1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000, 2001, 2002, 2003, 2004 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 2, 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 COPYING. If not, write to the Free
18 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 [1] "Branch Prediction for Free"
24 Ball and Larus; PLDI '93.
25 [2] "Static Branch Frequency and Program Profile Analysis"
26 Wu and Larus; MICRO-27.
27 [3] "Corpus-based Static Branch Prediction"
28 Calder, Grunwald, Lindsay, Martin, Mozer, and Zorn; PLDI '95. */
33 #include "coretypes.h"
38 #include "hard-reg-set.h"
39 #include "basic-block.h"
40 #include "insn-config.h"
55 #include "tree-flow.h"
57 #include "tree-dump.h"
58 #include "tree-pass.h"
60 #include "tree-scalar-evolution.h"
63 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
64 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
65 static sreal real_zero
, real_one
, real_almost_one
, real_br_prob_base
,
66 real_inv_br_prob_base
, real_one_half
, real_bb_freq_max
;
68 /* Random guesstimation given names. */
69 #define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 10 - 1)
70 #define PROB_EVEN (REG_BR_PROB_BASE / 2)
71 #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
72 #define PROB_ALWAYS (REG_BR_PROB_BASE)
74 static void combine_predictions_for_insn (rtx
, basic_block
);
75 static void dump_prediction (FILE *, enum br_predictor
, int, basic_block
, int);
76 static void estimate_loops_at_level (struct loop
*loop
);
77 static void propagate_freq (struct loop
*);
78 static void estimate_bb_frequencies (struct loops
*);
79 static int counts_to_freqs (void);
80 static bool last_basic_block_p (basic_block
);
81 static void compute_function_frequency (void);
82 static void choose_function_section (void);
83 static bool can_predict_insn_p (rtx
);
85 /* Information we hold about each branch predictor.
86 Filled using information from predict.def. */
90 const char *const name
; /* Name used in the debugging dumps. */
91 const int hitrate
; /* Expected hitrate used by
92 predict_insn_def call. */
96 /* Use given predictor without Dempster-Shaffer theory if it matches
97 using first_match heuristics. */
98 #define PRED_FLAG_FIRST_MATCH 1
100 /* Recompute hitrate in percent to our representation. */
102 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
104 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
105 static const struct predictor_info predictor_info
[]= {
106 #include "predict.def"
108 /* Upper bound on predictors. */
113 /* Return true in case BB can be CPU intensive and should be optimized
114 for maximal performance. */
117 maybe_hot_bb_p (basic_block bb
)
119 if (profile_info
&& flag_branch_probabilities
121 < profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
)))
123 if (bb
->frequency
< BB_FREQ_MAX
/ PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
))
128 /* Return true in case BB is cold and should be optimized for size. */
131 probably_cold_bb_p (basic_block bb
)
133 if (profile_info
&& flag_branch_probabilities
135 < profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
)))
137 if (bb
->frequency
< BB_FREQ_MAX
/ PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
))
142 /* Return true in case BB is probably never executed. */
144 probably_never_executed_bb_p (basic_block bb
)
146 if (profile_info
&& flag_branch_probabilities
)
147 return ((bb
->count
+ profile_info
->runs
/ 2) / profile_info
->runs
) == 0;
151 /* Return true if the one of outgoing edges is already predicted by
155 rtl_predicted_by_p (basic_block bb
, enum br_predictor predictor
)
158 if (!INSN_P (BB_END (bb
)))
160 for (note
= REG_NOTES (BB_END (bb
)); note
; note
= XEXP (note
, 1))
161 if (REG_NOTE_KIND (note
) == REG_BR_PRED
162 && INTVAL (XEXP (XEXP (note
, 0), 0)) == (int)predictor
)
167 /* Return true if the one of outgoing edges is already predicted by
171 tree_predicted_by_p (basic_block bb
, enum br_predictor predictor
)
173 struct edge_prediction
*i
= bb_ann (bb
)->predictions
;
174 for (i
= bb_ann (bb
)->predictions
; i
; i
= i
->next
)
175 if (i
->predictor
== predictor
)
181 predict_insn (rtx insn
, enum br_predictor predictor
, int probability
)
183 if (!any_condjump_p (insn
))
185 if (!flag_guess_branch_prob
)
189 = gen_rtx_EXPR_LIST (REG_BR_PRED
,
190 gen_rtx_CONCAT (VOIDmode
,
191 GEN_INT ((int) predictor
),
192 GEN_INT ((int) probability
)),
196 /* Predict insn by given predictor. */
199 predict_insn_def (rtx insn
, enum br_predictor predictor
,
200 enum prediction taken
)
202 int probability
= predictor_info
[(int) predictor
].hitrate
;
205 probability
= REG_BR_PROB_BASE
- probability
;
207 predict_insn (insn
, predictor
, probability
);
210 /* Predict edge E with given probability if possible. */
213 rtl_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
216 last_insn
= BB_END (e
->src
);
218 /* We can store the branch prediction information only about
219 conditional jumps. */
220 if (!any_condjump_p (last_insn
))
223 /* We always store probability of branching. */
224 if (e
->flags
& EDGE_FALLTHRU
)
225 probability
= REG_BR_PROB_BASE
- probability
;
227 predict_insn (last_insn
, predictor
, probability
);
230 /* Predict edge E with the given PROBABILITY. */
232 tree_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
234 struct edge_prediction
*i
= ggc_alloc (sizeof (struct edge_prediction
));
236 i
->next
= bb_ann (e
->src
)->predictions
;
237 bb_ann (e
->src
)->predictions
= i
;
238 i
->probability
= probability
;
239 i
->predictor
= predictor
;
243 /* Return true when we can store prediction on insn INSN.
