for (son
= first_dom_son (CDI_POST_DOMINATORS
, cur
);
2414 son
= next_dom_son (CDI_POST_DOMINATORS
, son
))
2415 predict_paths_for_bb (son
, bb
, pred
, taken
, visited
);
2418 /* Sets branch probabilities according to PREDiction and
2422 predict_paths_leading_to (basic_block bb
, enum br_predictor pred
,
2423 enum prediction taken
)
2425 bitmap visited
= BITMAP_ALLOC (NULL
);
2426 predict_paths_for_bb (bb
, bb
, pred
, taken
, visited
);
2427 BITMAP_FREE (visited
);
2430 /* Like predict_paths_leading_to but take edge instead of basic block. */
2433 predict_paths_leading_to_edge (edge e
, enum br_predictor pred
,
2434 enum prediction taken
)
2436 bool has_nonloop_edge
= false;
2440 basic_block bb
= e
->src
;
2441 FOR_EACH_EDGE (e2
, ei
, bb
->succs
)
2442 if (e2
->dest
!= e
->src
&& e2
->dest
!= e
->dest
2443 && !(e
->flags
& (EDGE_EH
| EDGE_FAKE
))
2444 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e2
->dest
))
2446 has_nonloop_edge
= true;
2449 if (!has_nonloop_edge
)
2451 bitmap visited
= BITMAP_ALLOC (NULL
);
2452 predict_paths_for_bb (bb
, bb
, pred
, taken
, visited
);
2453 BITMAP_FREE (visited
);
2456 predict_edge_def (e
, pred
, taken
);
2459 /* This is used to carry information about basic blocks. It is
2460 attached to the AUX field of the standard CFG block. */
2464 /* Estimated frequency of execution of basic_block. */
2467 /* To keep queue of basic blocks to process. */
2470 /* Number of predecessors we need to visit first. */
2474 /* Similar information for edges. */
2475 struct edge_prob_info
2477 /* In case edge is a loopback edge, the probability edge will be reached
2478 in case header is. Estimated number of iterations of the loop can be
2479 then computed as 1 / (1 - back_edge_prob). */
2480 sreal back_edge_prob
;
2481 /* True if the edge is a loopback edge in the natural loop. */
2482 unsigned int back_edge
:1;
2485 #define BLOCK_INFO(B) ((block_info *) (B)->aux)
2487 #define EDGE_INFO(E) ((edge_prob_info *) (E)->aux)
2489 /* Helper function for estimate_bb_frequencies.
2490 Propagate the frequencies in blocks marked in
2491 TOVISIT, starting in HEAD. */
2494 propagate_freq (basic_block head
, bitmap tovisit
)
2503 /* For each basic block we need to visit count number of his predecessors
2504 we need to visit first. */
2505 EXECUTE_IF_SET_IN_BITMAP (tovisit
, 0, i
, bi
)
2510 bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
2512 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2514 bool visit
= bitmap_bit_p (tovisit
, e
->src
->index
);
2516 if (visit
&& !(e
->flags
& EDGE_DFS_BACK
))
2518 else if (visit
&& dump_file
&& !EDGE_INFO (e
)->back_edge
)
2520 "Irreducible region hit, ignoring edge to %i->%i\n",
2521 e
->src
->index
, bb
->index
);
2523 BLOCK_INFO (bb
)->npredecessors
= count
;
2524 /* When function never returns, we will never process exit block. */
2525 if (!count
&& bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
2526 bb
->count
= bb
->frequency
= 0;
2529 BLOCK_INFO (head
)->frequency
= real_one
;
2531 for (bb
= head
; bb
; bb
= nextbb
)
2534 sreal cyclic_probability
= real_zero
;
2535 sreal frequency
= real_zero
;
2537 nextbb
= BLOCK_INFO (bb
)->next
;
2538 BLOCK_INFO (bb
)->next
= NULL
;
2540 /* Compute frequency of basic block. */
2543 #ifdef ENABLE_CHECKING
2544 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2545 gcc_assert (!bitmap_bit_p (tovisit
, e
->src
->index
)
2546 || (e
->flags
& EDGE_DFS_BACK
));
2549 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2550 if (EDGE_INFO (e
)->back_edge
)
