1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
24 [1] "Branch Prediction for Free"
25 Ball and Larus; PLDI '93.
26 [2] "Static Branch Frequency and Program Profile Analysis"
27 Wu and Larus; MICRO-27.
28 [3] "Corpus-based Static Branch Prediction"
29 Calder, Grunwald, Lindsay, Martin, Mozer, and Zorn; PLDI '95. */
34 #include "coretypes.h"
39 #include "hard-reg-set.h"
40 #include "basic-block.h"
41 #include "insn-config.h"
56 #include "tree-flow.h"
58 #include "tree-dump.h"
59 #include "tree-pass.h"
61 #include "tree-scalar-evolution.h"
64 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
65 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
66 static sreal real_zero
, real_one
, real_almost_one
, real_br_prob_base
,
67 real_inv_br_prob_base
, real_one_half
, real_bb_freq_max
;
69 /* Random guesstimation given names. */
70 #define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 100 - 1)
71 #define PROB_EVEN (REG_BR_PROB_BASE / 2)
72 #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
73 #define PROB_ALWAYS (REG_BR_PROB_BASE)
75 static void combine_predictions_for_insn (rtx
, basic_block
);
76 static void dump_prediction (FILE *, enum br_predictor
, int, basic_block
, int);
77 static void estimate_loops_at_level (struct loop
*, bitmap
);
78 static void propagate_freq (struct loop
*, bitmap
);
79 static void estimate_bb_frequencies (struct loops
*);
80 static void predict_paths_leading_to (basic_block
, int *, enum br_predictor
, enum prediction
);
81 static bool last_basic_block_p (basic_block
);
82 static void compute_function_frequency (void);
83 static void choose_function_section (void);
84 static bool can_predict_insn_p (rtx
);
86 /* Information we hold about each branch predictor.
87 Filled using information from predict.def. */
91 const char *const name
; /* Name used in the debugging dumps. */
92 const int hitrate
; /* Expected hitrate used by
93 predict_insn_def call. */
97 /* Use given predictor without Dempster-Shaffer theory if it matches
98 using first_match heuristics. */
99 #define PRED_FLAG_FIRST_MATCH 1
101 /* Recompute hitrate in percent to our representation. */
103 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
105 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
106 static const struct predictor_info predictor_info
[]= {
107 #include "predict.def"
109 /* Upper bound on predictors. */
114 /* Return true in case BB can be CPU intensive and should be optimized
115 for maximal performance. */
118 maybe_hot_bb_p (basic_block bb
)
120 if (profile_info
&& flag_branch_probabilities
122 < profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
)))
124 if (bb
->frequency
< BB_FREQ_MAX
/ PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
))
129 /* Return true in case BB is cold and should be optimized for size. */
132 probably_cold_bb_p (basic_block bb
)
134 if (profile_info
&& flag_branch_probabilities
136 < profile_info
->sum_max
/ PARAM_VALUE (HOT_BB_COUNT_FRACTION
)))
138 if (bb
->frequency
< BB_FREQ_MAX
/ PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION
))
143 /* Return true in case BB is probably never executed. */
145 probably_never_executed_bb_p (basic_block bb
)
147 if (profile_info
&& flag_branch_probabilities
)
148 return ((bb
->count
+ profile_info
->runs
/ 2) / profile_info
->runs
) == 0;
152 /* Return true if the one of outgoing edges is already predicted by
156 rtl_predicted_by_p (basic_block bb
, enum br_predictor predictor
)
159 if (!INSN_P (BB_END (bb
)))
161 for (note
= REG_NOTES (BB_END (bb
)); note
; note
= XEXP (note
, 1))
162 if (REG_NOTE_KIND (note
) == REG_BR_PRED
163 && INTVAL (XEXP (XEXP (note
, 0), 0)) == (int)predictor
)
168 /* Return true if the one of outgoing edges is already predicted by
172 tree_predicted_by_p (basic_block bb
, enum br_predictor predictor
)
174 struct edge_prediction
*i
;
175 for (i
= bb
->predictions
; i
; i
= i
->ep_next
)
176 if (i
->ep_predictor
== predictor
)
181 /* Return true when the probability of edge is reliable.
183 The profile guessing code is good at predicting branch outcome (ie.
184 taken/not taken), that is predicted right slightly over 75% of time.
185 It is however notoriously poor on predicting the probability itself.
186 In general the profile appear a lot flatter (with probabilities closer
187 to 50%) than the reality so it is bad idea to use it to drive optimization
188 such as those disabling dynamic branch prediction for well predictable
191 There are two exceptions - edges leading to noreturn edges and edges
192 predicted by number of iterations heuristics are predicted well. This macro
193 should be able to distinguish those, but at the moment it simply check for
194 noreturn heuristic that is only one giving probability over 99% or bellow
195 1%. In future we might want to propagate reliability information across the
196 CFG if we find this information useful on multiple places. */
198 probability_reliable_p (int prob
)
200 return (profile_status
== PROFILE_READ
201 || (profile_status
== PROFILE_GUESSED
202 && (prob
<= HITRATE (1) || prob
>= HITRATE (99))));
205 /* Same predicate as above, working on edges. */
207 edge_probability_reliable_p (edge e
)
209 return probability_reliable_p (e
->probability
);
212 /* Same predicate as edge_probability_reliable_p, working on notes. */
214 br_prob_note_reliable_p (rtx note
)
216 gcc_assert (REG_NOTE_KIND (note
) == REG_BR_PROB
);
217 return probability_reliable_p (INTVAL (XEXP (note
, 0)));
221 predict_insn (rtx insn
, enum br_predictor predictor
, int probability
)
223 gcc_assert (any_condjump_p (insn
));
224 if (!flag_guess_branch_prob
)
228 = gen_rtx_EXPR_LIST (REG_BR_PRED
,
229 gen_rtx_CONCAT (VOIDmode
,
230 GEN_INT ((int) predictor
),
231 GEN_INT ((int) probability
)),
235 /* Predict insn by given predictor. */
238 predict_insn_def (rtx insn
, enum br_predictor predictor
,
239 enum prediction taken
)
241 int probability
= predictor_info
[(int) predictor
].hitrate
;
244 probability
= REG_BR_PROB_BASE
- probability
;
246 predict_insn (insn
, predictor
, probability
);
249 /* Predict edge E with given probability if possible. */
252 rtl_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
255 last_insn
= BB_END (e
->src
);
257 /* We can store the branch prediction information only about
258 conditional jumps. */
259 if (!any_condjump_p (last_insn
))
262 /* We always store probability of branching. */
263 if (e
->flags
& EDGE_FALLTHRU
)
264 probability
= REG_BR_PROB_BASE
- probability
;
266 predict_insn (last_insn
, predictor
, probability
);
269 /* Predict edge E with the given PROBABILITY. */
271 tree_predict_edge (edge e
, enum br_predictor predictor
, int probability
)
273 gcc_assert (profile_status
!= PROFILE_GUESSED
);
274 if ((e
->src
!= ENTRY_BLOCK_PTR
&& EDGE_COUNT (e
->src
->succs
) > 1)
275 && flag_guess_branch_prob
&& optimize
)
277 struct edge_prediction
*i
= ggc_alloc (sizeof (struct edge_prediction
));
279 i
->ep_next
= e
->src
->predictions
;
280 e
->src
->predictions
= i
;
281 i
->ep_probability
= probability
;
282 i
->ep_predictor
= predictor
;
287 /* Remove all predictions on given basic block that are attached
290 remove_predictions_associated_with_edge (edge e
)
292 if (e
->src
->predictions
)
294 struct edge_prediction
**prediction
= &e
->src
->predictions
;
297 if ((*prediction
)->ep_edge
== e
)
298 *prediction
= (*prediction
)->ep_next
;
300 prediction
= &((*prediction
)->ep_next
);
305 /* Return true when we can store prediction on insn INSN.
