1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000-2023 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 "tree-pass.h"
44 #include "diagnostic-core.h"
45 #include "gimple-predict.h"
46 #include "fold-const.h"
52 #include "gimple-iterator.h"
54 #include "tree-ssa-loop-niter.h"
55 #include "tree-ssa-loop.h"
56 #include "tree-scalar-evolution.h"
57 #include "ipa-utils.h"
58 #include "gimple-pretty-print.h"
61 #include "stringpool.h"
64 /* Enum with reasons why a predictor is ignored. */
70 REASON_SINGLE_EDGE_DUPLICATE
,
71 REASON_EDGE_PAIR_DUPLICATE
74 /* String messages for the aforementioned enum. */
76 static const char *reason_messages
[] = {"", " (ignored)",
77 " (single edge duplicate)", " (edge pair duplicate)"};
80 static void combine_predictions_for_insn (rtx_insn
*, basic_block
);
81 static void dump_prediction (FILE *, enum br_predictor
, int, basic_block
,
82 enum predictor_reason
, edge
);
83 static void predict_paths_leading_to (basic_block
, enum br_predictor
,
85 class loop
*in_loop
= NULL
);
86 static void predict_paths_leading_to_edge (edge
, enum br_predictor
,
88 class loop
*in_loop
= NULL
);
89 static bool can_predict_insn_p (const rtx_insn
*);
90 static HOST_WIDE_INT
get_predictor_value (br_predictor
, HOST_WIDE_INT
);
91 static void determine_unlikely_bbs ();
93 /* Information we hold about each branch predictor.
94 Filled using information from predict.def. */
98 const char *const name
; /* Name used in the debugging dumps. */
99 const int hitrate
; /* Expected hitrate used by
100 predict_insn_def call. */
104 /* Use given predictor without Dempster-Shaffer theory if it matches
105 using first_match heuristics. */
106 #define PRED_FLAG_FIRST_MATCH 1
108 /* Recompute hitrate in percent to our representation. */
110 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
112 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
113 static const struct predictor_info predictor_info
[]= {
114 #include "predict.def"
116 /* Upper bound on predictors. */
121 static gcov_type min_count
= -1;
123 /* Determine the threshold for hot BB counts. */
126 get_hot_bb_threshold ()
130 const int hot_frac
= param_hot_bb_count_fraction
;
131 const gcov_type min_hot_count
133 ? profile_info
->sum_max
/ hot_frac
134 : (gcov_type
)profile_count::max_count
;
135 set_hot_bb_threshold (min_hot_count
);
137 fprintf (dump_file
, "Setting hotness threshold to %" PRId64
".\n",
143 /* Set the threshold for hot BB counts. */
146 set_hot_bb_threshold (gcov_type min
)
151 /* Return TRUE if COUNT is considered to be hot in function FUN. */
154 maybe_hot_count_p (struct function
*fun
, profile_count count
)
156 if (!count
.initialized_p ())
158 if (count
.ipa () == profile_count::zero ())
162 struct cgraph_node
*node
= cgraph_node::get (fun
->decl
);
163 if (!profile_info
|| profile_status_for_fn (fun
) != PROFILE_READ
)
165 if (node
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
)
167 if (node
->frequency
== NODE_FREQUENCY_HOT
)
170 if (profile_status_for_fn (fun
) == PROFILE_ABSENT
)
172 if (node
->frequency
== NODE_FREQUENCY_EXECUTED_ONCE
173 && count
< (ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
.apply_scale (2, 3)))
175 if (count
* param_hot_bb_frequency_fraction
176 < ENTRY_BLOCK_PTR_FOR_FN (fun
)->count
)
180 /* Code executed at most once is not hot. */
181 if (count
<= MAX (profile_info
? profile_info
->runs
: 1, 1))
183 return (count
>= get_hot_bb_threshold ());
186 /* Return true if basic block BB of function FUN can be CPU intensive
187 and should thus be optimized for maximum performance. */
190 maybe_hot_bb_p (struct function
*fun
, const_basic_block bb
)
192 gcc_checking_assert (fun
);
193 return maybe_hot_count_p (fun
, bb
->count
);
196 /* Return true if edge E can be CPU intensive and should thus be optimized
197 for maximum performance. */
200 maybe_hot_edge_p (edge e
)
202 return maybe_hot_count_p (cfun
, e
->count ());
205 /* Return true if COUNT is considered to be never executed in function FUN
206 or if function FUN is considered so in the static profile. */
209 probably_never_executed (struct function
*fun
, profile_count count
)
211 gcc_checking_assert (fun
);
212 if (count
.ipa () == profile_count::zero ())
214 /* Do not trust adjusted counts. This will make us to drop int cold section
215 code with low execution count as a result of inlining. These low counts
216 are not safe even with read profile and may lead us to dropping
217 code which actually gets executed into cold section of binary that is not
219 if (count
.precise_p () && profile_status_for_fn (fun
) == PROFILE_READ
)
221 const int unlikely_frac
= param_unlikely_bb_count_fraction
;
222 if (count
* unlikely_frac
>= profile_info
->runs
)
226 if ((!profile_info
|| profile_status_for_fn (fun
) != PROFILE_READ
)
227 && (cgraph_node::get (fun
->decl
)->frequency
228 == NODE_FREQUENCY_UNLIKELY_EXECUTED
))
233 /* Return true if basic block BB of function FUN is probably never executed. */
236 probably_never_executed_bb_p (struct function
*fun
, const_basic_block bb
)
238 return probably_never_executed (fun
, bb
->count
);
241 /* Return true if edge E is unlikely executed for obvious reasons. */
244 unlikely_executed_edge_p (edge e
)
246 return (e
->src
->count
== profile_count::zero ()
247 || e
->probability
== profile_probability::never ())
248 || (e
->flags
& (EDGE_EH
| EDGE_FAKE
));
251 /* Return true if edge E of function FUN is probably never executed. */
254 probably_never_executed_edge_p (struct function
*fun
, edge e
)
256 if (unlikely_executed_edge_p (e
))
258 return probably_never_executed (fun
, e
->count ());
261 /* Return true if function FUN should always be optimized for size. */
264 optimize_function_for_size_p (struct function
*fun
)
266 if (!fun
|| !fun
->decl
)
267 return optimize_size
? OPTIMIZE_SIZE_MAX
: OPTIMIZE_SIZE_NO
;
268 cgraph_node
*n
= cgraph_node::get (fun
->decl
);
270 return n
->optimize_for_size_p ();
271 return OPTIMIZE_SIZE_NO
;
274 /* Return true if function FUN should always be optimized for speed. */
277 optimize_function_for_speed_p (struct function
*fun
)
279 return !optimize_function_for_size_p (fun
);
282 /* Return the optimization type that should be used for function FUN. */
285 function_optimization_type (struct function
*fun
)
287 return (optimize_function_for_speed_p (fun
)
289 : OPTIMIZE_FOR_SIZE
);
292 /* Return TRUE if basic block BB should be optimized for size. */
295 optimize_bb_for_size_p (const_basic_block bb
)
297 enum optimize_size_level ret
= optimize_function_for_size_p (cfun
);
299 if (bb
&& ret
< OPTIMIZE_SIZE_MAX
&& bb
->count
== profile_count::zero ())
300 ret
= OPTIMIZE_SIZE_MAX
;
301 if (bb
&& ret
< OPTIMIZE_SIZE_BALANCED
&& !maybe_hot_bb_p (cfun
, bb
))
302 ret
= OPTIMIZE_SIZE_BALANCED
;
306 /* Return TRUE if basic block 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 the optimization type that should be used for basic block BB. */
317 bb_optimization_type (const_basic_block bb
)
319 return (optimize_bb_for_speed_p (bb
)
321 : OPTIMIZE_FOR_SIZE
);
324 /* Return TRUE if edge E should be optimized for size. */
327 optimize_edge_for_size_p (edge e
)
329 enum optimize_size_level ret
= optimize_function_for_size_p (cfun
);
331 if (ret
< OPTIMIZE_SIZE_MAX
&& unlikely_executed_edge_p (e
))
332 ret
= OPTIMIZE_SIZE_MAX
;
333 if (ret
< OPTIMIZE_SIZE_BALANCED
&& !maybe_hot_edge_p (e
))
334 ret
= OPTIMIZE_SIZE_BALANCED
;
338 /* Return TRUE if edge E should be optimized for speed. */
341 optimize_edge_for_speed_p (edge e
)
343 return !optimize_edge_for_size_p (e
);
346 /* Return TRUE if the current function is optimized for size. */
349 optimize_insn_for_size_p (void)
351 enum optimize_size_level ret
= optimize_function_for_size_p (cfun
);
352 if (ret
< OPTIMIZE_SIZE_BALANCED
&& !crtl
->maybe_hot_insn_p
)
353 ret
= OPTIMIZE_SIZE_BALANCED
;
357 /* Return TRUE if the current function is optimized for speed. */
360 optimize_insn_for_speed_p (void)
362 return !optimize_insn_for_size_p ();
365 /* Return the optimization type that should be used for the current
369 insn_optimization_type ()
371 return (optimize_insn_for_speed_p ()
373 : OPTIMIZE_FOR_SIZE
);
376 /* Return TRUE if LOOP should be optimized for size. */
379 optimize_loop_for_size_p (class loop
*loop
)
381 return optimize_bb_for_size_p (loop
->header
);
384 /* Return TRUE if LOOP should be optimized for speed. */
387 optimize_loop_for_speed_p (class loop
*loop
)
389 return optimize_bb_for_speed_p (loop
->header
);
392 /* Return TRUE if nest rooted at LOOP should be optimized for speed. */
395 optimize_loop_nest_for_speed_p (class loop
*loop
)
397 class loop
*l
= loop
;
398 if (optimize_loop_for_speed_p (loop
))
401 while (l
&& l
!= loop
)
403 if (optimize_loop_for_speed_p (l
))
411 while (l
!= loop
&& !l
->next
)
420 /* Return TRUE if nest rooted at LOOP should be optimized for size. */
423 optimize_loop_nest_for_size_p (class loop
*loop
)
425 enum optimize_size_level ret
= optimize_loop_for_size_p (loop
);
426 class loop
*l
= loop
;
429 while (l
&& l
!= loop
)
431 if (ret
== OPTIMIZE_SIZE_NO
)
433 ret
= MIN (optimize_loop_for_size_p (l
), ret
);
440 while (l
!= loop
&& !l
->next
)
449 /* Return true if edge E is likely to be well predictable by branch
453 predictable_edge_p (edge e
)
455 if (!e
->probability
.initialized_p ())
457 if ((e
->probability
.to_reg_br_prob_base ()
458 <= param_predictable_branch_outcome
* REG_BR_PROB_BASE
/ 100)
459 || (REG_BR_PROB_BASE
- e
->probability
.to_reg_br_prob_base ()
460 <= param_predictable_branch_outcome
* REG_BR_PROB_BASE
/ 100))
466 /* Set RTL expansion for BB profile. */
469 rtl_profile_for_bb (basic_block bb
)
471 crtl
->maybe_hot_insn_p
= maybe_hot_bb_p (cfun
, bb
);
474 /* Set RTL expansion for edge profile. */
477 rtl_profile_for_edge (edge e
)
479 crtl
->maybe_hot_insn_p
= maybe_hot_edge_p (e
);
482 /* Set RTL expansion to default mode (i.e. when profile info is not known). */
484 default_rtl_profile (void)
486 crtl
->maybe_hot_insn_p
= true;
489 /* Return true if the one of outgoing edges is already predicted by
493 rtl_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
496 if (!INSN_P (BB_END (bb
)))
498 for (note
= REG_NOTES (BB_END (bb
)); note
; note
= XEXP (note
, 1))
499 if (REG_NOTE_KIND (note
) == REG_BR_PRED
500 && INTVAL (XEXP (XEXP (note
, 0), 0)) == (int)predictor
)
505 /* Structure representing predictions in tree level. */
507 struct edge_prediction
{
508 struct edge_prediction
*ep_next
;
510 enum br_predictor ep_predictor
;
514 /* This map contains for a basic block the list of predictions for the
517 static hash_map
<const_basic_block
, edge_prediction
*> *bb_predictions
;
519 /* Return true if the one of outgoing edges is already predicted by
523 gimple_predicted_by_p (const_basic_block bb
, enum br_predictor predictor
)
525 struct edge_prediction
*i
;
526 edge_prediction
**preds
= bb_predictions
->get (bb
);
531 for (i
= *preds
; i
; i
= i
->ep_next
)
532 if (i
->ep_predictor
== predictor
)
537 /* Return true if the one of outgoing edges is already predicted by
538 PREDICTOR for edge E predicted as TAKEN. */
541 edge_predicted_by_p (edge e
, enum br_predictor predictor
, bool taken
)
543 struct edge_prediction
*i
;
544 basic_block bb
= e
->src
;
545 edge_prediction
**preds
= bb_predictions
->get (bb
);
549 int probability
= predictor_info
[(int) predictor
].hitrate
;
552 probability
= REG_BR_PROB_BASE
- probability
;
554 for (i
= *preds
; i
; i
= i
->ep_next
)
555 if (i
->ep_predictor
== predictor
557 && i
->ep_probability
== probability
)
562 /* Same predicate as above, working on edges. */
564 edge_probability_reliable_p (const_edge e
)
566 return e
->probability
.probably_reliable_p ();
569 /* Same predicate as edge_probability_reliable_p, working on notes. */
571 br_prob_note_reliable_p (const_rtx note
)
573 gcc_assert (REG_NOTE_KIND (note
) == REG_BR_PROB
);
574 return profile_probability::from_reg_br_prob_note
575 (XINT (note
, 0)).probably_reliable_p ();
579 predict_insn (rtx_insn
*insn
, enum br_predictor predictor
, int probability
)
581 gcc_assert (any_condjump_p (insn
));
582 if (!flag_guess_branch_prob
)
585 add_reg_note (insn
, REG_BR_PRED
,
586 gen_rtx_CONCAT (VOIDmode
,
587 GEN_INT ((int) predictor
),
588 GEN_INT ((int) probability
)));
591 /* Predict insn by given predictor. */
594 predict_insn_def (rtx_insn
*insn
, enum br_predictor predictor
,
595 enum prediction taken
)
597 int probability
= predictor_info
[(int) predictor
].hitrate
;
598 gcc_assert (probability
!= PROB_UNINITIALIZED
);
601 probability
= REG_BR_PROB_BASE
- probability
;
603 predict_insn (insn
, predictor
, probability
);
606 /* Predict edge E with given probability if possible. */
609 rtl_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
612 last_insn
= BB_END (e
->src
);
614 /* We can store the branch prediction information only about
615 conditional jumps. */
616 if (!any_condjump_p (last_insn
))
619 /* We always store probability of branching. */
620 if (e
->flags
& EDGE_FALLTHRU
)
621 probability
= REG_BR_PROB_BASE
- probability
;
623 predict_insn (last_insn
, predictor
, probability
);
626 /* Predict edge E with the given PROBABILITY. */
628 gimple_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
630 if (e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
)
631 && EDGE_COUNT (e
->src
->succs
) > 1
632 && flag_guess_branch_prob
635 struct edge_prediction
*i
= XNEW (struct edge_prediction
);
636 edge_prediction
*&preds
= bb_predictions
->get_or_insert (e
->src
);
640 i
->ep_probability
= probability
;
641 i
->ep_predictor
= predictor
;
646 /* Filter edge predictions PREDS by a function FILTER: if FILTER return false
647 the prediction is removed.
648 DATA are passed to the filter function. */
651 filter_predictions (edge_prediction
**preds
,
652 bool (*filter
) (edge_prediction
*, void *), void *data
)
659 struct edge_prediction
**prediction
= preds
;
660 struct edge_prediction
*next
;
664 if ((*filter
) (*prediction
, data
))
665 prediction
= &((*prediction
)->ep_next
);
668 next
= (*prediction
)->ep_next
;
676 /* Filter function predicate that returns true for a edge predicate P
677 if its edge is equal to DATA. */
680 not_equal_edge_p (edge_prediction
*p
, void *data
)
682 return p
->ep_edge
!= (edge
)data
;
685 /* Remove all predictions on given basic block that are attached
688 remove_predictions_associated_with_edge (edge e
)
693 edge_prediction
**preds
= bb_predictions
->get (e
->src
);
694 filter_predictions (preds
, not_equal_edge_p
, e
);
697 /* Clears the list of predictions stored for BB. */
700 clear_bb_predictions (basic_block bb
)
702 edge_prediction
**preds
= bb_predictions
->get (bb
);
703 struct edge_prediction
*pred
, *next
;
708 for (pred
= *preds
; pred
; pred
= next
)
710 next
= pred
->ep_next
;
716 /* Return true when we can store prediction on insn INSN.
