1 /* Branch prediction routines for the GNU compiler.
2 Copyright (C) 2000, 2001, 2002 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. */
36 #include "hard-reg-set.h"
37 #include "basic-block.h"
38 #include "insn-config.h"
53 /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE, 0.5,
55 static REAL_VALUE_TYPE real_zero, real_one, real_almost_one, real_br_prob_base,
56 real_one_half, real_bb_freq_max;
58 /* Random guesstimation given names. */
59 #define PROB_NEVER (0)
60 #define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 10 - 1)
61 #define PROB_UNLIKELY (REG_BR_PROB_BASE * 4 / 10 - 1)
62 #define PROB_EVEN (REG_BR_PROB_BASE / 2)
63 #define PROB_LIKELY (REG_BR_PROB_BASE - PROB_UNLIKELY)
64 #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
65 #define PROB_ALWAYS (REG_BR_PROB_BASE)
67 static bool predicted_by_p PARAMS ((basic_block,
69 static void combine_predictions_for_insn PARAMS ((rtx, basic_block));
70 static void dump_prediction PARAMS ((enum br_predictor, int,
72 static void estimate_loops_at_level PARAMS ((struct loop *loop));
73 static void propagate_freq PARAMS ((basic_block));
74 static void estimate_bb_frequencies PARAMS ((struct loops *));
75 static void counts_to_freqs PARAMS ((void));
76 static void process_note_predictions PARAMS ((basic_block, int *, int *,
78 static void process_note_prediction PARAMS ((basic_block, int *, int *,
79 sbitmap *, int, int));
80 static bool last_basic_block_p PARAMS ((basic_block));
81 static void compute_function_frequency PARAMS ((void));
82 static void choose_function_section PARAMS ((void));
84 /* Information we hold about each branch predictor.
85 Filled using information from predict.def. */
89 const char *const name; /* Name used in the debugging dumps. */
90 const int hitrate; /* Expected hitrate used by
91 predict_insn_def call. */
95 /* Use given predictor without Dempster-Shaffer theory if it matches
96 using first_match heuristics. */
97 #define PRED_FLAG_FIRST_MATCH 1
99 /* Recompute hitrate in percent to our representation. */
101 #define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
103 #define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
104 static const struct predictor_info predictor_info[]= {
105 #include "predict.def"
107 /* Upper bound on predictors. */
112 /* Return true in case BB can be CPU intensive and should be optimized
113 for maximal perofmrance. */
119 if (profile_info.count_profiles_merged
120 && flag_branch_probabilities
122 < profile_info.max_counter_in_program
123 / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
125 if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
130 /* Return true in case BB is cold and should be optimized for size. */
133 probably_cold_bb_p (bb)
136 if (profile_info.count_profiles_merged
137 && flag_branch_probabilities
139 < profile_info.max_counter_in_program
140 / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
142 if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
147 /* Return true in case BB is probably never executed. */
149 probably_never_executed_bb_p (bb)
152 if (profile_info.count_profiles_merged
153 && flag_branch_probabilities)
154 return ((bb->count + profile_info.count_profiles_merged / 2)
155 / profile_info.count_profiles_merged) == 0;
159 /* Return true if the one of outgoing edges is already predicted by
163 predicted_by_p (bb, predictor)
165 enum br_predictor predictor;
168 if (!INSN_P (bb->end))
170 for (note = REG_NOTES (bb->end); note; note = XEXP (note, 1))
171 if (REG_NOTE_KIND (note) == REG_BR_PRED
172 && INTVAL (XEXP (XEXP (note, 0), 0)) == (int)predictor)
178 predict_insn (insn, predictor, probability)
181 enum br_predictor predictor;
183 if (!any_condjump_p (insn))
187 = gen_rtx_EXPR_LIST (REG_BR_PRED,
188 gen_rtx_CONCAT (VOIDmode,
189 GEN_INT ((int) predictor),
190 GEN_INT ((int) probability)),
194 /* Predict insn by given predictor. */
197 predict_insn_def (insn, predictor, taken)
199 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 predict_edge (e, predictor, probability)
216 enum br_predictor predictor;
219 last_insn = e->src->end;
221 /* We can store the branch prediction information only about
222 conditional jumps. */
223 if (!any_condjump_p (last_insn))
226 /* We always store probability of branching. */
227 if (e->flags & EDGE_FALLTHRU)