306 At the moment we represent predictions only on conditional
307 jumps, not at computed jump or other complicated cases. */
309 can_predict_insn_p (rtx insn
)
311 return (JUMP_P (insn
)
312 && any_condjump_p (insn
)
313 && EDGE_COUNT (BLOCK_FOR_INSN (insn
)->succs
) >= 2);
316 /* Predict edge E by given predictor if possible. */
319 predict_edge_def (edge e
, enum br_predictor predictor
,
320 enum prediction taken
)
322 int probability
= predictor_info
[(int) predictor
].hitrate
;
325 probability
= REG_BR_PROB_BASE
- probability
;
327 predict_edge (e
, predictor
, probability
);
330 /* Invert all branch predictions or probability notes in the INSN. This needs
331 to be done each time we invert the condition used by the jump. */
334 invert_br_probabilities (rtx insn
)
338 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
339 if (REG_NOTE_KIND (note
) == REG_BR_PROB
)
340 XEXP (note
, 0) = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (note
, 0)));
341 else if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
342 XEXP (XEXP (note
, 0), 1)
343 = GEN_INT (REG_BR_PROB_BASE
- INTVAL (XEXP (XEXP (note
, 0), 1)));
346 /* Dump information about the branch prediction to the output file. */
349 dump_prediction (FILE *file
, enum br_predictor predictor
, int probability
,
350 basic_block bb
, int used
)
358 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
359 if (! (e
->flags
& EDGE_FALLTHRU
))
362 fprintf (file
, " %s heuristics%s: %.1f%%",
363 predictor_info
[predictor
].name
,
364 used
? "" : " (ignored)", probability
* 100.0 / REG_BR_PROB_BASE
);
368 fprintf (file
, " exec ");
369 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, bb
->count
);
372 fprintf (file
, " hit ");
373 fprintf (file
, HOST_WIDEST_INT_PRINT_DEC
, e
->count
);
374 fprintf (file
, " (%.1f%%)", e
->count
* 100.0 / bb
->count
);
378 fprintf (file
, "\n");
381 /* We can not predict the probabilities of outgoing edges of bb. Set them
382 evenly and hope for the best. */
384 set_even_probabilities (basic_block bb
)
390 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
391 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
393 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
394 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
395 e
->probability
= (REG_BR_PROB_BASE
+ nedges
/ 2) / nedges
;
400 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
401 note if not already present. Remove now useless REG_BR_PRED notes. */
404 combine_predictions_for_insn (rtx insn
, basic_block bb
)
409 int best_probability
= PROB_EVEN
;
410 int best_predictor
= END_PREDICTORS
;
411 int combined_probability
= REG_BR_PROB_BASE
/ 2;
413 bool first_match
= false;
416 if (!can_predict_insn_p (insn
))
418 set_even_probabilities (bb
);
422 prob_note
= find_reg_note (insn
, REG_BR_PROB
, 0);
423 pnote
= ®_NOTES (insn
);
425 fprintf (dump_file
, "Predictions for insn %i bb %i\n", INSN_UID (insn
),
428 /* We implement "first match" heuristics and use probability guessed
429 by predictor with smallest index. */
430 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
431 if (REG_NOTE_KIND (note
) == REG_BR_PRED
)
433 int predictor
= INTVAL (XEXP (XEXP (note
, 0), 0));
434 int probability
= INTVAL (XEXP (XEXP (note
, 0), 1));
437 if (best_predictor
> predictor
)
438 best_probability
= probability
, best_predictor
= predictor
;
440 d
= (combined_probability
* probability
441 + (REG_BR_PROB_BASE
- combined_probability
)
442 * (REG_BR_PROB_BASE
- probability
));
444 /* Use FP math to avoid overflows of 32bit integers. */
446 /* If one probability is 0% and one 100%, avoid division by zero. */
447 combined_probability
= REG_BR_PROB_BASE
/ 2;
449 combined_probability
= (((double) combined_probability
) * probability
450 * REG_BR_PROB_BASE
/ d
+ 0.5);
453 /* Decide which heuristic to use. In case we didn't match anything,
454 use no_prediction heuristic, in case we did match, use either
455 first match or Dempster-Shaffer theory depending on the flags. */
457 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
461 dump_prediction (dump_file
, PRED_NO_PREDICTION
,
462 combined_probability
, bb
, true);
465 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
,
467 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
,
472 combined_probability
= best_probability
;
473 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
477 if (REG_NOTE_KIND (*pnote
) == REG_BR_PRED
)
479 int predictor
= INTVAL (XEXP (XEXP (*pnote
, 0), 0));
480 int probability
= INTVAL (XEXP (XEXP (*pnote
, 0), 1));
482 dump_prediction (dump_file
, predictor
, probability
, bb
,
483 !first_match
|| best_predictor
== predictor
);
484 *pnote
= XEXP (*pnote
, 1);
487 pnote
= &XEXP (*pnote
, 1);
493 = gen_rtx_EXPR_LIST (REG_BR_PROB
,
494 GEN_INT (combined_probability
), REG_NOTES (insn
));
496 /* Save the prediction into CFG in case we are seeing non-degenerated
498 if (!single_succ_p (bb
))
500 BRANCH_EDGE (bb
)->probability
= combined_probability
;
501 FALLTHRU_EDGE (bb
)->probability
502 = REG_BR_PROB_BASE
- combined_probability
;
505 else if (!single_succ_p (bb
))
507 int prob
= INTVAL (XEXP (prob_note
, 0));
509 BRANCH_EDGE (bb
)->probability
= prob
;
510 FALLTHRU_EDGE (bb
)->probability
= REG_BR_PROB_BASE
- prob
;
513 single_succ_edge (bb
)->probability
= REG_BR_PROB_BASE
;
516 /* Combine predictions into single probability and store them into CFG.
