1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 [1] "Branch Prediction for Free"
24 Ball and Larus; PLDI '93.
25 [2] "Static Branch Frequency and Program Profile Analysis"
26 Wu and Larus; MICRO-27.
27 [3] "Corpus-based Static Branch Prediction"
28 Calder, Grunwald, Lindsay, Martin, Mozer, and Zorn; PLDI '95. */
33 #include "coretypes.h"
38 #include "hard-reg-set.h"
39 #include "basic-block.h"
40 #include "insn-config.h"
46 #include "diagnostic-core.h"
55 #include "tree-flow.h"
57 #include "tree-dump.h"
58 #include "tree-pass.h"
60 #include "tree-scalar-evolution.h"
62 #include "pointer-set.h"
64 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
65 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
66 static sreal real_zero
, real_one
, real_almost_one
, real_br_prob_base
,
67 real_inv_br_prob_base
, real_one_half
, real_bb_freq_max
;
69 /* Random guesstimation given names.
70 PROV_VERY_UNLIKELY should be small enough so basic block predicted
71 by it gets bellow HOT_BB_FREQUENCY_FRANCTION. */
72 #define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 2000 - 1)
73 #define PROB_EVEN (REG_BR_PROB_BASE / 2)
74 #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
75 #define PROB_ALWAYS (REG_BR_PROB_BASE)
77 static void combine_predictions_for_insn (rtx
, basic_block
);
78 static void dump_prediction (FILE *, enum br_predictor
, int, basic_block
, int);
79 static void predict_paths_leading_to (basic_block
, enum br_predictor
, enum prediction
);
80 static void predict_paths_leading_to_edge (edge
, enum br_predictor
, enum prediction
);
81 static bool can_predict_insn_p (const_rtx
);
83 /* Information we hold about each branch predictor.
84 Filled using information from predict.def. */
88 const char *const name
; /* Name used in the debugging dumps. */
89 const int hitrate
; /* Expected hitrate used by
90 predict_insn_def call. */
94 /* Use given predictor without Dempster-Shaffer theory if it matches
95 using first_match heuristics. */
96 #define PRED_FLAG_FIRST_MATCH 1
98 /* Recompute hitrate in percent to our representation. */
100 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
102 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
103 static const struct predictor_info predictor_info
[]= {
104 #include "predict.def"
106 /* Upper bound on predictors. */
111 /* Return TRUE if frequency FREQ is considered to be hot. */
114 maybe_hot_frequency_p (int freq
)
116 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
117 if (!profile_info
|| !flag_branch_probabilities
)
119 if (node
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
)
121 if (node
->frequency
== NODE_FREQUENCY_HOT
)
124 if (profile_status
== PROFILE_ABSENT
)
126 if (node
->frequency
== NODE_FREQUENCY_EXECUTED_ONCE
127 && freq
< (ENTRY_BLOCK_PTR
->frequency
* 2 / 3))
129 if (freq
< ENTRY_BLOCK_PTR
->frequency
/ PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
))
134 /* Return TRUE if frequency FREQ is considered to be hot. */
137 maybe_hot_count_p (gcov_type count
)
139 if (profile_status
!= PROFILE_READ
)
141 /* Code executed at most once is not hot. */
142 if (profile_info
->runs
>= count
)
145 > profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
));
148 /* Return true in case BB can be CPU intensive and should be optimized
149 for maximal performance. */
152 maybe_hot_bb_p (const_basic_block bb
)
154 if (profile_status
== PROFILE_READ
)
155 return maybe_hot_count_p (bb
->count
);
156 return maybe_hot_frequency_p (bb
->frequency
);
159 /* Return true if the call can be hot. */
162 cgraph_maybe_hot_edge_p (struct cgraph_edge
*edge
)
164 if (profile_info
&& flag_branch_probabilities
166 <= profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
)))
168 if (edge
->caller
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
169 || edge
->callee
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
)
171 if (edge
->caller
->frequency
> NODE_FREQUENCY_UNLIKELY_EXECUTED
172 && edge
->callee
->frequency
<= NODE_FREQUENCY_EXECUTED_ONCE
)
176 if (edge
->caller
->frequency
== NODE_FREQUENCY_HOT
)
178 if (edge
->caller
->frequency
== NODE_FREQUENCY_EXECUTED_ONCE
179 && edge
->frequency
< CGRAPH_FREQ_BASE
* 3 / 2)
181 if (flag_guess_branch_prob
182 && edge
->frequency
<= (CGRAPH_FREQ_BASE
183 / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
)))
188 /* Return true in case BB can be CPU intensive and should be optimized
189 for maximal performance. */
192 maybe_hot_edge_p (edge e
)
194 if (profile_status
== PROFILE_READ
)
195 return maybe_hot_count_p (e
->count
);
196 return maybe_hot_frequency_p (EDGE_FREQUENCY (e
));
200 /* Return true in case BB is probably never executed. */
203 probably_never_executed_bb_p (const_basic_block bb
)
205 if (profile_info
&& flag_branch_probabilities
)
206 return ((bb
->count
+ profile_info
->runs
/ 2) / profile_info
->runs
) == 0;
207 if ((!profile_info
|| !flag_branch_probabilities
)
208 && (cgraph_get_node (current_function_decl
)->frequency
209 == NODE_FREQUENCY_UNLIKELY_EXECUTED
))
214 /* Return true if NODE should be optimized for size. */
217 cgraph_optimize_for_size_p (struct cgraph_node
*node
)
221 if (node
&& (node
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
))
227 /* Return true when current function should always be optimized for size. */
230 optimize_function_for_size_p (struct function
*fun
)
234 if (!fun
|| !fun
->decl
)
236 return cgraph_optimize_for_size_p (cgraph_get_node (fun
->decl
));
239 /* Return true when current function should always be optimized for speed. */
242 optimize_function_for_speed_p (struct function
*fun
)
244 return !optimize_function_for_size_p (fun
);
247 /* Return TRUE when BB should be optimized for size. */
250 optimize_bb_for_size_p (const_basic_block bb
)
252 return optimize_function_for_size_p (cfun
) || !maybe_hot_bb_p (bb
);
255 /* Return TRUE when BB should be optimized for speed. */
258 optimize_bb_for_speed_p (const_basic_block bb
)
260 return !optimize_bb_for_size_p (bb
);
263 /* Return TRUE when BB should be optimized for size. */
266 optimize_edge_for_size_p (edge e
)
268 return optimize_function_for_size_p (cfun
) || !maybe_hot_edge_p (e
);
271 /* Return TRUE when BB should be optimized for speed. */
274 optimize_edge_for_speed_p (edge e
)
276 return !optimize_edge_for_size_p (e
);
279 /* Return TRUE when BB should be optimized for size. */
282 optimize_insn_for_size_p (void)
284 return optimize_function_for_size_p (cfun
) || !crtl
->maybe_hot_insn_p
;
287 /* Return TRUE when BB should be optimized for speed. */
290 optimize_insn_for_speed_p (void)
292 return !optimize_insn_for_size_p ();
295 /* Return TRUE when LOOP should be optimized for size. */
298 optimize_loop_for_size_p (struct loop
*loop
)
300 return optimize_bb_for_size_p (loop
->header
);
303 /* Return TRUE when LOOP should be optimized for speed. */
306 optimize_loop_for_speed_p (struct loop
*loop
)
308 return optimize_bb_for_speed_p (loop
->header
);
311 /* Return TRUE when LOOP nest should be optimized for speed. */
314 optimize_loop_nest_for_speed_p (struct loop
*loop
)
316 struct loop
*l
= loop
;
317 if (optimize_loop_for_speed_p (loop
))
320 while (l
&& l
!= loop
)
322 if (optimize_loop_for_speed_p (l
))
330 while (l
!= loop
&& !l
->next
)
339 /* Return TRUE when LOOP nest should be optimized for size. */
342 optimize_loop_nest_for_size_p (struct loop
*loop
)
344 return !optimize_loop_nest_for_speed_p (loop
);
347 /* Return true when edge E is likely to be well predictable by branch
351 predictable_edge_p (edge e
)
353 if (profile_status
== PROFILE_ABSENT
)
356 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME
) * REG_BR_PROB_BASE
/ 100)
357 || (REG_BR_PROB_BASE
- e
->probability
358 <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME
) * REG_BR_PROB_BASE
/ 100))
364 /* Set RTL expansion for BB profile. */
367 rtl_profile_for_bb (basic_block bb
)
369 crtl
->maybe_hot_insn_p
= maybe_hot_bb_p (bb
);
372 /* Set RTL expansion for edge profile. */
375 rtl_profile_for_edge (edge e
)
377 crtl
->maybe_hot_insn_p
= maybe_hot_edge_p (e
);
380 /* Set RTL expansion to default mode (i.e. when profile info is not known). */
382 default_rtl_profile (void)
384 crtl
->maybe_hot_insn_p
= true;
387 /* Return true if the one of outgoing edges is already predicted by
391 rtl_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
394 if (!INSN_P (BB_END (bb
)))
396 for (note
= REG_NOTES (BB_END (bb
)); note
; note
= XEXP (note
, 1))
397 if (REG_NOTE_KIND (note
) == REG_BR_PRED
398 && INTVAL (XEXP (XEXP (note
, 0), 0)) == (int)predictor
)
403 /* This map contains for a basic block the list of predictions for the
406 static struct pointer_map_t
*bb_predictions
;
408 /* Structure representing predictions in tree level. */
410 struct edge_prediction
{
411 struct edge_prediction
*ep_next
;
413 enum br_predictor ep_predictor
;
417 /* Return true if the one of outgoing edges is already predicted by
421 gimple_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
423 struct edge_prediction
*i
;
424 void **preds
= pointer_map_contains (bb_predictions
, bb
);
429 for (i
= (struct edge_prediction
*) *preds
; i
; i
= i
->ep_next
)
430 if (i
->ep_predictor
== predictor
)
435 /* Return true when the probability of edge is reliable.