717 At the moment we represent predictions only on conditional
718 jumps, not at computed jump or other complicated cases. */
720 can_predict_insn_p (const rtx_insn
*insn
)
722 return (JUMP_P (insn
)
723 && any_condjump_p (insn
)
724 && EDGE_COUNT (BLOCK_FOR_INSN (insn
)->succs
) >= 2);
727 /* Predict edge E by given predictor if possible. */
730 predict_edge_def (edge e
, enum br_predictor predictor
,
731 enum prediction taken
)
733 int probability
= predictor_info
[(int) predictor
].hitrate
;
736 probability
= REG_BR_PROB_BASE
- probability
;
738 predict_edge (e
, predictor
, probability
);
741 /* Invert all branch predictions or probability notes in the INSN. This needs
742 to be done each time we invert the condition used by the jump. */
745 invert_br_probabilities (rtx insn
)
749 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
750 if (REG_NOTE_KIND (note
) == REG_BR_PROB
)
751 XINT (note
, 0) = profile_probability::from_reg_br_prob_note
752 (XINT (note
, 0)).invert ().to_reg_br_prob_note ();
753 else if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
754 XEXP (XEXP (note
, 0), 1)
755 = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (XEXP (note
, 0), 1)));
758 /* Dump information about the branch prediction to the output file. */
761 dump_prediction (FILE *file
, enum br_predictor predictor
, int probability
,
762 basic_block bb
, enum predictor_reason reason
= REASON_NONE
,
772 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
773 if (! (e
->flags
& EDGE_FALLTHRU
))
776 char edge_info_str
[128];
778 sprintf (edge_info_str
, " of edge %d->%d", ep_edge
->src
->index
,
779 ep_edge
->dest
->index
);
781 edge_info_str
[0] = '\0';
783 fprintf (file
, " %s heuristics%s%s: %.2f%%",
784 predictor_info
[predictor
].name
,
785 edge_info_str
, reason_messages
[reason
],
786 probability
* 100.0 / REG_BR_PROB_BASE
);
788 if (bb
->count
.initialized_p ())
790 fprintf (file
, " exec ");
791 bb
->count
.dump (file
);
794 fprintf (file
, " hit ");
795 e
->count ().dump (file
);
796 fprintf (file
, " (%.1f%%)", e
->count ().to_gcov_type() * 100.0
797 / bb
->count
.to_gcov_type ());
801 fprintf (file
, "\n");
803 /* Print output that be easily read by analyze_brprob.py script. We are
804 interested only in counts that are read from GCDA files. */
805 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
806 && bb
->count
.precise_p ()
807 && reason
== REASON_NONE
)
809 fprintf (file
, ";;heuristics;%s;%" PRId64
";%" PRId64
";%.1f;\n",
810 predictor_info
[predictor
].name
,
811 bb
->count
.to_gcov_type (), e
->count ().to_gcov_type (),
812 probability
* 100.0 / REG_BR_PROB_BASE
);
816 /* Return true if STMT is known to be unlikely executed. */
819 unlikely_executed_stmt_p (gimple
*stmt
)
821 if (!is_gimple_call (stmt
))
823 /* NORETURN attribute alone is not strong enough: exit() may be quite
824 likely executed once during program run. */
825 if (gimple_call_fntype (stmt
)
826 && lookup_attribute ("cold",
827 TYPE_ATTRIBUTES (gimple_call_fntype (stmt
)))
828 && !lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
)))
830 tree decl
= gimple_call_fndecl (stmt
);
833 if (lookup_attribute ("cold", DECL_ATTRIBUTES (decl
))
834 && !lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
)))
837 cgraph_node
*n
= cgraph_node::get (decl
);
842 n
= n
->ultimate_alias_target (&avail
);
843 if (avail
< AVAIL_AVAILABLE
)
846 || n
->decl
== current_function_decl
)
848 return n
->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED
;
851 /* Return true if BB is unlikely executed. */
854 unlikely_executed_bb_p (basic_block bb
)
856 if (bb
->count
== profile_count::zero ())
858 if (bb
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) || bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
860 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
861 !gsi_end_p (gsi
); gsi_next (&gsi
))
863 if (unlikely_executed_stmt_p (gsi_stmt (gsi
)))
865 if (stmt_can_terminate_bb_p (gsi_stmt (gsi
)))
871 /* We cannot predict the probabilities of outgoing edges of bb. Set them
872 evenly and hope for the best. If UNLIKELY_EDGES is not null, distribute
873 even probability for all edges not mentioned in the set. These edges
874 are given PROB_VERY_UNLIKELY probability. Similarly for LIKELY_EDGES,
875 if we have exactly one likely edge, make the other edges predicted
879 set_even_probabilities (basic_block bb
,
880 hash_set
<edge
> *unlikely_edges
= NULL
,
881 hash_set
<edge_prediction
*> *likely_edges
= NULL
)
883 unsigned nedges
= 0, unlikely_count
= 0;
886 profile_probability all
= profile_probability::always ();
888 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
889 if (e
->probability
.initialized_p ())
890 all
-= e
->probability
;
891 else if (!unlikely_executed_edge_p (e
))
894 if (unlikely_edges
!= NULL
&& unlikely_edges
->contains (e
))
896 all
-= profile_probability::very_unlikely ();
901 /* Make the distribution even if all edges are unlikely. */
902 unsigned likely_count
= likely_edges
? likely_edges
->elements () : 0;
903 if (unlikely_count
== nedges
)
905 unlikely_edges
= NULL
;
909 /* If we have one likely edge, then use its probability and distribute
910 remaining probabilities as even. */
911 if (likely_count
== 1)
913 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
914 if (e
->probability
.initialized_p ())
916 else if (!unlikely_executed_edge_p (e
))
918 edge_prediction
*prediction
= *likely_edges
->begin ();
919 int p
= prediction
->ep_probability
;
920 profile_probability prob
921 = profile_probability::from_reg_br_prob_base (p
);
923 if (prediction
->ep_edge
== e
)
924 e
->probability
= prob
;
925 else if (unlikely_edges
!= NULL
&& unlikely_edges
->contains (e
))
926 e
->probability
= profile_probability::very_unlikely ();
929 profile_probability remainder
= prob
.invert ();
930 remainder
-= (profile_probability::very_unlikely ()
932 int count
= nedges
- unlikely_count
- 1;
933 gcc_assert (count
>= 0);
935 e
->probability
= remainder
/ count
;
939 e
->probability
= profile_probability::never ();
943 /* Make all unlikely edges unlikely and the rest will have even
945 unsigned scale
= nedges
- unlikely_count
;
946 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
947 if (e
->probability
.initialized_p ())
949 else if (!unlikely_executed_edge_p (e
))
951 if (unlikely_edges
!= NULL
&& unlikely_edges
->contains (e
))
952 e
->probability
= profile_probability::very_unlikely ();
954 e
->probability
= all
/ scale
;
957 e
->probability
= profile_probability::never ();
961 /* Add REG_BR_PROB note to JUMP with PROB. */
964 add_reg_br_prob_note (rtx_insn
*jump
, profile_probability prob
)
966 gcc_checking_assert (JUMP_P (jump
) && !find_reg_note (jump
, REG_BR_PROB
, 0));
967 add_int_reg_note (jump
, REG_BR_PROB
, prob
.to_reg_br_prob_note ());
970 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
971 note if not already present. Remove now useless REG_BR_PRED notes. */
974 combine_predictions_for_insn (rtx_insn
*insn
, basic_block bb
)
979 int best_probability
= PROB_EVEN
;
980 enum br_predictor best_predictor
= END_PREDICTORS
;
981 int combined_probability
= REG_BR_PROB_BASE
/ 2;
983 bool first_match
= false;
986 if (!can_predict_insn_p (insn
))
988 set_even_probabilities (bb
);
992 prob_note
= find_reg_note (insn
, REG_BR_PROB
, 0);
993 pnote
= ®_NOTES (insn
);
995 fprintf (dump_file
, "Predictions for insn %i bb %i\n", INSN_UID (insn
),
998 /* We implement "first match" heuristics and use probability guessed
999 by predictor with smallest index. */
1000 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
1001 if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
1003 enum br_predictor predictor
= ((enum br_predictor
)
1004 INTVAL (XEXP (XEXP (note
, 0), 0)));
1005 int probability
= INTVAL (XEXP (XEXP (note
, 0), 1));
1008 if (best_predictor
> predictor
1009 && predictor_info
[predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
1010 best_probability
= probability
, best_predictor
= predictor
;
1012 d
= (combined_probability
* probability
1013 + (REG_BR_PROB_BASE
- combined_probability
)
1014 * (REG_BR_PROB_BASE
- probability
));
1016 /* Use FP math to avoid overflows of 32bit integers. */
1018 /* If one probability is 0% and one 100%, avoid division by zero. */
1019 combined_probability
= REG_BR_PROB_BASE
/ 2;
1021 combined_probability
= (((double) combined_probability
) * probability
1022 * REG_BR_PROB_BASE
/ d
+ 0.5);
1025 /* Decide which heuristic to use. In case we didn't match anything,
1026 use no_prediction heuristic, in case we did match, use either
1027 first match or Dempster-Shaffer theory depending on the flags. */
1029 if (best_predictor
!= END_PREDICTORS
)
1033 dump_prediction (dump_file
, PRED_NO_PREDICTION
,
1034 combined_probability
, bb
);
1038 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
,
1039 bb
, !first_match
? REASON_NONE
: REASON_IGNORED
);
1041 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
,
1042 bb
, first_match
? REASON_NONE
: REASON_IGNORED
);
1046 combined_probability
= best_probability
;
1047 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
);
1051 if (REG_NOTE_KIND (*pnote
) == REG_BR_PRED
)
1053 enum br_predictor predictor
= ((enum br_predictor
)
1054 INTVAL (XEXP (XEXP (*pnote
, 0), 0)));
1055 int probability
= INTVAL (XEXP (XEXP (*pnote
, 0), 1));
1057 dump_prediction (dump_file
, predictor
, probability
, bb
,
1058 (!first_match
|| best_predictor
== predictor
)
1059 ? REASON_NONE
: REASON_IGNORED
);
1060 *pnote
= XEXP (*pnote
, 1);
1063 pnote
= &XEXP (*pnote
, 1);
1068 profile_probability p
1069 = profile_probability::from_reg_br_prob_base (combined_probability
);
1070 add_reg_br_prob_note (insn
, p
);
1072 /* Save the prediction into CFG in case we are seeing non-degenerated
1073 conditional jump. */
1074 if (!single_succ_p (bb
))
1076 BRANCH_EDGE (bb
)->probability
= p
;
1077 FALLTHRU_EDGE (bb
)->probability
1078 = BRANCH_EDGE (bb
)->probability
.invert ();
1081 else if (!single_succ_p (bb
))
1083 profile_probability prob
= profile_probability::from_reg_br_prob_note
1084 (XINT (prob_note
, 0));
1086 BRANCH_EDGE (bb
)->probability
= prob
;
1087 FALLTHRU_EDGE (bb
)->probability
= prob
.invert ();
1090 single_succ_edge (bb
)->probability
= profile_probability::always ();
1093 /* Edge prediction hash traits. */
1095 struct predictor_hash
: pointer_hash
<edge_prediction
>
1098 static inline hashval_t
hash (const edge_prediction
*);
1099 static inline bool equal (const edge_prediction
*, const edge_prediction
*);
1102 /* Calculate hash value of an edge prediction P based on predictor and
1103 normalized probability. */
1106 predictor_hash::hash (const edge_prediction
*p
)
1108 inchash::hash hstate
;
1109 hstate
.add_int (p
->ep_predictor
);
1111 int prob
= p
->ep_probability
;
1112 if (prob
> REG_BR_PROB_BASE
/ 2)
1113 prob
= REG_BR_PROB_BASE
- prob
;
1115 hstate
.add_int (prob
);
1117 return hstate
.end ();
1120 /* Return true whether edge predictions P1 and P2 use the same predictor and
1121 have equal (or opposed probability). */
1124 predictor_hash::equal (const edge_prediction
*p1
, const edge_prediction
*p2
)
1126 return (p1
->ep_predictor
== p2
->ep_predictor
1127 && (p1
->ep_probability
== p2
->ep_probability
1128 || p1
->ep_probability
== REG_BR_PROB_BASE
- p2
->ep_probability
));
1131 struct predictor_hash_traits
: predictor_hash
,
1132 typed_noop_remove
<edge_prediction
*> {};
1134 /* Return true if edge prediction P is not in DATA hash set. */
1137 not_removed_prediction_p (edge_prediction
*p
, void *data
)
1139 hash_set
<edge_prediction
*> *remove
= (hash_set
<edge_prediction
*> *) data
;
1140 return !remove
->contains (p
);
1143 /* Prune predictions for a basic block BB. Currently we do following
1146 1) remove duplicate prediction that is guessed with the same probability
1147 (different than 1/2) to both edge
1148 2) remove duplicates for a prediction that belongs with the same probability
1154 prune_predictions_for_bb (basic_block bb
)
1156 edge_prediction
**preds
= bb_predictions
->get (bb
);
1160 hash_table
<predictor_hash_traits
> s (13);
1161 hash_set
<edge_prediction
*> remove
;
1163 /* Step 1: identify predictors that should be removed. */
1164 for (edge_prediction
*pred
= *preds
; pred
; pred
= pred
->ep_next
)
1166 edge_prediction
*existing
= s
.find (pred
);
1169 if (pred
->ep_edge
== existing
->ep_edge
1170 && pred
->ep_probability
== existing
->ep_probability
)
1172 /* Remove a duplicate predictor. */
1173 dump_prediction (dump_file
, pred
->ep_predictor
,
1174 pred
->ep_probability
, bb
,
1175 REASON_SINGLE_EDGE_DUPLICATE
, pred
->ep_edge
);
1179 else if (pred
->ep_edge
!= existing
->ep_edge
1180 && pred
->ep_probability
== existing
->ep_probability
1181 && pred
->ep_probability
!= REG_BR_PROB_BASE
/ 2)
1183 /* Remove both predictors as they predict the same
1185 dump_prediction (dump_file
, existing
->ep_predictor
,
1186 pred
->ep_probability
, bb
,
1187 REASON_EDGE_PAIR_DUPLICATE
,
1189 dump_prediction (dump_file
, pred
->ep_predictor
,
1190 pred
->ep_probability
, bb
,
1191 REASON_EDGE_PAIR_DUPLICATE
,
1194 remove
.add (existing
);
1199 edge_prediction
**slot2
= s
.find_slot (pred
, INSERT
);
1203 /* Step 2: Remove predictors. */
1204 filter_predictions (preds
, not_removed_prediction_p
, &remove
);
1208 /* Combine predictions into single probability and store them into CFG.
1209 Remove now useless prediction entries.
1210 If DRY_RUN is set, only produce dumps and do not modify profile. */
1213 combine_predictions_for_bb (basic_block bb
, bool dry_run
)
1215 int best_probability
= PROB_EVEN
;
1216 enum br_predictor best_predictor
= END_PREDICTORS
;
1217 int combined_probability
= REG_BR_PROB_BASE
/ 2;
1219 bool first_match
= false;
1221 struct edge_prediction
*pred
;
1223 edge e
, first
= NULL
, second
= NULL
;
1228 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1230 if (!unlikely_executed_edge_p (e
))
1233 if (first
&& !second
)
1238 else if (!e
->probability
.initialized_p ())
1239 e
->probability
= profile_probability::never ();
1240 if (!e
->probability
.initialized_p ())
1242 else if (e
->probability
== profile_probability::never ())
1246 /* When there is no successor or only one choice, prediction is easy.
1248 When we have a basic block with more than 2 successors, the situation
1249 is more complicated as DS theory cannot be used literally.
1250 More precisely, let's assume we predicted edge e1 with probability p1,
1251 thus: m1({b1}) = p1. As we're going to combine more than 2 edges, we
1252 need to find probability of e.g. m1({b2}), which we don't know.
1253 The only approximation is to equally distribute 1-p1 to all edges
1256 According to numbers we've got from SPEC2006 benchark, there's only
1257 one interesting reliable predictor (noreturn call), which can be
1258 handled with a bit easier approach. */
1261 hash_set
<edge
> unlikely_edges (4);
1262 hash_set
<edge_prediction
*> likely_edges (4);
1264 /* Identify all edges that have a probability close to very unlikely.
1265 Doing the approach for very unlikely doesn't worth for doing as
1266 there's no such probability in SPEC2006 benchmark. */
1267 edge_prediction
**preds
= bb_predictions
->get (bb
);
1269 for (pred
= *preds
; pred
; pred
= pred
->ep_next
)
1271 if (pred
->ep_probability
<= PROB_VERY_UNLIKELY
1272 || pred
->ep_predictor
== PRED_COLD_LABEL
)
1273 unlikely_edges
.add (pred
->ep_edge
);
1274 else if (pred
->ep_probability
>= PROB_VERY_LIKELY
1275 || pred
->ep_predictor
== PRED_BUILTIN_EXPECT
1276 || pred
->ep_predictor
== PRED_HOT_LABEL
)
1277 likely_edges
.add (pred
);
1280 /* It can happen that an edge is both in likely_edges and unlikely_edges.