244 At the moment we represent predictions only on conditional
245 jumps, not at computed jump or other complicated cases. */
247 can_predict_insn_p (rtx insn
)
249 return (JUMP_P (insn
)
250 && any_condjump_p (insn
)
251 && BLOCK_FOR_INSN (insn
)->succ
->succ_next
);
254 /* Predict edge E by given predictor if possible. */
257 predict_edge_def (edge e
, enum br_predictor predictor
,
258 enum prediction taken
)
260 int probability
= predictor_info
[(int) predictor
].hitrate
;
263 probability
= REG_BR_PROB_BASE
- probability
;
265 predict_edge (e
, predictor
, probability
);
268 /* Invert all branch predictions or probability notes in the INSN. This needs
269 to be done each time we invert the condition used by the jump. */
272 invert_br_probabilities (rtx insn
)
276 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
277 if (REG_NOTE_KIND (note
) == REG_BR_PROB
)
278 XEXP (note
, 0) = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (note
, 0)));
279 else if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
280 XEXP (XEXP (note
, 0), 1)
281 = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (XEXP (note
, 0), 1)));
284 /* Dump information about the branch prediction to the output file. */
287 dump_prediction (FILE *file
, enum br_predictor predictor
, int probability
,
288 basic_block bb
, int used
)
295 while (e
&& (e
->flags
& EDGE_FALLTHRU
))
298 fprintf (file
, " %s heuristics%s: %.1f%%",
299 predictor_info
[predictor
].name
,
300 used
? "" : " (ignored)", probability
* 100.0 / REG_BR_PROB_BASE
);
304 fprintf (file
, " exec ");
305 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, bb
->count
);
308 fprintf (file
, " hit ");
309 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, e
->count
);
310 fprintf (file
, " (%.1f%%)", e
->count
* 100.0 / bb
->count
);
314 fprintf (file
, "\n");
317 /* We can not predict the probabilities of outgoing edges of bb. Set them
318 evenly and hope for the best. */
320 set_even_probabilities (basic_block bb
)
325 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
326 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
328 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
329 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
330 e
->probability
= (REG_BR_PROB_BASE
+ nedges
/ 2) / nedges
;
335 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
336 note if not already present. Remove now useless REG_BR_PRED notes. */
339 combine_predictions_for_insn (rtx insn
, basic_block bb
)
344 int best_probability
= PROB_EVEN
;
345 int best_predictor
= END_PREDICTORS
;
346 int combined_probability
= REG_BR_PROB_BASE
/ 2;
348 bool first_match
= false;
351 if (!can_predict_insn_p (insn
))
353 set_even_probabilities (bb
);
357 prob_note
= find_reg_note (insn
, REG_BR_PROB
, 0);
358 pnote
= ®_NOTES (insn
);
360 fprintf (dump_file
, "Predictions for insn %i bb %i\n", INSN_UID (insn
),
363 /* We implement "first match" heuristics and use probability guessed
364 by predictor with smallest index. */
365 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
366 if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
368 int predictor
= INTVAL (XEXP (XEXP (note
, 0), 0));
369 int probability
= INTVAL (XEXP (XEXP (note
, 0), 1));
372 if (best_predictor
> predictor
)
373 best_probability
= probability
, best_predictor
= predictor
;
375 d
= (combined_probability
* probability
376 + (REG_BR_PROB_BASE
- combined_probability
)
377 * (REG_BR_PROB_BASE
- probability
));
379 /* Use FP math to avoid overflows of 32bit integers. */
381 /* If one probability is 0% and one 100%, avoid division by zero. */
382 combined_probability
= REG_BR_PROB_BASE
/ 2;
384 combined_probability
= (((double) combined_probability
) * probability
385 * REG_BR_PROB_BASE
/ d
+ 0.5);
388 /* Decide which heuristic to use. In case we didn't match anything,
389 use no_prediction heuristic, in case we did match, use either
390 first match or Dempster-Shaffer theory depending on the flags. */
392 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
396 dump_prediction (dump_file
, PRED_NO_PREDICTION
,
397 combined_probability
, bb
, true);
400 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
,
402 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
,
407 combined_probability
= best_probability
;
408 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
412 if (REG_NOTE_KIND (*pnote
) == REG_BR_PRED
)
414 int predictor
= INTVAL (XEXP (XEXP (*pnote
, 0), 0));
415 int probability
= INTVAL (XEXP (XEXP (*pnote
, 0), 1));
417 dump_prediction (dump_file
, predictor
, probability
, bb
,
418 !first_match
|| best_predictor
== predictor
);
419 *pnote
= XEXP (*pnote
, 1);
422 pnote
= &XEXP (*pnote
, 1);
428 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
429 GEN_INT (combined_probability
), REG_NOTES (insn
));
431 /* Save the prediction into CFG in case we are seeing non-degenerated
433 if (bb
->succ
->succ_next
)
435 BRANCH_EDGE (bb
)->probability
= combined_probability
;
436 FALLTHRU_EDGE (bb
)->probability
437 = REG_BR_PROB_BASE
- combined_probability
;
442 /* Combine predictions into single probability and store them into CFG.
443 Remove now useless prediction entries. */
446 combine_predictions_for_bb (FILE *file
, basic_block bb
)
448 int best_probability
= PROB_EVEN
;
449 int best_predictor
= END_PREDICTORS
;
450 int combined_probability
= REG_BR_PROB_BASE
/ 2;
452 bool first_match
= false;
454 struct edge_prediction
*pred
;
456 edge e
, first
= NULL
, second
= NULL
;
458 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
459 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
462 if (first
&& !second
)
468 /* When there is no successor or only one choice, prediction is easy.