437 The profile guessing code is good at predicting branch outcome (ie.
438 taken/not taken), that is predicted right slightly over 75% of time.
439 It is however notoriously poor on predicting the probability itself.
440 In general the profile appear a lot flatter (with probabilities closer
441 to 50%) than the reality so it is bad idea to use it to drive optimization
442 such as those disabling dynamic branch prediction for well predictable
445 There are two exceptions - edges leading to noreturn edges and edges
446 predicted by number of iterations heuristics are predicted well. This macro
447 should be able to distinguish those, but at the moment it simply check for
448 noreturn heuristic that is only one giving probability over 99% or bellow
449 1%. In future we might want to propagate reliability information across the
450 CFG if we find this information useful on multiple places. */
452 probability_reliable_p (int prob
)
454 return (profile_status
== PROFILE_READ
455 || (profile_status
== PROFILE_GUESSED
456 && (prob
<= HITRATE (1) || prob
>= HITRATE (99))));
459 /* Same predicate as above, working on edges. */
461 edge_probability_reliable_p (const_edge e
)
463 return probability_reliable_p (e
->probability
);
466 /* Same predicate as edge_probability_reliable_p, working on notes. */
468 br_prob_note_reliable_p (const_rtx note
)
470 gcc_assert (REG_NOTE_KIND (note
) == REG_BR_PROB
);
471 return probability_reliable_p (INTVAL (XEXP (note
, 0)));
475 predict_insn (rtx insn
, enum br_predictor predictor
, int probability
)
477 gcc_assert (any_condjump_p (insn
));
478 if (!flag_guess_branch_prob
)
481 add_reg_note (insn
, REG_BR_PRED
,
482 gen_rtx_CONCAT (VOIDmode
,
483 GEN_INT ((int) predictor
),
484 GEN_INT ((int) probability
)));
487 /* Predict insn by given predictor. */
490 predict_insn_def (rtx insn
, enum br_predictor predictor
,
491 enum prediction taken
)
493 int probability
= predictor_info
[(int) predictor
].hitrate
;
496 probability
= REG_BR_PROB_BASE
- probability
;
498 predict_insn (insn
, predictor
, probability
);
501 /* Predict edge E with given probability if possible. */
504 rtl_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
507 last_insn
= BB_END (e
->src
);
509 /* We can store the branch prediction information only about
510 conditional jumps. */
511 if (!any_condjump_p (last_insn
))
514 /* We always store probability of branching. */
515 if (e
->flags
& EDGE_FALLTHRU
)
516 probability
= REG_BR_PROB_BASE
- probability
;
518 predict_insn (last_insn
, predictor
, probability
);
521 /* Predict edge E with the given PROBABILITY. */
523 gimple_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
525 gcc_assert (profile_status
!= PROFILE_GUESSED
);
526 if ((e
->src
!= ENTRY_BLOCK_PTR
&& EDGE_COUNT (e
->src
->succs
) > 1)
527 && flag_guess_branch_prob
&& optimize
)
529 struct edge_prediction
*i
= XNEW (struct edge_prediction
);
530 void **preds
= pointer_map_insert (bb_predictions
, e
->src
);
532 i
->ep_next
= (struct edge_prediction
*) *preds
;
534 i
->ep_probability
= probability
;
535 i
->ep_predictor
= predictor
;
540 /* Remove all predictions on given basic block that are attached
543 remove_predictions_associated_with_edge (edge e
)
550 preds
= pointer_map_contains (bb_predictions
, e
->src
);
554 struct edge_prediction
**prediction
= (struct edge_prediction
**) preds
;
555 struct edge_prediction
*next
;
559 if ((*prediction
)->ep_edge
== e
)
561 next
= (*prediction
)->ep_next
;
566 prediction
= &((*prediction
)->ep_next
);
571 /* Clears the list of predictions stored for BB. */
574 clear_bb_predictions (basic_block bb
)
576 void **preds
= pointer_map_contains (bb_predictions
, bb
);
577 struct edge_prediction
*pred
, *next
;
582 for (pred
= (struct edge_prediction
*) *preds
; pred
; pred
= next
)
584 next
= pred
->ep_next
;
590 /* Return true when we can store prediction on insn INSN.