517 Remove now useless prediction entries. */
520 combine_predictions_for_bb (basic_block bb
)
522 int best_probability
= PROB_EVEN
;
523 int best_predictor
= END_PREDICTORS
;
524 int combined_probability
= REG_BR_PROB_BASE
/ 2;
526 bool first_match
= false;
528 struct edge_prediction
*pred
;
530 edge e
, first
= NULL
, second
= NULL
;
533 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
534 if (!(e
->flags
& (EDGE_EH
| EDGE_FAKE
)))
537 if (first
&& !second
)
543 /* When there is no successor or only one choice, prediction is easy.
545 We are lazy for now and predict only basic blocks with two outgoing
546 edges. It is possible to predict generic case too, but we have to
547 ignore first match heuristics and do more involved combining. Implement
552 set_even_probabilities (bb
);
553 bb
->predictions
= NULL
;
555 fprintf (dump_file
, "%i edges in bb %i predicted to even probabilities\n",
561 fprintf (dump_file
, "Predictions for bb %i\n", bb
->index
);
563 /* We implement "first match" heuristics and use probability guessed
564 by predictor with smallest index. */
565 for (pred
= bb
->predictions
; pred
; pred
= pred
->ep_next
)
567 int predictor
= pred
->ep_predictor
;
568 int probability
= pred
->ep_probability
;
570 if (pred
->ep_edge
!= first
)
571 probability
= REG_BR_PROB_BASE
- probability
;
574 if (best_predictor
> predictor
)
575 best_probability
= probability
, best_predictor
= predictor
;
577 d
= (combined_probability
* probability
578 + (REG_BR_PROB_BASE
- combined_probability
)
579 * (REG_BR_PROB_BASE
- probability
));
581 /* Use FP math to avoid overflows of 32bit integers. */
583 /* If one probability is 0% and one 100%, avoid division by zero. */
584 combined_probability
= REG_BR_PROB_BASE
/ 2;
586 combined_probability
= (((double) combined_probability
) * probability
587 * REG_BR_PROB_BASE
/ d
+ 0.5);
590 /* Decide which heuristic to use. In case we didn't match anything,
591 use no_prediction heuristic, in case we did match, use either
592 first match or Dempster-Shaffer theory depending on the flags. */
594 if (predictor_info
[best_predictor
].flags
& PRED_FLAG_FIRST_MATCH
)
598 dump_prediction (dump_file
, PRED_NO_PREDICTION
, combined_probability
, bb
, true);
601 dump_prediction (dump_file
, PRED_DS_THEORY
, combined_probability
, bb
,
603 dump_prediction (dump_file
, PRED_FIRST_MATCH
, best_probability
, bb
,
608 combined_probability
= best_probability
;
609 dump_prediction (dump_file
, PRED_COMBINED
, combined_probability
, bb
, true);
611 for (pred
= bb
->predictions
; pred
; pred
= pred
->ep_next
)
613 int predictor
= pred
->ep_predictor
;
614 int probability
= pred
->ep_probability
;
616 if (pred
->ep_edge
!= EDGE_SUCC (bb
, 0))
617 probability
= REG_BR_PROB_BASE
- probability
;
618 dump_prediction (dump_file
, predictor
, probability
, bb
,
619 !first_match
|| best_predictor
== predictor
);
621 bb
->predictions
= NULL
;
625 first
->probability
= combined_probability
;
626 second
->probability
= REG_BR_PROB_BASE
- combined_probability
;
630 /* Predict edge probabilities by exploiting loop structure.
631 When RTLSIMPLELOOPS is set, attempt to count number of iterations by analyzing
632 RTL otherwise use tree based approach. */
634 predict_loops (struct loops
*loops_info
)
638 scev_initialize (loops_info
);
640 /* Try to predict out blocks in a loop that are not part of a
642 for (i
= 1; i
< loops_info
->num
; i
++)
644 basic_block bb
, *bbs
;
647 struct loop
*loop
= loops_info
->parray
[i
];
649 struct tree_niter_desc niter_desc
;
651 exits
= get_loop_exit_edges (loop
, &n_exits
);
654 for (j
= 0; j
< n_exits
; j
++)
658 if (number_of_iterations_exit (loop
, exits
[j
], &niter_desc
, false))
659 niter
= niter_desc
.niter
;
660 if (!niter
|| TREE_CODE (niter_desc
.niter
) != INTEGER_CST
)
661 niter
= loop_niter_by_eval (loop
, exits
[j
]);
663 if (TREE_CODE (niter
) == INTEGER_CST
)
666 int max
= PARAM_VALUE (PARAM_MAX_PREDICTED_ITERATIONS
);
667 if (host_integerp (niter
, 1)
668 && tree_int_cst_lt (niter
,
669 build_int_cstu (NULL_TREE
, max
- 1)))
671 HOST_WIDE_INT nitercst
= tree_low_cst (niter
, 1) + 1;
672 probability
= ((REG_BR_PROB_BASE
+ nitercst
/ 2)
676 probability
= ((REG_BR_PROB_BASE
+ max
/ 2) / max
);
678 predict_edge (exits
[j
], PRED_LOOP_ITERATIONS
, probability
);
683 bbs
= get_loop_body (loop
);
685 for (j
= 0; j
< loop
->num_nodes
; j
++)
687 int header_found
= 0;
693 /* Bypass loop heuristics on continue statement. These
694 statements construct loops via "non-loop" constructs
695 in the source language and are better to be handled
697 if (predicted_by_p (bb
, PRED_CONTINUE
))
700 /* Loop branch heuristics - predict an edge back to a
701 loop's head as taken. */
702 if (bb
== loop
->latch
)
704 e
= find_edge (loop
->latch
, loop
->header
);
708 predict_edge_def (e
, PRED_LOOP_BRANCH
, TAKEN
);
712 /* Loop exit heuristics - predict an edge exiting the loop if the
713 conditional has no loop header successors as not taken. */
716 /* For loop with many exits we don't want to predict all exits
717 with the pretty large probability, because if all exits are
718 considered in row, the loop would be predicted to iterate
719 almost never. The code to divide probability by number of
720 exits is very rough. It should compute the number of exits
721 taken in each patch through function (not the overall number
722 of exits that might be a lot higher for loops with wide switch
723 statements in them) and compute n-th square root.
725 We limit the minimal probability by 2% to avoid
726 EDGE_PROBABILITY_RELIABLE from trusting the branch prediction
727 as this was causing regression in perl benchmark containing such
730 int probability
= ((REG_BR_PROB_BASE
731 - predictor_info
[(int) PRED_LOOP_EXIT
].hitrate
)
733 if (probability
< HITRATE (2))
734 probability
= HITRATE (2);
735 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
736 if (e
->dest
->index
< NUM_FIXED_BLOCKS
737 || !flow_bb_inside_loop_p (loop
, e
->dest
))
738 predict_edge (e
, PRED_LOOP_EXIT
, probability
);
742 /* Free basic blocks from get_loop_body. */
747 current_loops
= NULL
;
750 /* Attempt to predict probabilities of BB outgoing edges using local
753 bb_estimate_probability_locally (basic_block bb
)
755 rtx last_insn
= BB_END (bb
);
758 if (! can_predict_insn_p (last_insn
))
760 cond
= get_condition (last_insn
, NULL
, false, false);
764 /* Try "pointer heuristic."