470 We are lazy for now and predict only basic blocks with two outgoing
471 edges. It is possible to predict generic case too, but we have to
472 ignore first match heuristics and do more involved combining. Implement
477 set_even_probabilities (bb
);
478 bb_ann (bb
)->predictions
= NULL
;
480 fprintf (file
, "%i edges in bb %i predicted to even probabilities\n",
486 fprintf (file
, "Predictions for bb %i\n", bb
->index
);
488 /* We implement "first match" heuristics and use probability guessed
489 by predictor with smallest index. */
490 for (pred
= bb_ann (bb
)->predictions
; pred
; pred
= pred
->next
)
492 int predictor
= pred
->predictor
;
493 int probability
= pred
->probability
;
495 if (pred
->edge
!= first
)
496 probability
= REG_BR_PROB_BASE
- probability
;
499 if (best_predictor
> predictor
)
500 best_probability
= probability
, best_predictor
= predictor
;
502 d
= (combined_probability
* probability
503 + (REG_BR_PROB_BASE
- combined_probability
)
504 * (REG_BR_PROB_BASE
- probability
));
506 /* Use FP math to avoid overflows of 32bit integers. */
508 /* If one probability is 0% and one 100%, avoid division by zero. */
509 combined_probability
= REG_BR_PROB_BASE
/ 2;
511 combined_probability
= (((double) combined_probability
) * probability
512 * REG_BR_PROB_BASE
/ d
+ 0.5);
515 /* Decide which heuristic to use. In case we didn't match anything,
516 use no_prediction heuristic, in case we did match, use either
517 first match or Dempster-Shaffer theory depending on the flags. */
519 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
523 dump_prediction (file
, PRED_NO_PREDICTION
, combined_probability
, bb
, true);
526 dump_prediction (file
, PRED_DS_THEORY
, combined_probability
, bb
,
528 dump_prediction (file
, PRED_FIRST_MATCH
, best_probability
, bb
,
533 combined_probability
= best_probability
;
534 dump_prediction (file
, PRED_COMBINED
, combined_probability
, bb
, true);
536 for (pred
= bb_ann (bb
)->predictions
; pred
; pred
= pred
->next
)
538 int predictor
= pred
->predictor
;
539 int probability
= pred
->probability
;
541 if (pred
->edge
!= bb
->succ
)
542 probability
= REG_BR_PROB_BASE
- probability
;
543 dump_prediction (file
, predictor
, probability
, bb
,
544 !first_match
|| best_predictor
== predictor
);
546 bb_ann (bb
)->predictions
= NULL
;
550 first
->probability
= combined_probability
;
551 second
->probability
= REG_BR_PROB_BASE
- combined_probability
;
555 /* Predict edge probabilities by exploiting loop structure.
556 When RTLSIMPLELOOPS is set, attempt to count number of iterations by analyzing
557 RTL otherwise use tree based approach. */
559 predict_loops (struct loops
*loops_info
, bool rtlsimpleloops
)
564 scev_initialize (loops_info
);
566 /* Try to predict out blocks in a loop that are not part of a
568 for (i
= 1; i
< loops_info
->num
; i
++)
570 basic_block bb
, *bbs
;
573 struct loop
*loop
= loops_info
->parray
[i
];
574 struct niter_desc desc
;
575 unsigned HOST_WIDE_INT niter
;
577 flow_loop_scan (loop
, LOOP_EXIT_EDGES
);
578 exits
= loop
->num_exits
;
582 iv_analysis_loop_init (loop
);
583 find_simple_exit (loop
, &desc
);
585 if (desc
.simple_p
&& desc
.const_iter
)
588 niter
= desc
.niter
+ 1;
589 if (niter
== 0) /* We might overflow here. */
592 prob
= (REG_BR_PROB_BASE
593 - (REG_BR_PROB_BASE
+ niter
/2) / niter
);
594 /* Branch prediction algorithm gives 0 frequency for everything
595 after the end of loop for loop having 0 probability to finish. */
596 if (prob
== REG_BR_PROB_BASE
)
597 prob
= REG_BR_PROB_BASE
- 1;
598 predict_edge (desc
.in_edge
, PRED_LOOP_ITERATIONS
,
606 struct tree_niter_desc niter_desc
;
608 exits
= get_loop_exit_edges (loop
, &n_exits
);
609 for (j
= 0; j
< n_exits
; j
++)
613 if (number_of_iterations_exit (loop
, exits
[j
], &niter_desc
))
614 niter
= niter_desc
.niter
;
615 if (!niter
|| TREE_CODE (niter_desc
.niter
) != INTEGER_CST
)
616 niter
= loop_niter_by_eval (loop
, exits
[j
]);
618 if (TREE_CODE (niter
) == INTEGER_CST
)
621 if (host_integerp (niter
, 1)
622 && tree_int_cst_lt (niter
,
623 build_int_cstu (NULL_TREE
,
624 REG_BR_PROB_BASE
- 1)))
626 HOST_WIDE_INT nitercst
= tree_low_cst (niter
, 1) + 1;
627 probability
= (REG_BR_PROB_BASE
+ nitercst
/ 2) / nitercst
;
632 predict_edge (exits
[j
], PRED_LOOP_ITERATIONS
, probability
);
639 bbs
= get_loop_body (loop
);
641 for (j
= 0; j
< loop
->num_nodes
; j
++)
643 int header_found
= 0;
648 /* Bypass loop heuristics on continue statement. These
649 statements construct loops via "non-loop" constructs
650 in the source language and are better to be handled
652 if ((rtlsimpleloops
&& !can_predict_insn_p (BB_END (bb
)))
653 || predicted_by_p (bb
, PRED_CONTINUE
))
656 /* Loop branch heuristics - predict an edge back to a
657 loop's head as taken. */
658 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
659 if (e
->dest
== loop
->header
660 && e
->src
== loop
->latch
)
663 predict_edge_def (e
, PRED_LOOP_BRANCH
, TAKEN
);
666 /* Loop exit heuristics - predict an edge exiting the loop if the
667 conditional has no loop header successors as not taken. */
669 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
670 if (e
->dest
->index
< 0
671 || !flow_bb_inside_loop_p (loop
, e
->dest
))
675 - predictor_info
[(int) PRED_LOOP_EXIT
].hitrate
)
679 /* Free basic blocks from get_loop_body. */
687 /* Attempt to predict probabilities of BB outgoing edges using local
690 bb_estimate_probability_locally (basic_block bb
)
692 rtx last_insn
= BB_END (bb
);
695 if (! can_predict_insn_p (last_insn
))
697 cond
= get_condition (last_insn
, NULL
, false, false);
701 /* Try "pointer heuristic."