2552 cyclic_probability
+= EDGE_INFO (e
)->back_edge_prob
;
2554 else if (!(e
->flags
& EDGE_DFS_BACK
))
2556 /* frequency += (e->probability
2557 * BLOCK_INFO (e->src)->frequency /
2558 REG_BR_PROB_BASE); */
2560 sreal
tmp (e
->probability
, 0);
2561 tmp
*= BLOCK_INFO (e
->src
)->frequency
;
2562 tmp
*= real_inv_br_prob_base
;
2566 if (cyclic_probability
== real_zero
)
2568 BLOCK_INFO (bb
)->frequency
= frequency
;
2572 if (cyclic_probability
> real_almost_one
)
2573 cyclic_probability
= real_almost_one
;
2575 /* BLOCK_INFO (bb)->frequency = frequency
2576 / (1 - cyclic_probability) */
2578 cyclic_probability
= real_one
- cyclic_probability
;
2579 BLOCK_INFO (bb
)->frequency
= frequency
/ cyclic_probability
;
2583 bitmap_clear_bit (tovisit
, bb
->index
);
2585 e
= find_edge (bb
, head
);
2588 /* EDGE_INFO (e)->back_edge_prob
2589 = ((e->probability * BLOCK_INFO (bb)->frequency)
2590 / REG_BR_PROB_BASE); */
2592 sreal
tmp (e
->probability
, 0);
2593 tmp
*= BLOCK_INFO (bb
)->frequency
;
2594 EDGE_INFO (e
)->back_edge_prob
= tmp
* real_inv_br_prob_base
;
2597 /* Propagate to successor blocks. */
2598 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2599 if (!(e
->flags
& EDGE_DFS_BACK
)
2600 && BLOCK_INFO (e
->dest
)->npredecessors
)
2602 BLOCK_INFO (e
->dest
)->npredecessors
--;
2603 if (!BLOCK_INFO (e
->dest
)->npredecessors
)
2608 BLOCK_INFO (last
)->next
= e
->dest
;
2616 /* Estimate frequencies in loops at same nest level. */
2619 estimate_loops_at_level (struct loop
*first_loop
)
2623 for (loop
= first_loop
; loop
; loop
= loop
->next
)
2628 bitmap tovisit
= BITMAP_ALLOC (NULL
);
2630 estimate_loops_at_level (loop
->inner
);
2632 /* Find current loop back edge and mark it. */
2633 e
= loop_latch_edge (loop
);
2634 EDGE_INFO (e
)->back_edge
= 1;
2636 bbs
= get_loop_body (loop
);
2637 for (i
= 0; i
< loop
->num_nodes
; i
++)
2638 bitmap_set_bit (tovisit
, bbs
[i
]->index
);
2640 propagate_freq (loop
->header
, tovisit
);
2641 BITMAP_FREE (tovisit
);
2645 /* Propagates frequencies through structure of loops. */
2648 estimate_loops (void)
2650 bitmap tovisit
= BITMAP_ALLOC (NULL
);
2653 /* Start by estimating the frequencies in the loops. */
2654 if (number_of_loops (cfun
) > 1)
2655 estimate_loops_at_level (current_loops
->tree_root
->inner
);
2657 /* Now propagate the frequencies through all the blocks. */
2658 FOR_ALL_BB_FN (bb
, cfun
)
2660 bitmap_set_bit (tovisit
, bb
->index
);
2662 propagate_freq (ENTRY_BLOCK_PTR_FOR_FN (cfun
), tovisit
);
2663 BITMAP_FREE (tovisit
);
2666 /* Drop the profile for NODE to guessed, and update its frequency based on
2667 whether it is expected to be hot given the CALL_COUNT. */
2670 drop_profile (struct cgraph_node
*node
, gcov_type call_count
)
2672 struct function
*fn
= DECL_STRUCT_FUNCTION (node
->decl
);
2673 /* In the case where this was called by another function with a
2674 dropped profile, call_count will be 0. Since there are no
2675 non-zero call counts to this function, we don't know for sure
2676 whether it is hot, and therefore it will be marked normal below. */
2677 bool hot
= maybe_hot_count_p (NULL
, call_count
);
2681 "Dropping 0 profile for %s/%i. %s based on calls.\n",
2682 node
->name (), node
->order
,
2683 hot
? "Function is hot" : "Function is normal");
2684 /* We only expect to miss profiles for functions that are reached
2685 via non-zero call edges in cases where the function may have
2686 been linked from another module or library (COMDATs and extern
2687 templates). See the comments below for handle_missing_profiles.