228 probability = REG_BR_PROB_BASE - probability;
230 predict_insn (last_insn, predictor, probability);
233 /* Predict edge E by given predictor if possible. */
236 predict_edge_def (e, predictor, taken)
238 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_edge (e, predictor, probability);
249 /* Invert all branch predictions or probability notes in the INSN. This needs
250 to be done each time we invert the condition used by the jump. */
253 invert_br_probabilities (insn)
258 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
259 if (REG_NOTE_KIND (note) == REG_BR_PROB)
260 XEXP (note, 0) = GEN_INT (REG_BR_PROB_BASE - INTVAL (XEXP (note, 0)));
261 else if (REG_NOTE_KIND (note) == REG_BR_PRED)
262 XEXP (XEXP (note, 0), 1)
263 = GEN_INT (REG_BR_PROB_BASE - INTVAL (XEXP (XEXP (note, 0), 1)));
266 /* Dump information about the branch prediction to the output file. */
269 dump_prediction (predictor, probability, bb, used)
270 enum br_predictor predictor;
280 while (e && (e->flags & EDGE_FALLTHRU))
283 fprintf (rtl_dump_file, " %s heuristics%s: %.1f%%",
284 predictor_info[predictor].name,
285 used ? "" : " (ignored)", probability * 100.0 / REG_BR_PROB_BASE);
289 fprintf (rtl_dump_file, " exec ");
290 fprintf (rtl_dump_file, HOST_WIDEST_INT_PRINT_DEC, bb->count);
293 fprintf (rtl_dump_file, " hit ");
294 fprintf (rtl_dump_file, HOST_WIDEST_INT_PRINT_DEC, e->count);
295 fprintf (rtl_dump_file, " (%.1f%%)", e->count * 100.0 / bb->count);
299 fprintf (rtl_dump_file, "\n");
302 /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB
303 note if not already present. Remove now useless REG_BR_PRED notes. */
306 combine_predictions_for_insn (insn, bb)
310 rtx prob_note = find_reg_note (insn, REG_BR_PROB, 0);
311 rtx *pnote = ®_NOTES (insn);
313 int best_probability = PROB_EVEN;
314 int best_predictor = END_PREDICTORS;
315 int combined_probability = REG_BR_PROB_BASE / 2;
317 bool first_match = false;
321 fprintf (rtl_dump_file, "Predictions for insn %i bb %i\n", INSN_UID (insn),
324 /* We implement "first match" heuristics and use probability guessed
325 by predictor with smallest index. In the future we will use better
326 probability combination techniques. */
327 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
328 if (REG_NOTE_KIND (note) == REG_BR_PRED)
330 int predictor = INTVAL (XEXP (XEXP (note, 0), 0));
331 int probability = INTVAL (XEXP (XEXP (note, 0), 1));
334 if (best_predictor > predictor)
335 best_probability = probability, best_predictor = predictor;
337 d = (combined_probability * probability
338 + (REG_BR_PROB_BASE - combined_probability)
339 * (REG_BR_PROB_BASE - probability));
341 /* Use FP math to avoid overflows of 32bit integers. */
343 /* If one probability is 0% and one 100%, avoid division by zero. */
344 combined_probability = REG_BR_PROB_BASE / 2;
346 combined_probability = (((double) combined_probability) * probability
347 * REG_BR_PROB_BASE / d + 0.5);
350 /* Decide which heuristic to use. In case we didn't match anything,
351 use no_prediction heuristic, in case we did match, use either
352 first match or Dempster-Shaffer theory depending on the flags. */
354 if (predictor_info [best_predictor].flags & PRED_FLAG_FIRST_MATCH)
358 dump_prediction (PRED_NO_PREDICTION, combined_probability, bb, true);
361 dump_prediction (PRED_DS_THEORY, combined_probability, bb, !first_match);
362 dump_prediction (PRED_FIRST_MATCH, best_probability, bb, first_match);
366 combined_probability = best_probability;
367 dump_prediction (PRED_COMBINED, combined_probability, bb, true);
371 if (REG_NOTE_KIND (*pnote) == REG_BR_PRED)
373 int predictor = INTVAL (XEXP (XEXP (*pnote, 0), 0));
374 int probability = INTVAL (XEXP (XEXP (*pnote, 0), 1));
376 dump_prediction (predictor, probability, bb,
377 !first_match || best_predictor == predictor);
378 *pnote = XEXP (*pnote, 1);
381 pnote = &XEXP (*pnote, 1);
387 = gen_rtx_EXPR_LIST (REG_BR_PROB,
388 GEN_INT (combined_probability), REG_NOTES (insn));
390 /* Save the prediction into CFG in case we are seeing non-degenerated
392 if (bb->succ->succ_next)
394 BRANCH_EDGE (bb)->probability = combined_probability;
395 FALLTHRU_EDGE (bb)->probability
396 = REG_BR_PROB_BASE - combined_probability;
401 /* Statically estimate the probability that a branch will be taken.