702 A comparison ptr == 0 is predicted as false.
703 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
704 if (COMPARISON_P (cond
)
705 && ((REG_P (XEXP (cond
, 0)) && REG_POINTER (XEXP (cond
, 0)))
706 || (REG_P (XEXP (cond
, 1)) && REG_POINTER (XEXP (cond
, 1)))))
708 if (GET_CODE (cond
) == EQ
)
709 predict_insn_def (last_insn
, PRED_POINTER
, NOT_TAKEN
);
710 else if (GET_CODE (cond
) == NE
)
711 predict_insn_def (last_insn
, PRED_POINTER
, TAKEN
);
715 /* Try "opcode heuristic."
716 EQ tests are usually false and NE tests are usually true. Also,
717 most quantities are positive, so we can make the appropriate guesses
718 about signed comparisons against zero. */
719 switch (GET_CODE (cond
))
722 /* Unconditional branch. */
723 predict_insn_def (last_insn
, PRED_UNCONDITIONAL
,
724 cond
== const0_rtx
? NOT_TAKEN
: TAKEN
);
729 /* Floating point comparisons appears to behave in a very
730 unpredictable way because of special role of = tests in
732 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
734 /* Comparisons with 0 are often used for booleans and there is
735 nothing useful to predict about them. */
736 else if (XEXP (cond
, 1) == const0_rtx
737 || XEXP (cond
, 0) == const0_rtx
)
740 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, NOT_TAKEN
);
745 /* Floating point comparisons appears to behave in a very
746 unpredictable way because of special role of = tests in
748 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
750 /* Comparisons with 0 are often used for booleans and there is
751 nothing useful to predict about them. */
752 else if (XEXP (cond
, 1) == const0_rtx
753 || XEXP (cond
, 0) == const0_rtx
)
756 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, TAKEN
);
760 predict_insn_def (last_insn
, PRED_FPOPCODE
, TAKEN
);
764 predict_insn_def (last_insn
, PRED_FPOPCODE
, NOT_TAKEN
);
769 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
770 || XEXP (cond
, 1) == constm1_rtx
)
771 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, NOT_TAKEN
);
776 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
777 || XEXP (cond
, 1) == constm1_rtx
)
778 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, TAKEN
);
786 /* Statically estimate the probability that a branch will be taken and produce
787 estimated profile. When profile feedback is present never executed portions
788 of function gets estimated. */
791 estimate_probability (struct loops
*loops_info
)
795 connect_infinite_loops_to_exit ();
796 calculate_dominance_info (CDI_DOMINATORS
);
797 calculate_dominance_info (CDI_POST_DOMINATORS
);
799 predict_loops (loops_info
, true);
803 /* Attempt to predict conditional jumps using a number of heuristics. */
806 rtx last_insn
= BB_END (bb
);
809 if (! can_predict_insn_p (last_insn
))
812 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
814 /* Predict early returns to be probable, as we've already taken
815 care for error returns and other are often used for fast paths
817 if ((e
->dest
== EXIT_BLOCK_PTR
818 || (e
->dest
->succ
&& !e
->dest
->succ
->succ_next
819 && e
->dest
->succ
->dest
== EXIT_BLOCK_PTR
))
820 && !predicted_by_p (bb
, PRED_NULL_RETURN
)
821 && !predicted_by_p (bb
, PRED_CONST_RETURN
)
822 && !predicted_by_p (bb
, PRED_NEGATIVE_RETURN
)
823 && !last_basic_block_p (e
->dest
))
824 predict_edge_def (e
, PRED_EARLY_RETURN
, TAKEN
);
826 /* Look for block we are guarding (i.e. we dominate it,
827 but it doesn't postdominate us). */
828 if (e
->dest
!= EXIT_BLOCK_PTR
&& e
->dest
!= bb
829 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
830 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
834 /* The call heuristic claims that a guarded function call
835 is improbable. This is because such calls are often used
836 to signal exceptional situations such as printing error
838 for (insn
= BB_HEAD (e
->dest
); insn
!= NEXT_INSN (BB_END (e
->dest
));
839 insn
= NEXT_INSN (insn
))
841 /* Constant and pure calls are hardly used to signalize
842 something exceptional. */
843 && ! CONST_OR_PURE_CALL_P (insn
))
845 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
850 bb_estimate_probability_locally (bb
);
853 /* Attach the combined probability to each conditional jump. */
855 combine_predictions_for_insn (BB_END (bb
), bb
);
857 remove_fake_edges ();
858 estimate_bb_frequencies (loops_info
);
859 free_dominance_info (CDI_POST_DOMINATORS
);
860 if (profile_status
== PROFILE_ABSENT
)
861 profile_status
= PROFILE_GUESSED
;
864 /* Set edge->probability for each successor edge of BB. */
866 guess_outgoing_edge_probabilities (basic_block bb
)
868 bb_estimate_probability_locally (bb
);
869 combine_predictions_for_insn (BB_END (bb
), bb
);
872 /* Return constant EXPR will likely have at execution time, NULL if unknown.
873 The function is used by builtin_expect branch predictor so the evidence
874 must come from this construct and additional possible constant folding.