591 At the moment we represent predictions only on conditional
592 jumps, not at computed jump or other complicated cases. */
594 can_predict_insn_p (const_rtx insn
)
596 return (JUMP_P (insn
)
597 && any_condjump_p (insn
)
598 && EDGE_COUNT (BLOCK_FOR_INSN (insn
)->succs
) >= 2);
601 /* Predict edge E by given predictor if possible. */
604 predict_edge_def (edge e
, enum br_predictor predictor
,
605 enum prediction taken
)
607 int probability
= predictor_info
[(int) predictor
].hitrate
;
610 probability
= REG_BR_PROB_BASE
- probability
;
612 predict_edge (e
, predictor
, probability
);
615 /* Invert all branch predictions or probability notes in the INSN. This needs
616 to be done each time we invert the condition used by the jump. */
619 invert_br_probabilities (rtx insn
)
623 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
624 if (REG_NOTE_KIND (note
) == REG_BR_PROB
)
625 XEXP (note
, 0) = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (note
, 0)));
626 else if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
627 XEXP (XEXP (note
, 0), 1)
628 = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (XEXP (note
, 0), 1)));
631 /* Dump information about the branch prediction to the output file. */
634 dump_prediction (FILE *file
, enum br_predictor predictor
, int probability
,
635 basic_block bb
, int used
)
643 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
644 if (! (e
->flags
& EDGE_FALLTHRU
))
647 fprintf (file
, " %s heuristics%s: %.1f%%",
648 predictor_info
[predictor
].name
,
649 used
? "" : " (ignored)", probability
* 100.0 / REG_BR_PROB_BASE
);
653 fprintf (file
, " exec ");
654 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, bb
->count
);
657 fprintf (file
, " hit ");
658 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, e
->count
);
659 fprintf (file
, " (%.1f%%)", e
->count
* 100.0 / bb
->count
);
663 fprintf (file
, "\n");
666 /* We can not predict the probabilities of outgoing edges of bb. Set them
667 evenly and hope for the best. */
669 set_even_probabilities (basic_block bb
)
675 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
676 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
678 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
679 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
680 e
->probability
= (REG_BR_PROB_BASE
+ nedges
/ 2) / nedges
;
685 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
686 note if not already present. Remove now useless REG_BR_PRED notes. */
689 combine_predictions_for_insn (rtx insn
, basic_block bb
)
694 int best_probability
= PROB_EVEN
;
695 enum br_predictor best_predictor
= END_PREDICTORS
;
696 int combined_probability
= REG_BR_PROB_BASE
/ 2;
698 bool first_match
= false;
701 if (!can_predict_insn_p (insn
))
703 set_even_probabilities (bb
);
707 prob_note
= find_reg_note (insn
, REG_BR_PROB
, 0);
708 pnote
= ®_NOTES (insn
);
710 fprintf (dump_file
, "Predictions for insn %i bb %i\n", INSN_UID (insn
),
713 /* We implement "first match" heuristics and use probability guessed
714 by predictor with smallest index. */
715 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
716 if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
718 enum br_predictor predictor
= ((enum br_predictor
)
719 INTVAL (XEXP (XEXP (note
, 0), 0)));
720 int probability
= INTVAL (XEXP (XEXP (note
, 0), 1));
723 if (best_predictor
> predictor
)
724 best_probability
= probability
, best_predictor
= predictor
;
726 d
= (combined_probability
* probability
727 + (REG_BR_PROB_BASE
- combined_probability
)
728 * (REG_BR_PROB_BASE
- probability
));
730 /* Use FP math to avoid overflows of 32bit integers. */
732 /* If one probability is 0% and one 100%, avoid division by zero. */
733 combined_probability
= REG_BR_PROB_BASE
/ 2;
735 combined_probability
= (((double) combined_probability
) * probability
736 * REG_BR_PROB_BASE
/ d
+ 0.5);
739 /* Decide which heuristic to use. In case we didn't match anything,
740 use no_prediction heuristic, in case we did match, use either
741 first match or Dempster-Shaffer theory depending on the flags. */
743 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
747 dump_prediction (dump_file
, PRED_NO_PREDICTION
,
748 combined_probability
, bb
, true);
751 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
,
753 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
,
758 combined_probability
= best_probability
;
759 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
763 if (REG_NOTE_KIND (*pnote
) == REG_BR_PRED
)
765 enum br_predictor predictor
= ((enum br_predictor
)
766 INTVAL (XEXP (XEXP (*pnote
, 0), 0)));
767 int probability
= INTVAL (XEXP (XEXP (*pnote
, 0), 1));
769 dump_prediction (dump_file
, predictor
, probability
, bb
,
770 !first_match
|| best_predictor
== predictor
);
771 *pnote
= XEXP (*pnote
, 1);
774 pnote
= &XEXP (*pnote
, 1);
779 add_reg_note (insn
, REG_BR_PROB
, GEN_INT (combined_probability
));
781 /* Save the prediction into CFG in case we are seeing non-degenerated
783 if (!single_succ_p (bb
))
785 BRANCH_EDGE (bb
)->probability
= combined_probability
;
786 FALLTHRU_EDGE (bb
)->probability
787 = REG_BR_PROB_BASE
- combined_probability
;
790 else if (!single_succ_p (bb
))
792 int prob
= INTVAL (XEXP (prob_note
, 0));
794 BRANCH_EDGE (bb
)->probability
= prob
;
795 FALLTHRU_EDGE (bb
)->probability
= REG_BR_PROB_BASE
- prob
;
798 single_succ_edge (bb
)->probability
= REG_BR_PROB_BASE
;
801 /* Combine predictions into single probability and store them into CFG.
802 Remove now useless prediction entries. */
805 combine_predictions_for_bb (basic_block bb
)
807 int best_probability
= PROB_EVEN
;
808 enum br_predictor best_predictor
= END_PREDICTORS
;
809 int combined_probability
= REG_BR_PROB_BASE
/ 2;
811 bool first_match
= false;
813 struct edge_prediction
*pred
;
815 edge e
, first
= NULL
, second
= NULL
;
819 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
820 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
823 if (first
&& !second
)
829 /* When there is no successor or only one choice, prediction is easy.
831 We are lazy for now and predict only basic blocks with two outgoing
832 edges. It is possible to predict generic case too, but we have to
833 ignore first match heuristics and do more involved combining. Implement
838 set_even_probabilities (bb
);
839 clear_bb_predictions (bb
);
841 fprintf (dump_file
, "%i edges in bb %i predicted to even probabilities\n",
847 fprintf (dump_file
, "Predictions for bb %i\n", bb
->index
);
849 preds
= pointer_map_contains (bb_predictions
, bb
);
852 /* We implement "first match" heuristics and use probability guessed
853 by predictor with smallest index. */
854 for (pred
= (struct edge_prediction
*) *preds
; pred
; pred
= pred
->ep_next
)
856 enum br_predictor predictor
= pred
->ep_predictor
;
857 int probability
= pred
->ep_probability
;
859 if (pred
->ep_edge
!= first
)
860 probability
= REG_BR_PROB_BASE
- probability
;
863 /* First match heuristics would be widly confused if we predicted
865 if (best_predictor
> predictor
)
867 struct edge_prediction
*pred2
;
868 int prob
= probability
;
870 for (pred2
= (struct edge_prediction
*) *preds
; pred2
; pred2
= pred2
->ep_next
)
871 if (pred2
!= pred
&& pred2
->ep_predictor
== pred
->ep_predictor
)
873 int probability2
= pred
->ep_probability
;
875 if (pred2
->ep_edge
!= first
)
876 probability2
= REG_BR_PROB_BASE
- probability2
;
878 if ((probability
< REG_BR_PROB_BASE
/ 2) !=
879 (probability2
< REG_BR_PROB_BASE
/ 2))
882 /* If the same predictor later gave better result, go for it! */
883 if ((probability
>= REG_BR_PROB_BASE
/ 2 && (probability2
> probability
))
884 || (probability
<= REG_BR_PROB_BASE
/ 2 && (probability2
< probability
)))
888 best_probability
= prob
, best_predictor
= predictor
;
891 d
= (combined_probability
* probability
892 + (REG_BR_PROB_BASE
- combined_probability
)
893 * (REG_BR_PROB_BASE
- probability
));
895 /* Use FP math to avoid overflows of 32bit integers. */
897 /* If one probability is 0% and one 100%, avoid division by zero. */
898 combined_probability
= REG_BR_PROB_BASE
/ 2;
900 combined_probability
= (((double) combined_probability
)
902 * REG_BR_PROB_BASE
/ d
+ 0.5);
906 /* Decide which heuristic to use. In case we didn't match anything,
907 use no_prediction heuristic, in case we did match, use either
908 first match or Dempster-Shaffer theory depending on the flags. */
910 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
914 dump_prediction (dump_file
, PRED_NO_PREDICTION
, combined_probability
, bb
, true);
917 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
, bb
,
919 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
, bb
,
924 combined_probability
= best_probability
;
925 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
929 for (pred
= (struct edge_prediction
*) *preds
; pred
; pred
= pred
->ep_next
)
931 enum br_predictor predictor
= pred
->ep_predictor
;
932 int probability
= pred
->ep_probability
;
934 if (pred
->ep_edge
!= EDGE_SUCC (bb
, 0))
935 probability
= REG_BR_PROB_BASE
- probability
;
936 dump_prediction (dump_file
, predictor
, probability
, bb
,
937 !first_match
|| best_predictor
== predictor
);
940 clear_bb_predictions (bb
);
944 first
->probability
= combined_probability
;
945 second
->probability
= REG_BR_PROB_BASE
- combined_probability
;
949 /* Predict edge probabilities by exploiting loop structure. */
957 /* Try to predict out blocks in a loop that are not part of a