1281 Clear both sets in that situation. */
1282 for (hash_set
<edge_prediction
*>::iterator it
= likely_edges
.begin ();
1283 it
!= likely_edges
.end (); ++it
)
1284 if (unlikely_edges
.contains ((*it
)->ep_edge
))
1286 likely_edges
.empty ();
1287 unlikely_edges
.empty ();
1292 set_even_probabilities (bb
, &unlikely_edges
, &likely_edges
);
1293 clear_bb_predictions (bb
);
1296 fprintf (dump_file
, "Predictions for bb %i\n", bb
->index
);
1297 if (unlikely_edges
.is_empty ())
1299 "%i edges in bb %i predicted to even probabilities\n",
1304 "%i edges in bb %i predicted with some unlikely edges\n",
1306 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1307 if (!unlikely_executed_edge_p (e
))
1308 dump_prediction (dump_file
, PRED_COMBINED
,
1309 e
->probability
.to_reg_br_prob_base (), bb
, REASON_NONE
, e
);
1316 fprintf (dump_file
, "Predictions for bb %i\n", bb
->index
);
1318 prune_predictions_for_bb (bb
);
1320 edge_prediction
**preds
= bb_predictions
->get (bb
);
1324 /* We implement "first match" heuristics and use probability guessed
1325 by predictor with smallest index. */
1326 for (pred
= *preds
; pred
; pred
= pred
->ep_next
)
1328 enum br_predictor predictor
= pred
->ep_predictor
;
1329 int probability
= pred
->ep_probability
;
1331 if (pred
->ep_edge
!= first
)
1332 probability
= REG_BR_PROB_BASE
- probability
;
1335 /* First match heuristics would be widly confused if we predicted
1337 if (best_predictor
> predictor
1338 && predictor_info
[predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
1340 struct edge_prediction
*pred2
;
1341 int prob
= probability
;
1343 for (pred2
= (struct edge_prediction
*) *preds
;
1344 pred2
; pred2
= pred2
->ep_next
)
1345 if (pred2
!= pred
&& pred2
->ep_predictor
== pred
->ep_predictor
)
1347 int probability2
= pred2
->ep_probability
;
1349 if (pred2
->ep_edge
!= first
)
1350 probability2
= REG_BR_PROB_BASE
- probability2
;
1352 if ((probability
< REG_BR_PROB_BASE
/ 2) !=
1353 (probability2
< REG_BR_PROB_BASE
/ 2))
1356 /* If the same predictor later gave better result, go for it! */
1357 if ((probability
>= REG_BR_PROB_BASE
/ 2 && (probability2
> probability
))
1358 || (probability
<= REG_BR_PROB_BASE
/ 2 && (probability2
< probability
)))
1359 prob
= probability2
;
1362 best_probability
= prob
, best_predictor
= predictor
;
1365 d
= (combined_probability
* probability
1366 + (REG_BR_PROB_BASE
- combined_probability
)
1367 * (REG_BR_PROB_BASE
- probability
));
1369 /* Use FP math to avoid overflows of 32bit integers. */
1371 /* If one probability is 0% and one 100%, avoid division by zero. */
1372 combined_probability
= REG_BR_PROB_BASE
/ 2;
1374 combined_probability
= (((double) combined_probability
)
1376 * REG_BR_PROB_BASE
/ d
+ 0.5);
1380 /* Decide which heuristic to use. In case we didn't match anything,
1381 use no_prediction heuristic, in case we did match, use either
1382 first match or Dempster-Shaffer theory depending on the flags. */
1384 if (best_predictor
!= END_PREDICTORS
)
1388 dump_prediction (dump_file
, PRED_NO_PREDICTION
, combined_probability
, bb
);
1392 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
, bb
,
1393 !first_match
? REASON_NONE
: REASON_IGNORED
);
1395 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
, bb
,
1396 first_match
? REASON_NONE
: REASON_IGNORED
);
1400 combined_probability
= best_probability
;
1401 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
);
1405 for (pred
= (struct edge_prediction
*) *preds
; pred
; pred
= pred
->ep_next
)
1407 enum br_predictor predictor
= pred
->ep_predictor
;
1408 int probability
= pred
->ep_probability
;
1410 dump_prediction (dump_file
, predictor
, probability
, bb
,
1411 (!first_match
|| best_predictor
== predictor
)
1412 ? REASON_NONE
: REASON_IGNORED
, pred
->ep_edge
);
1415 clear_bb_predictions (bb
);
1418 /* If we have only one successor which is unknown, we can compute missing
1422 profile_probability prob
= profile_probability::always ();
1423 edge missing
= NULL
;
1425 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1426 if (e
->probability
.initialized_p ())
1427 prob
-= e
->probability
;
1428 else if (missing
== NULL
)
1432 missing
->probability
= prob
;
1434 /* If nothing is unknown, we have nothing to update. */
1435 else if (!nunknown
&& nzero
!= (int)EDGE_COUNT (bb
->succs
))
1440 = profile_probability::from_reg_br_prob_base (combined_probability
);
1441 second
->probability
= first
->probability
.invert ();
1445 /* Check if T1 and T2 satisfy the IV_COMPARE condition.
1446 Return the SSA_NAME if the condition satisfies, NULL otherwise.
1448 T1 and T2 should be one of the following cases:
1449 1. T1 is SSA_NAME, T2 is NULL
1450 2. T1 is SSA_NAME, T2 is INTEGER_CST between [-4, 4]
1451 3. T2 is SSA_NAME, T1 is INTEGER_CST between [-4, 4] */
1454 strips_small_constant (tree t1
, tree t2
)
1461 else if (TREE_CODE (t1
) == SSA_NAME
)
1463 else if (tree_fits_shwi_p (t1
))
1464 value
= tree_to_shwi (t1
);
1470 else if (tree_fits_shwi_p (t2
))
1471 value
= tree_to_shwi (t2
);
1472 else if (TREE_CODE (t2
) == SSA_NAME
)
1480 if (value
<= 4 && value
>= -4)
1486 /* Return the SSA_NAME in T or T's operands.
1487 Return NULL if SSA_NAME cannot be found. */
1490 get_base_value (tree t
)
1492 if (TREE_CODE (t
) == SSA_NAME
)
1495 if (!BINARY_CLASS_P (t
))
1498 switch (TREE_OPERAND_LENGTH (t
))
1501 return strips_small_constant (TREE_OPERAND (t
, 0), NULL
);
1503 return strips_small_constant (TREE_OPERAND (t
, 0),
1504 TREE_OPERAND (t
, 1));
1510 /* Check the compare STMT in LOOP. If it compares an induction
1511 variable to a loop invariant, return true, and save
1512 LOOP_INVARIANT, COMPARE_CODE and LOOP_STEP.
1513 Otherwise return false and set LOOP_INVAIANT to NULL. */
1516 is_comparison_with_loop_invariant_p (gcond
*stmt
, class loop
*loop
,
1517 tree
*loop_invariant
,
1518 enum tree_code
*compare_code
,
1522 tree op0
, op1
, bound
, base
;
1524 enum tree_code code
;
1527 code
= gimple_cond_code (stmt
);
1528 *loop_invariant
= NULL
;
1544 op0
= gimple_cond_lhs (stmt
);
1545 op1
= gimple_cond_rhs (stmt
);
1547 if ((TREE_CODE (op0
) != SSA_NAME
&& TREE_CODE (op0
) != INTEGER_CST
)
1548 || (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op1
) != INTEGER_CST
))
1550 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op0
, &iv0
, true))
1552 if (!simple_iv (loop
, loop_containing_stmt (stmt
), op1
, &iv1
, true))
1554 if (TREE_CODE (iv0
.step
) != INTEGER_CST
1555 || TREE_CODE (iv1
.step
) != INTEGER_CST
)
1557 if ((integer_zerop (iv0
.step
) && integer_zerop (iv1
.step
))
1558 || (!integer_zerop (iv0
.step
) && !integer_zerop (iv1
.step
)))
1561 if (integer_zerop (iv0
.step
))
1563 if (code
!= NE_EXPR
&& code
!= EQ_EXPR
)
1564 code
= invert_tree_comparison (code
, false);
1567 if (tree_fits_shwi_p (iv1
.step
))
1576 if (tree_fits_shwi_p (iv0
.step
))
1582 if (TREE_CODE (bound
) != INTEGER_CST
)
1583 bound
= get_base_value (bound
);
1586 if (TREE_CODE (base
) != INTEGER_CST
)
1587 base
= get_base_value (base
);
1591 *loop_invariant
= bound
;
1592 *compare_code
= code
;
1594 *loop_iv_base
= base
;
1598 /* Compare two SSA_NAMEs: returns TRUE if T1 and T2 are value coherent. */
1601 expr_coherent_p (tree t1
, tree t2
)
1604 tree ssa_name_1
= NULL
;
1605 tree ssa_name_2
= NULL
;
1607 gcc_assert (TREE_CODE (t1
) == SSA_NAME
|| TREE_CODE (t1
) == INTEGER_CST
);
1608 gcc_assert (TREE_CODE (t2
) == SSA_NAME
|| TREE_CODE (t2
) == INTEGER_CST
);
1613 if (TREE_CODE (t1
) == INTEGER_CST
&& TREE_CODE (t2
) == INTEGER_CST
)
1615 if (TREE_CODE (t1
) == INTEGER_CST
|| TREE_CODE (t2
) == INTEGER_CST
)
1618 /* Check to see if t1 is expressed/defined with t2. */
1619 stmt
= SSA_NAME_DEF_STMT (t1
);
1620 gcc_assert (stmt
!= NULL
);
1621 if (is_gimple_assign (stmt
))
1623 ssa_name_1
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
1624 if (ssa_name_1
&& ssa_name_1
== t2
)
1628 /* Check to see if t2 is expressed/defined with t1. */
1629 stmt
= SSA_NAME_DEF_STMT (t2
);
1630 gcc_assert (stmt
!= NULL
);
1631 if (is_gimple_assign (stmt
))
1633 ssa_name_2
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_USE
);
1634 if (ssa_name_2
&& ssa_name_2
== t1
)
1638 /* Compare if t1 and t2's def_stmts are identical. */
1639 if (ssa_name_2
!= NULL
&& ssa_name_1
== ssa_name_2
)
1645 /* Return true if E is predicted by one of loop heuristics. */
1648 predicted_by_loop_heuristics_p (basic_block bb
)
1650 struct edge_prediction
*i
;
1651 edge_prediction
**preds
= bb_predictions
->get (bb
);
1656 for (i
= *preds
; i
; i
= i
->ep_next
)
1657 if (i
->ep_predictor
== PRED_LOOP_ITERATIONS_GUESSED
1658 || i
->ep_predictor
== PRED_LOOP_ITERATIONS_MAX
1659 || i
->ep_predictor
== PRED_LOOP_ITERATIONS
1660 || i
->ep_predictor
== PRED_LOOP_EXIT
1661 || i
->ep_predictor
== PRED_LOOP_EXIT_WITH_RECURSION
1662 || i
->ep_predictor
== PRED_LOOP_EXTRA_EXIT
)
1667 /* Predict branch probability of BB when BB contains a branch that compares
1668 an induction variable in LOOP with LOOP_IV_BASE_VAR to LOOP_BOUND_VAR. The
1669 loop exit is compared using LOOP_BOUND_CODE, with step of LOOP_BOUND_STEP.
1672 for (int i = 0; i < bound; i++) {
1679 In this loop, we will predict the branch inside the loop to be taken. */
1682 predict_iv_comparison (class loop
*loop
, basic_block bb
,
1683 tree loop_bound_var
,
1684 tree loop_iv_base_var
,
1685 enum tree_code loop_bound_code
,
1686 int loop_bound_step
)
1689 tree compare_var
, compare_base
;
1690 enum tree_code compare_code
;
1691 tree compare_step_var
;
1695 if (predicted_by_loop_heuristics_p (bb
))
1698 stmt
= last_stmt (bb
);
1699 if (!stmt
|| gimple_code (stmt
) != GIMPLE_COND
)
1701 if (!is_comparison_with_loop_invariant_p (as_a
<gcond
*> (stmt
),
1708 /* Find the taken edge. */
1709 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
1710 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1713 /* When comparing an IV to a loop invariant, NE is more likely to be
1714 taken while EQ is more likely to be not-taken. */
1715 if (compare_code
== NE_EXPR
)
1717 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1720 else if (compare_code
== EQ_EXPR
)
1722 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1726 if (!expr_coherent_p (loop_iv_base_var
, compare_base
))
1729 /* If loop bound, base and compare bound are all constants, we can
1730 calculate the probability directly. */
1731 if (tree_fits_shwi_p (loop_bound_var
)
1732 && tree_fits_shwi_p (compare_var
)
1733 && tree_fits_shwi_p (compare_base
))
1736 wi::overflow_type overflow
;
1737 bool overall_overflow
= false;
1738 widest_int compare_count
, tem
;
1740 /* (loop_bound - base) / compare_step */
1741 tem
= wi::sub (wi::to_widest (loop_bound_var
),
1742 wi::to_widest (compare_base
), SIGNED
, &overflow
);
1743 overall_overflow
|= overflow
;
1744 widest_int loop_count
= wi::div_trunc (tem
,
1745 wi::to_widest (compare_step_var
),
1747 overall_overflow
|= overflow
;
1749 if (!wi::neg_p (wi::to_widest (compare_step_var
))
1750 ^ (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1752 /* (loop_bound - compare_bound) / compare_step */
1753 tem
= wi::sub (wi::to_widest (loop_bound_var
),
1754 wi::to_widest (compare_var
), SIGNED
, &overflow
);
1755 overall_overflow
|= overflow
;
1756 compare_count
= wi::div_trunc (tem
, wi::to_widest (compare_step_var
),
1758 overall_overflow
|= overflow
;
1762 /* (compare_bound - base) / compare_step */
1763 tem
= wi::sub (wi::to_widest (compare_var
),
1764 wi::to_widest (compare_base
), SIGNED
, &overflow
);
1765 overall_overflow
|= overflow
;
1766 compare_count
= wi::div_trunc (tem
, wi::to_widest (compare_step_var
),
1768 overall_overflow
|= overflow
;
1770 if (compare_code
== LE_EXPR
|| compare_code
== GE_EXPR
)
1772 if (loop_bound_code
== LE_EXPR
|| loop_bound_code
== GE_EXPR
)
1774 if (wi::neg_p (compare_count
))
1776 if (wi::neg_p (loop_count
))
1778 if (loop_count
== 0)
1780 else if (wi::cmps (compare_count
, loop_count
) == 1)
1781 probability
= REG_BR_PROB_BASE
;
1784 tem
= compare_count
* REG_BR_PROB_BASE
;
1785 tem
= wi::udiv_trunc (tem
, loop_count
);
1786 probability
= tem
.to_uhwi ();
1789 /* FIXME: The branch prediction seems broken. It has only 20% hitrate. */
1790 if (!overall_overflow
)
1791 predict_edge (then_edge
, PRED_LOOP_IV_COMPARE
, probability
);
1796 if (expr_coherent_p (loop_bound_var
, compare_var
))
1798 if ((loop_bound_code
== LT_EXPR
|| loop_bound_code
== LE_EXPR
)
1799 && (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1800 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1801 else if ((loop_bound_code
== GT_EXPR
|| loop_bound_code
== GE_EXPR
)
1802 && (compare_code
== GT_EXPR
|| compare_code
== GE_EXPR
))
1803 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1804 else if (loop_bound_code
== NE_EXPR
)
1806 /* If the loop backedge condition is "(i != bound)", we do
1807 the comparison based on the step of IV:
1808 * step < 0 : backedge condition is like (i > bound)
1809 * step > 0 : backedge condition is like (i < bound) */
1810 gcc_assert (loop_bound_step
!= 0);
1811 if (loop_bound_step
> 0
1812 && (compare_code
== LT_EXPR
1813 || compare_code
== LE_EXPR
))
1814 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1815 else if (loop_bound_step
< 0
1816 && (compare_code
== GT_EXPR
1817 || compare_code
== GE_EXPR
))
1818 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1820 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1823 /* The branch is predicted not-taken if loop_bound_code is
1824 opposite with compare_code. */
1825 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1827 else if (expr_coherent_p (loop_iv_base_var
, compare_var
))
1830 for (i = s; i < h; i++)
1832 The branch should be predicted taken. */
1833 if (loop_bound_step
> 0
1834 && (compare_code
== GT_EXPR
|| compare_code
== GE_EXPR
))
1835 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1836 else if (loop_bound_step
< 0
1837 && (compare_code
== LT_EXPR
|| compare_code
== LE_EXPR
))
1838 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, TAKEN
);
1840 predict_edge_def (then_edge
, PRED_LOOP_IV_COMPARE_GUESS
, NOT_TAKEN
);
1844 /* Predict for extra loop exits that will lead to EXIT_EDGE. The extra loop
1845 exits are resulted from short-circuit conditions that will generate an
1848 if (foo() || global > 10)
1851 This will be translated into:
1856 if foo() goto BB6 else goto BB5
1858 if global > 10 goto BB6 else goto BB7
1862 iftmp = (PHI 0(BB5), 1(BB6))
1863 if iftmp == 1 goto BB8 else goto BB3
1865 outside of the loop...