876 We may want to implement more involved value guess (such as value range
877 propagation based prediction), but such tricks shall go to new
881 expr_expected_value (tree expr
, bitmap visited
)
883 if (TREE_CONSTANT (expr
))
885 else if (TREE_CODE (expr
) == SSA_NAME
)
887 tree def
= SSA_NAME_DEF_STMT (expr
);
889 /* If we were already here, break the infinite cycle. */
890 if (bitmap_bit_p (visited
, SSA_NAME_VERSION (expr
)))
892 bitmap_set_bit (visited
, SSA_NAME_VERSION (expr
));
894 if (TREE_CODE (def
) == PHI_NODE
)
896 /* All the arguments of the PHI node must have the same constant
899 tree val
= NULL
, new_val
;
901 for (i
= 0; i
< PHI_NUM_ARGS (def
); i
++)
903 tree arg
= PHI_ARG_DEF (def
, i
);
905 /* If this PHI has itself as an argument, we cannot
906 determine the string length of this argument. However,
907 if we can find a expected constant value for the other
908 PHI args then we can still be sure that this is
909 likely a constant. So be optimistic and just
910 continue with the next argument. */
911 if (arg
== PHI_RESULT (def
))
914 new_val
= expr_expected_value (arg
, visited
);
919 else if (!operand_equal_p (val
, new_val
, false))
924 if (TREE_CODE (def
) != MODIFY_EXPR
|| TREE_OPERAND (def
, 0) != expr
)
926 return expr_expected_value (TREE_OPERAND (def
, 1), visited
);
928 else if (TREE_CODE (expr
) == CALL_EXPR
)
930 tree decl
= get_callee_fndecl (expr
);
933 if (DECL_BUILT_IN (decl
) && DECL_FUNCTION_CODE (decl
) == BUILT_IN_EXPECT
)
935 tree arglist
= TREE_OPERAND (expr
, 1);
938 if (arglist
== NULL_TREE
939 || TREE_CHAIN (arglist
) == NULL_TREE
)
941 val
= TREE_VALUE (TREE_CHAIN (TREE_OPERAND (expr
, 1)));
942 if (TREE_CONSTANT (val
))
944 return TREE_VALUE (TREE_CHAIN (TREE_OPERAND (expr
, 1)));
947 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
950 op0
= expr_expected_value (TREE_OPERAND (expr
, 0), visited
);
953 op1
= expr_expected_value (TREE_OPERAND (expr
, 1), visited
);
956 res
= fold (build (TREE_CODE (expr
), TREE_TYPE (expr
), op0
, op1
));
957 if (TREE_CONSTANT (res
))
961 if (UNARY_CLASS_P (expr
))
964 op0
= expr_expected_value (TREE_OPERAND (expr
, 0), visited
);
967 res
= fold (build1 (TREE_CODE (expr
), TREE_TYPE (expr
), op0
));
968 if (TREE_CONSTANT (res
))
975 /* Get rid of all builtin_expect calls we no longer need. */
977 strip_builtin_expect (void)
982 block_stmt_iterator bi
;
983 for (bi
= bsi_start (bb
); !bsi_end_p (bi
); bsi_next (&bi
))
985 tree stmt
= bsi_stmt (bi
);
989 if (TREE_CODE (stmt
) == MODIFY_EXPR
990 && TREE_CODE (TREE_OPERAND (stmt
, 1)) == CALL_EXPR
991 && (fndecl
= get_callee_fndecl (TREE_OPERAND (stmt
, 1)))
992 && DECL_BUILT_IN (fndecl
)
993 && DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_EXPECT
994 && (arglist
= TREE_OPERAND (TREE_OPERAND (stmt
, 1), 1))
995 && TREE_CHAIN (arglist
))
997 TREE_OPERAND (stmt
, 1) = TREE_VALUE (arglist
);
1004 /* Predict using opcode of the last statement in basic block. */
1006 tree_predict_by_opcode (basic_block bb
)
1008 tree stmt
= last_stmt (bb
);
1016 if (!stmt
|| TREE_CODE (stmt
) != COND_EXPR
)
1018 for (then_edge
= bb
->succ
; then_edge
; then_edge
= then_edge
->succ_next
)
1019 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1021 cond
= TREE_OPERAND (stmt
, 0);
1022 if (!COMPARISON_CLASS_P (cond
))
1024 op0
= TREE_OPERAND (cond
, 0);
1025 type
= TREE_TYPE (op0
);
1026 visited
= BITMAP_XMALLOC ();
1027 val
= expr_expected_value (cond
, visited
);
1028 BITMAP_XFREE (visited
);
1031 if (integer_zerop (val
))
1032 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, NOT_TAKEN
);
1034 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, TAKEN
);
1037 /* Try "pointer heuristic."
1038 A comparison ptr == 0 is predicted as false.
1039 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1040 if (POINTER_TYPE_P (type
))
1042 if (TREE_CODE (cond
) == EQ_EXPR
)
1043 predict_edge_def (then_edge
, PRED_TREE_POINTER
, NOT_TAKEN
);
1044 else if (TREE_CODE (cond
) == NE_EXPR
)
1045 predict_edge_def (then_edge
, PRED_TREE_POINTER
, TAKEN
);
1049 /* Try "opcode heuristic."
1050 EQ tests are usually false and NE tests are usually true. Also,
1051 most quantities are positive, so we can make the appropriate guesses
1052 about signed comparisons against zero. */
1053 switch (TREE_CODE (cond
))
1057 /* Floating point comparisons appears to behave in a very
1058 unpredictable way because of special role of = tests in
1060 if (FLOAT_TYPE_P (type
))
1062 /* Comparisons with 0 are often used for booleans and there is
1063 nothing useful to predict about them. */
1064 else if (integer_zerop (op0
)
1065 || integer_zerop (TREE_OPERAND (cond
, 1)))
1068 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, NOT_TAKEN
);
1073 /* Floating point comparisons appears to behave in a very
1074 unpredictable way because of special role of = tests in
1076 if (FLOAT_TYPE_P (type
))
1078 /* Comparisons with 0 are often used for booleans and there is
1079 nothing useful to predict about them. */
1080 else if (integer_zerop (op0
)
1081 || integer_zerop (TREE_OPERAND (cond
, 1)))
1084 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, TAKEN
);
1088 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, TAKEN
);
1091 case UNORDERED_EXPR
:
1092 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, NOT_TAKEN
);
1097 if (integer_zerop (TREE_OPERAND (cond
, 1))
1098 || integer_onep (TREE_OPERAND (cond
, 1))
1099 || integer_all_onesp (TREE_OPERAND (cond
, 1))
1100 || real_zerop (TREE_OPERAND (cond
, 1))
1101 || real_onep (TREE_OPERAND (cond
, 1))
1102 || real_minus_onep (TREE_OPERAND (cond
, 1)))
1103 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, NOT_TAKEN
);
1108 if (integer_zerop (TREE_OPERAND (cond
, 1))
1109 || integer_onep (TREE_OPERAND (cond
, 1))
1110 || integer_all_onesp (TREE_OPERAND (cond
, 1))
1111 || real_zerop (TREE_OPERAND (cond
, 1))
1112 || real_onep (TREE_OPERAND (cond
, 1))
1113 || real_minus_onep (TREE_OPERAND (cond