959 FOR_EACH_LOOP (li
, loop
, 0)
961 basic_block bb
, *bbs
;
963 VEC (edge
, heap
) *exits
;
964 struct tree_niter_desc niter_desc
;
967 exits
= get_loop_exit_edges (loop
);
968 n_exits
= VEC_length (edge
, exits
);
970 FOR_EACH_VEC_ELT (edge
, exits
, j
, ex
)
973 HOST_WIDE_INT nitercst
;
974 int max
= PARAM_VALUE (PARAM_MAX_PREDICTED_ITERATIONS
);
976 enum br_predictor predictor
;
978 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false))
979 niter
= niter_desc
.niter
;
980 if (!niter
|| TREE_CODE (niter_desc
.niter
) != INTEGER_CST
)
981 niter
= loop_niter_by_eval (loop
, ex
);
983 if (TREE_CODE (niter
) == INTEGER_CST
)
985 if (host_integerp (niter
, 1)
987 && compare_tree_int (niter
, max
- 1) == -1)
988 nitercst
= tree_low_cst (niter
, 1) + 1;
991 predictor
= PRED_LOOP_ITERATIONS
;
993 /* If we have just one exit and we can derive some information about
994 the number of iterations of the loop from the statements inside
995 the loop, use it to predict this exit. */
996 else if (n_exits
== 1)
998 nitercst
= max_stmt_executions_int (loop
, false);
1004 predictor
= PRED_LOOP_ITERATIONS_GUESSED
;
1009 /* If the prediction for number of iterations is zero, do not
1010 predict the exit edges. */
1014 probability
= ((REG_BR_PROB_BASE
+ nitercst
/ 2) / nitercst
);
1015 predict_edge (ex
, predictor
, probability
);
1017 VEC_free (edge
, heap
, exits
);
1019 bbs
= get_loop_body (loop
);
1021 for (j
= 0; j
< loop
->num_nodes
; j
++)
1023 int header_found
= 0;
1029 /* Bypass loop heuristics on continue statement. These
1030 statements construct loops via "non-loop" constructs
1031 in the source language and are better to be handled
1033 if (predicted_by_p (bb
, PRED_CONTINUE
))
1036 /* Loop branch heuristics - predict an edge back to a
1037 loop's head as taken. */
1038 if (bb
== loop
->latch
)
1040 e
= find_edge (loop
->latch
, loop
->header
);
1044 predict_edge_def (e
, PRED_LOOP_BRANCH
, TAKEN
);
1048 /* Loop exit heuristics - predict an edge exiting the loop if the
1049 conditional has no loop header successors as not taken. */
1051 /* If we already used more reliable loop exit predictors, do not
1052 bother with PRED_LOOP_EXIT. */
1053 && !predicted_by_p (bb
, PRED_LOOP_ITERATIONS_GUESSED
)
1054 && !predicted_by_p (bb
, PRED_LOOP_ITERATIONS
))
1056 /* For loop with many exits we don't want to predict all exits
1057 with the pretty large probability, because if all exits are
1058 considered in row, the loop would be predicted to iterate
1059 almost never. The code to divide probability by number of
1060 exits is very rough. It should compute the number of exits
1061 taken in each patch through function (not the overall number
1062 of exits that might be a lot higher for loops with wide switch
1063 statements in them) and compute n-th square root.
1065 We limit the minimal probability by 2% to avoid
1066 EDGE_PROBABILITY_RELIABLE from trusting the branch prediction
1067 as this was causing regression in perl benchmark containing such
1070 int probability
= ((REG_BR_PROB_BASE
1071 - predictor_info
[(int) PRED_LOOP_EXIT
].hitrate
)
1073 if (probability
< HITRATE (2))
1074 probability
= HITRATE (2);
1075 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1076 if (e
->dest
->index
< NUM_FIXED_BLOCKS
1077 || !flow_bb_inside_loop_p (loop
, e
->dest
))
1078 predict_edge (e
, PRED_LOOP_EXIT
, probability
);
1082 /* Free basic blocks from get_loop_body. */
1087 /* Attempt to predict probabilities of BB outgoing edges using local
1090 bb_estimate_probability_locally (basic_block bb
)
1092 rtx last_insn
= BB_END (bb
);
1095 if (! can_predict_insn_p (last_insn
))
1097 cond
= get_condition (last_insn
, NULL
, false, false);
1101 /* Try "pointer heuristic."
1102 A comparison ptr == 0 is predicted as false.
1103 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1104 if (COMPARISON_P (cond
)
1105 && ((REG_P (XEXP (cond
, 0)) && REG_POINTER (XEXP (cond
, 0)))
1106 || (REG_P (XEXP (cond
, 1)) && REG_POINTER (XEXP (cond
, 1)))))
1108 if (GET_CODE (cond
) == EQ
)
1109 predict_insn_def (last_insn
, PRED_POINTER
, NOT_TAKEN
);
1110 else if (GET_CODE (cond
) == NE
)
1111 predict_insn_def (last_insn
, PRED_POINTER
, TAKEN
);
1115 /* Try "opcode heuristic."
1116 EQ tests are usually false and NE tests are usually true. Also,
1117 most quantities are positive, so we can make the appropriate guesses
1118 about signed comparisons against zero. */
1119 switch (GET_CODE (cond
))
1122 /* Unconditional branch. */
1123 predict_insn_def (last_insn
, PRED_UNCONDITIONAL
,
1124 cond
== const0_rtx
? NOT_TAKEN
: TAKEN
);
1129 /* Floating point comparisons appears to behave in a very
1130 unpredictable way because of special role of = tests in
1132 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
1134 /* Comparisons with 0 are often used for booleans and there is
1135 nothing useful to predict about them. */
1136 else if (XEXP (cond
, 1) == const0_rtx
1137 || XEXP (cond
, 0) == const0_rtx
)
1140 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, NOT_TAKEN
);
1145 /* Floating point comparisons appears to behave in a very
1146 unpredictable way because of special role of = tests in
1148 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
1150 /* Comparisons with 0 are often used for booleans and there is
1151 nothing useful to predict about them. */
1152 else if (XEXP (cond
, 1) == const0_rtx
1153 || XEXP (cond
, 0) == const0_rtx
)
1156 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, TAKEN
);
1160 predict_insn_def (last_insn
, PRED_FPOPCODE
, TAKEN
);
1164 predict_insn_def (last_insn
, PRED_FPOPCODE
, NOT_TAKEN
);
1169 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
1170 || XEXP (cond
, 1) == constm1_rtx
)
1171 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, NOT_TAKEN
);
1176 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
1177 || XEXP (cond
, 1) == constm1_rtx
)
1178 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, TAKEN
);
1186 /* Set edge->probability for each successor edge of BB. */
1188 guess_outgoing_edge_probabilities (basic_block bb
)
1190 bb_estimate_probability_locally (bb
);
1191 combine_predictions_for_insn (BB_END (bb
), bb
);
1194 static tree
expr_expected_value (tree
, bitmap
);
1196 /* Helper function for expr_expected_value. */
1199 expr_expected_value_1 (tree type
, tree op0
, enum tree_code code
,
1200 tree op1
, bitmap visited
)
1204 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
1206 if (TREE_CONSTANT (op0
))
1209 if (code
!= SSA_NAME
)
1212 def
= SSA_NAME_DEF_STMT (op0
);
1214 /* If we were already here, break the infinite cycle. */
1215 if (!bitmap_set_bit (visited
, SSA_NAME_VERSION (op0
)))
1218 if (gimple_code (def
) == GIMPLE_PHI
)
1220 /* All the arguments of the PHI node must have the same constant
1222 int i
, n
= gimple_phi_num_args (def
);
1223 tree val
= NULL
, new_val
;
1225 for (i
= 0; i
< n
; i
++)
1227 tree arg
= PHI_ARG_DEF (def
, i
);
1229 /* If this PHI has itself as an argument, we cannot
1230 determine the string length of this argument. However,
1231 if we can find an expected constant value for the other
1232 PHI args then we can still be sure that this is
1233 likely a constant. So be optimistic and just
1234 continue with the next argument. */
1235 if (arg
== PHI_RESULT (def
))
1238 new_val
= expr_expected_value (arg
, visited
);
1243 else if (!operand_equal_p (val
, new_val
, false))
1248 if (is_gimple_assign (def
))
1250 if (gimple_assign_lhs (def
) != op0
)
1253 return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def
)),
1254 gimple_assign_rhs1 (def
),
1255 gimple_assign_rhs_code (def
),
1256 gimple_assign_rhs2 (def
),
1260 if (is_gimple_call (def
))
1262 tree decl
= gimple_call_fndecl (def
);
1265 if (DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_NORMAL
)
1266 switch (DECL_FUNCTION_CODE (decl
))
1268 case BUILT_IN_EXPECT
:
1271 if (gimple_call_num_args (def
) != 2)
1273 val
= gimple_call_arg (def
, 0);
1274 if (TREE_CONSTANT (val
))
1276 return gimple_call_arg (def
, 1);
1279 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N
:
1280 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1
:
1281 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2
:
1282 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4
:
1283 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8
:
1284 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16
:
1285 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE
:
1286 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N
:
1287 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1
:
1288 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2
:
1289 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4
:
1290 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8
:
1291 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16
:
1292 /* Assume that any given atomic operation has low contention,
1293 and thus the compare-and-swap operation succeeds. */
1294 return boolean_true_node
;
1301 if (get_gimple_rhs_class (code
) == GIMPLE_BINARY_RHS
)
1304 op0
= expr_expected_value (op0
, visited
);
1307 op1
= expr_expected_value (op1
, visited
);
1310 res
= fold_build2 (code
, type
, op0
, op1
);
1311 if (TREE_CONSTANT (res
))
1315 if (get_gimple_rhs_class (code
) == GIMPLE_UNARY_RHS
)
1318 op0
= expr_expected_value (op0
, visited
);
1321 res
= fold_build1 (code
, type
, op0
);
1322 if (TREE_CONSTANT (res
))
1329 /* Return constant EXPR will likely have at execution time, NULL if unknown.