402 ??? In the next revision there will be a number of other predictors added
403 from the above references. Further, each heuristic will be factored out
404 into its own function for clarity (and to facilitate the combination of
408 estimate_probability (loops_info)
409 struct loops *loops_info;
411 sbitmap *dominators, *post_dominators;
414 dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
415 post_dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
416 calculate_dominance_info (NULL, dominators, CDI_DOMINATORS);
417 calculate_dominance_info (NULL, post_dominators, CDI_POST_DOMINATORS);
419 /* Try to predict out blocks in a loop that are not part of a
421 for (i = 0; i < loops_info->num; i++)
425 struct loop *loop = &loops_info->array[i];
427 flow_loop_scan (loops_info, loop, LOOP_EXIT_EDGES);
428 exits = loop->num_exits;
430 for (j = loop->first->index; j <= loop->last->index; ++j)
431 if (TEST_BIT (loop->nodes, j))
433 int header_found = 0;
436 /* Bypass loop heuristics on continue statement. These
437 statements construct loops via "non-loop" constructs
438 in the source language and are better to be handled
440 if (predicted_by_p (BASIC_BLOCK (j), PRED_CONTINUE))
443 /* Loop branch heuristics - predict an edge back to a
444 loop's head as taken. */
445 for (e = BASIC_BLOCK(j)->succ; e; e = e->succ_next)
446 if (e->dest == loop->header
447 && e->src == loop->latch)
450 predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN);
453 /* Loop exit heuristics - predict an edge exiting the loop if the
454 conditinal has no loop header successors as not taken. */
456 for (e = BASIC_BLOCK(j)->succ; e; e = e->succ_next)
457 if (e->dest->index < 0
458 || !TEST_BIT (loop->nodes, e->dest->index))
462 - predictor_info [(int) PRED_LOOP_EXIT].hitrate)
467 /* Attempt to predict conditional jumps using a number of heuristics. */
468 for (i = 0; i < n_basic_blocks; i++)
470 basic_block bb = BASIC_BLOCK (i);
471 rtx last_insn = bb->end;
475 if (GET_CODE (last_insn) != JUMP_INSN || ! any_condjump_p (last_insn))
478 for (e = bb->succ; e; e = e->succ_next)
480 /* Predict early returns to be probable, as we've already taken
481 care for error returns and other are often used for fast paths
483 if ((e->dest == EXIT_BLOCK_PTR
484 || (e->dest->succ && !e->dest->succ->succ_next
485 && e->dest->succ->dest == EXIT_BLOCK_PTR))
486 && !predicted_by_p (bb, PRED_NULL_RETURN)
487 && !predicted_by_p (bb, PRED_CONST_RETURN)
488 && !predicted_by_p (bb, PRED_NEGATIVE_RETURN)
489 && !last_basic_block_p (e->dest))
490 predict_edge_def (e, PRED_EARLY_RETURN, TAKEN);
492 /* Look for block we are guarding (ie we dominate it,
493 but it doesn't postdominate us). */
494 if (e->dest != EXIT_BLOCK_PTR && e->dest != bb
495 && TEST_BIT (dominators[e->dest->index], e->src->index)
496 && !TEST_BIT (post_dominators[e->src->index], e->dest->index))
500 /* The call heuristic claims that a guarded function call
501 is improbable. This is because such calls are often used
502 to signal exceptional situations such as printing error
504 for (insn = e->dest->head; insn != NEXT_INSN (e->dest->end);
505 insn = NEXT_INSN (insn))
506 if (GET_CODE (insn) == CALL_INSN
507 /* Constant and pure calls are hardly used to signalize
508 something exceptional. */
509 && ! CONST_OR_PURE_CALL_P (insn))
511 predict_edge_def (e, PRED_CALL, NOT_TAKEN);
517 cond = get_condition (last_insn, &earliest);
521 /* Try "pointer heuristic."