, 1)))
1114 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, TAKEN
);
1122 /* Predict branch probabilities and estimate profile of the tree CFG. */
1124 tree_estimate_probability (void)
1127 struct loops loops_info
;
1129 flow_loops_find (&loops_info
, LOOP_TREE
);
1130 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1131 flow_loops_dump (&loops_info
, dump_file
, NULL
, 0);
1133 connect_infinite_loops_to_exit ();
1134 calculate_dominance_info (CDI_DOMINATORS
);
1135 calculate_dominance_info (CDI_POST_DOMINATORS
);
1137 predict_loops (&loops_info
, false);
1143 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
1145 /* Predict early returns to be probable, as we've already taken
1146 care for error returns and other are often used for fast paths
1147 trought function. */
1148 if ((e
->dest
== EXIT_BLOCK_PTR
1149 || (e
->dest
->succ
&& !e
->dest
->succ
->succ_next
1150 && e
->dest
->succ
->dest
== EXIT_BLOCK_PTR
))
1151 && !predicted_by_p (bb
, PRED_NULL_RETURN
)
1152 && !predicted_by_p (bb
, PRED_CONST_RETURN
)
1153 && !predicted_by_p (bb
, PRED_NEGATIVE_RETURN
)
1154 && !last_basic_block_p (e
->dest
))
1155 predict_edge_def (e
, PRED_EARLY_RETURN
, TAKEN
);
1157 /* Look for block we are guarding (i.e. we dominate it,
1158 but it doesn't postdominate us). */
1159 if (e
->dest
!= EXIT_BLOCK_PTR
&& e
->dest
!= bb
1160 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
1161 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
1163 block_stmt_iterator bi
;
1165 /* The call heuristic claims that a guarded function call
1166 is improbable. This is because such calls are often used
1167 to signal exceptional situations such as printing error
1169 for (bi
= bsi_start (e
->dest
); !bsi_end_p (bi
);
1172 tree stmt
= bsi_stmt (bi
);
1173 if ((TREE_CODE (stmt
) == CALL_EXPR
1174 || (TREE_CODE (stmt
) == MODIFY_EXPR
1175 && TREE_CODE (TREE_OPERAND (stmt
, 1)) == CALL_EXPR
))
1176 /* Constant and pure calls are hardly used to signalize
1177 something exceptional. */
1178 && TREE_SIDE_EFFECTS (stmt
))
1180 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
1186 tree_predict_by_opcode (bb
);
1189 combine_predictions_for_bb (dump_file
, bb
);
1191 if (0) /* FIXME: Enable once we are pass down the profile to RTL level. */
1192 strip_builtin_expect ();
1193 estimate_bb_frequencies (&loops_info
);
1194 free_dominance_info (CDI_POST_DOMINATORS
);
1195 remove_fake_exit_edges ();
1196 flow_loops_free (&loops_info
);
1197 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1198 dump_tree_cfg (dump_file
, dump_flags
);
1199 if (profile_status
== PROFILE_ABSENT
)
1200 profile_status
= PROFILE_GUESSED
;
1203 /* __builtin_expect dropped tokens into the insn stream describing expected
1204 values of registers. Generate branch probabilities based off these
1208 expected_value_to_br_prob (void)
1210 rtx insn
, cond
, ev
= NULL_RTX
, ev_reg
= NULL_RTX
;
1212 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
1214 switch (GET_CODE (insn
))
1217 /* Look for expected value notes. */
1218 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_EXPECTED_VALUE
)
1220 ev
= NOTE_EXPECTED_VALUE (insn
);
1221 ev_reg
= XEXP (ev
, 0);
1227 /* Never propagate across labels. */
1232 /* Look for simple conditional branches. If we haven't got an
1233 expected value yet, no point going further. */
1234 if (!JUMP_P (insn
) || ev
== NULL_RTX
1235 || ! any_condjump_p (insn
))
1240 /* Look for insns that clobber the EV register. */
1241 if (ev
&& reg_set_p (ev_reg
, insn
))
1246 /* Collect the branch condition, hopefully relative to EV_REG. */
1247 /* ??? At present we'll miss things like
1248 (expected_value (eq r70 0))
1250 (set r80 (lt r70 r71))
1251 (set pc (if_then_else (ne r80 0) ...))
1252 as canonicalize_condition will render this to us as
1254 Could use cselib to try and reduce this further. */
1255 cond
= XEXP (SET_SRC (pc_set (insn
)), 0);
1256 cond
= canonicalize_condition (insn
, cond
, 0, NULL
, ev_reg
,
1258 if (! cond
|| XEXP (cond
, 0) != ev_reg
1259 || GET_CODE (XEXP (cond
, 1)) != CONST_INT
)
1262 /* Substitute and simplify. Given that the expression we're
1263 building involves two constants, we should wind up with either
1265 cond
= gen_rtx_fmt_ee (GET_CODE (cond
), VOIDmode
,
1266 XEXP (ev
, 1), XEXP (cond
, 1));
1267 cond
= simplify_rtx (cond
);
1269 /* Turn the condition into a scaled branch probability. */
1270 if (cond
!= const_true_rtx
&& cond
!= const0_rtx
)
1272 predict_insn_def (insn
, PRED_BUILTIN_EXPECT
,
1273 cond
== const_true_rtx
? TAKEN
: NOT_TAKEN
);
1277 /* Check whether this is the last basic block of function. Commonly
1278 there is one extra common cleanup block. */
1280 last_basic_block_p (basic_block bb
)
1282 if (bb
== EXIT_BLOCK_PTR
)
1285 return (bb
->next_bb
== EXIT_BLOCK_PTR
1286 || (bb
->next_bb
->next_bb
== EXIT_BLOCK_PTR
1287 && bb
->succ
&& !bb
->succ
->succ_next
1288 && bb
->succ
->dest
->next_bb
== EXIT_BLOCK_PTR
));
1291 /* This is used to carry information about basic blocks. It is
1292 attached to the AUX field of the standard CFG block. */
1294 typedef struct block_info_def
1296 /* Estimated frequency of execution of basic_block. */
1299 /* To keep queue of basic blocks to process. */
1302 /* True if block needs to be visited in propagate_freq. */
1303 unsigned int tovisit
:1;
1305 /* Number of predecessors we need to visit first. */
1309 /* Similar information for edges. */
1310 typedef struct edge_info_def
1312 /* In case edge is an loopback edge, the probability edge will be reached
1313 in case header is. Estimated number of iterations of the loop can be
1314 then computed as 1 / (1 - back_edge_prob). */
1315 sreal back_edge_prob
;
1316 /* True if the edge is an loopback edge in the natural loop. */
1317 unsigned int back_edge
:1;
1320 #define BLOCK_INFO(B) ((block_info) (B)->aux)
1321 #define EDGE_INFO(E) ((edge_info) (E)->aux)
1323 /* Helper function for estimate_bb_frequencies.