1330 The function is used by builtin_expect branch predictor so the evidence
1331 must come from this construct and additional possible constant folding.
1333 We may want to implement more involved value guess (such as value range
1334 propagation based prediction), but such tricks shall go to new
1338 expr_expected_value (tree expr
, bitmap visited
)
1340 enum tree_code code
;
1343 if (TREE_CONSTANT (expr
))
1346 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
1347 return expr_expected_value_1 (TREE_TYPE (expr
),
1348 op0
, code
, op1
, visited
);
1352 /* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements
1353 we no longer need. */
1355 strip_predict_hints (void)
1363 gimple_stmt_iterator bi
;
1364 for (bi
= gsi_start_bb (bb
); !gsi_end_p (bi
);)
1366 gimple stmt
= gsi_stmt (bi
);
1368 if (gimple_code (stmt
) == GIMPLE_PREDICT
)
1370 gsi_remove (&bi
, true);
1373 else if (gimple_code (stmt
) == GIMPLE_CALL
)
1375 tree fndecl
= gimple_call_fndecl (stmt
);
1378 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
1379 && DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_EXPECT
1380 && gimple_call_num_args (stmt
) == 2)
1382 var
= gimple_call_lhs (stmt
);
1386 = gimple_build_assign (var
, gimple_call_arg (stmt
, 0));
1387 gsi_replace (&bi
, ass_stmt
, true);
1391 gsi_remove (&bi
, true);
1402 /* Predict using opcode of the last statement in basic block. */
1404 tree_predict_by_opcode (basic_block bb
)
1406 gimple stmt
= last_stmt (bb
);
1415 if (!stmt
|| gimple_code (stmt
) != GIMPLE_COND
)
1417 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
1418 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1420 op0
= gimple_cond_lhs (stmt
);
1421 op1
= gimple_cond_rhs (stmt
);
1422 cmp
= gimple_cond_code (stmt
);
1423 type
= TREE_TYPE (op0
);
1424 visited
= BITMAP_ALLOC (NULL
);
1425 val
= expr_expected_value_1 (boolean_type_node
, op0
, cmp
, op1
, visited
);
1426 BITMAP_FREE (visited
);
1429 if (integer_zerop (val
))
1430 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, NOT_TAKEN
);
1432 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, TAKEN
);
1435 /* Try "pointer heuristic."
1436 A comparison ptr == 0 is predicted as false.
1437 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1438 if (POINTER_TYPE_P (type
))
1441 predict_edge_def (then_edge
, PRED_TREE_POINTER
, NOT_TAKEN
);
1442 else if (cmp
== NE_EXPR
)
1443 predict_edge_def (then_edge
, PRED_TREE_POINTER
, TAKEN
);
1447 /* Try "opcode heuristic."
1448 EQ tests are usually false and NE tests are usually true. Also,
1449 most quantities are positive, so we can make the appropriate guesses
1450 about signed comparisons against zero. */
1455 /* Floating point comparisons appears to behave in a very
1456 unpredictable way because of special role of = tests in
1458 if (FLOAT_TYPE_P (type
))
1460 /* Comparisons with 0 are often used for booleans and there is
1461 nothing useful to predict about them. */
1462 else if (integer_zerop (op0
) || integer_zerop (op1
))
1465 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, NOT_TAKEN
);
1470 /* Floating point comparisons appears to behave in a very
1471 unpredictable way because of special role of = tests in
1473 if (FLOAT_TYPE_P (type
))
1475 /* Comparisons with 0 are often used for booleans and there is
1476 nothing useful to predict about them. */
1477 else if (integer_zerop (op0
)
1478 || integer_zerop (op1
))
1481 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, TAKEN
);
1485 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, TAKEN
);
1488 case UNORDERED_EXPR
:
1489 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, NOT_TAKEN
);
1494 if (integer_zerop (op1
)
1495 || integer_onep (op1
)
1496 || integer_all_onesp (op1
)
1499 || real_minus_onep (op1
))
1500 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, NOT_TAKEN
);
1505 if (integer_zerop (op1
)
1506 || integer_onep (op1
)
1507 || integer_all_onesp (op1
)
1510 || real_minus_onep (op1
))
1511 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, TAKEN
);
1519 /* Try to guess whether the value of return means error code. */
1521 static enum br_predictor
1522 return_prediction (tree val
, enum prediction
*prediction
)
1526 return PRED_NO_PREDICTION
;
1527 /* Different heuristics for pointers and scalars. */
1528 if (POINTER_TYPE_P (TREE_TYPE (val
)))
1530 /* NULL is usually not returned. */
1531 if (integer_zerop (val
))
1533 *prediction
= NOT_TAKEN
;
1534 return PRED_NULL_RETURN
;
1537 else if (INTEGRAL_TYPE_P (TREE_TYPE (val
)))
1539 /* Negative return values are often used to indicate
1541 if (TREE_CODE (val
) == INTEGER_CST
1542 && tree_int_cst_sgn (val
) < 0)
1544 *prediction
= NOT_TAKEN
;
1545 return PRED_NEGATIVE_RETURN
;
1547 /* Constant return values seems to be commonly taken.