2688 Also, only warn in cases where the missing counts exceed the
2689 number of training runs. In certain cases with an execv followed
2690 by a no-return call the profile for the no-return call is not
2691 dumped and there can be a mismatch. */
2692 if (!DECL_COMDAT (node
->decl
) && !DECL_EXTERNAL (node
->decl
)
2693 && call_count
> profile_info
->runs
)
2695 if (flag_profile_correction
)
2699 "Missing counts for called function %s/%i\n",
2700 node
->name (), node
->order
);
2703 warning (0, "Missing counts for called function %s/%i",
2704 node
->name (), node
->order
);
2707 profile_status_for_fn (fn
)
2708 = (flag_guess_branch_prob
? PROFILE_GUESSED
: PROFILE_ABSENT
);
2710 = hot
? NODE_FREQUENCY_HOT
: NODE_FREQUENCY_NORMAL
;
2713 /* In the case of COMDAT routines, multiple object files will contain the same
2714 function and the linker will select one for the binary. In that case
2715 all the other copies from the profile instrument binary will be missing
2716 profile counts. Look for cases where this happened, due to non-zero
2717 call counts going to 0-count functions, and drop the profile to guessed
2718 so that we can use the estimated probabilities and avoid optimizing only
2721 The other case where the profile may be missing is when the routine
2722 is not going to be emitted to the object file, e.g. for "extern template"
2723 class methods. Those will be marked DECL_EXTERNAL. Emit a warning in
2724 all other cases of non-zero calls to 0-count functions. */
2727 handle_missing_profiles (void)
2729 struct cgraph_node
*node
;
2730 int unlikely_count_fraction
= PARAM_VALUE (UNLIKELY_BB_COUNT_FRACTION
);
2731 vec
<struct cgraph_node
*> worklist
;
2732 worklist
.create (64);
2734 /* See if 0 count function has non-0 count callers. In this case we
2735 lost some profile. Drop its function profile to PROFILE_GUESSED. */
2736 FOR_EACH_DEFINED_FUNCTION (node
)
2738 struct cgraph_edge
*e
;
2739 gcov_type call_count
= 0;
2740 gcov_type max_tp_first_run
= 0;
2741 struct function
*fn
= DECL_STRUCT_FUNCTION (node
->decl
);
2745 for (e
= node
->callers
; e
; e
= e
->next_caller
)
2747 call_count
+= e
->count
;
2749 if (e
->caller
->tp_first_run
> max_tp_first_run
)
2750 max_tp_first_run
= e
->caller
->tp_first_run
;
2753 /* If time profile is missing, let assign the maximum that comes from
2754 caller functions. */
2755 if (!node
->tp_first_run
&& max_tp_first_run
)
2756 node
->tp_first_run
= max_tp_first_run
+ 1;
2760 && (call_count
* unlikely_count_fraction
>= profile_info
->runs
))
2762 drop_profile (node
, call_count
);
2763 worklist
.safe_push (node
);
2767 /* Propagate the profile dropping to other 0-count COMDATs that are
2768 potentially called by COMDATs we already dropped the profile on. */
2769 while (worklist
.length () > 0)
2771 struct cgraph_edge
*e
;
2773 node
= worklist
.pop ();
2774 for (e
= node
->callees
; e
; e
= e
->next_caller
)
2776 struct cgraph_node
*callee
= e
->callee
;
2777 struct function
*fn
= DECL_STRUCT_FUNCTION (callee
->decl
);
2779 if (callee
->count
> 0)
2781 if (DECL_COMDAT (callee
->decl
) && fn
&& fn
->cfg
2782 && profile_status_for_fn (fn
) == PROFILE_READ
)
2784 drop_profile (node
, 0);
2785 worklist
.safe_push (callee
);
2789 worklist
.release ();
2792 /* Convert counts measured by profile driven feedback to frequencies.