765 A comparison ptr == 0 is predicted as false.
766 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
767 if (COMPARISON_P (cond
)
768 && ((REG_P (XEXP (cond
, 0)) && REG_POINTER (XEXP (cond
, 0)))
769 || (REG_P (XEXP (cond
, 1)) && REG_POINTER (XEXP (cond
, 1)))))
771 if (GET_CODE (cond
) == EQ
)
772 predict_insn_def (last_insn
, PRED_POINTER
, NOT_TAKEN
);
773 else if (GET_CODE (cond
) == NE
)
774 predict_insn_def (last_insn
, PRED_POINTER
, TAKEN
);
778 /* Try "opcode heuristic."
779 EQ tests are usually false and NE tests are usually true. Also,
780 most quantities are positive, so we can make the appropriate guesses
781 about signed comparisons against zero. */
782 switch (GET_CODE (cond
))
785 /* Unconditional branch. */
786 predict_insn_def (last_insn
, PRED_UNCONDITIONAL
,
787 cond
== const0_rtx
? NOT_TAKEN
: TAKEN
);
792 /* Floating point comparisons appears to behave in a very
793 unpredictable way because of special role of = tests in
795 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
797 /* Comparisons with 0 are often used for booleans and there is
798 nothing useful to predict about them. */
799 else if (XEXP (cond
, 1) == const0_rtx
800 || XEXP (cond
, 0) == const0_rtx
)
803 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, NOT_TAKEN
);
808 /* Floating point comparisons appears to behave in a very
809 unpredictable way because of special role of = tests in
811 if (FLOAT_MODE_P (GET_MODE (XEXP (cond
, 0))))
813 /* Comparisons with 0 are often used for booleans and there is
814 nothing useful to predict about them. */
815 else if (XEXP (cond
, 1) == const0_rtx
816 || XEXP (cond
, 0) == const0_rtx
)
819 predict_insn_def (last_insn
, PRED_OPCODE_NONEQUAL
, TAKEN
);
823 predict_insn_def (last_insn
, PRED_FPOPCODE
, TAKEN
);
827 predict_insn_def (last_insn
, PRED_FPOPCODE
, NOT_TAKEN
);
832 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
833 || XEXP (cond
, 1) == constm1_rtx
)
834 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, NOT_TAKEN
);
839 if (XEXP (cond
, 1) == const0_rtx
|| XEXP (cond
, 1) == const1_rtx
840 || XEXP (cond
, 1) == constm1_rtx
)
841 predict_insn_def (last_insn
, PRED_OPCODE_POSITIVE
, TAKEN
);
849 /* Set edge->probability for each successor edge of BB. */
851 guess_outgoing_edge_probabilities (basic_block bb
)
853 bb_estimate_probability_locally (bb
);
854 combine_predictions_for_insn (BB_END (bb
), bb
);
857 /* Return constant EXPR will likely have at execution time, NULL if unknown.
858 The function is used by builtin_expect branch predictor so the evidence
859 must come from this construct and additional possible constant folding.
861 We may want to implement more involved value guess (such as value range
862 propagation based prediction), but such tricks shall go to new
866 expr_expected_value (tree expr
, bitmap visited
)
868 if (TREE_CONSTANT (expr
))
870 else if (TREE_CODE (expr
) == SSA_NAME
)
872 tree def
= SSA_NAME_DEF_STMT (expr
);
874 /* If we were already here, break the infinite cycle. */
875 if (bitmap_bit_p (visited
, SSA_NAME_VERSION (expr
)))
877 bitmap_set_bit (visited
, SSA_NAME_VERSION (expr
));
879 if (TREE_CODE (def
) == PHI_NODE
)
881 /* All the arguments of the PHI node must have the same constant
884 tree val
= NULL
, new_val
;
886 for (i
= 0; i
< PHI_NUM_ARGS (def
); i
++)
888 tree arg
= PHI_ARG_DEF (def
, i
);
890 /* If this PHI has itself as an argument, we cannot
891 determine the string length of this argument. However,
892 if we can find an expected constant value for the other
893 PHI args then we can still be sure that this is
894 likely a constant. So be optimistic and just
895 continue with the next argument. */
896 if (arg
== PHI_RESULT (def
))
899 new_val
= expr_expected_value (arg
, visited
);
904 else if (!operand_equal_p (val
, new_val
, false))
909 if (TREE_CODE (def
) != MODIFY_EXPR
|| TREE_OPERAND (def
, 0) != expr
)
911 return expr_expected_value (TREE_OPERAND (def
, 1), visited
);
913 else if (TREE_CODE (expr
) == CALL_EXPR
)
915 tree decl
= get_callee_fndecl (expr
);
918 if (DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_NORMAL
919 && DECL_FUNCTION_CODE (decl
) == BUILT_IN_EXPECT
)
921 tree arglist
= TREE_OPERAND (expr
, 1);
924 if (arglist
== NULL_TREE
925 || TREE_CHAIN (arglist
) == NULL_TREE
)
927 val
= TREE_VALUE (TREE_CHAIN (TREE_OPERAND (expr
, 1)));
928 if (TREE_CONSTANT (val
))
930 return TREE_VALUE (TREE_CHAIN (TREE_OPERAND (expr
, 1)));
933 if (BINARY_CLASS_P (expr
) || COMPARISON_CLASS_P (expr
))
936 op0
= expr_expected_value (TREE_OPERAND (expr
, 0), visited
);
939 op1
= expr_expected_value (TREE_OPERAND (expr
, 1), visited
);
942 res
= fold_build2 (TREE_CODE (expr
), TREE_TYPE (expr
), op0
, op1
);
943 if (TREE_CONSTANT (res
))
947 if (UNARY_CLASS_P (expr
))
950 op0
= expr_expected_value (TREE_OPERAND (expr
, 0), visited
);
953 res
= fold_build1 (TREE_CODE (expr
), TREE_TYPE (expr
), op0
);
954 if (TREE_CONSTANT (res
))
961 /* Get rid of all builtin_expect calls we no longer need. */
963 strip_builtin_expect (void)
968 block_stmt_iterator bi
;
969 for (bi
= bsi_start (bb
); !bsi_end_p (bi
); bsi_next (&bi
))
971 tree stmt
= bsi_stmt (bi
);
975 if (TREE_CODE (stmt
) == MODIFY_EXPR
976 && TREE_CODE (TREE_OPERAND (stmt
, 1)) == CALL_EXPR
977 && (fndecl
= get_callee_fndecl (TREE_OPERAND (stmt
, 1)))
978 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
979 && DECL_FUNCTION_CODE (fndecl
) == BUILT_IN_EXPECT
980 && (arglist
= TREE_OPERAND (TREE_OPERAND (stmt
, 1), 1))
981 && TREE_CHAIN (arglist
))
983 TREE_OPERAND (stmt
, 1) = TREE_VALUE (arglist
);
990 /* Predict using opcode of the last statement in basic block. */
992 tree_predict_by_opcode (basic_block bb
)
994 tree stmt
= last_stmt (bb
);
1003 if (!stmt
|| TREE_CODE (stmt
) != COND_EXPR
)
1005 FOR_EACH_EDGE (then_edge
, ei
, bb
->succs
)
1006 if (then_edge
->flags
& EDGE_TRUE_VALUE
)
1008 cond
= TREE_OPERAND (stmt
, 0);
1009 if (!COMPARISON_CLASS_P (cond
))
1011 op0
= TREE_OPERAND (cond
, 0);
1012 type
= TREE_TYPE (op0
);
1013 visited
= BITMAP_ALLOC (NULL
);
1014 val
= expr_expected_value (cond
, visited
);
1015 BITMAP_FREE (visited
);
1018 if (integer_zerop (val
))
1019 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, NOT_TAKEN
);
1021 predict_edge_def (then_edge
, PRED_BUILTIN_EXPECT
, TAKEN
);
1024 /* Try "pointer heuristic."