1548 Zero/one often represent booleans so exclude them from the
1550 if (TREE_CONSTANT (val
)
1551 && (!integer_zerop (val
) && !integer_onep (val
)))
1553 *prediction
= TAKEN
;
1554 return PRED_CONST_RETURN
;
1557 return PRED_NO_PREDICTION
;
1560 /* Find the basic block with return expression and look up for possible
1561 return value trying to apply RETURN_PREDICTION heuristics. */
1563 apply_return_prediction (void)
1565 gimple return_stmt
= NULL
;
1569 int phi_num_args
, i
;
1570 enum br_predictor pred
;
1571 enum prediction direction
;
1574 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
1576 return_stmt
= last_stmt (e
->src
);
1578 && gimple_code (return_stmt
) == GIMPLE_RETURN
)
1583 return_val
= gimple_return_retval (return_stmt
);
1586 if (TREE_CODE (return_val
) != SSA_NAME
1587 || !SSA_NAME_DEF_STMT (return_val
)
1588 || gimple_code (SSA_NAME_DEF_STMT (return_val
)) != GIMPLE_PHI
)
1590 phi
= SSA_NAME_DEF_STMT (return_val
);
1591 phi_num_args
= gimple_phi_num_args (phi
);
1592 pred
= return_prediction (PHI_ARG_DEF (phi
, 0), &direction
);
1594 /* Avoid the degenerate case where all return values form the function
1595 belongs to same category (ie they are all positive constants)
1596 so we can hardly say something about them. */
1597 for (i
= 1; i
< phi_num_args
; i
++)
1598 if (pred
!= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
))
1600 if (i
!= phi_num_args
)
1601 for (i
= 0; i
< phi_num_args
; i
++)
1603 pred
= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
);
1604 if (pred
!= PRED_NO_PREDICTION
)
1605 predict_paths_leading_to_edge (gimple_phi_arg_edge (phi
, i
), pred
,
1610 /* Look for basic block that contains unlikely to happen events
1611 (such as noreturn calls) and mark all paths leading to execution
1612 of this basic blocks as unlikely. */
1615 tree_bb_level_predictions (void)
1618 bool has_return_edges
= false;
1622 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
1623 if (!(e
->flags
& (EDGE_ABNORMAL
| EDGE_FAKE
| EDGE_EH
)))
1625 has_return_edges
= true;
1629 apply_return_prediction ();
1633 gimple_stmt_iterator gsi
;
1635 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1637 gimple stmt
= gsi_stmt (gsi
);
1640 if (is_gimple_call (stmt
))
1642 if ((gimple_call_flags (stmt
) & ECF_NORETURN
)
1643 && has_return_edges
)
1644 predict_paths_leading_to (bb
, PRED_NORETURN
,
1646 decl
= gimple_call_fndecl (stmt
);
1648 && lookup_attribute ("cold",
1649 DECL_ATTRIBUTES (decl
)))
1650 predict_paths_leading_to (bb
, PRED_COLD_FUNCTION
,
1653 else if (gimple_code (stmt
) == GIMPLE_PREDICT
)
1655 predict_paths_leading_to (bb
, gimple_predict_predictor (stmt
),
1656 gimple_predict_outcome (stmt
));
1657 /* Keep GIMPLE_PREDICT around so early inlining will propagate
1658 hints to callers. */
1664 #ifdef ENABLE_CHECKING
1666 /* Callback for pointer_map_traverse, asserts that the pointer map is
1670 assert_is_empty (const void *key ATTRIBUTE_UNUSED
, void **value
,
1671 void *data ATTRIBUTE_UNUSED
)
1673 gcc_assert (!*value
);
1678 /* Predict branch probabilities and estimate profile for basic block BB. */
1681 tree_estimate_probability_bb (basic_block bb
)
1687 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1689 /* Predict early returns to be probable, as we've already taken
1690 care for error returns and other cases are often used for
1691 fast paths through function.
1693 Since we've already removed the return statements, we are
1694 looking for CFG like:
1704 if (e
->dest
!= bb
->next_bb
1705 && e
->dest
!= EXIT_BLOCK_PTR
1706 && single_succ_p (e
->dest
)
1707 && single_succ_edge (e
->dest
)->dest
== EXIT_BLOCK_PTR
1708 && (last
= last_stmt (e
->dest
)) != NULL
1709 && gimple_code (last
) == GIMPLE_RETURN
)
1714 if (single_succ_p (bb
))
1716 FOR_EACH_EDGE (e1
, ei1
, bb
->preds
)
1717 if (!predicted_by_p (e1
->src
, PRED_NULL_RETURN
)
1718 && !predicted_by_p (e1
->src
, PRED_CONST_RETURN
)
1719 && !predicted_by_p (e1
->src
, PRED_NEGATIVE_RETURN
))
1720 predict_edge_def (e1
, PRED_TREE_EARLY_RETURN
, NOT_TAKEN
);
1723 if (!predicted_by_p (e
->src
, PRED_NULL_RETURN
)
1724 && !predicted_by_p (e
->src
, PRED_CONST_RETURN
)
1725 && !predicted_by_p (e
->src
, PRED_NEGATIVE_RETURN
))
1726 predict_edge_def (e
, PRED_TREE_EARLY_RETURN
, NOT_TAKEN
);
1729 /* Look for block we are guarding (ie we dominate it,
1730 but it doesn't postdominate us). */
1731 if (e
->dest
!= EXIT_BLOCK_PTR
&& e
->dest
!= bb
1732 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
1733 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
1735 gimple_stmt_iterator bi
;
1737 /* The call heuristic claims that a guarded function call
1738 is improbable. This is because such calls are often used
1739 to signal exceptional situations such as printing error
1741 for (bi
= gsi_start_bb (e
->dest
); !gsi_end_p (bi
);
1744 gimple stmt
= gsi_stmt (bi
);
1745 if (is_gimple_call (stmt
)
1746 /* Constant and pure calls are hardly used to signalize
1747 something exceptional. */
1748 && gimple_has_side_effects (stmt
))
1750 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
1756 tree_predict_by_opcode (bb
);
1759 /* Predict branch probabilities and estimate profile of the tree CFG.
1760 This function can be called from the loop optimizers to recompute
1761 the profile information. */
1764 tree_estimate_probability (void)
1768 add_noreturn_fake_exit_edges ();
1769 connect_infinite_loops_to_exit ();
1770 /* We use loop_niter_by_eval, which requires that the loops have
1772 create_preheaders (CP_SIMPLE_PREHEADERS
);
1773 calculate_dominance_info (CDI_POST_DOMINATORS
);
1775 bb_predictions
= pointer_map_create ();
1776 tree_bb_level_predictions ();
1777 record_loop_exits ();
1779 if (number_of_loops () > 1)
1783 tree_estimate_probability_bb (bb
);
1786 combine_predictions_for_bb (bb
);
1788 #ifdef ENABLE_CHECKING
1789 pointer_map_traverse (bb_predictions
, assert_is_empty
, NULL
);
1791 pointer_map_destroy (bb_predictions
);
1792 bb_predictions
= NULL
;
1794 estimate_bb_frequencies ();
1795 free_dominance_info (CDI_POST_DOMINATORS
);
1796 remove_fake_exit_edges ();
1799 /* Predict branch probabilities and estimate profile of the tree CFG.