2793 Return nonzero iff there was any nonzero execution count. */
2796 counts_to_freqs (void)
2798 gcov_type count_max
, true_count_max
= 0;
2801 /* Don't overwrite the estimated frequencies when the profile for
2802 the function is missing. We may drop this function PROFILE_GUESSED
2803 later in drop_profile (). */
2804 if (!flag_auto_profile
&& !ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
)
2807 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
2808 true_count_max
= MAX (bb
->count
, true_count_max
);
2810 count_max
= MAX (true_count_max
, 1);
2811 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
2812 bb
->frequency
= (bb
->count
* BB_FREQ_MAX
+ count_max
/ 2) / count_max
;
2814 return true_count_max
;
2817 /* Return true if function is likely to be expensive, so there is no point to
2818 optimize performance of prologue, epilogue or do inlining at the expense
2819 of code size growth. THRESHOLD is the limit of number of instructions
2820 function can execute at average to be still considered not expensive. */
2823 expensive_function_p (int threshold
)
2825 unsigned int sum
= 0;
2829 /* We can not compute accurately for large thresholds due to scaled
2831 gcc_assert (threshold
<= BB_FREQ_MAX
);
2833 /* Frequencies are out of range. This either means that function contains
2834 internal loop executing more than BB_FREQ_MAX times or profile feedback
2835 is available and function has not been executed at all. */
2836 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
== 0)
2839 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
2840 limit
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->frequency
* threshold
;
2841 FOR_EACH_BB_FN (bb
, cfun
)
2845 FOR_BB_INSNS (bb
, insn
)
2846 if (active_insn_p (insn
))
2848 sum
+= bb
->frequency
;
2857 /* Estimate and propagate basic block frequencies using the given branch
2858 probabilities. If FORCE is true, the frequencies are used to estimate
2859 the counts even when there are already non-zero profile counts. */
2862 estimate_bb_frequencies (bool force
)
2867 if (force
|| profile_status_for_fn (cfun
) != PROFILE_READ
|| !counts_to_freqs ())
2869 static int real_values_initialized
= 0;
2871 if (!real_values_initialized
)
2873 real_values_initialized
= 1;
2874 real_zero
= sreal (0, 0);
2875 real_one
= sreal (1, 0);
2876 real_br_prob_base
= sreal (REG_BR_PROB_BASE
, 0);
2877 real_bb_freq_max
= sreal (BB_FREQ_MAX
, 0);
2878 real_one_half
= sreal (1, -1);
2879 real_inv_br_prob_base
= real_one
/ real_br_prob_base
;
2880 real_almost_one
= real_one
- real_inv_br_prob_base
;
2883 mark_dfs_back_edges ();
2885 single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
))->probability
=
2888 /* Set up block info for each basic block. */
2889 alloc_aux_for_blocks (sizeof (block_info
));
2890 alloc_aux_for_edges (sizeof (edge_prob_info
));
2891 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
2896 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2898 EDGE_INFO (e
)->back_edge_prob
= sreal (e
->probability
, 0);
2899 EDGE_INFO (e
)->back_edge_prob
*= real_inv_br_prob_base
;
2903 /* First compute frequencies locally for each loop from innermost
2904 to outermost to examine frequencies for back edges. */
2907 freq_max
= real_zero
;
2908 FOR_EACH_BB_FN (bb
, cfun
)
2909 if (freq_max
< BLOCK_INFO (bb
)->frequency
)
2910 freq_max
= BLOCK_INFO (bb
)->frequency
;
2912 freq_max
= real_bb_freq_max
/ freq_max
;
2913 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
2915 sreal tmp
= BLOCK_INFO (bb