1025 A comparison ptr == 0 is predicted as false.
1026 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
1027 if (POINTER_TYPE_P (type
))
1029 if (TREE_CODE (cond
) == EQ_EXPR
)
1030 predict_edge_def (then_edge
, PRED_TREE_POINTER
, NOT_TAKEN
);
1031 else if (TREE_CODE (cond
) == NE_EXPR
)
1032 predict_edge_def (then_edge
, PRED_TREE_POINTER
, TAKEN
);
1036 /* Try "opcode heuristic."
1037 EQ tests are usually false and NE tests are usually true. Also,
1038 most quantities are positive, so we can make the appropriate guesses
1039 about signed comparisons against zero. */
1040 switch (TREE_CODE (cond
))
1044 /* Floating point comparisons appears to behave in a very
1045 unpredictable way because of special role of = tests in
1047 if (FLOAT_TYPE_P (type
))
1049 /* Comparisons with 0 are often used for booleans and there is
1050 nothing useful to predict about them. */
1051 else if (integer_zerop (op0
)
1052 || integer_zerop (TREE_OPERAND (cond
, 1)))
1055 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, NOT_TAKEN
);
1060 /* Floating point comparisons appears to behave in a very
1061 unpredictable way because of special role of = tests in
1063 if (FLOAT_TYPE_P (type
))
1065 /* Comparisons with 0 are often used for booleans and there is
1066 nothing useful to predict about them. */
1067 else if (integer_zerop (op0
)
1068 || integer_zerop (TREE_OPERAND (cond
, 1)))
1071 predict_edge_def (then_edge
, PRED_TREE_OPCODE_NONEQUAL
, TAKEN
);
1075 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, TAKEN
);
1078 case UNORDERED_EXPR
:
1079 predict_edge_def (then_edge
, PRED_TREE_FPOPCODE
, NOT_TAKEN
);
1084 if (integer_zerop (TREE_OPERAND (cond
, 1))
1085 || integer_onep (TREE_OPERAND (cond
, 1))
1086 || integer_all_onesp (TREE_OPERAND (cond
, 1))
1087 || real_zerop (TREE_OPERAND (cond
, 1))
1088 || real_onep (TREE_OPERAND (cond
, 1))
1089 || real_minus_onep (TREE_OPERAND (cond
, 1)))
1090 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, NOT_TAKEN
);
1095 if (integer_zerop (TREE_OPERAND (cond
, 1))
1096 || integer_onep (TREE_OPERAND (cond
, 1))
1097 || integer_all_onesp (TREE_OPERAND (cond
, 1))
1098 || real_zerop (TREE_OPERAND (cond
, 1))
1099 || real_onep (TREE_OPERAND (cond
, 1))
1100 || real_minus_onep (TREE_OPERAND (cond
, 1)))
1101 predict_edge_def (then_edge
, PRED_TREE_OPCODE_POSITIVE
, TAKEN
);
1109 /* Try to guess whether the value of return means error code. */
1110 static enum br_predictor
1111 return_prediction (tree val
, enum prediction
*prediction
)
1115 return PRED_NO_PREDICTION
;
1116 /* Different heuristics for pointers and scalars. */
1117 if (POINTER_TYPE_P (TREE_TYPE (val
)))
1119 /* NULL is usually not returned. */
1120 if (integer_zerop (val
))
1122 *prediction
= NOT_TAKEN
;
1123 return PRED_NULL_RETURN
;
1126 else if (INTEGRAL_TYPE_P (TREE_TYPE (val
)))
1128 /* Negative return values are often used to indicate
1130 if (TREE_CODE (val
) == INTEGER_CST
1131 && tree_int_cst_sgn (val
) < 0)
1133 *prediction
= NOT_TAKEN
;
1134 return PRED_NEGATIVE_RETURN
;
1136 /* Constant return values seems to be commonly taken.