1867 The edge BB7->BB8 is loop exit because BB8 is outside of the loop.
1868 From the dataflow, we can infer that BB4->BB6 and BB5->BB6 are also loop
1869 exits. This function takes BB7->BB8 as input, and finds out the extra loop
1870 exits to predict them using PRED_LOOP_EXTRA_EXIT. */
1873 predict_extra_loop_exits (class loop
*loop
, edge exit_edge
)
1876 bool check_value_one
;
1877 gimple
*lhs_def_stmt
;
1879 tree cmp_rhs
, cmp_lhs
;
1883 last
= last_stmt (exit_edge
->src
);
1886 cmp_stmt
= dyn_cast
<gcond
*> (last
);
1890 cmp_rhs
= gimple_cond_rhs (cmp_stmt
);
1891 cmp_lhs
= gimple_cond_lhs (cmp_stmt
);
1892 if (!TREE_CONSTANT (cmp_rhs
)
1893 || !(integer_zerop (cmp_rhs
) || integer_onep (cmp_rhs
)))
1895 if (TREE_CODE (cmp_lhs
) != SSA_NAME
)
1898 /* If check_value_one is true, only the phi_args with value '1' will lead
1899 to loop exit. Otherwise, only the phi_args with value '0' will lead to
1901 check_value_one
= (((integer_onep (cmp_rhs
))
1902 ^ (gimple_cond_code (cmp_stmt
) == EQ_EXPR
))
1903 ^ ((exit_edge
->flags
& EDGE_TRUE_VALUE
) != 0));
1905 lhs_def_stmt
= SSA_NAME_DEF_STMT (cmp_lhs
);
1909 phi_stmt
= dyn_cast
<gphi
*> (lhs_def_stmt
);
1913 for (i
= 0; i
< gimple_phi_num_args (phi_stmt
); i
++)
1917 tree val
= gimple_phi_arg_def (phi_stmt
, i
);
1918 edge e
= gimple_phi_arg_edge (phi_stmt
, i
);
1920 if (!TREE_CONSTANT (val
) || !(integer_zerop (val
) || integer_onep (val
)))
1922 if ((check_value_one
^ integer_onep (val
)) == 1)
1924 if (EDGE_COUNT (e
->src
->succs
) != 1)
1926 predict_paths_leading_to_edge (e
, PRED_LOOP_EXTRA_EXIT
, NOT_TAKEN
,
1931 FOR_EACH_EDGE (e1
, ei
, e
->src
->preds
)
1932 predict_paths_leading_to_edge (e1
, PRED_LOOP_EXTRA_EXIT
, NOT_TAKEN
,
1938 /* Predict edge probabilities by exploiting loop structure. */
1941 predict_loops (void)
1944 hash_set
<class loop
*> with_recursion(10);
1946 FOR_EACH_BB_FN (bb
, cfun
)
1948 gimple_stmt_iterator gsi
;
1951 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1952 if (is_gimple_call (gsi_stmt (gsi
))
1953 && (decl
= gimple_call_fndecl (gsi_stmt (gsi
))) != NULL
1954 && recursive_call_p (current_function_decl
, decl
))
1956 class loop
*loop
= bb
->loop_father
;
1957 while (loop
&& !with_recursion
.add (loop
))
1958 loop
= loop_outer (loop
);
1962 /* Try to predict out blocks in a loop that are not part of a
1964 for (auto loop
: loops_list (cfun
, LI_FROM_INNERMOST
))
1966 basic_block bb
, *bbs
;
1967 unsigned j
, n_exits
= 0;
1968 class tree_niter_desc niter_desc
;
1970 class nb_iter_bound
*nb_iter
;
1971 enum tree_code loop_bound_code
= ERROR_MARK
;
1972 tree loop_bound_step
= NULL
;
1973 tree loop_bound_var
= NULL
;
1974 tree loop_iv_base
= NULL
;
1976 bool recursion
= with_recursion
.contains (loop
);
1978 auto_vec
<edge
> exits
= get_loop_exit_edges (loop
);
1979 FOR_EACH_VEC_ELT (exits
, j
, ex
)
1980 if (!unlikely_executed_edge_p (ex
) && !(ex
->flags
& EDGE_ABNORMAL_CALL
))
1985 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1986 fprintf (dump_file
, "Predicting loop %i%s with %i exits.\n",
1987 loop
->num
, recursion
? " (with recursion)":"", n_exits
);
1988 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
1989 && max_loop_iterations_int (loop
) >= 0)
1992 "Loop %d iterates at most %i times.\n", loop
->num
,
1993 (int)max_loop_iterations_int (loop
));
1995 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
1996 && likely_max_loop_iterations_int (loop
) >= 0)
1998 fprintf (dump_file
, "Loop %d likely iterates at most %i times.\n",
1999 loop
->num
, (int)likely_max_loop_iterations_int (loop
));
2002 FOR_EACH_VEC_ELT (exits
, j
, ex
)
2005 HOST_WIDE_INT nitercst
;
2006 int max
= param_max_predicted_iterations
;
2008 enum br_predictor predictor
;
2011 if (unlikely_executed_edge_p (ex
)
2012 || (ex
->flags
& EDGE_ABNORMAL_CALL
))
2014 /* Loop heuristics do not expect exit conditional to be inside
2015 inner loop. We predict from innermost to outermost loop. */
2016 if (predicted_by_loop_heuristics_p (ex
->src
))
2018 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2019 fprintf (dump_file
, "Skipping exit %i->%i because "
2020 "it is already predicted.\n",
2021 ex
->src
->index
, ex
->dest
->index
);
2024 predict_extra_loop_exits (loop
, ex
);
2026 if (number_of_iterations_exit (loop
, ex
, &niter_desc
, false, false))
2027 niter
= niter_desc
.niter
;
2028 if (!niter
|| TREE_CODE (niter_desc
.niter
) != INTEGER_CST
)
2029 niter
= loop_niter_by_eval (loop
, ex
);
2030 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
2031 && TREE_CODE (niter
) == INTEGER_CST
)
2033 fprintf (dump_file
, "Exit %i->%i %d iterates ",
2034 ex
->src
->index
, ex
->dest
->index
,
2036 print_generic_expr (dump_file
, niter
, TDF_SLIM
);
2037 fprintf (dump_file
, " times.\n");
2040 if (TREE_CODE (niter
) == INTEGER_CST
)
2042 if (tree_fits_uhwi_p (niter
)
2044 && compare_tree_int (niter
, max
- 1) == -1)
2045 nitercst
= tree_to_uhwi (niter
) + 1;
2048 predictor
= PRED_LOOP_ITERATIONS
;
2050 /* If we have just one exit and we can derive some information about
2051 the number of iterations of the loop from the statements inside
2052 the loop, use it to predict this exit. */
2053 else if (n_exits
== 1
2054 && estimated_stmt_executions (loop
, &nit
))
2056 if (wi::gtu_p (nit
, max
))
2059 nitercst
= nit
.to_shwi ();
2060 predictor
= PRED_LOOP_ITERATIONS_GUESSED
;
2062 /* If we have likely upper bound, trust it for very small iteration
2063 counts. Such loops would otherwise get mispredicted by standard
2064 LOOP_EXIT heuristics. */
2065 else if (n_exits
== 1
2066 && likely_max_stmt_executions (loop
, &nit
)
2068 RDIV (REG_BR_PROB_BASE
,
2072 ? PRED_LOOP_EXIT_WITH_RECURSION
2073 : PRED_LOOP_EXIT
].hitrate
)))
2075 nitercst
= nit
.to_shwi ();
2076 predictor
= PRED_LOOP_ITERATIONS_MAX
;
2080 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2081 fprintf (dump_file
, "Nothing known about exit %i->%i.\n",
2082 ex
->src
->index
, ex
->dest
->index
);
2086 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2087 fprintf (dump_file
, "Recording prediction to %i iterations by %s.\n",
2088 (int)nitercst
, predictor_info
[predictor
].name
);
2089 /* If the prediction for number of iterations is zero, do not
2090 predict the exit edges. */
2094 probability
= RDIV (REG_BR_PROB_BASE
, nitercst
);
2095 predict_edge (ex
, predictor
, probability
);
2098 /* Find information about loop bound variables. */
2099 for (nb_iter
= loop
->bounds
; nb_iter
;
2100 nb_iter
= nb_iter
->next
)
2102 && gimple_code (nb_iter
->stmt
) == GIMPLE_COND
)
2104 stmt
= as_a
<gcond
*> (nb_iter
->stmt
);
2107 if (!stmt
&& last_stmt (loop
->header
)
2108 && gimple_code (last_stmt (loop
->header
)) == GIMPLE_COND
)
2109 stmt
= as_a
<gcond
*> (last_stmt (loop
->header
));
2111 is_comparison_with_loop_invariant_p (stmt
, loop
,
2117 bbs
= get_loop_body (loop
);
2119 for (j
= 0; j
< loop
->num_nodes
; j
++)
2126 /* Bypass loop heuristics on continue statement. These
2127 statements construct loops via "non-loop" constructs
2128 in the source language and are better to be handled
2130 if (predicted_by_p (bb
, PRED_CONTINUE
))
2132 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2133 fprintf (dump_file
, "BB %i predicted by continue.\n",
2138 /* If we already used more reliable loop exit predictors, do not
2139 bother with PRED_LOOP_EXIT. */
2140 if (!predicted_by_loop_heuristics_p (bb
))
2142 /* For loop with many exits we don't want to predict all exits
2143 with the pretty large probability, because if all exits are
2144 considered in row, the loop would be predicted to iterate
2145 almost never. The code to divide probability by number of
2146 exits is very rough. It should compute the number of exits
2147 taken in each patch through function (not the overall number
2148 of exits that might be a lot higher for loops with wide switch
2149 statements in them) and compute n-th square root.
2151 We limit the minimal probability by 2% to avoid
2152 EDGE_PROBABILITY_RELIABLE from trusting the branch prediction
2153 as this was causing regression in perl benchmark containing such
2156 int probability
= ((REG_BR_PROB_BASE
2159 ? PRED_LOOP_EXIT_WITH_RECURSION
2160 : PRED_LOOP_EXIT
].hitrate
)
2162 if (probability
< HITRATE (2))
2163 probability
= HITRATE (2);
2164 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2165 if (e
->dest
->index
< NUM_FIXED_BLOCKS
2166 || !flow_bb_inside_loop_p (loop
, e
->dest
))
2168 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2170 "Predicting exit %i->%i with prob %i.\n",
2171 e
->src
->index
, e
->dest
->index
, probability
);
2173 recursion
? PRED_LOOP_EXIT_WITH_RECURSION
2174 : PRED_LOOP_EXIT
, probability
);
2178 predict_iv_comparison (loop
, bb
, loop_bound_var
, loop_iv_base
,
2180 tree_to_shwi (loop_bound_step
));
2183 /* In the following code
2188 guess that cond is unlikely. */
2189 if (loop_outer (loop
)->num
)
2191 basic_block bb
= NULL
;
2192 edge preheader_edge
= loop_preheader_edge (loop
);
2194 if (single_pred_p (preheader_edge
->src
)
2195 && single_succ_p (preheader_edge
->src
))
2196 preheader_edge
= single_pred_edge (preheader_edge
->src
);
2198 gimple
*stmt
= last_stmt (preheader_edge
->src
);
2199 /* Pattern match fortran loop preheader:
2200 _16 = BUILTIN_EXPECT (_15, 1, PRED_FORTRAN_LOOP_PREHEADER);
2201 _17 = (logical(kind=4)) _16;
2207 Loop guard branch prediction says nothing about duplicated loop
2208 headers produced by fortran frontend and in this case we want
2209 to predict paths leading to this preheader. */
2212 && gimple_code (stmt
) == GIMPLE_COND
2213 && gimple_cond_code (stmt
) == NE_EXPR
2214 && TREE_CODE (gimple_cond_lhs (stmt
)) == SSA_NAME
2215 && integer_zerop (gimple_cond_rhs (stmt
)))
2217 gimple
*call_stmt
= SSA_NAME_DEF_STMT (gimple_cond_lhs (stmt
));
2218 if (gimple_code (call_stmt
) == GIMPLE_ASSIGN
2219 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (call_stmt
))
2220 && TREE_CODE (gimple_assign_rhs1 (call_stmt
)) == SSA_NAME
)
2221 call_stmt
= SSA_NAME_DEF_STMT (gimple_assign_rhs1 (call_stmt
));
2222 if (gimple_call_internal_p (call_stmt
, IFN_BUILTIN_EXPECT
)
2223 && TREE_CODE (gimple_call_arg (call_stmt
, 2)) == INTEGER_CST
2224 && tree_fits_uhwi_p (gimple_call_arg (call_stmt
, 2))
2225 && tree_to_uhwi (gimple_call_arg (call_stmt
, 2))
2226 == PRED_FORTRAN_LOOP_PREHEADER
)
2227 bb
= preheader_edge
->src
;
2231 if (!dominated_by_p (CDI_DOMINATORS
,
2232 loop_outer (loop
)->latch
, loop
->header
))
2233 predict_paths_leading_to_edge (loop_preheader_edge (loop
),
2235 ? PRED_LOOP_GUARD_WITH_RECURSION
2242 if (!dominated_by_p (CDI_DOMINATORS
,
2243 loop_outer (loop
)->latch
, bb
))
2244 predict_paths_leading_to (bb
,
2246 ? PRED_LOOP_GUARD_WITH_RECURSION
2253 /* Free basic blocks from get_loop_body. */
2258 /* Attempt to predict probabilities of BB outgoing edges using local
2261 bb_estimate_probability_locally (basic_block bb
)
2263 rtx_insn
*last_insn
= BB_END (bb
);
2266 if (! can_predict_insn_p (last_insn
))
2268 cond
= get_condition (last_insn
, NULL
, false, false);
2272 /* Try "pointer heuristic."
2273 A comparison ptr == 0 is predicted as false.
2274 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
2275 if (COMPARISON_P (cond
)
2276 && ((REG_P (XEXP (cond
, 0)) && REG_POINTER (XEXP (cond
, 0)))
2277 || (REG_P (XEXP (cond
, 1)) && REG_POINTER (XEXP (cond
, 1)))))
2279 if (GET_CODE (cond
) == EQ
)
2280 predict_insn_def (last_insn
, PRED_POINTER
, NOT_TAKEN
);
2281 else if (GET_CODE (cond
) == NE
)
2282 predict_insn_def (last_insn
, PRED_POINTER
, TAKEN
);
2286 /* Try "opcode heuristic."