)->frequency
* freq_max
+ real_one_half
;
2916 bb
->frequency
= tmp
.to_int ();
2919 free_aux_for_blocks ();
2920 free_aux_for_edges ();
2922 compute_function_frequency ();
2925 /* Decide whether function is hot, cold or unlikely executed. */
2927 compute_function_frequency (void)
2930 struct cgraph_node
*node
= cgraph_node::get (current_function_decl
);
2932 if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
2933 || MAIN_NAME_P (DECL_NAME (current_function_decl
)))
2934 node
->only_called_at_startup
= true;
2935 if (DECL_STATIC_DESTRUCTOR (current_function_decl
))
2936 node
->only_called_at_exit
= true;
2938 if (profile_status_for_fn (cfun
) != PROFILE_READ
)
2940 int flags
= flags_from_decl_or_type (current_function_decl
);
2941 if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
))
2943 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
2944 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl
))
2946 node
->frequency
= NODE_FREQUENCY_HOT
;
2947 else if (flags
& ECF_NORETURN
)
2948 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2949 else if (MAIN_NAME_P (DECL_NAME (current_function_decl
)))
2950 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2951 else if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
2952 || DECL_STATIC_DESTRUCTOR (current_function_decl
))
2953 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2957 /* Only first time try to drop function into unlikely executed.
2958 After inlining the roundoff errors may confuse us.
2959 Ipa-profile pass will drop functions only called from unlikely
2960 functions to unlikely and that is most of what we care about. */
2961 if (!cfun
->after_inlining
)
2962 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
2963 FOR_EACH_BB_FN (bb
, cfun
)
2965 if (maybe_hot_bb_p (cfun
, bb
))
2967 node
->frequency
= NODE_FREQUENCY_HOT
;
2970 if (!probably_never_executed_bb_p (cfun
, bb
))
2971 node
->frequency
= NODE_FREQUENCY_NORMAL
;
2975 /* Build PREDICT_EXPR. */
2977 build_predict_expr (enum br_predictor predictor
, enum prediction taken
)
2979 tree t
= build1 (PREDICT_EXPR
, void_type_node
,
2980 build_int_cst (integer_type_node
, predictor
));
2981 SET_PREDICT_EXPR_OUTCOME (t
, taken
);
2986 predictor_name (enum br_predictor predictor
)
2988 return predictor_info
[predictor
].name
;
2991 /* Predict branch probabilities and estimate profile of the tree CFG. */
2995 const pass_data pass_data_profile
=
2997 GIMPLE_PASS
, /* type */
2998 "profile_estimate", /* name */
2999 OPTGROUP_NONE
, /* optinfo_flags */
3000 TV_BRANCH_PROB
, /* tv_id */
3001 PROP_cfg
, /* properties_required */
3002 0, /* properties_provided */
3003 0, /* properties_destroyed */
3004 0, /* todo_flags_start */
3005 0, /* todo_flags_finish */
3008 class pass_profile
: public gimple_opt_pass
3011 pass_profile (gcc::context
*ctxt
)
3012 : gimple_opt_pass (pass_data_profile
, ctxt
)
3015 /* opt_pass methods: */
3016 virtual bool gate (function
*) { return flag_guess_branch_prob
; }
3017 virtual unsigned int execute (function
*);
3019 }; // class pass_profile
3022 pass_profile::execute (function
*fun
)
3026 loop_optimizer_init (LOOPS_NORMAL
);
3027 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3028 flow_loops_dump (dump_file
, NULL
, 0);
3030 mark_irreducible_loops ();
3032 nb_loops
= number_of_loops (fun
);
3036 tree_estimate_probability ();
3041 loop_optimizer_finalize ();
3042 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3043 gimple_dump_cfg (dump_file
, dump_flags
);
3044 if (profile_status_for_fn (fun