1137 Zero/one often represent booleans so exclude them from the
1139 if (TREE_CONSTANT (val
)
1140 && (!integer_zerop (val
) && !integer_onep (val
)))
1142 *prediction
= TAKEN
;
1143 return PRED_NEGATIVE_RETURN
;
1146 return PRED_NO_PREDICTION
;
1149 /* Find the basic block with return expression and look up for possible
1150 return value trying to apply RETURN_PREDICTION heuristics. */
1152 apply_return_prediction (int *heads
)
1154 tree return_stmt
= NULL
;
1158 int phi_num_args
, i
;
1159 enum br_predictor pred
;
1160 enum prediction direction
;
1163 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
1165 return_stmt
= last_stmt (e
->src
);
1166 if (TREE_CODE (return_stmt
) == RETURN_EXPR
)
1171 return_val
= TREE_OPERAND (return_stmt
, 0);
1174 if (TREE_CODE (return_val
) == MODIFY_EXPR
)
1175 return_val
= TREE_OPERAND (return_val
, 1);
1176 if (TREE_CODE (return_val
) != SSA_NAME
1177 || !SSA_NAME_DEF_STMT (return_val
)
1178 || TREE_CODE (SSA_NAME_DEF_STMT (return_val
)) != PHI_NODE
)
1180 for (phi
= SSA_NAME_DEF_STMT (return_val
); phi
; phi
= PHI_CHAIN (phi
))
1181 if (PHI_RESULT (phi
) == return_val
)
1185 phi_num_args
= PHI_NUM_ARGS (phi
);
1186 pred
= return_prediction (PHI_ARG_DEF (phi
, 0), &direction
);
1188 /* Avoid the degenerate case where all return values form the function
1189 belongs to same category (ie they are all positive constants)
1190 so we can hardly say something about them. */
1191 for (i
= 1; i
< phi_num_args
; i
++)
1192 if (pred
!= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
))
1194 if (i
!= phi_num_args
)
1195 for (i
= 0; i
< phi_num_args
; i
++)
1197 pred
= return_prediction (PHI_ARG_DEF (phi
, i
), &direction
);
1198 if (pred
!= PRED_NO_PREDICTION
)
1199 predict_paths_leading_to (PHI_ARG_EDGE (phi
, i
)->src
, heads
, pred
,
1204 /* Look for basic block that contains unlikely to happen events
1205 (such as noreturn calls) and mark all paths leading to execution
1206 of this basic blocks as unlikely. */
1209 tree_bb_level_predictions (void)
1214 heads
= XCNEWVEC (int, last_basic_block
);
1215 heads
[ENTRY_BLOCK_PTR
->next_bb
->index
] = last_basic_block
;
1217 apply_return_prediction (heads
);
1221 block_stmt_iterator bsi
= bsi_last (bb
);
1223 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1225 tree stmt
= bsi_stmt (bsi
);
1226 switch (TREE_CODE (stmt
))
1229 if (TREE_CODE (TREE_OPERAND (stmt
, 1)) == CALL_EXPR
)
1231 stmt
= TREE_OPERAND (stmt
, 1);
1237 if (call_expr_flags (stmt
) & ECF_NORETURN
)
1238 predict_paths_leading_to (bb
, heads
, PRED_NORETURN
,
1250 /* Predict branch probabilities and estimate profile of the tree CFG. */
1252 tree_estimate_probability (void)
1255 struct loops loops_info
;
1257 flow_loops_find (&loops_info
);
1258 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1259 flow_loops_dump (&loops_info
, dump_file
, NULL
, 0);
1261 add_noreturn_fake_exit_edges ();
1262 connect_infinite_loops_to_exit ();
1263 calculate_dominance_info (CDI_DOMINATORS
);
1264 calculate_dominance_info (CDI_POST_DOMINATORS
);
1266 tree_bb_level_predictions ();
1268 mark_irreducible_loops (&loops_info
);
1269 predict_loops (&loops_info
);
1276 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1278 /* Predict early returns to be probable, as we've already taken
1279 care for error returns and other cases are often used for
1280 fast paths through function. */
1281 if (e
->dest
== EXIT_BLOCK_PTR
1282 && TREE_CODE (last_stmt (bb
)) == RETURN_EXPR
1283 && !single_pred_p (bb
))
1288 FOR_EACH_EDGE (e1
, ei1
, bb
->preds
)
1289 if (!predicted_by_p (e1
->src
, PRED_NULL_RETURN
)
1290 && !predicted_by_p (e1
->src
, PRED_CONST_RETURN
)
1291 && !predicted_by_p (e1
->src
, PRED_NEGATIVE_RETURN
)
1292 && !last_basic_block_p (e1
->src
))
1293 predict_edge_def (e1
, PRED_TREE_EARLY_RETURN
, NOT_TAKEN
);
1296 /* Look for block we are guarding (ie we dominate it,
1297 but it doesn't postdominate us). */
1298 if (e
->dest
!= EXIT_BLOCK_PTR
&& e
->dest
!= bb
1299 && dominated_by_p (CDI_DOMINATORS
, e
->dest
, e
->src
)
1300 && !dominated_by_p (CDI_POST_DOMINATORS
, e
->src
, e
->dest
))
1302 block_stmt_iterator bi
;
1304 /* The call heuristic claims that a guarded function call
1305 is improbable. This is because such calls are often used
1306 to signal exceptional situations such as printing error
1308 for (bi
= bsi_start (e
->dest
); !bsi_end_p (bi
);
1311 tree stmt
= bsi_stmt (bi
);
1312 if ((TREE_CODE (stmt
) == CALL_EXPR
1313 || (TREE_CODE (stmt
) == MODIFY_EXPR
1314 && TREE_CODE (TREE_OPERAND (stmt
, 1)) == CALL_EXPR
))
1315 /* Constant and pure calls are hardly used to signalize
1316 something exceptional. */
1317 && TREE_SIDE_EFFECTS (stmt
))
1319 predict_edge_def (e
, PRED_CALL
, NOT_TAKEN
);
1325 tree_predict_by_opcode (bb
);
1328 combine_predictions_for_bb (bb
);
1330 strip_builtin_expect ();
1331 estimate_bb_frequencies (&loops_info
);
1332 free_dominance_info (CDI_POST_DOMINATORS
);
1333 remove_fake_exit_edges ();
1334 flow_loops_free (&loops_info
);
1335 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1336 dump_tree_cfg (dump_file
, dump_flags
);
1337 if (profile_status
== PROFILE_ABSENT
)
1338 profile_status
= PROFILE_GUESSED
;
1342 /* Check whether this is the last basic block of function. Commonly
1343 there is one extra common cleanup block. */
1345 last_basic_block_p (basic_block bb
)
1347 if (bb
== EXIT_BLOCK_PTR
)
1350 return (bb
->next_bb
== EXIT_BLOCK_PTR
1351 || (bb
->next_bb
->next_bb
== EXIT_BLOCK_PTR
1352 && single_succ_p (bb
)
1353 && single_succ (bb
)->next_bb
== EXIT_BLOCK_PTR
));
1356 /* Sets branch probabilities according to PREDiction and
1357 FLAGS. HEADS[bb->index] should be index of basic block in that we
1358 need to alter branch predictions (i.e. the first of our dominators
1359 such that we do not post-dominate it) (but we fill this information
1360 on demand, so -1 may be there in case this was not needed yet). */
1363 predict_paths_leading_to (basic_block bb
, int *heads
, enum br_predictor pred
,
1364 enum prediction taken
)
1370 if (heads
[bb
->index
] == ENTRY_BLOCK
)
1372 /* This is first time we need this field in heads array; so
1373 find first dominator that we do not post-dominate (we are
1374 using already known members of heads array). */
1375 basic_block ai
= bb
;
1376 basic_block next_ai
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
1379 while (heads
[next_ai
->index
] == ENTRY_BLOCK
)
1381 if (!dominated_by_p (CDI_POST_DOMINATORS
, next_ai
, bb
))
1383 heads
[next_ai
->index
] = ai
->index
;
1385 next_ai
= get_immediate_dominator (CDI_DOMINATORS
, next_ai
);
1387 if (!dominated_by_p (CDI_POST_DOMINATORS
, next_ai
, bb
))
1388 head
= next_ai
->index
;
1390 head
= heads
[next_ai
->index
];
1391 while (next_ai
!= bb
)
1394 ai
= BASIC_BLOCK (heads
[ai
->index
]);
1395 heads
[next_ai
->index
] = head
;
1398 y
= heads
[bb
->index
];
1400 /* Now find the edge that leads to our branch and aply the prediction. */
1402 if (y
== last_basic_block
)
1404 FOR_EACH_EDGE (e
, ei
, BASIC_BLOCK (y
)->succs
)
1405 if (e
->dest
->index
>= NUM_FIXED_BLOCKS
1406 && dominated_by_p (CDI_POST_DOMINATORS
, e
->dest
, bb
))
1407 predict_edge_def (e
, pred
, taken
);
1410 /* This is used to carry information about basic blocks. It is
1411 attached to the AUX field of the standard CFG block. */
1413 typedef struct block_info_def
1415 /* Estimated frequency of execution of basic_block. */
1418 /* To keep queue of basic blocks to process. */
1421 /* Number of predecessors we need to visit first. */
1425 /* Similar information for edges. */
1426 typedef struct edge_info_def
1428 /* In case edge is a loopback edge, the probability edge will be reached
1429 in case header is. Estimated number of iterations of the loop can be
1430 then computed as 1 / (1 - back_edge_prob). */
1431 sreal back_edge_prob
;
1432 /* True if the edge is a loopback edge in the natural loop. */
1433 unsigned int back_edge
:1;
1436 #define BLOCK_INFO(B) ((block_info) (B)->aux)
1437 #define EDGE_INFO(E) ((edge_info) (E)->aux)
1439 /* Helper function for estimate_bb_frequencies.