2287 EQ tests are usually false and NE tests are usually true. Also,
2288 most quantities are positive, so we can make the appropriate guesses
2289 about signed comparisons against zero. */
2290 switch (GET_CODE (cond
))
2293 /* Unconditional branch. */
2294 predict_insn_def (last_insn
, PRED_UNCONDITIONAL
,
2295 cond
== const0_rtx
? NOT_TAKEN
: TAKEN
);
2300 /* Floating point comparisons appears to behave in a very
2301 unpredictable way because of special role of = tests in
2303 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
2305 /* Comparisons with 0 are often used for booleans and there is
2306 nothing useful to predict about them. */
2307 else if (XEXP (cond
, 1) == const0_rtx
2308 || XEXP (cond
, 0) == const0_rtx
)
2311 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, NOT_TAKEN
);
2316 /* Floating point comparisons appears to behave in a very
2317 unpredictable way because of special role of = tests in
2319 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
2321 /* Comparisons with 0 are often used for booleans and there is
2322 nothing useful to predict about them. */
2323 else if (XEXP (cond
, 1) == const0_rtx
2324 || XEXP (cond
, 0) == const0_rtx
)
2327 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, TAKEN
);
2331 predict_insn_def (last_insn
, PRED_FPOPCODE
, TAKEN
);
2335 predict_insn_def (last_insn
, PRED_FPOPCODE
, NOT_TAKEN
);
2340 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
2341 || XEXP (cond
, 1) == constm1_rtx
)
2342 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, NOT_TAKEN
);
2347 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
2348 || XEXP (cond
, 1) == constm1_rtx
)
2349 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, TAKEN
);
2357 /* Set edge->probability for each successor edge of BB. */
2359 guess_outgoing_edge_probabilities (basic_block bb
)
2361 bb_estimate_probability_locally (bb
);
2362 combine_predictions_for_insn (BB_END (bb
), bb
);
2365 static tree
expr_expected_value (tree
, bitmap
, enum br_predictor
*predictor
,
2366 HOST_WIDE_INT
*probability
);
2368 /* Helper function for expr_expected_value. */
2371 expr_expected_value_1 (tree type
, tree op0
, enum tree_code code
,
2372 tree op1
, bitmap visited
, enum br_predictor
*predictor
,
2373 HOST_WIDE_INT
*probability
)
2377 /* Reset returned probability value. */
2379 *predictor
= PRED_UNCONDITIONAL
;
2381 if (get_gimple_rhs_class (code
) == GIMPLE_SINGLE_RHS
)
2383 if (TREE_CONSTANT (op0
))
2386 if (code
== IMAGPART_EXPR
)
2388 if (TREE_CODE (TREE_OPERAND (op0
, 0)) == SSA_NAME
)
2390 def
= SSA_NAME_DEF_STMT (TREE_OPERAND (op0
, 0));
2391 if (is_gimple_call (def
)
2392 && gimple_call_internal_p (def
)
2393 && (gimple_call_internal_fn (def
)
2394 == IFN_ATOMIC_COMPARE_EXCHANGE
))
2396 /* Assume that any given atomic operation has low contention,
2397 and thus the compare-and-swap operation succeeds. */
2398 *predictor
= PRED_COMPARE_AND_SWAP
;
2399 return build_one_cst (TREE_TYPE (op0
));
2404 if (code
!= SSA_NAME
)
2407 def
= SSA_NAME_DEF_STMT (op0
);
2409 /* If we were already here, break the infinite cycle. */
2410 if (!bitmap_set_bit (visited
, SSA_NAME_VERSION (op0
)))
2413 if (gimple_code (def
) == GIMPLE_PHI
)
2415 /* All the arguments of the PHI node must have the same constant
2417 int i
, n
= gimple_phi_num_args (def
);
2418 tree val
= NULL
, new_val
;
2420 for (i
= 0; i
< n
; i
++)
2422 tree arg
= PHI_ARG_DEF (def
, i
);
2423 enum br_predictor predictor2
;
2425 /* If this PHI has itself as an argument, we cannot
2426 determine the string length of this argument. However,
2427 if we can find an expected constant value for the other
2428 PHI args then we can still be sure that this is
2429 likely a constant. So be optimistic and just
2430 continue with the next argument. */
2431 if (arg
== PHI_RESULT (def
))
2434 HOST_WIDE_INT probability2
;
2435 new_val
= expr_expected_value (arg
, visited
, &predictor2
,
2438 /* It is difficult to combine value predictors. Simply assume
2439 that later predictor is weaker and take its prediction. */
2440 if (*predictor
< predictor2
)
2442 *predictor
= predictor2
;
2443 *probability
= probability2
;
2449 else if (!operand_equal_p (val
, new_val
, false))
2454 if (is_gimple_assign (def
))
2456 if (gimple_assign_lhs (def
) != op0
)
2459 return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def
)),
2460 gimple_assign_rhs1 (def
),
2461 gimple_assign_rhs_code (def
),
2462 gimple_assign_rhs2 (def
),
2463 visited
, predictor
, probability
);
2466 if (is_gimple_call (def
))
2468 tree decl
= gimple_call_fndecl (def
);
2471 if (gimple_call_internal_p (def
)
2472 && gimple_call_internal_fn (def
) == IFN_BUILTIN_EXPECT
)
2474 gcc_assert (gimple_call_num_args (def
) == 3);
2475 tree val
= gimple_call_arg (def
, 0);
2476 if (TREE_CONSTANT (val
))
2478 tree val2
= gimple_call_arg (def
, 2);
2479 gcc_assert (TREE_CODE (val2
) == INTEGER_CST
2480 && tree_fits_uhwi_p (val2
)
2481 && tree_to_uhwi (val2
) < END_PREDICTORS
);
2482 *predictor
= (enum br_predictor
) tree_to_uhwi (val2
);
2483 if (*predictor
== PRED_BUILTIN_EXPECT
)
2485 = HITRATE (param_builtin_expect_probability
);
2486 return gimple_call_arg (def
, 1);
2491 if (DECL_IS_MALLOC (decl
) || DECL_IS_OPERATOR_NEW_P (decl
))
2494 *predictor
= PRED_MALLOC_NONNULL
;
2495 return boolean_true_node
;
2498 if (DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_NORMAL
)
2499 switch (DECL_FUNCTION_CODE (decl
))
2501 case BUILT_IN_EXPECT
:
2504 if (gimple_call_num_args (def
) != 2)
2506 val
= gimple_call_arg (def
, 0);
2507 if (TREE_CONSTANT (val
))
2509 *predictor
= PRED_BUILTIN_EXPECT
;
2511 = HITRATE (param_builtin_expect_probability
);
2512 return gimple_call_arg (def
, 1);
2514 case BUILT_IN_EXPECT_WITH_PROBABILITY
:
2517 if (gimple_call_num_args (def
) != 3)
2519 val
= gimple_call_arg (def
, 0);
2520 if (TREE_CONSTANT (val
))
2522 /* Compute final probability as:
2523 probability * REG_BR_PROB_BASE. */
2524 tree prob
= gimple_call_arg (def
, 2);
2525 tree t
= TREE_TYPE (prob
);
2526 tree base
= build_int_cst (integer_type_node
,
2528 base
= build_real_from_int_cst (t
, base
);
2529 tree r
= fold_build2_initializer_loc (UNKNOWN_LOCATION
,
2530 MULT_EXPR
, t
, prob
, base
);
2531 if (TREE_CODE (r
) != REAL_CST
)
2533 error_at (gimple_location (def
),
2534 "probability %qE must be "
2535 "constant floating-point expression", prob
);
2539 = real_to_integer (TREE_REAL_CST_PTR (r
));
2540 if (probi
>= 0 && probi
<= REG_BR_PROB_BASE
)
2542 *predictor
= PRED_BUILTIN_EXPECT_WITH_PROBABILITY
;
2543 *probability
= probi
;
2546 error_at (gimple_location (def
),
2547 "probability %qE is outside "
2548 "the range [0.0, 1.0]", prob
);
2550 return gimple_call_arg (def
, 1);
2553 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N
:
2554 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1
:
2555 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2
:
2556 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4
:
2557 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8
:
2558 case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16
:
2559 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE
:
2560 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N
:
2561 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1
:
2562 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2
:
2563 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4
:
2564 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8
:
2565 case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16
:
2566 /* Assume that any given atomic operation has low contention,
2567 and thus the compare-and-swap operation succeeds. */
2568 *predictor
= PRED_COMPARE_AND_SWAP
;
2569 return boolean_true_node
;
2570 case BUILT_IN_REALLOC
:
2572 *predictor
= PRED_MALLOC_NONNULL
;
2573 return boolean_true_node
;
2582 if (get_gimple_rhs_class (code
) == GIMPLE_BINARY_RHS
)
2585 enum br_predictor predictor2
;
2586 HOST_WIDE_INT probability2
;
2587 op0
= expr_expected_value (op0
, visited
, predictor
, probability
);
2590 op1
= expr_expected_value (op1
, visited
, &predictor2
, &probability2
);
2593 res
= fold_build2 (code
, type
, op0
, op1
);
2594 if (TREE_CODE (res
) == INTEGER_CST
2595 && TREE_CODE (op0
) == INTEGER_CST
2596 && TREE_CODE (op1
) == INTEGER_CST
)
2598 /* Combine binary predictions. */
2599 if (*probability
!= -1 || probability2
!= -1)
2601 HOST_WIDE_INT p1
= get_predictor_value (*predictor
, *probability
);
2602 HOST_WIDE_INT p2
= get_predictor_value (predictor2
, probability2
);
2603 *probability
= RDIV (p1
* p2
, REG_BR_PROB_BASE
);
2606 if (*predictor
< predictor2
)
2607 *predictor
= predictor2
;
2613 if (get_gimple_rhs_class (code
) == GIMPLE_UNARY_RHS
)
2616 op0
= expr_expected_value (op0
, visited
, predictor
, probability
);
2619 res
= fold_build1 (code
, type
, op0
);
2620 if (TREE_CONSTANT (res
))
2627 /* Return constant EXPR will likely have at execution time, NULL if unknown.
2628 The function is used by builtin_expect branch predictor so the evidence
2629 must come from this construct and additional possible constant folding.
2631 We may want to implement more involved value guess (such as value range
2632 propagation based prediction), but such tricks shall go to new
2636 expr_expected_value (tree expr
, bitmap visited
,
2637 enum br_predictor
*predictor
,
2638 HOST_WIDE_INT
*probability
)
2640 enum tree_code code
;
2643 if (TREE_CONSTANT (expr
))
2645 *predictor
= PRED_UNCONDITIONAL
;
2650 extract_ops_from_tree (expr
, &code
, &op0
, &op1
);
2651 return expr_expected_value_1 (TREE_TYPE (expr
),
2652 op0
, code
, op1
, visited
, predictor
,
2657 /* Return probability of a PREDICTOR. If the predictor has variable
2658 probability return passed PROBABILITY. */
2660 static HOST_WIDE_INT
2661 get_predictor_value (br_predictor predictor
, HOST_WIDE_INT probability
)
2665 case PRED_BUILTIN_EXPECT
:
2666 case PRED_BUILTIN_EXPECT_WITH_PROBABILITY
:
2667 gcc_assert (probability
!= -1);
2670 gcc_assert (probability
== -1);
2671 return predictor_info
[(int) predictor
].hitrate
;
2675 /* Predict using opcode of the last statement in basic block. */
2677 tree_predict_by_opcode (basic_block bb
)
2679 gimple
*stmt
= last_stmt (bb
);
2686 enum br_predictor predictor
;
2687 HOST_WIDE_INT probability
;
2692 if (gswitch
*sw
= dyn_cast
<gswitch
*> (stmt
))
2694 tree index
= gimple_switch_index (sw
);
2695 tree val
= expr_expected_value (index
, auto_bitmap (),
2696 &predictor
, &probability
);
2697 if (val
&& TREE_CODE (val
) == INTEGER_CST
)
2699 edge e
= find_taken_edge_switch_expr (sw
, val
);
2700 if (predictor
== PRED_BUILTIN_EXPECT
)
2702 int percent
= param_builtin_expect_probability
;
2703 gcc_assert (percent
>= 0 && percent
<= 100);
2704 predict_edge (e
, PRED_BUILTIN_EXPECT
,
2708 predict_edge_def (e
, predictor
, TAKEN
);
2712 if (gimple_code (stmt
) != GIMPLE_COND
)
2714 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
2715 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
2717 op0
= gimple_cond_lhs (stmt
);
2718 op1
= gimple_cond_rhs (stmt
);
2719 cmp
= gimple_cond_code (stmt
);
2720 type
= TREE_TYPE (op0
);
2721 val
= expr_expected_value_1 (boolean_type_node
, op0
, cmp
, op1
, auto_bitmap (),
2722 &predictor
, &probability
);
2723 if (val
&& TREE_CODE (val
) == INTEGER_CST
)
2725 HOST_WIDE_INT prob
= get_predictor_value (predictor
, probability
);
2726 if (integer_zerop (val
))
2727 prob
= REG_BR_PROB_BASE
- prob
;
2728 predict_edge (then_edge
, predictor
, prob
);
2730 /* Try "pointer heuristic."
2731 A comparison ptr == 0 is predicted as false.
2732 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
2733 if (POINTER_TYPE_P (type
))
2736 predict_edge_def (then_edge
, PRED_TREE_POINTER
, NOT_TAKEN
);
2737 else if (cmp
== NE_EXPR
)
2738 predict_edge_def (then_edge
, PRED_TREE_POINTER
, TAKEN
);
2742 /* Try "opcode heuristic."
2743 EQ tests are usually false and NE tests are usually true. Also,
2744 most quantities are positive, so we can make the appropriate guesses
2745 about signed comparisons against zero. */
2750 /* Floating point comparisons appears to behave in a very
2751 unpredictable way because of special role of = tests in
2753 if (FLOAT_TYPE_P (type
))
2755 /* Comparisons with 0 are often used for booleans and there is
2756 nothing useful to predict about them. */
2757 else if (integer_zerop (op0
) || integer_zerop (op1
))
2760 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, NOT_TAKEN
);
2765 /* Floating point comparisons appears to behave in a very
2766 unpredictable way because of special role of = tests in
2768 if (FLOAT_TYPE_P (type
))
2770 /* Comparisons with 0 are often used for booleans and there is
2771 nothing useful to predict about them. */
2772 else if (integer_zerop (op0
)
2773 || integer_zerop (op1
))
2776 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, TAKEN
);
2780 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, TAKEN
);
2783 case UNORDERED_EXPR
:
2784 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, NOT_TAKEN
);
2789 if (integer_zerop (op1
)
2790 || integer_onep (op1
)
2791 || integer_all_onesp (op1
)
2794 || real_minus_onep (op1
))
2795 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, NOT_TAKEN
);
2800 if (integer_zerop (op1
)
2801 || integer_onep (op1
)
2802 || integer_all_onesp (op1
)
2805 || real_minus_onep (op1
))
2806 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, TAKEN
);
2814 /* Returns TRUE if the STMT is exit(0) like statement. */
2817 is_exit_with_zero_arg (const gimple
*stmt
)
2819 /* This is not exit, _exit or _Exit. */
2820 if (!gimple_call_builtin_p (stmt
, BUILT_IN_EXIT
)
2821 && !gimple_call_builtin_p (stmt
, BUILT_IN__EXIT
)
2822 && !gimple_call_builtin_p (stmt
, BUILT_IN__EXIT2
))
2825 /* Argument is an interger zero. */
2826 return integer_zerop (gimple_call_arg (stmt
, 0));
2829 /* Try to guess whether the value of return means error code. */
2831 static enum br_predictor
2832 return_prediction (tree val
, enum prediction
*prediction
)
2836 return PRED_NO_PREDICTION
;
2837 /* Different heuristics for pointers and scalars. */
2838 if (POINTER_TYPE_P (TREE_TYPE (val
)))
2840 /* NULL is usually not returned. */
2841 if (integer_zerop (val
))
2843 *prediction
= NOT_TAKEN
;
2844 return PRED_NULL_RETURN
;
2847 else if (INTEGRAL_TYPE_P (TREE_TYPE (val
)))
2849 /* Negative return values are often used to indicate
2851 if (TREE_CODE (val
) == INTEGER_CST
2852 && tree_int_cst_sgn (val
) < 0)
2854 *prediction
= NOT_TAKEN
;
2855 return PRED_NEGATIVE_RETURN
;
2857 /* Constant return values seems to be commonly taken.
2858 Zero/one often represent booleans so exclude them from the
2860 if (TREE_CONSTANT (val
)
2861 && (!integer_zerop (val
) && !integer_onep (val
)))
2863 *prediction
= NOT_TAKEN
;
2864 return PRED_CONST_RETURN
;
2867 return PRED_NO_PREDICTION
;
2870 /* Return zero if phi result could have values other than -1, 0 or 1,
2871 otherwise return a bitmask, with bits 0, 1 and 2 set if -1, 0 and 1
2872 values are used or likely. */
2875 zero_one_minusone (gphi
*phi
, int limit
)
2877 int phi_num_args
= gimple_phi_num_args (phi
);
2879 for (int i
= 0; i
< phi_num_args
; i
++)
2881 tree t
= PHI_ARG_DEF (phi
, i
);
2882 if (TREE_CODE (t
) != INTEGER_CST
)
2884 wide_int w
= wi::to_wide (t
);
2894 for (int i
= 0; i
< phi_num_args
; i
++)
2896 tree t
= PHI_ARG_DEF (phi
, i
);
2897 if (TREE_CODE (t
) == INTEGER_CST
)
2899 if (TREE_CODE (t
) != SSA_NAME
)
2901 gimple
*g
= SSA_NAME_DEF_STMT (t
);
2902 if (gimple_code (g
) == GIMPLE_PHI
&& limit
> 0)
2903 if (int r
= zero_one_minusone (as_a
<gphi
*> (g
), limit
- 1))
2908 if (!is_gimple_assign (g
))
2910 if (gimple_assign_cast_p (g
))
2912 tree rhs1
= gimple_assign_rhs1 (g
);
2913 if (TREE_CODE (rhs1
) != SSA_NAME
2914 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1
))
2915 || TYPE_PRECISION (TREE_TYPE (rhs1
)) != 1
2916 || !TYPE_UNSIGNED (TREE_TYPE (rhs1
)))
2921 if (TREE_CODE_CLASS (gimple_assign_rhs_code (g
)) != tcc_comparison
)
2928 /* Find the basic block with return expression and look up for possible
2929 return value trying to apply RETURN_PREDICTION heuristics. */
2931 apply_return_prediction (void)
2933 greturn
*return_stmt
= NULL
;
2937 int phi_num_args
, i
;
2938 enum br_predictor pred
;
2939 enum prediction direction
;
2942 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
2944 gimple
*last
= last_stmt (e
->src
);
2946 && gimple_code (last
) == GIMPLE_RETURN
)
2948 return_stmt
= as_a
<greturn
*> (last
);
2954 return_val
= gimple_return_retval (return_stmt
);
2957 if (TREE_CODE (return_val
) != SSA_NAME
2958 || !SSA_NAME_DEF_STMT (return_val
)
2959 || gimple_code (SSA_NAME_DEF_STMT (return_val
)) != GIMPLE_PHI
)
2961 phi
= as_a
<gphi
*> (SSA_NAME_DEF_STMT (return_val
));
2962 phi_num_args
= gimple_phi_num_args (phi
);
2963 pred
= return_prediction (PHI_ARG_DEF (phi
, 0), &direction
);
2965 /* Avoid the case where the function returns -1, 0 and 1 values and
2966 nothing else. Those could be qsort etc. comparison functions
2967 where the negative return isn't less probable than positive.