1324 Propagate the frequencies for LOOP. */
1327 propagate_freq (struct loop
*loop
)
1329 basic_block head
= loop
->header
;
1335 /* For each basic block we need to visit count number of his predecessors
1336 we need to visit first. */
1337 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1339 if (BLOCK_INFO (bb
)->tovisit
)
1343 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
1344 if (BLOCK_INFO (e
->src
)->tovisit
&& !(e
->flags
& EDGE_DFS_BACK
))
1346 else if (BLOCK_INFO (e
->src
)->tovisit
1347 && dump_file
&& !EDGE_INFO (e
)->back_edge
)
1349 "Irreducible region hit, ignoring edge to %i->%i\n",
1350 e
->src
->index
, bb
->index
);
1351 BLOCK_INFO (bb
)->npredecessors
= count
;
1355 memcpy (&BLOCK_INFO (head
)->frequency
, &real_one
, sizeof (real_one
));
1357 for (bb
= head
; bb
; bb
= nextbb
)
1359 sreal cyclic_probability
, frequency
;
1361 memcpy (&cyclic_probability
, &real_zero
, sizeof (real_zero
));
1362 memcpy (&frequency
, &real_zero
, sizeof (real_zero
));
1364 nextbb
= BLOCK_INFO (bb
)->next
;
1365 BLOCK_INFO (bb
)->next
= NULL
;
1367 /* Compute frequency of basic block. */
1370 #ifdef ENABLE_CHECKING
1371 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
1372 if (BLOCK_INFO (e
->src
)->tovisit
&& !(e
->flags
& EDGE_DFS_BACK
))
1376 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
1377 if (EDGE_INFO (e
)->back_edge
)
1379 sreal_add (&cyclic_probability
, &cyclic_probability
,
1380 &EDGE_INFO (e
)->back_edge_prob
);
1382 else if (!(e
->flags
& EDGE_DFS_BACK
))
1386 /* frequency += (e->probability
1387 * BLOCK_INFO (e->src)->frequency /
1388 REG_BR_PROB_BASE); */
1390 sreal_init (&tmp
, e
->probability
, 0);
1391 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (e
->src
)->frequency
);
1392 sreal_mul (&tmp
, &tmp
, &real_inv_br_prob_base
);
1393 sreal_add (&frequency
, &frequency
, &tmp
);
1396 if (sreal_compare (&cyclic_probability
, &real_zero
) == 0)
1398 memcpy (&BLOCK_INFO (bb
)->frequency
, &frequency
,
1399 sizeof (frequency
));
1403 if (sreal_compare (&cyclic_probability
, &real_almost_one
) > 0)
1405 memcpy (&cyclic_probability
, &real_almost_one
,
1406 sizeof (real_almost_one
));
1409 /* BLOCK_INFO (bb)->frequency = frequency
1410 / (1 - cyclic_probability) */
1412 sreal_sub (&cyclic_probability
, &real_one
, &cyclic_probability
);
1413 sreal_div (&BLOCK_INFO (bb
)->frequency
,
1414 &frequency
, &cyclic_probability
);
1418 BLOCK_INFO (bb
)->tovisit
= 0;
1420 /* Compute back edge frequencies. */
1421 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
1422 if (e
->dest
== head
)
1426 /* EDGE_INFO (e)->back_edge_prob
1427 = ((e->probability * BLOCK_INFO (bb)->frequency)
1428 / REG_BR_PROB_BASE); */
1430 sreal_init (&tmp
, e
->probability
, 0);
1431 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (bb
)->frequency
);
1432 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
1433 &tmp
, &real_inv_br_prob_base
);
1436 /* Propagate to successor blocks. */
1437 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
1438 if (!(e
->flags
& EDGE_DFS_BACK
)
1439 && BLOCK_INFO (e
->dest
)->npredecessors
)
1441 BLOCK_INFO (e
->dest
)->npredecessors
--;
1442 if (!BLOCK_INFO (e
->dest
)->npredecessors
)
1447 BLOCK_INFO (last
)->next
= e
->dest
;
1455 /* Estimate probabilities of loopback edges in loops at same nest level. */
1458 estimate_loops_at_level (struct loop
*first_loop
)
1462 for (loop
= first_loop
; loop
; loop
= loop
->next
)
1468 estimate_loops_at_level (loop
->inner
);
1470 if (loop
->latch
->succ
) /* Do not do this for dummy function loop. */
1472 /* Find current loop back edge and mark it. */
1473 e
= loop_latch_edge (loop
);
1474 EDGE_INFO (e
)->back_edge
= 1;
1477 bbs
= get_loop_body (loop
);
1478 for (i
= 0; i
< loop
->num_nodes
; i
++)
1479 BLOCK_INFO (bbs
[i
])->tovisit
= 1;
1481 propagate_freq (loop
);
1485 /* Convert counts measured by profile driven feedback to frequencies.