) == PROFILE_ABSENT
)
3045 profile_status_for_fn (fun
) = PROFILE_GUESSED
;
3052 make_pass_profile (gcc::context
*ctxt
)
3054 return new pass_profile (ctxt
);
3059 const pass_data pass_data_strip_predict_hints
=
3061 GIMPLE_PASS
, /* type */
3062 "*strip_predict_hints", /* name */
3063 OPTGROUP_NONE
, /* optinfo_flags */
3064 TV_BRANCH_PROB
, /* tv_id */
3065 PROP_cfg
, /* properties_required */
3066 0, /* properties_provided */
3067 0, /* properties_destroyed */
3068 0, /* todo_flags_start */
3069 0, /* todo_flags_finish */
3072 class pass_strip_predict_hints
: public gimple_opt_pass
3075 pass_strip_predict_hints (gcc::context
*ctxt
)
3076 : gimple_opt_pass (pass_data_strip_predict_hints
, ctxt
)
3079 /* opt_pass methods: */
3080 opt_pass
* clone () { return new pass_strip_predict_hints (m_ctxt
); }
3081 virtual unsigned int execute (function
*);
3083 }; // class pass_strip_predict_hints
3085 /* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements
3086 we no longer need. */
3088 pass_strip_predict_hints::execute (function
*fun
)
3094 FOR_EACH_BB_FN (bb
, fun
)
3096 gimple_stmt_iterator bi
;
3097 for (bi
= gsi_start_bb (bb
); !gsi_end_p (bi
);)
3099 gimple stmt
= gsi_stmt (bi
);
3101 if (gimple_code (stmt
) == GIMPLE_PREDICT
)
3103 gsi_remove (&bi
, true);
3106 else if (is_gimple_call (stmt
))
3108 tree fndecl
= gimple_call_fndecl (stmt
);
3111 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
3112 && DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_EXPECT
3113 && gimple_call_num_args (stmt
) == 2)
3114 || (gimple_call_internal_p (stmt
)
3115 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
))
3117 var
= gimple_call_lhs (stmt
);
3121 = gimple_build_assign (var
, gimple_call_arg (stmt
, 0));
3122 gsi_replace (&bi
, ass_stmt
, true);
3126 gsi_remove (&bi
, true);
3140 make_pass_strip_predict_hints (gcc::context
*ctxt
)
3142 return new pass_strip_predict_hints (ctxt
);
3145 /* Rebuild function frequencies. Passes are in general expected to
3146 maintain profile by hand, however in some cases this is not possible:
3147 for example when inlining several functions with loops freuqencies might run
3148 out of scale and thus needs to be recomputed. */
3151 rebuild_frequencies (void)
3153 timevar_push (TV_REBUILD_FREQUENCIES
);
3155 /* When the max bb count in the function is small, there is a higher
3156 chance that there were truncation errors in the integer scaling
3157 of counts by inlining and other optimizations. This could lead
3158 to incorrect classification of code as being cold when it isn't.
3159 In that case, force the estimation of bb counts/frequencies from the
3160 branch probabilities, rather than computing frequencies from counts,
3161 which may also lead to frequencies incorrectly reduced to 0. There
3162 is less precision in the probabilities, so we only do this for small
3164 gcov_type count_max
= 0;
3166 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
3167 count_max
= MAX (bb
->count
, count_max
);
3169 if (profile_status_for_fn (cfun
) == PROFILE_GUESSED
3170 || (!flag_auto_profile
&& profile_status_for_fn (cfun
) == PROFILE_READ
3171 && count_max
< REG_BR_PROB_BASE
/10))
3173 loop_optimizer_init (0);
3174 add_noreturn_fake_exit_edges ();
3175 mark_irreducible_loops ();
3176 connect_infinite_loops_to_exit ();
3177 estimate_bb_frequencies (true);
3178 remove_fake_exit_edges ();
3179 loop_optimizer_finalize ();
3181 else if (profile_status_for_fn (cfun
) == PROFILE_READ
)
3185 timevar_pop (TV_REBUILD_FREQUENCIES
);