2968 For this require that the function returns at least -1 or 1
2969 or -1 and a boolean value or comparison result, so that functions
2970 returning just -1 and 0 are treated as if -1 represents error value. */
2971 if (INTEGRAL_TYPE_P (TREE_TYPE (return_val
))
2972 && !TYPE_UNSIGNED (TREE_TYPE (return_val
))
2973 && TYPE_PRECISION (TREE_TYPE (return_val
)) > 1)
2974 if (int r
= zero_one_minusone (phi
, 3))
2975 if ((r
& (1 | 4)) == (1 | 4))
2978 /* Avoid the degenerate case where all return values form the function
2979 belongs to same category (ie they are all positive constants)
2980 so we can hardly say something about them. */
2981 for (i
= 1; i
< phi_num_args
; i
++)
2982 if (pred
!= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
))
2984 if (i
!= phi_num_args
)
2985 for (i
= 0; i
< phi_num_args
; i
++)
2987 pred
= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
);
2988 if (pred
!= PRED_NO_PREDICTION
)
2989 predict_paths_leading_to_edge (gimple_phi_arg_edge (phi
, i
), pred
,
2994 /* Look for basic block that contains unlikely to happen events
2995 (such as noreturn calls) and mark all paths leading to execution
2996 of this basic blocks as unlikely. */
2999 tree_bb_level_predictions (void)
3002 bool has_return_edges
= false;
3006 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
3007 if (!unlikely_executed_edge_p (e
) && !(e
->flags
& EDGE_ABNORMAL_CALL
))
3009 has_return_edges
= true;
3013 apply_return_prediction ();
3015 FOR_EACH_BB_FN (bb
, cfun
)
3017 gimple_stmt_iterator gsi
;
3019 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
3021 gimple
*stmt
= gsi_stmt (gsi
);
3024 if (is_gimple_call (stmt
))
3026 if (gimple_call_noreturn_p (stmt
)
3028 && !is_exit_with_zero_arg (stmt
))
3029 predict_paths_leading_to (bb
, PRED_NORETURN
,
3031 decl
= gimple_call_fndecl (stmt
);
3033 && lookup_attribute ("cold",
3034 DECL_ATTRIBUTES (decl
)))
3035 predict_paths_leading_to (bb
, PRED_COLD_FUNCTION
,
3037 if (decl
&& recursive_call_p (current_function_decl
, decl
))
3038 predict_paths_leading_to (bb
, PRED_RECURSIVE_CALL
,
3041 else if (gimple_code (stmt
) == GIMPLE_PREDICT
)
3043 predict_paths_leading_to (bb
, gimple_predict_predictor (stmt
),
3044 gimple_predict_outcome (stmt
));
3045 /* Keep GIMPLE_PREDICT around so early inlining will propagate
3046 hints to callers. */
3052 /* Callback for hash_map::traverse, asserts that the pointer map is
3056 assert_is_empty (const_basic_block
const &, edge_prediction
*const &value
,
3059 gcc_assert (!value
);
3063 /* Predict branch probabilities and estimate profile for basic block BB.
3064 When LOCAL_ONLY is set do not use any global properties of CFG. */
3067 tree_estimate_probability_bb (basic_block bb
, bool local_only
)
3072 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3074 /* Look for block we are guarding (ie we dominate it,
3075 but it doesn't postdominate us). */
3076 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
) && e
->dest
!= bb
3078 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
3079 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
3081 gimple_stmt_iterator bi
;
3083 /* The call heuristic claims that a guarded function call
3084 is improbable. This is because such calls are often used
3085 to signal exceptional situations such as printing error
3087 for (bi
= gsi_start_bb (e
->dest
); !gsi_end_p (bi
);
3090 gimple
*stmt
= gsi_stmt (bi
);
3091 if (is_gimple_call (stmt
)
3092 && !gimple_inexpensive_call_p (as_a
<gcall
*> (stmt
))
3093 /* Constant and pure calls are hardly used to signalize
3094 something exceptional. */
3095 && gimple_has_side_effects (stmt
))
3097 if (gimple_call_fndecl (stmt
))
3098 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
3099 else if (virtual_method_call_p (gimple_call_fn (stmt
)))
3100 predict_edge_def (e
, PRED_POLYMORPHIC_CALL
, NOT_TAKEN
);
3102 predict_edge_def (e
, PRED_INDIR_CALL
, TAKEN
);
3108 tree_predict_by_opcode (bb
);
3111 /* Predict branch probabilities and estimate profile of the tree CFG.
3112 This function can be called from the loop optimizers to recompute
3113 the profile information.
3114 If DRY_RUN is set, do not modify CFG and only produce dump files. */
3117 tree_estimate_probability (bool dry_run
)
3121 connect_infinite_loops_to_exit ();
3122 /* We use loop_niter_by_eval, which requires that the loops have
3124 create_preheaders (CP_SIMPLE_PREHEADERS
);
3125 calculate_dominance_info (CDI_POST_DOMINATORS
);
3126 /* Decide which edges are known to be unlikely. This improves later
3127 branch prediction. */
3128 determine_unlikely_bbs ();
3130 bb_predictions
= new hash_map
<const_basic_block
, edge_prediction
*>;
3131 tree_bb_level_predictions ();
3132 record_loop_exits ();
3134 if (number_of_loops (cfun
) > 1)
3137 FOR_EACH_BB_FN (bb
, cfun
)
3138 tree_estimate_probability_bb (bb
, false);
3140 FOR_EACH_BB_FN (bb
, cfun
)
3141 combine_predictions_for_bb (bb
, dry_run
);
3144 bb_predictions
->traverse
<void *, assert_is_empty
> (NULL
);
3146 delete bb_predictions
;
3147 bb_predictions
= NULL
;
3150 estimate_bb_frequencies (false);
3151 free_dominance_info (CDI_POST_DOMINATORS
);
3152 remove_fake_exit_edges ();
3155 /* Set edge->probability for each successor edge of BB. */
3157 tree_guess_outgoing_edge_probabilities (basic_block bb
)
3159 bb_predictions
= new hash_map
<const_basic_block
, edge_prediction
*>;
3160 tree_estimate_probability_bb (bb
, true);
3161 combine_predictions_for_bb (bb
, false);
3163 bb_predictions
->traverse
<void *, assert_is_empty
> (NULL
);
3164 delete bb_predictions
;
3165 bb_predictions
= NULL
;
3168 /* Filter function predicate that returns true for a edge predicate P
3169 if its edge is equal to DATA. */
3172 not_loop_guard_equal_edge_p (edge_prediction
*p
, void *data
)
3174 return p
->ep_edge
!= (edge
)data
|| p
->ep_predictor
!= PRED_LOOP_GUARD
;
3177 /* Predict edge E with PRED unless it is already predicted by some predictor
3178 considered equivalent. */
3181 maybe_predict_edge (edge e
, enum br_predictor pred
, enum prediction taken
)
3183 if (edge_predicted_by_p (e
, pred
, taken
))
3185 if (pred
== PRED_LOOP_GUARD
3186 && edge_predicted_by_p (e
, PRED_LOOP_GUARD_WITH_RECURSION
, taken
))
3188 /* Consider PRED_LOOP_GUARD_WITH_RECURSION superrior to LOOP_GUARD. */
3189 if (pred
== PRED_LOOP_GUARD_WITH_RECURSION
)
3191 edge_prediction
**preds
= bb_predictions
->get (e
->src
);
3193 filter_predictions (preds
, not_loop_guard_equal_edge_p
, e
);
3195 predict_edge_def (e
, pred
, taken
);
3197 /* Predict edges to successors of CUR whose sources are not postdominated by
3198 BB by PRED and recurse to all postdominators. */
3201 predict_paths_for_bb (basic_block cur
, basic_block bb
,
3202 enum br_predictor pred
,
3203 enum prediction taken
,
3204 bitmap visited
, class loop
*in_loop
= NULL
)
3210 /* If we exited the loop or CUR is unconditional in the loop, there is
3213 && (!flow_bb_inside_loop_p (in_loop
, cur
)
3214 || dominated_by_p (CDI_DOMINATORS
, in_loop
->latch
, cur
)))
3217 /* We are looking for all edges forming edge cut induced by
3218 set of all blocks postdominated by BB. */
3219 FOR_EACH_EDGE (e
, ei
, cur
->preds
)
3220 if (e
->src
->index
>= NUM_FIXED_BLOCKS
3221 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, bb
))
3227 /* Ignore fake edges and eh, we predict them as not taken anyway. */
3228 if (unlikely_executed_edge_p (e
))
3230 gcc_assert (bb
== cur
|| dominated_by_p (CDI_POST_DOMINATORS
, cur
, bb
));
3232 /* See if there is an edge from e->src that is not abnormal
3233 and does not lead to BB and does not exit the loop. */
3234 FOR_EACH_EDGE (e2
, ei2
, e
->src
->succs
)
3236 && !unlikely_executed_edge_p (e2
)
3237 && !dominated_by_p (CDI_POST_DOMINATORS
, e2
->dest
, bb
)
3238 && (!in_loop
|| !loop_exit_edge_p (in_loop
, e2
)))
3244 /* If there is non-abnormal path leaving e->src, predict edge
3245 using predictor. Otherwise we need to look for paths
3248 The second may lead to infinite loop in the case we are predicitng
3249 regions that are only reachable by abnormal edges. We simply
3250 prevent visiting given BB twice. */
3252 maybe_predict_edge (e
, pred
, taken
);
3253 else if (bitmap_set_bit (visited
, e
->src
->index
))
3254 predict_paths_for_bb (e
->src
, e
->src
, pred
, taken
, visited
, in_loop
);
3256 for (son
= first_dom_son (CDI_POST_DOMINATORS
, cur
);
3258 son
= next_dom_son (CDI_POST_DOMINATORS
, son
))
3259 predict_paths_for_bb (son
, bb
, pred
, taken
, visited
, in_loop
);
3262 /* Sets branch probabilities according to PREDiction and
3266 predict_paths_leading_to (basic_block bb
, enum br_predictor pred
,
3267 enum prediction taken
, class loop
*in_loop
)
3269 predict_paths_for_bb (bb
, bb
, pred
, taken
, auto_bitmap (), in_loop
);
3272 /* Like predict_paths_leading_to but take edge instead of basic block. */
3275 predict_paths_leading_to_edge (edge e
, enum br_predictor pred
,
3276 enum prediction taken
, class loop
*in_loop
)
3278 bool has_nonloop_edge
= false;
3282 basic_block bb
= e
->src
;
3283 FOR_EACH_EDGE (e2
, ei
, bb
->succs
)
3284 if (e2
->dest
!= e
->src
&& e2
->dest
!= e
->dest
3285 && !unlikely_executed_edge_p (e2
)
3286 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e2
->dest
))
3288 has_nonloop_edge
= true;
3292 if (!has_nonloop_edge
)
3293 predict_paths_for_bb (bb
, bb
, pred
, taken
, auto_bitmap (), in_loop
);
3295 maybe_predict_edge (e
, pred
, taken
);
3298 /* This is used to carry information about basic blocks. It is
3299 attached to the AUX field of the standard CFG block. */
3304 /* Estimated frequency of execution of basic_block. */
3307 /* To keep queue of basic blocks to process. */
3310 /* Number of predecessors we need to visit first. */
3314 /* Similar information for edges. */
3315 class edge_prob_info
3318 /* In case edge is a loopback edge, the probability edge will be reached
3319 in case header is. Estimated number of iterations of the loop can be
3320 then computed as 1 / (1 - back_edge_prob). */
3321 sreal back_edge_prob
;
3322 /* True if the edge is a loopback edge in the natural loop. */
3323 unsigned int back_edge
:1;
3326 #define BLOCK_INFO(B) ((block_info *) (B)->aux)
3328 #define EDGE_INFO(E) ((edge_prob_info *) (E)->aux)
3330 /* Helper function for estimate_bb_frequencies.
3331 Propagate the frequencies in blocks marked in
3332 TOVISIT, starting in HEAD. */
3335 propagate_freq (basic_block head
, bitmap tovisit
,
3336 sreal max_cyclic_prob
)
3345 /* For each basic block we need to visit count number of his predecessors
3346 we need to visit first. */
3347 EXECUTE_IF_SET_IN_BITMAP (tovisit
, 0, i
, bi
)
3352 bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
3354 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3356 bool visit
= bitmap_bit_p (tovisit
, e
->src
->index
);
3358 if (visit
&& !(e
->flags
& EDGE_DFS_BACK
))
3360 else if (visit
&& dump_file
&& !EDGE_INFO (e
)->back_edge
)
3362 "Irreducible region hit, ignoring edge to %i->%i\n",
3363 e
->src
->index
, bb
->index
);
3365 BLOCK_INFO (bb
)->npredecessors
= count
;
3366 /* When function never returns, we will never process exit block. */
3367 if (!count
&& bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
3368 bb
->count
= profile_count::zero ();
3371 BLOCK_INFO (head
)->frequency
= 1;
3373 for (bb
= head
; bb
; bb
= nextbb
)
3376 sreal cyclic_probability
= 0;
3377 sreal frequency
= 0;
3379 nextbb
= BLOCK_INFO (bb
)->next
;
3380 BLOCK_INFO (bb
)->next
= NULL
;
3382 /* Compute frequency of basic block. */
3386 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3387 gcc_assert (!bitmap_bit_p (tovisit
, e
->src
->index
)
3388 || (e
->flags
& EDGE_DFS_BACK
));
3390 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3391 if (EDGE_INFO (e
)->back_edge
)
3392 cyclic_probability
+= EDGE_INFO (e
)->back_edge_prob
;
3393 else if (!(e
->flags
& EDGE_DFS_BACK
))
3395 /* FIXME: Graphite is producing edges with no profile. Once
3396 this is fixed, drop this. */
3397 sreal tmp
= e
->probability
.initialized_p () ?
3398 e
->probability
.to_sreal () : 0;
3399 frequency
+= tmp
* BLOCK_INFO (e
->src
)->frequency
;
3402 if (cyclic_probability
== 0)
3404 BLOCK_INFO (bb
)->frequency
= frequency
;
3408 if (cyclic_probability
> max_cyclic_prob
)
3412 "cyclic probability of bb %i is %f (capped to %f)"
3413 "; turning freq %f",
3414 bb
->index
, cyclic_probability
.to_double (),
3415 max_cyclic_prob
.to_double (),
3416 frequency
.to_double ());
3418 cyclic_probability
= max_cyclic_prob
;
3422 "cyclic probability of bb %i is %f; turning freq %f",
3423 bb
->index
, cyclic_probability
.to_double (),
3424 frequency
.to_double ());
3426 BLOCK_INFO (bb
)->frequency
= frequency
3427 / (sreal (1) - cyclic_probability
);
3429 fprintf (dump_file
, " to %f\n",
3430 BLOCK_INFO (bb
)->frequency
.to_double ());
3434 bitmap_clear_bit (tovisit
, bb
->index
);
3436 e
= find_edge (bb
, head
);
3439 /* FIXME: Graphite is producing edges with no profile. Once
3440 this is fixed, drop this. */
3441 sreal tmp
= e
->probability
.initialized_p () ?