1440 Propagate the frequencies for LOOP. */
1443 propagate_freq (struct loop
*loop
, bitmap tovisit
)
1445 basic_block head
= loop
->header
;
1453 /* For each basic block we need to visit count number of his predecessors
1454 we need to visit first. */
1455 EXECUTE_IF_SET_IN_BITMAP (tovisit
, 0, i
, bi
)
1460 /* The outermost "loop" includes the exit block, which we can not
1461 look up via BASIC_BLOCK. Detect this and use EXIT_BLOCK_PTR
1462 directly. Do the same for the entry block. */
1463 bb
= BASIC_BLOCK (i
);
1465 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1467 bool visit
= bitmap_bit_p (tovisit
, e
->src
->index
);
1469 if (visit
&& !(e
->flags
& EDGE_DFS_BACK
))
1471 else if (visit
&& dump_file
&& !EDGE_INFO (e
)->back_edge
)
1473 "Irreducible region hit, ignoring edge to %i->%i\n",
1474 e
->src
->index
, bb
->index
);
1476 BLOCK_INFO (bb
)->npredecessors
= count
;
1479 memcpy (&BLOCK_INFO (head
)->frequency
, &real_one
, sizeof (real_one
));
1481 for (bb
= head
; bb
; bb
= nextbb
)
1484 sreal cyclic_probability
, frequency
;
1486 memcpy (&cyclic_probability
, &real_zero
, sizeof (real_zero
));
1487 memcpy (&frequency
, &real_zero
, sizeof (real_zero
));
1489 nextbb
= BLOCK_INFO (bb
)->next
;
1490 BLOCK_INFO (bb
)->next
= NULL
;
1492 /* Compute frequency of basic block. */
1495 #ifdef ENABLE_CHECKING
1496 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1497 gcc_assert (!bitmap_bit_p (tovisit
, e
->src
->index
)
1498 || (e
->flags
& EDGE_DFS_BACK
));
1501 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1502 if (EDGE_INFO (e
)->back_edge
)
1504 sreal_add (&cyclic_probability
, &cyclic_probability
,
1505 &EDGE_INFO (e
)->back_edge_prob
);
1507 else if (!(e
->flags
& EDGE_DFS_BACK
))
1511 /* frequency += (e->probability
1512 * BLOCK_INFO (e->src)->frequency /
1513 REG_BR_PROB_BASE); */
1515 sreal_init (&tmp
, e
->probability
, 0);
1516 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (e
->src
)->frequency
);
1517 sreal_mul (&tmp
, &tmp
, &real_inv_br_prob_base
);
1518 sreal_add (&frequency
, &frequency
, &tmp
);
1521 if (sreal_compare (&cyclic_probability
, &real_zero
) == 0)
1523 memcpy (&BLOCK_INFO (bb
)->frequency
, &frequency
,
1524 sizeof (frequency
));
1528 if (sreal_compare (&cyclic_probability
, &real_almost_one
) > 0)
1530 memcpy (&cyclic_probability
, &real_almost_one
,
1531 sizeof (real_almost_one
));
1534 /* BLOCK_INFO (bb)->frequency = frequency
1535 / (1 - cyclic_probability) */
1537 sreal_sub (&cyclic_probability
, &real_one
, &cyclic_probability
);
1538 sreal_div (&BLOCK_INFO (bb
)->frequency
,
1539 &frequency
, &cyclic_probability
);
1543 bitmap_clear_bit (tovisit
, bb
->index
);
1545 e
= find_edge (bb
, head
);
1550 /* EDGE_INFO (e)->back_edge_prob
1551 = ((e->probability * BLOCK_INFO (bb)->frequency)
1552 / REG_BR_PROB_BASE); */
1554 sreal_init (&tmp
, e
->probability
, 0);
1555 sreal_mul (&tmp
, &tmp
, &BLOCK_INFO (bb
)->frequency
);
1556 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
1557 &tmp
, &real_inv_br_prob_base
);
1560 /* Propagate to successor blocks. */
1561 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1562 if (!(e
->flags
& EDGE_DFS_BACK
)
1563 && BLOCK_INFO (e
->dest
)->npredecessors
)
1565 BLOCK_INFO (e
->dest
)->npredecessors
--;
1566 if (!BLOCK_INFO (e
->dest
)->npredecessors
)
1571 BLOCK_INFO (last
)->next
= e
->dest
;
1579 /* Estimate probabilities of loopback edges in loops at same nest level. */
1582 estimate_loops_at_level (struct loop
*first_loop
, bitmap tovisit
)
1586 for (loop
= first_loop
; loop
; loop
= loop
->next
)
1592 estimate_loops_at_level (loop
->inner
, tovisit
);
1594 /* Do not do this for dummy function loop. */
1595 if (EDGE_COUNT (loop
->latch
->succs
) > 0)
1597 /* Find current loop back edge and mark it. */
1598 e
= loop_latch_edge (loop
);
1599 EDGE_INFO (e
)->back_edge
= 1;
1602 bbs
= get_loop_body (loop
);
1603 for (i
= 0; i
< loop
->num_nodes
; i
++)
1604 bitmap_set_bit (tovisit
, bbs
[i
]->index
);
1606 propagate_freq (loop
, tovisit
);
1610 /* Convert counts measured by profile driven feedback to frequencies.