1486 Return nonzero iff there was any nonzero execution count. */
1489 counts_to_freqs (void)
1491 gcov_type count_max
, true_count_max
= 0;
1495 true_count_max
= MAX (bb
->count
, true_count_max
);
1497 count_max
= MAX (true_count_max
, 1);
1498 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1499 bb
->frequency
= (bb
->count
* BB_FREQ_MAX
+ count_max
/ 2) / count_max
;
1500 return true_count_max
;
1503 /* Return true if function is likely to be expensive, so there is no point to
1504 optimize performance of prologue, epilogue or do inlining at the expense
1505 of code size growth. THRESHOLD is the limit of number of instructions
1506 function can execute at average to be still considered not expensive. */
1509 expensive_function_p (int threshold
)
1511 unsigned int sum
= 0;
1515 /* We can not compute accurately for large thresholds due to scaled
1517 if (threshold
> BB_FREQ_MAX
)
1520 /* Frequencies are out of range. This either means that function contains
1521 internal loop executing more than BB_FREQ_MAX times or profile feedback
1522 is available and function has not been executed at all. */
1523 if (ENTRY_BLOCK_PTR
->frequency
== 0)
1526 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
1527 limit
= ENTRY_BLOCK_PTR
->frequency
* threshold
;
1532 for (insn
= BB_HEAD (bb
); insn
!= NEXT_INSN (BB_END (bb
));
1533 insn
= NEXT_INSN (insn
))
1534 if (active_insn_p (insn
))
1536 sum
+= bb
->frequency
;
1545 /* Estimate basic blocks frequency by given branch probabilities. */
1548 estimate_bb_frequencies (struct loops
*loops
)
1553 if (!flag_branch_probabilities
|| !counts_to_freqs ())
1555 static int real_values_initialized
= 0;
1557 if (!real_values_initialized
)
1559 real_values_initialized
= 1;
1560 sreal_init (&real_zero
, 0, 0);
1561 sreal_init (&real_one
, 1, 0);
1562 sreal_init (&real_br_prob_base
, REG_BR_PROB_BASE
, 0);
1563 sreal_init (&real_bb_freq_max
, BB_FREQ_MAX
, 0);
1564 sreal_init (&real_one_half
, 1, -1);
1565 sreal_div (&real_inv_br_prob_base
, &real_one
, &real_br_prob_base
);
1566 sreal_sub (&real_almost_one
, &real_one
, &real_inv_br_prob_base
);
1569 mark_dfs_back_edges ();
1571 ENTRY_BLOCK_PTR
->succ
->probability
= REG_BR_PROB_BASE
;
1573 /* Set up block info for each basic block. */
1574 alloc_aux_for_blocks (sizeof (struct block_info_def
));
1575 alloc_aux_for_edges (sizeof (struct edge_info_def
));
1576 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1580 BLOCK_INFO (bb
)->tovisit
= 0;
1581 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
1583 sreal_init (&EDGE_INFO (e
)->back_edge_prob
, e
->probability
, 0);
1584 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
1585 &EDGE_INFO (e
)->back_edge_prob
,
1586 &real_inv_br_prob_base
);
1590 /* First compute probabilities locally for each loop from innermost
1591 to outermost to examine probabilities for back edges. */
1592 estimate_loops_at_level (loops
->tree_root
);
1594 memcpy (&freq_max
, &real_zero
, sizeof (real_zero
));
1596 if (sreal_compare (&freq_max
, &BLOCK_INFO (bb
)->frequency
) < 0)
1597 memcpy (&freq_max
, &BLOCK_INFO (bb
)->frequency
, sizeof (freq_max
));
1599 sreal_div (&freq_max
, &real_bb_freq_max
, &freq_max
);
1600 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1604 sreal_mul (&tmp
, &BLOCK_INFO (bb
)->frequency
, &freq_max
);
1605 sreal_add (&tmp
, &tmp
, &real_one_half
);
1606 bb
->frequency
= sreal_to_int (&tmp
);
1609 free_aux_for_blocks ();
1610 free_aux_for_edges ();
1612 compute_function_frequency ();
1613 if (flag_reorder_functions
)
1614 choose_function_section ();
1617 /* Decide whether function is hot, cold or unlikely executed. */
1619 compute_function_frequency (void)
1623 if (!profile_info
|| !flag_branch_probabilities
)
1625 cfun
->function_frequency
= FUNCTION_FREQUENCY_UNLIKELY_EXECUTED
;
1628 if (maybe_hot_bb_p (bb
))
1630 cfun
->function_frequency
= FUNCTION_FREQUENCY_HOT
;
1633 if (!probably_never_executed_bb_p (bb
))
1634 cfun
->function_frequency
= FUNCTION_FREQUENCY_NORMAL
;
1638 /* Choose appropriate section for the function. */
1640 choose_function_section (void)
1642 if (DECL_SECTION_NAME (current_function_decl
)
1643 || !targetm
.have_named_sections
1644 /* Theoretically we can split the gnu.linkonce text section too,
1645 but this requires more work as the frequency needs to match
1646 for all generated objects so we need to merge the frequency
1647 of all instances. For now just never set frequency for these. */
1648 || DECL_ONE_ONLY (current_function_decl
))
1651 /* If we are doing the partitioning optimization, let the optimization
1652 choose the correct section into which to put things. */
1654 if (flag_reorder_blocks_and_partition
)
1657 if (cfun
->function_frequency
== FUNCTION_FREQUENCY_HOT
)
1658 DECL_SECTION_NAME (current_function_decl
) =
1659 build_string (strlen (HOT_TEXT_SECTION_NAME
), HOT_TEXT_SECTION_NAME
);
1660 if (cfun
->function_frequency
== FUNCTION_FREQUENCY_UNLIKELY_EXECUTED
)
1661 DECL_SECTION_NAME (current_function_decl
) =
1662 build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME
),
1663 UNLIKELY_EXECUTED_TEXT_SECTION_NAME
);
1667 struct tree_opt_pass pass_profile
=
1669 "profile", /* name */
1671 tree_estimate_probability
, /* execute */
1674 0, /* static_pass_number */
1675 TV_BRANCH_PROB
, /* tv_id */
1676 PROP_cfg
, /* properties_required */
1677 0, /* properties_provided */
1678 0, /* properties_destroyed */
1679 0, /* todo_flags_start */
1680 TODO_ggc_collect
| TODO_verify_ssa
, /* todo_flags_finish */