1800 This is the driver function for PASS_PROFILE. */
1803 tree_estimate_probability_driver (void)
1807 loop_optimizer_init (0);
1808 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1809 flow_loops_dump (dump_file
, NULL
, 0);
1811 mark_irreducible_loops ();
1813 nb_loops
= number_of_loops ();
1817 tree_estimate_probability ();
1822 loop_optimizer_finalize ();
1823 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1824 gimple_dump_cfg (dump_file
, dump_flags
);
1825 if (profile_status
== PROFILE_ABSENT
)
1826 profile_status
= PROFILE_GUESSED
;
1830 /* Predict edges to successors of CUR whose sources are not postdominated by
1831 BB by PRED and recurse to all postdominators. */
1834 predict_paths_for_bb (basic_block cur
, basic_block bb
,
1835 enum br_predictor pred
,
1836 enum prediction taken
,
1843 /* We are looking for all edges forming edge cut induced by
1844 set of all blocks postdominated by BB. */
1845 FOR_EACH_EDGE (e
, ei
, cur
->preds
)
1846 if (e
->src
->index
>= NUM_FIXED_BLOCKS
1847 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, bb
))
1853 /* Ignore fake edges and eh, we predict them as not taken anyway. */
1854 if (e
->flags
& (EDGE_EH
| EDGE_FAKE
))
1856 gcc_assert (bb
== cur
|| dominated_by_p (CDI_POST_DOMINATORS
, cur
, bb
));
1858 /* See if there is an edge from e->src that is not abnormal
1859 and does not lead to BB. */
1860 FOR_EACH_EDGE (e2
, ei2
, e
->src
->succs
)
1862 && !(e2
->flags
& (EDGE_EH
| EDGE_FAKE
))
1863 && !dominated_by_p (CDI_POST_DOMINATORS
, e2
->dest
, bb
))
1869 /* If there is non-abnormal path leaving e->src, predict edge
1870 using predictor. Otherwise we need to look for paths
1873 The second may lead to infinite loop in the case we are predicitng
1874 regions that are only reachable by abnormal edges. We simply
1875 prevent visiting given BB twice. */
1877 predict_edge_def (e
, pred
, taken
);
1878 else if (bitmap_set_bit (visited
, e
->src
->index
))
1879 predict_paths_for_bb (e
->src
, e
->src
, pred
, taken
, visited
);
1881 for (son
= first_dom_son (CDI_POST_DOMINATORS
, cur
);
1883 son
= next_dom_son (CDI_POST_DOMINATORS
, son
))
1884 predict_paths_for_bb (son
, bb
, pred
, taken
, visited
);
1887 /* Sets branch probabilities according to PREDiction and
1891 predict_paths_leading_to (basic_block bb
, enum br_predictor pred
,
1892 enum prediction taken
)
1894 bitmap visited
= BITMAP_ALLOC (NULL
);
1895 predict_paths_for_bb (bb
, bb
, pred
, taken
, visited
);
1896 BITMAP_FREE (visited
);
1899 /* Like predict_paths_leading_to but take edge instead of basic block. */
1902 predict_paths_leading_to_edge (edge e
, enum br_predictor pred
,
1903 enum prediction taken
)
1905 bool has_nonloop_edge
= false;
1909 basic_block bb
= e
->src
;
1910 FOR_EACH_EDGE (e2
, ei
, bb
->succs
)
1911 if (e2
->dest
!= e
->src
&& e2
->dest
!= e
->dest
1912 && !(e
->flags
& (EDGE_EH
| EDGE_FAKE
))
1913 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e2
->dest
))
1915 has_nonloop_edge
= true;
1918 if (!has_nonloop_edge
)
1920 bitmap visited
= BITMAP_ALLOC (NULL
);
1921 predict_paths_for_bb (bb
, bb
, pred
, taken
, visited
);
1922 BITMAP_FREE (visited
);
1925 predict_edge_def (e
, pred
, taken
);
1928 /* This is used to carry information about basic blocks. It is
1929 attached to the AUX field of the standard CFG block. */
1931 typedef struct block_info_def
1933 /* Estimated frequency of execution of basic_block. */
1936 /* To keep queue of basic blocks to process. */
1939 /* Number of predecessors we need to visit first. */
1943 /* Similar information for edges. */
1944 typedef struct edge_info_def
1946 /* In case edge is a loopback edge, the probability edge will be reached
1947 in case header is. Estimated number of iterations of the loop can be
1948 then computed as 1 / (1 - back_edge_prob). */
1949 sreal back_edge_prob
;
1950 /* True if the edge is a loopback edge in the natural loop. */
1951 unsigned int back_edge
:1;
1954 #define BLOCK_INFO(B) ((block_info) (B)->aux)
1955 #define EDGE_INFO(E) ((edge_info) (E)->aux)
1957 /* Helper function for estimate_bb_frequencies.
1958 Propagate the frequencies in blocks marked in
1959 TOVISIT, starting in HEAD. */
1962 propagate_freq (basic_block head
, bitmap tovisit
)
1971 /* For each basic block we need to visit count number of his predecessors
1972 we need to visit first. */
1973 EXECUTE_IF_SET_IN_BITMAP (tovisit
, 0, i
, bi
)
1978 bb
= BASIC_BLOCK (i
);
1980 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1982 bool visit
= bitmap_bit_p (tovisit
, e
->src
->index
);
1984 if (visit
&& !(e
->flags
& EDGE_DFS_BACK
))
1986 else if (visit
&& dump_file
&& !EDGE_INFO (e
)->back_edge
)
1988 "Irreducible region hit, ignoring edge to %i->%i\n",
1989 e
->src
->index
, bb
->index
);
1991 BLOCK_INFO (bb
)->npredecessors
= count
;
1992 /* When function never returns, we will never process exit block. */
1993 if (!count
&& bb
== EXIT_BLOCK_PTR
)
1994 bb
->count
= bb
->frequency
= 0;
1997 memcpy (&BLOCK_INFO (head
)->frequency
, &real_one
, sizeof (real_one
));
1999 for (bb
= head
; bb
; bb
= nextbb
)
2002 sreal cyclic_probability
, frequency
;
2004 memcpy (&cyclic_probability
, &real_zero
, sizeof (real_zero
));
2005 memcpy (&frequency
, &real_zero
, sizeof (real_zero
));
2007 nextbb
= BLOCK_INFO (bb
)->next
;
2008 BLOCK_INFO (bb
)->next
= NULL
;
2010 /* Compute frequency of basic block. */
2013 #ifdef ENABLE_CHECKING
2014 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2015 gcc_assert (!bitmap_bit_p (tovisit
, e
->src
->index
)
2016 || (e
->flags
& EDGE_DFS_BACK
));
2019 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2020 if (EDGE_INFO (e
)->back_edge
)
2022 sreal_add (&cyclic_probability
, &cyclic_probability
,
2023 &EDGE_INFO (e
)->back_edge_prob
);
2025 else if (!(e
->flags
& EDGE_DFS_BACK
))
2029 /* frequency += (e->probability
2030 * BLOCK_INFO (e->src)->frequency /
2031 REG_BR_PROB_BASE); */
2033 sreal_init (&tmp
, e
->probability
, 0);
2034 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (e
->src
)->frequency
);
2035 sreal_mul (&tmp
, &tmp
, &real_inv_br_prob_base
);
2036 sreal_add (&frequency
, &frequency
, &tmp
);
2039 if (sreal_compare (&cyclic_probability
, &real_zero
) == 0)
2041 memcpy (&BLOCK_INFO (bb
)->frequency
, &frequency
,
2042 sizeof (frequency
));
2046 if (sreal_compare (&cyclic_probability
, &real_almost_one
) > 0)
2048 memcpy (&cyclic_probability
, &real_almost_one
,
2049 sizeof (real_almost_one
));
2052 /* BLOCK_INFO (bb)->frequency = frequency
2053 / (1 - cyclic_probability) */
2055 sreal_sub (&cyclic_probability
, &real_one
, &cyclic_probability
);
2056 sreal_div (&BLOCK_INFO (bb
)->frequency
,
2057 &frequency
, &cyclic_probability
);
2061 bitmap_clear_bit (tovisit
, bb
->index
);
2063 e
= find_edge (bb
, head
);
2068 /* EDGE_INFO (e)->back_edge_prob
2069 = ((e->probability * BLOCK_INFO (bb)->frequency)
2070 / REG_BR_PROB_BASE); */
2072 sreal_init (&tmp
, e
->probability
, 0);
2073 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (bb
)->frequency
);
2074 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
2075 &tmp
, &real_inv_br_prob_base
);
2078 /* Propagate to successor blocks. */
2079 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2080 if (!(e
->flags
& EDGE_DFS_BACK
)
2081 && BLOCK_INFO (e
->dest
)->npredecessors
)
2083 BLOCK_INFO (e
->dest
)->npredecessors
--;
2084 if (!BLOCK_INFO (e
->dest
)->npredecessors
)
2089 BLOCK_INFO (last
)->next
= e
->dest
;
2097 /* Estimate probabilities of loopback edges in loops at same nest level. */
2100 estimate_loops_at_level (struct loop
*first_loop
)
2104 for (loop
= first_loop
; loop
; loop
= loop
->next
)
2109 bitmap tovisit
= BITMAP_ALLOC (NULL
);
2111 estimate_loops_at_level (loop
->inner
);
2113 /* Find current loop back edge and mark it. */
2114 e
= loop_latch_edge (loop
);
2115 EDGE_INFO (e
)->back_edge
= 1;
2117 bbs
= get_loop_body (loop
);
2118 for (i
= 0; i
< loop
->num_nodes
; i
++)
2119 bitmap_set_bit (tovisit
, bbs
[i
]->index
);
2121 propagate_freq (loop
->header
, tovisit
);
2122 BITMAP_FREE (tovisit
);
2126 /* Propagates frequencies through structure of loops. */
2129 estimate_loops (void)
2131 bitmap tovisit
= BITMAP_ALLOC (NULL
);
2134 /* Start by estimating the frequencies in the loops. */
2135 if (number_of_loops () > 1)
2136 estimate_loops_at_level (current_loops
->tree_root
->inner
);
2138 /* Now propagate the frequencies through all the blocks. */
2141 bitmap_set_bit (tovisit
, bb
->index
);
2143 propagate_freq (ENTRY_BLOCK_PTR
, tovisit
);
2144 BITMAP_FREE (tovisit
);
2147 /* Convert counts measured by profile driven feedback to frequencies.