3442 e
->probability
.to_sreal () : 0;
3443 EDGE_INFO (e
)->back_edge_prob
= tmp
* BLOCK_INFO (bb
)->frequency
;
3446 /* Propagate to successor blocks. */
3447 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3448 if (!(e
->flags
& EDGE_DFS_BACK
)
3449 && BLOCK_INFO (e
->dest
)->npredecessors
)
3451 BLOCK_INFO (e
->dest
)->npredecessors
--;
3452 if (!BLOCK_INFO (e
->dest
)->npredecessors
)
3457 BLOCK_INFO (last
)->next
= e
->dest
;
3465 /* Estimate frequencies in loops at same nest level. */
3468 estimate_loops_at_level (class loop
*first_loop
, sreal max_cyclic_prob
)
3472 for (loop
= first_loop
; loop
; loop
= loop
->next
)
3477 auto_bitmap tovisit
;
3479 estimate_loops_at_level (loop
->inner
, max_cyclic_prob
);
3481 /* Find current loop back edge and mark it. */
3482 e
= loop_latch_edge (loop
);
3483 EDGE_INFO (e
)->back_edge
= 1;
3485 bbs
= get_loop_body (loop
);
3486 for (i
= 0; i
< loop
->num_nodes
; i
++)
3487 bitmap_set_bit (tovisit
, bbs
[i
]->index
);
3489 propagate_freq (loop
->header
, tovisit
, max_cyclic_prob
);
3493 /* Propagates frequencies through structure of loops. */
3496 estimate_loops (void)
3498 auto_bitmap tovisit
;
3500 sreal max_cyclic_prob
= (sreal
)1
3501 - (sreal
)1 / (param_max_predicted_iterations
+ 1);
3503 /* Start by estimating the frequencies in the loops. */
3504 if (number_of_loops (cfun
) > 1)
3505 estimate_loops_at_level (current_loops
->tree_root
->inner
, max_cyclic_prob
);
3507 /* Now propagate the frequencies through all the blocks. */
3508 FOR_ALL_BB_FN (bb
, cfun
)
3510 bitmap_set_bit (tovisit
, bb
->index
);
3512 propagate_freq (ENTRY_BLOCK_PTR_FOR_FN (cfun
), tovisit
, max_cyclic_prob
);
3515 /* Drop the profile for NODE to guessed, and update its frequency based on
3516 whether it is expected to be hot given the CALL_COUNT. */
3519 drop_profile (struct cgraph_node
*node
, profile_count call_count
)
3521 struct function
*fn
= DECL_STRUCT_FUNCTION (node
->decl
);
3522 /* In the case where this was called by another function with a
3523 dropped profile, call_count will be 0. Since there are no
3524 non-zero call counts to this function, we don't know for sure
3525 whether it is hot, and therefore it will be marked normal below. */
3526 bool hot
= maybe_hot_count_p (NULL
, call_count
);
3530 "Dropping 0 profile for %s. %s based on calls.\n",
3532 hot
? "Function is hot" : "Function is normal");
3533 /* We only expect to miss profiles for functions that are reached
3534 via non-zero call edges in cases where the function may have
3535 been linked from another module or library (COMDATs and extern
3536 templates). See the comments below for handle_missing_profiles.
3537 Also, only warn in cases where the missing counts exceed the
3538 number of training runs. In certain cases with an execv followed
3539 by a no-return call the profile for the no-return call is not
3540 dumped and there can be a mismatch. */
3541 if (!DECL_COMDAT (node
->decl
) && !DECL_EXTERNAL (node
->decl
)
3542 && call_count
> profile_info
->runs
)
3544 if (flag_profile_correction
)
3548 "Missing counts for called function %s\n",
3549 node
->dump_name ());
3552 warning (0, "Missing counts for called function %s",
3553 node
->dump_name ());
3557 if (opt_for_fn (node
->decl
, flag_guess_branch_prob
))
3560 = !ENTRY_BLOCK_PTR_FOR_FN (fn
)->count
.nonzero_p ();
3561 FOR_ALL_BB_FN (bb
, fn
)
3562 if (clear_zeros
|| !(bb
->count
== profile_count::zero ()))
3563 bb
->count
= bb
->count
.guessed_local ();
3564 fn
->cfg
->count_max
= fn
->cfg
->count_max
.guessed_local ();
3568 FOR_ALL_BB_FN (bb
, fn
)
3569 bb
->count
= profile_count::uninitialized ();
3570 fn
->cfg
->count_max
= profile_count::uninitialized ();
3573 struct cgraph_edge
*e
;
3574 for (e
= node
->callees
; e
; e
= e
->next_callee
)
3575 e
->count
= gimple_bb (e
->call_stmt
)->count
;
3576 for (e
= node
->indirect_calls
; e
; e
= e
->next_callee
)
3577 e
->count
= gimple_bb (e
->call_stmt
)->count
;
3578 node
->count
= ENTRY_BLOCK_PTR_FOR_FN (fn
)->count
;
3580 profile_status_for_fn (fn
)
3581 = (flag_guess_branch_prob
? PROFILE_GUESSED
: PROFILE_ABSENT
);
3583 = hot
? NODE_FREQUENCY_HOT
: NODE_FREQUENCY_NORMAL
;
3586 /* In the case of COMDAT routines, multiple object files will contain the same
3587 function and the linker will select one for the binary. In that case
3588 all the other copies from the profile instrument binary will be missing
3589 profile counts. Look for cases where this happened, due to non-zero
3590 call counts going to 0-count functions, and drop the profile to guessed
3591 so that we can use the estimated probabilities and avoid optimizing only
3594 The other case where the profile may be missing is when the routine
3595 is not going to be emitted to the object file, e.g. for "extern template"
3596 class methods. Those will be marked DECL_EXTERNAL. Emit a warning in
3597 all other cases of non-zero calls to 0-count functions. */
3600 handle_missing_profiles (void)
3602 const int unlikely_frac
= param_unlikely_bb_count_fraction
;
3603 struct cgraph_node
*node
;
3604 auto_vec
<struct cgraph_node
*, 64> worklist
;
3606 /* See if 0 count function has non-0 count callers. In this case we
3607 lost some profile. Drop its function profile to PROFILE_GUESSED. */
3608 FOR_EACH_DEFINED_FUNCTION (node
)
3610 struct cgraph_edge
*e
;
3611 profile_count call_count
= profile_count::zero ();
3612 gcov_type max_tp_first_run
= 0;
3613 struct function
*fn
= DECL_STRUCT_FUNCTION (node
->decl
);
3615 if (node
->count
.ipa ().nonzero_p ())
3617 for (e
= node
->callers
; e
; e
= e
->next_caller
)
3618 if (e
->count
.ipa ().initialized_p () && e
->count
.ipa () > 0)
3620 call_count
= call_count
+ e
->count
.ipa ();
3622 if (e
->caller
->tp_first_run
> max_tp_first_run
)
3623 max_tp_first_run
= e
->caller
->tp_first_run
;
3626 /* If time profile is missing, let assign the maximum that comes from
3627 caller functions. */
3628 if (!node
->tp_first_run
&& max_tp_first_run
)
3629 node
->tp_first_run
= max_tp_first_run
+ 1;
3633 && call_count
* unlikely_frac
>= profile_info
->runs
)
3635 drop_profile (node
, call_count
);
3636 worklist
.safe_push (node
);
3640 /* Propagate the profile dropping to other 0-count COMDATs that are
3641 potentially called by COMDATs we already dropped the profile on. */
3642 while (worklist
.length () > 0)
3644 struct cgraph_edge
*e
;
3646 node
= worklist
.pop ();
3647 for (e
= node
->callees
; e
; e
= e
->next_caller
)
3649 struct cgraph_node
*callee
= e
->callee
;
3650 struct function
*fn
= DECL_STRUCT_FUNCTION (callee
->decl
);
3652 if (!(e
->count
.ipa () == profile_count::zero ())
3653 && callee
->count
.ipa ().nonzero_p ())
3655 if ((DECL_COMDAT (callee
->decl
) || DECL_EXTERNAL (callee
->decl
))
3657 && profile_status_for_fn (fn
) == PROFILE_READ
)
3659 drop_profile (node
, profile_count::zero ());
3660 worklist
.safe_push (callee
);
3666 /* Convert counts measured by profile driven feedback to frequencies.
3667 Return nonzero iff there was any nonzero execution count. */
3670 update_max_bb_count (void)
3672 profile_count true_count_max
= profile_count::uninitialized ();
3675 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
3676 true_count_max
= true_count_max
.max (bb
->count
);
3678 cfun
->cfg
->count_max
= true_count_max
;
3680 return true_count_max
.ipa ().nonzero_p ();
3683 /* Return true if function is likely to be expensive, so there is no point to
3684 optimize performance of prologue, epilogue or do inlining at the expense
3685 of code size growth. THRESHOLD is the limit of number of instructions
3686 function can execute at average to be still considered not expensive. */
3689 expensive_function_p (int threshold
)
3693 /* If profile was scaled in a way entry block has count 0, then the function
3694 is deifnitly taking a lot of time. */
3695 if (!ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.nonzero_p ())
3698 profile_count limit
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
* threshold
;
3699 profile_count sum
= profile_count::zero ();
3700 FOR_EACH_BB_FN (bb
, cfun
)
3704 if (!bb
->count
.initialized_p ())
3707 fprintf (dump_file
, "Function is considered expensive because"
3708 " count of bb %i is not initialized\n", bb
->index
);
3712 FOR_BB_INSNS (bb
, insn
)
3713 if (active_insn_p (insn
))
3724 /* All basic blocks that are reachable only from unlikely basic blocks are
3728 propagate_unlikely_bbs_forward (void)
3730 auto_vec
<basic_block
, 64> worklist
;
3735 if (!(ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
== profile_count::zero ()))
3737 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->aux
= (void *)(size_t) 1;
3738 worklist
.safe_push (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
3740 while (worklist
.length () > 0)
3742 bb
= worklist
.pop ();
3743 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3744 if (!(e
->count () == profile_count::zero ())
3745 && !(e
->dest
->count
== profile_count::zero ())
3748 e
->dest
->aux
= (void *)(size_t) 1;
3749 worklist
.safe_push (e
->dest
);
3754 FOR_ALL_BB_FN (bb
, cfun
)
3758 if (!(bb
->count
== profile_count::zero ())
3759 && (dump_file
&& (dump_flags
& TDF_DETAILS
)))
3761 "Basic block %i is marked unlikely by forward prop\n",
3763 bb
->count
= profile_count::zero ();
3770 /* Determine basic blocks/edges that are known to be unlikely executed and set
3771 their counters to zero.
3772 This is done with first identifying obviously unlikely BBs/edges and then
3773 propagating in both directions. */
3776 determine_unlikely_bbs ()
3779 auto_vec
<basic_block
, 64> worklist
;
3783 FOR_EACH_BB_FN (bb
, cfun
)
3785 if (!(bb
->count
== profile_count::zero ())
3786 && unlikely_executed_bb_p (bb
))
3788 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3789 fprintf (dump_file
, "Basic block %i is locally unlikely\n",
3791 bb
->count
= profile_count::zero ();
3794 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3795 if (!(e
->probability
== profile_probability::never ())
3796 && unlikely_executed_edge_p (e
))
3798 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3799 fprintf (dump_file
, "Edge %i->%i is locally unlikely\n",
3800 bb
->index
, e
->dest
->index
);
3801 e
->probability
= profile_probability::never ();
3804 gcc_checking_assert (!bb
->aux
);
3806 propagate_unlikely_bbs_forward ();
3808 auto_vec
<int, 64> nsuccs
;
3809 nsuccs
.safe_grow_cleared (last_basic_block_for_fn (cfun
), true);
3810 FOR_ALL_BB_FN (bb
, cfun
)
3811 if (!(bb
->count
== profile_count::zero ())
3812 && bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
3814 nsuccs
[bb
->index
] = 0;
3815 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3816 if (!(e
->probability
== profile_probability::never ())
3817 && !(e
->dest
->count
== profile_count::zero ()))
3818 nsuccs
[bb
->index
]++;
3819 if (!nsuccs
[bb
->index
])
3820 worklist
.safe_push (bb
);
3822 while (worklist
.length () > 0)
3824 bb
= worklist
.pop ();
3825 if (bb
->count
== profile_count::zero ())
3827 if (bb
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
3830 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
3831 !gsi_end_p (gsi
); gsi_next (&gsi
))
3832 if (stmt_can_terminate_bb_p (gsi_stmt (gsi
))
3833 /* stmt_can_terminate_bb_p special cases noreturns because it
3834 assumes that fake edges are created. We want to know that
3835 noreturn alone does not imply BB to be unlikely. */
3836 || (is_gimple_call (gsi_stmt (gsi
))
3837 && (gimple_call_flags (gsi_stmt (gsi
)) & ECF_NORETURN
)))
3845 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3847 "Basic block %i is marked unlikely by backward prop\n",
3849 bb
->count
= profile_count::zero ();
3850 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3851 if (!(e
->probability
== profile_probability::never ()))
3853 if (!(e
->src
->count
== profile_count::zero ()))
3855 gcc_checking_assert (nsuccs
[e
->src
->index
] > 0);
3856 nsuccs
[e
->src
->index
]--;
3857 if (!nsuccs
[e
->src
->index
])
3858 worklist
.safe_push (e
->src
);
3862 /* Finally all edges from non-0 regions to 0 are unlikely. */
3863 FOR_ALL_BB_FN (bb
, cfun
)
3865 if (!(bb
->count
== profile_count::zero ()))
3866 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3867 if (!(e
->probability
== profile_probability::never ())
3868 && e
->dest
->count
== profile_count::zero ())
3870 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3871 fprintf (dump_file
, "Edge %i->%i is unlikely because "
3872 "it enters unlikely block\n",
3873 bb
->index
, e
->dest
->index
);
3874 e
->probability
= profile_probability::never ();
3879 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3880 if (e
->probability
== profile_probability::never ())
3890 && !(other
->probability
== profile_probability::always ()))
3892 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3893 fprintf (dump_file
, "Edge %i->%i is locally likely\n",
3894 bb
->index
, other
->dest
->index
);
3895 other
->probability
= profile_probability::always ();
3898 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
== profile_count::zero ())
3899 cgraph_node::get (current_function_decl
)->count
= profile_count::zero ();
3902 /* Estimate and propagate basic block frequencies using the given branch
3903 probabilities. If FORCE is true, the frequencies are used to estimate
3904 the counts even when there are already non-zero profile counts. */
3907 estimate_bb_frequencies (bool force
)
3912 determine_unlikely_bbs ();
3914 if (force
|| profile_status_for_fn (cfun
) != PROFILE_READ
3915 || !update_max_bb_count ())
3918 mark_dfs_back_edges ();
3920 single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
))->probability
=
3921 profile_probability::always ();
3923 /* Set up block info for each basic block. */
3924 alloc_aux_for_blocks (sizeof (block_info
));
3925 alloc_aux_for_edges (sizeof (edge_prob_info
));
3926 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
3931 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3933 /* FIXME: Graphite is producing edges with no profile. Once
3934 this is fixed, drop this. */
3935 if (e
->probability
.initialized_p ())
3936 EDGE_INFO (e
)->back_edge_prob
3937 = e
->probability
.to_sreal ();
3939 /* back_edge_prob = 0.5 */
3940 EDGE_INFO (e
)->back_edge_prob
= sreal (1, -1);
3944 /* First compute frequencies locally for each loop from innermost
3945 to outermost to examine frequencies for back edges. */
3949 FOR_EACH_BB_FN (bb
, cfun
)
3950 if (freq_max
< BLOCK_INFO (bb
)->frequency
)
3951 freq_max
= BLOCK_INFO (bb
)->frequency
;
3953 /* Scaling frequencies up to maximal profile count may result in
3954 frequent overflows especially when inlining loops.