1611 Return nonzero iff there was any nonzero execution count. */
1614 counts_to_freqs (void)
1616 gcov_type count_max
, true_count_max
= 0;
1620 true_count_max
= MAX (bb
->count
, true_count_max
);
1622 count_max
= MAX (true_count_max
, 1);
1623 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1624 bb
->frequency
= (bb
->count
* BB_FREQ_MAX
+ count_max
/ 2) / count_max
;
1625 return true_count_max
;
1628 /* Return true if function is likely to be expensive, so there is no point to
1629 optimize performance of prologue, epilogue or do inlining at the expense
1630 of code size growth. THRESHOLD is the limit of number of instructions
1631 function can execute at average to be still considered not expensive. */
1634 expensive_function_p (int threshold
)
1636 unsigned int sum
= 0;
1640 /* We can not compute accurately for large thresholds due to scaled
1642 gcc_assert (threshold
<= BB_FREQ_MAX
);
1644 /* Frequencies are out of range. This either means that function contains
1645 internal loop executing more than BB_FREQ_MAX times or profile feedback
1646 is available and function has not been executed at all. */
1647 if (ENTRY_BLOCK_PTR
->frequency
== 0)
1650 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
1651 limit
= ENTRY_BLOCK_PTR
->frequency
* threshold
;
1656 for (insn
= BB_HEAD (bb
); insn
!= NEXT_INSN (BB_END (bb
));
1657 insn
= NEXT_INSN (insn
))
1658 if (active_insn_p (insn
))
1660 sum
+= bb
->frequency
;
1669 /* Estimate basic blocks frequency by given branch probabilities. */
1672 estimate_bb_frequencies (struct loops
*loops
)
1677 if (!flag_branch_probabilities
|| !counts_to_freqs ())
1679 static int real_values_initialized
= 0;
1682 if (!real_values_initialized
)
1684 real_values_initialized
= 1;
1685 sreal_init (&real_zero
, 0, 0);
1686 sreal_init (&real_one
, 1, 0);
1687 sreal_init (&real_br_prob_base
, REG_BR_PROB_BASE
, 0);
1688 sreal_init (&real_bb_freq_max
, BB_FREQ_MAX
, 0);
1689 sreal_init (&real_one_half
, 1, -1);
1690 sreal_div (&real_inv_br_prob_base
, &real_one
, &real_br_prob_base
);
1691 sreal_sub (&real_almost_one
, &real_one
, &real_inv_br_prob_base
);
1694 mark_dfs_back_edges ();
1696 single_succ_edge (ENTRY_BLOCK_PTR
)->probability
= REG_BR_PROB_BASE
;
1698 /* Set up block info for each basic block. */
1699 tovisit
= BITMAP_ALLOC (NULL
);
1700 alloc_aux_for_blocks (sizeof (struct block_info_def
));
1701 alloc_aux_for_edges (sizeof (struct edge_info_def
));
1702 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1707 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1709 sreal_init (&EDGE_INFO (e
)->back_edge_prob
, e
->probability
, 0);
1710 sreal_mul (&EDGE_INFO (e
)->back_edge_prob
,
1711 &EDGE_INFO (e
)->back_edge_prob
,
1712 &real_inv_br_prob_base
);
1716 /* First compute probabilities locally for each loop from innermost
1717 to outermost to examine probabilities for back edges. */
1718 estimate_loops_at_level (loops
->tree_root
, tovisit
);
1720 memcpy (&freq_max
, &real_zero
, sizeof (real_zero
));
1722 if (sreal_compare (&freq_max
, &BLOCK_INFO (bb
)->frequency
) < 0)
1723 memcpy (&freq_max
, &BLOCK_INFO (bb
)->frequency
, sizeof (freq_max
));
1725 sreal_div (&freq_max
, &real_bb_freq_max
, &freq_max
);
1726 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, NULL
, next_bb
)
1730 sreal_mul (&tmp
, &BLOCK_INFO (bb
)->frequency
, &freq_max
);
1731 sreal_add (&tmp
, &tmp
, &real_one_half
);
1732 bb
->frequency
= sreal_to_int (&tmp
);
1735 free_aux_for_blocks ();
1736 free_aux_for_edges ();
1737 BITMAP_FREE (tovisit
);
1739 compute_function_frequency ();
1740 if (flag_reorder_functions
)
1741 choose_function_section ();
1744 /* Decide whether function is hot, cold or unlikely executed. */
1746 compute_function_frequency (void)
1750 if (!profile_info
|| !flag_branch_probabilities
)
1752 cfun
->function_frequency
= FUNCTION_FREQUENCY_UNLIKELY_EXECUTED
;
1755 if (maybe_hot_bb_p (bb
))
1757 cfun
->function_frequency
= FUNCTION_FREQUENCY_HOT
;
1760 if (!probably_never_executed_bb_p (bb
))
1761 cfun
->function_frequency
= FUNCTION_FREQUENCY_NORMAL
;
1765 /* Choose appropriate section for the function. */
1767 choose_function_section (void)
1769 if (DECL_SECTION_NAME (current_function_decl
)
1770 || !targetm
.have_named_sections
1771 /* Theoretically we can split the gnu.linkonce text section too,
1772 but this requires more work as the frequency needs to match
1773 for all generated objects so we need to merge the frequency
1774 of all instances. For now just never set frequency for these. */
1775 || DECL_ONE_ONLY (current_function_decl
))
1778 /* If we are doing the partitioning optimization, let the optimization
1779 choose the correct section into which to put things. */
1781 if (flag_reorder_blocks_and_partition
)
1784 if (cfun
->function_frequency
== FUNCTION_FREQUENCY_HOT
)
1785 DECL_SECTION_NAME (current_function_decl
) =
1786 build_string (strlen (HOT_TEXT_SECTION_NAME
), HOT_TEXT_SECTION_NAME
);
1787 if (cfun
->function_frequency
== FUNCTION_FREQUENCY_UNLIKELY_EXECUTED
)
1788 DECL_SECTION_NAME (current_function_decl
) =
1789 build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME
),
1790 UNLIKELY_EXECUTED_TEXT_SECTION_NAME
);
1794 gate_estimate_probability (void)
1796 return flag_guess_branch_prob
;
1799 struct tree_opt_pass pass_profile
=
1801 "profile", /* name */
1802 gate_estimate_probability
, /* gate */
1803 tree_estimate_probability
, /* execute */
1806 0, /* static_pass_number */
1807 TV_BRANCH_PROB
, /* tv_id */
1808 PROP_cfg
, /* properties_required */
1809 0, /* properties_provided */
1810 0, /* properties_destroyed */
1811 0, /* todo_flags_start */
1812 TODO_ggc_collect
| TODO_verify_ssa
, /* todo_flags_finish */