2148 Return nonzero iff there was any nonzero execution count. */
2151 counts_to_freqs (void)
2153 gcov_type count_max
, true_count_max
= 0;
2156 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
2157 true_count_max
= MAX (bb
->count
, true_count_max
);
2159 count_max
= MAX (true_count_max
, 1);
2160 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
2161 bb
->frequency
= (bb
->count
* BB_FREQ_MAX
+ count_max
/ 2) / count_max
;
2163 return true_count_max
;
2166 /* Return true if function is likely to be expensive, so there is no point to
2167 optimize performance of prologue, epilogue or do inlining at the expense
2168 of code size growth. THRESHOLD is the limit of number of instructions
2169 function can execute at average to be still considered not expensive. */
2172 expensive_function_p (int threshold
)
2174 unsigned int sum
= 0;
2178 /* We can not compute accurately for large thresholds due to scaled
2180 gcc_assert (threshold
<= BB_FREQ_MAX
);
2182 /* Frequencies are out of range. This either means that function contains
2183 internal loop executing more than BB_FREQ_MAX times or profile feedback
2184 is available and function has not been executed at all. */
2185 if (ENTRY_BLOCK_PTR
->frequency
== 0)
2188 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
2189 limit
= ENTRY_BLOCK_PTR
->frequency
* threshold
;
2194 for (insn
= BB_HEAD (bb
); insn
!= NEXT_INSN (BB_END (bb
));
2195 insn
= NEXT_INSN (insn
))
2196 if (active_insn_p (insn
))
2198 sum
+= bb
->frequency
;
2207 /* Estimate basic blocks frequency by given branch probabilities. */
2210 estimate_bb_frequencies (void)
2215 if (profile_status
!= PROFILE_READ
|| !counts_to_freqs ())
2217 static int real_values_initialized
= 0;
2219 if (!real_values_initialized
)
2221 real_values_initialized
= 1;
2222 sreal_init (&real_zero
, 0, 0);
2223 sreal_init (&real_one
, 1, 0);
2224 sreal_init (&real_br_prob_base
, REG_BR_PROB_BASE
, 0);
2225 sreal_init (&real_bb_freq_max
, BB_FREQ_MAX
, 0);
2226 sreal_init (&real_one_half
, 1, -1);
2227 sreal_div (&real_inv_br_prob_base
, &real_one
, &real_br_prob_base
);
2228 sreal_sub (&real_almost_one
, &real_one
, &real_inv_br_prob_base
);
2231 mark_dfs_back_edges ();
2233 single_succ_edge (ENTRY_BLOCK_PTR
)->probability
= REG_BR_PROB_BASE
;
2235 /* Set up block info for each basic block. */
2236 alloc_aux_for_blocks (sizeof (struct block_info_def
));
2237 alloc_aux_for_edges (sizeof (struct edge_info_def
));
2238 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
2243 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2245 sreal_init (&EDGE_INFO (e
)->back_edge_prob
, e
->probability
, 0);
2246 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
2247 &EDGE_INFO (e
)->back_edge_prob
,
2248 &real_inv_br_prob_base
);
2252 /* First compute probabilities locally for each loop from innermost
2253 to outermost to examine probabilities for back edges. */
2256 memcpy (&freq_max
, &real_zero
, sizeof (real_zero
));
2258 if (sreal_compare (&freq_max
, &BLOCK_INFO (bb
)->frequency
) < 0)
2259 memcpy (&freq_max
, &BLOCK_INFO (bb
)->frequency
, sizeof (freq_max
));
2261 sreal_div (&freq_max
, &real_bb_freq_max
, &freq_max
);
2262 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
2266 sreal_mul (&tmp
, &BLOCK_INFO (bb
)->frequency
, &freq_max
);
2267 sreal_add (&tmp
, &tmp
, &real_one_half
);
2268 bb
->frequency
= sreal_to_int (&tmp
);
2271 free_aux_for_blocks ();
2272 free_aux_for_edges ();
2274 compute_function_frequency ();
2277 /* Decide whether function is hot, cold or unlikely executed. */
2279 compute_function_frequency (void)
2282 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
2283 if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
2284 || MAIN_NAME_P (DECL_NAME (current_function_decl
)))
2285 node
->only_called_at_startup
= true;
2286 if (DECL_STATIC_DESTRUCTOR (current_function_decl
))
2287 node
->only_called_at_exit
= true;
2289 if (!profile_info
|| !flag_branch_probabilities
)
2291 int flags
= flags_from_decl_or_type (current_function_decl
);
2292 if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
))
2294 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
2295 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl
))
2297 node
->frequency
= NODE_FREQUENCY_HOT
;
2298 else if (flags
& ECF_NORETURN
)
2299 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2300 else if (MAIN_NAME_P (DECL_NAME (current_function_decl
)))
2301 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2302 else if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
2303 || DECL_STATIC_DESTRUCTOR (current_function_decl
))
2304 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
2307 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
2310 if (maybe_hot_bb_p (bb
))
2312 node
->frequency
= NODE_FREQUENCY_HOT
;
2315 if (!probably_never_executed_bb_p (bb
))
2316 node
->frequency
= NODE_FREQUENCY_NORMAL
;
2321 gate_estimate_probability (void)
2323 return flag_guess_branch_prob
;
2326 /* Build PREDICT_EXPR. */
2328 build_predict_expr (enum br_predictor predictor
, enum prediction taken
)
2330 tree t
= build1 (PREDICT_EXPR
, void_type_node
,
2331 build_int_cst (integer_type_node
, predictor
));
2332 SET_PREDICT_EXPR_OUTCOME (t
, taken
);
2337 predictor_name (enum br_predictor predictor
)
2339 return predictor_info
[predictor
].name
;
2342 struct gimple_opt_pass pass_profile
=
2346 "profile_estimate", /* name */
2347 gate_estimate_probability
, /* gate */
2348 tree_estimate_probability_driver
, /* execute */
2351 0, /* static_pass_number */
2352 TV_BRANCH_PROB
, /* tv_id */
2353 PROP_cfg
, /* properties_required */
2354 0, /* properties_provided */
2355 0, /* properties_destroyed */
2356 0, /* todo_flags_start */
2357 TODO_ggc_collect
| TODO_verify_ssa
/* todo_flags_finish */
2361 struct gimple_opt_pass pass_strip_predict_hints
=
2365 "*strip_predict_hints", /* name */
2367 strip_predict_hints
, /* execute */
2370 0, /* static_pass_number */
2371 TV_BRANCH_PROB
, /* tv_id */
2372 PROP_cfg
, /* properties_required */
2373 0, /* properties_provided */
2374 0, /* properties_destroyed */
2375 0, /* todo_flags_start */
2376 TODO_ggc_collect
| TODO_verify_ssa
/* todo_flags_finish */
2380 /* Rebuild function frequencies. Passes are in general expected to
2381 maintain profile by hand, however in some cases this is not possible:
2382 for example when inlining several functions with loops freuqencies might run
2383 out of scale and thus needs to be recomputed. */
2386 rebuild_frequencies (void)
2388 timevar_push (TV_REBUILD_FREQUENCIES
);
2389 if (profile_status
== PROFILE_GUESSED
)
2391 loop_optimizer_init (0);
2392 add_noreturn_fake_exit_edges ();
2393 mark_irreducible_loops ();
2394 connect_infinite_loops_to_exit ();
2395 estimate_bb_frequencies ();
2396 remove_fake_exit_edges ();
2397 loop_optimizer_finalize ();
2399 else if (profile_status
== PROFILE_READ
)
2403 timevar_pop (TV_REBUILD_FREQUENCIES
);