3955 Small scalling results in unnecesary precision loss. Stay in
3956 the half of the (exponential) range. */
3957 freq_max
= (sreal (1) << (profile_count::n_bits
/ 2)) / freq_max
;
3960 profile_count ipa_count
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.ipa ();
3961 cfun
->cfg
->count_max
= profile_count::uninitialized ();
3962 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
3964 sreal tmp
= BLOCK_INFO (bb
)->frequency
;
3967 gimple_stmt_iterator gsi
;
3970 /* Self recursive calls can not have frequency greater than 1
3971 or program will never terminate. This will result in an
3972 inconsistent bb profile but it is better than greatly confusing
3973 IPA cost metrics. */
3974 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
3975 if (is_gimple_call (gsi_stmt (gsi
))
3976 && (decl
= gimple_call_fndecl (gsi_stmt (gsi
))) != NULL
3977 && recursive_call_p (current_function_decl
, decl
))
3980 fprintf (dump_file
, "Dropping frequency of recursive call"
3981 " in bb %i from %f\n", bb
->index
,
3983 tmp
= (sreal
)9 / (sreal
)10;
3987 tmp
= tmp
* freq_max
+ sreal (1, -1);
3988 profile_count count
= profile_count::from_gcov_type (tmp
.to_int ());
3990 /* If we have profile feedback in which this function was never
3991 executed, then preserve this info. */
3992 if (!(bb
->count
== profile_count::zero ()))
3993 bb
->count
= count
.guessed_local ().combine_with_ipa_count (ipa_count
);
3994 cfun
->cfg
->count_max
= cfun
->cfg
->count_max
.max (bb
->count
);
3997 free_aux_for_blocks ();
3998 free_aux_for_edges ();
4000 compute_function_frequency ();
4003 /* Decide whether function is hot, cold or unlikely executed. */
4005 compute_function_frequency (void)
4008 struct cgraph_node
*node
= cgraph_node::get (current_function_decl
);
4010 if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
4011 || MAIN_NAME_P (DECL_NAME (current_function_decl
)))
4012 node
->only_called_at_startup
= true;
4013 if (DECL_STATIC_DESTRUCTOR (current_function_decl
))
4014 node
->only_called_at_exit
= true;
4016 if (!ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.ipa_p ())
4018 int flags
= flags_from_decl_or_type (current_function_decl
);
4019 if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
))
4021 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
4022 else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl
))
4024 node
->frequency
= NODE_FREQUENCY_HOT
;
4025 else if (flags
& ECF_NORETURN
)
4026 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
4027 else if (MAIN_NAME_P (DECL_NAME (current_function_decl
)))
4028 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
4029 else if (DECL_STATIC_CONSTRUCTOR (current_function_decl
)
4030 || DECL_STATIC_DESTRUCTOR (current_function_decl
))
4031 node
->frequency
= NODE_FREQUENCY_EXECUTED_ONCE
;
4035 node
->frequency
= NODE_FREQUENCY_UNLIKELY_EXECUTED
;
4036 if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl
))
4038 warn_function_cold (current_function_decl
);
4039 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
.ipa() == profile_count::zero ())
4041 FOR_EACH_BB_FN (bb
, cfun
)
4043 if (maybe_hot_bb_p (cfun
, bb
))
4045 node
->frequency
= NODE_FREQUENCY_HOT
;
4048 if (!probably_never_executed_bb_p (cfun
, bb
))
4049 node
->frequency
= NODE_FREQUENCY_NORMAL
;
4053 /* Build PREDICT_EXPR. */
4055 build_predict_expr (enum br_predictor predictor
, enum prediction taken
)
4057 tree t
= build1 (PREDICT_EXPR
, void_type_node
,
4058 build_int_cst (integer_type_node
, predictor
));
4059 SET_PREDICT_EXPR_OUTCOME (t
, taken
);
4064 predictor_name (enum br_predictor predictor
)
4066 return predictor_info
[predictor
].name
;
4069 /* Predict branch probabilities and estimate profile of the tree CFG. */
4073 const pass_data pass_data_profile
=
4075 GIMPLE_PASS
, /* type */
4076 "profile_estimate", /* name */
4077 OPTGROUP_NONE
, /* optinfo_flags */
4078 TV_BRANCH_PROB
, /* tv_id */
4079 PROP_cfg
, /* properties_required */
4080 0, /* properties_provided */
4081 0, /* properties_destroyed */
4082 0, /* todo_flags_start */
4083 0, /* todo_flags_finish */
4086 class pass_profile
: public gimple_opt_pass
4089 pass_profile (gcc::context
*ctxt
)
4090 : gimple_opt_pass (pass_data_profile
, ctxt
)
4093 /* opt_pass methods: */
4094 bool gate (function
*) final override
{ return flag_guess_branch_prob
; }
4095 unsigned int execute (function
*) final override
;
4097 }; // class pass_profile
4100 pass_profile::execute (function
*fun
)
4104 if (profile_status_for_fn (cfun
) == PROFILE_GUESSED
)
4107 loop_optimizer_init (LOOPS_NORMAL
);
4108 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4109 flow_loops_dump (dump_file
, NULL
, 0);
4111 nb_loops
= number_of_loops (fun
);
4115 tree_estimate_probability (false);
4120 loop_optimizer_finalize ();
4121 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4122 gimple_dump_cfg (dump_file
, dump_flags
);
4123 if (profile_status_for_fn (fun
) == PROFILE_ABSENT
)
4124 profile_status_for_fn (fun
) = PROFILE_GUESSED
;
4125 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4127 for (auto loop
: loops_list (cfun
, LI_FROM_INNERMOST
))
4128 if (loop
->header
->count
.initialized_p ())
4129 fprintf (dump_file
, "Loop got predicted %d to iterate %i times.\n",
4131 (int)expected_loop_iterations_unbounded (loop
));
4139 make_pass_profile (gcc::context
*ctxt
)
4141 return new pass_profile (ctxt
);
4144 /* Return true when PRED predictor should be removed after early
4145 tree passes. Most of the predictors are beneficial to survive
4146 as early inlining can also distribute then into caller's bodies. */
4149 strip_predictor_early (enum br_predictor pred
)
4153 case PRED_TREE_EARLY_RETURN
:
4160 /* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements
4161 we no longer need. EARLY is set to true when called from early
4165 strip_predict_hints (function
*fun
, bool early
)
4170 bool changed
= false;
4172 FOR_EACH_BB_FN (bb
, fun
)
4174 gimple_stmt_iterator bi
;
4175 for (bi
= gsi_start_bb (bb
); !gsi_end_p (bi
);)
4177 gimple
*stmt
= gsi_stmt (bi
);
4179 if (gimple_code (stmt
) == GIMPLE_PREDICT
)
4182 || strip_predictor_early (gimple_predict_predictor (stmt
)))
4184 gsi_remove (&bi
, true);
4189 else if (is_gimple_call (stmt
))
4191 tree fndecl
= gimple_call_fndecl (stmt
);
4194 && ((fndecl
!= NULL_TREE
4195 && fndecl_built_in_p (fndecl
, BUILT_IN_EXPECT
)
4196 && gimple_call_num_args (stmt
) == 2)
4197 || (fndecl
!= NULL_TREE
4198 && fndecl_built_in_p (fndecl
,
4199 BUILT_IN_EXPECT_WITH_PROBABILITY
)
4200 && gimple_call_num_args (stmt
) == 3)
4201 || (gimple_call_internal_p (stmt
)
4202 && gimple_call_internal_fn (stmt
) == IFN_BUILTIN_EXPECT
)))
4204 var
= gimple_call_lhs (stmt
);
4209 = gimple_build_assign (var
, gimple_call_arg (stmt
, 0));
4210 gsi_replace (&bi
, ass_stmt
, true);
4214 gsi_remove (&bi
, true);
4222 return changed
? TODO_cleanup_cfg
: 0;
4227 const pass_data pass_data_strip_predict_hints
=
4229 GIMPLE_PASS
, /* type */
4230 "*strip_predict_hints", /* name */
4231 OPTGROUP_NONE
, /* optinfo_flags */
4232 TV_BRANCH_PROB
, /* tv_id */
4233 PROP_cfg
, /* properties_required */
4234 0, /* properties_provided */
4235 0, /* properties_destroyed */
4236 0, /* todo_flags_start */
4237 0, /* todo_flags_finish */
4240 class pass_strip_predict_hints
: public gimple_opt_pass
4243 pass_strip_predict_hints (gcc::context
*ctxt
)
4244 : gimple_opt_pass (pass_data_strip_predict_hints
, ctxt
)
4247 /* opt_pass methods: */
4248 opt_pass
* clone () final override
4250 return new pass_strip_predict_hints (m_ctxt
);
4252 void set_pass_param (unsigned int n
, bool param
) final override
4254 gcc_assert (n
== 0);
4258 unsigned int execute (function
*) final override
;
4263 }; // class pass_strip_predict_hints
4266 pass_strip_predict_hints::execute (function
*fun
)
4268 return strip_predict_hints (fun
, early_p
);
4274 make_pass_strip_predict_hints (gcc::context
*ctxt
)
4276 return new pass_strip_predict_hints (ctxt
);
4279 /* Rebuild function frequencies. Passes are in general expected to
4280 maintain profile by hand, however in some cases this is not possible:
4281 for example when inlining several functions with loops freuqencies might run
4282 out of scale and thus needs to be recomputed. */
4285 rebuild_frequencies (void)
4287 timevar_push (TV_REBUILD_FREQUENCIES
);
4289 /* When the max bb count in the function is small, there is a higher
4290 chance that there were truncation errors in the integer scaling
4291 of counts by inlining and other optimizations. This could lead
4292 to incorrect classification of code as being cold when it isn't.
4293 In that case, force the estimation of bb counts/frequencies from the
4294 branch probabilities, rather than computing frequencies from counts,
4295 which may also lead to frequencies incorrectly reduced to 0. There
4296 is less precision in the probabilities, so we only do this for small
4298 cfun
->cfg
->count_max
= profile_count::uninitialized ();
4300 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), NULL
, next_bb
)
4301 cfun
->cfg
->count_max
= cfun
->cfg
->count_max
.max (bb
->count
);
4303 if (profile_status_for_fn (cfun
) == PROFILE_GUESSED
)
4305 loop_optimizer_init (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
);
4306 connect_infinite_loops_to_exit ();
4307 estimate_bb_frequencies (true);
4308 remove_fake_exit_edges ();
4309 loop_optimizer_finalize ();
4311 else if (profile_status_for_fn (cfun
) == PROFILE_READ
)
4312 update_max_bb_count ();
4313 else if (profile_status_for_fn (cfun
) == PROFILE_ABSENT
4314 && !flag_guess_branch_prob
)
4318 timevar_pop (TV_REBUILD_FREQUENCIES
);
4321 /* Perform a dry run of the branch prediction pass and report comparsion of
4322 the predicted and real profile into the dump file. */
4325 report_predictor_hitrates (void)
4329 loop_optimizer_init (LOOPS_NORMAL
);
4330 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4331 flow_loops_dump (dump_file
, NULL
, 0);
4333 nb_loops
= number_of_loops (cfun
);
4337 tree_estimate_probability (true);
4342 loop_optimizer_finalize ();
4345 /* Force edge E to be cold.
4346 If IMPOSSIBLE is true, for edge to have count and probability 0 otherwise
4347 keep low probability to represent possible error in a guess. This is used
4348 i.e. in case we predict loop to likely iterate given number of times but
4349 we are not 100% sure.
4351 This function locally updates profile without attempt to keep global
4352 consistency which cannot be reached in full generality without full profile
4353 rebuild from probabilities alone. Doing so is not necessarily a good idea
4354 because frequencies and counts may be more realistic then probabilities.
4356 In some cases (such as for elimination of early exits during full loop
4357 unrolling) the caller can ensure that profile will get consistent
4361 force_edge_cold (edge e
, bool impossible
)
4363 profile_count count_sum
= profile_count::zero ();
4364 profile_probability prob_sum
= profile_probability::never ();
4367 bool uninitialized_exit
= false;
4369 /* When branch probability guesses are not known, then do nothing. */
4370 if (!impossible
&& !e
->count ().initialized_p ())
4373 profile_probability goal
= (impossible
? profile_probability::never ()
4374 : profile_probability::very_unlikely ());
4376 /* If edge is already improbably or cold, just return. */
4377 if (e
->probability
<= goal
4378 && (!impossible
|| e
->count () == profile_count::zero ()))
4380 FOR_EACH_EDGE (e2
, ei
, e
->src
->succs
)
4383 if (e
->flags
& EDGE_FAKE
)
4385 if (e2
->count ().initialized_p ())
4386 count_sum
+= e2
->count ();
4387 if (e2
->probability
.initialized_p ())
4388 prob_sum
+= e2
->probability
;
4390 uninitialized_exit
= true;
4393 /* If we are not guessing profiles but have some other edges out,
4394 just assume the control flow goes elsewhere. */
4395 if (uninitialized_exit
)
4396 e
->probability
= goal
;
4397 /* If there are other edges out of e->src, redistribute probabilitity
4399 else if (prob_sum
> profile_probability::never ())
4401 if (!(e
->probability
< goal
))
4402 e
->probability
= goal
;
4404 profile_probability prob_comp
= prob_sum
/ e
->probability
.invert ();
4406 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4407 fprintf (dump_file
, "Making edge %i->%i %s by redistributing "
4408 "probability to other edges.\n",
4409 e
->src
->index
, e
->dest
->index
,
4410 impossible
? "impossible" : "cold");
4411 FOR_EACH_EDGE (e2
, ei
, e
->src
->succs
)
4414 e2
->probability
/= prob_comp
;
4416 if (current_ir_type () != IR_GIMPLE
4417 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
4418 update_br_prob_note (e
->src
);
4420 /* If all edges out of e->src are unlikely, the basic block itself
4424 if (prob_sum
== profile_probability::never ())
4425 e
->probability
= profile_probability::always ();
4429 e
->probability
= profile_probability::never ();
4430 /* If BB has some edges out that are not impossible, we cannot
4431 assume that BB itself is. */
4434 if (current_ir_type () != IR_GIMPLE
4435 && e
->src
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
))
4436 update_br_prob_note (e
->src
);
4437 if (e
->src
->count
== profile_count::zero ())
4439 if (count_sum
== profile_count::zero () && impossible
)
4442 if (e
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
4444 else if (current_ir_type () == IR_GIMPLE
)
4445 for (gimple_stmt_iterator gsi
= gsi_start_bb (e
->src
);
4446 !gsi_end_p (gsi
); gsi_next (&gsi
))
4448 if (stmt_can_terminate_bb_p (gsi_stmt (gsi
)))
4454 /* FIXME: Implement RTL path. */
4459 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4461 "Making bb %i impossible and dropping count to 0.\n",
4463 e
->src
->count
= profile_count::zero ();
4464 FOR_EACH_EDGE (e2
, ei
, e
->src
->preds
)
4465 force_edge_cold (e2
, impossible
);
4470 /* If we did not adjusting, the source basic block has no likely edeges
4471 leaving other direction. In that case force that bb cold, too.
4472 This in general is difficult task to do, but handle special case when
4473 BB has only one predecestor. This is common case when we are updating
4474 after loop transforms. */
4475 if (!(prob_sum
> profile_probability::never ())
4476 && count_sum
== profile_count::zero ()
4477 && single_pred_p (e
->src
) && e
->src
->count
.to_frequency (cfun
)
4478 > (impossible
? 0 : 1))
4480 int old_frequency
= e
->src
->count
.to_frequency (cfun
);
4481 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4482 fprintf (dump_file
, "Making bb %i %s.\n", e
->src
->index
,
4483 impossible
? "impossible" : "cold");
4484 int new_frequency
= MIN (e
->src
->count
.to_frequency (cfun
),
4485 impossible
? 0 : 1);
4487 e
->src
->count
= profile_count::zero ();
4489 e
->src
->count
= e
->count ().apply_scale (new_frequency
,
4491 force_edge_cold (single_pred_edge (e
->src
), impossible
);
4493 else if (dump_file
&& (dump_flags
& TDF_DETAILS
)
4494 && maybe_hot_bb_p (cfun
, e
->src
))
4495 fprintf (dump_file
, "Giving up on making bb %i %s.\n", e
->src
->index
,
4496 impossible
? "impossible" : "cold");
4500 /* Change E's probability to NEW_E_PROB, redistributing the probabilities
4501 of other outgoing edges proportionally.
4503 Note that this function does not change the profile counts of any
4504 basic blocks. The caller must do that instead, using whatever
4505 information it has about the region that needs updating. */
4508 change_edge_frequency (edge e
, profile_probability new_e_prob
)
4510 profile_probability old_e_prob
= e
->probability
;
4511 profile_probability old_other_prob
= old_e_prob
.invert ();
4512 profile_probability new_other_prob
= new_e_prob
.invert ();
4514 e
->probability
= new_e_prob
;
4515 profile_probability cumulative_prob
= new_e_prob
;
4517 unsigned int num_other
= EDGE_COUNT (e
->src
->succs
) - 1;
4520 FOR_EACH_EDGE (other_e
, ei
, e
->src
->succs
)
4525 /* Ensure that the probabilities add up to 1 without
4527 other_e
->probability
= cumulative_prob
.invert ();
4530 other_e
->probability
/= old_other_prob
;
4531 other_e
->probability
*= new_other_prob
;
4532 cumulative_prob
+= other_e
->probability
;
4539 namespace selftest
{
4541 /* Test that value range of predictor values defined in predict.def is
4542 within range (50, 100]. */
4544 struct branch_predictor
4550 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) { NAME, HITRATE },
4553 test_prediction_value_range ()
4555 branch_predictor predictors
[] = {
4556 #include "predict.def"
4557 { NULL
, PROB_UNINITIALIZED
}
4560 for (unsigned i
= 0; predictors
[i
].name
!= NULL
; i
++)
4562 if (predictors
[i
].probability
== PROB_UNINITIALIZED
)
4565 unsigned p
= 100 * predictors
[i
].probability
/ REG_BR_PROB_BASE
;
4566 ASSERT_TRUE (p
>= 50 && p
<= 100);
4570 #undef DEF_PREDICTOR
4572 /* Run all of the selfests within this file. */
4577 test_prediction_value_range ();
4580 } // namespace selftest
4581 #endif /* CHECKING_P. */