522 A comparison ptr == 0 is predicted as false.
523 Similarly, a comparison ptr1 == ptr2 is predicted as false. */
524 if (GET_RTX_CLASS (GET_CODE (cond)) == '<'
525 && ((REG_P (XEXP (cond, 0)) && REG_POINTER (XEXP (cond, 0)))
526 || (REG_P (XEXP (cond, 1)) && REG_POINTER (XEXP (cond, 1)))))
528 if (GET_CODE (cond) == EQ)
529 predict_insn_def (last_insn, PRED_POINTER, NOT_TAKEN);
530 else if (GET_CODE (cond) == NE)
531 predict_insn_def (last_insn, PRED_POINTER, TAKEN);
535 /* Try "opcode heuristic."
536 EQ tests are usually false and NE tests are usually true. Also,
537 most quantities are positive, so we can make the appropriate guesses
538 about signed comparisons against zero. */
539 switch (GET_CODE (cond))
542 /* Unconditional branch. */
543 predict_insn_def (last_insn, PRED_UNCONDITIONAL,
544 cond == const0_rtx ? NOT_TAKEN : TAKEN);
549 /* Floating point comparisons appears to behave in a very
550 inpredictable way because of special role of = tests in
552 if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0))))
554 /* Comparisons with 0 are often used for booleans and there is
555 nothing usefull to predict about them. */
556 else if (XEXP (cond, 1) == const0_rtx
557 || XEXP (cond, 0) == const0_rtx)
560 predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, NOT_TAKEN);
565 /* Floating point comparisons appears to behave in a very
566 inpredictable way because of special role of = tests in
568 if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0))))
570 /* Comparisons with 0 are often used for booleans and there is
571 nothing usefull to predict about them. */
572 else if (XEXP (cond, 1) == const0_rtx
573 || XEXP (cond, 0) == const0_rtx)
576 predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, TAKEN);
580 predict_insn_def (last_insn, PRED_FPOPCODE, TAKEN);
584 predict_insn_def (last_insn, PRED_FPOPCODE, NOT_TAKEN);
589 if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx
590 || XEXP (cond, 1) == constm1_rtx)
591 predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, NOT_TAKEN);
596 if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx
597 || XEXP (cond, 1) == constm1_rtx)
598 predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, TAKEN);
606 /* Attach the combined probability to each conditional jump. */
607 for (i = 0; i < n_basic_blocks; i++)
608 if (GET_CODE (BLOCK_END (i)) == JUMP_INSN
609 && any_condjump_p (BLOCK_END (i))
610 && BASIC_BLOCK (i)->succ->succ_next != NULL)
611 combine_predictions_for_insn (BLOCK_END (i), BASIC_BLOCK (i));
613 sbitmap_vector_free (post_dominators);
614 sbitmap_vector_free (dominators);
616 estimate_bb_frequencies (loops_info);
619 /* __builtin_expect dropped tokens into the insn stream describing expected
620 values of registers. Generate branch probabilities based off these
624 expected_value_to_br_prob ()
626 rtx insn, cond, ev = NULL_RTX, ev_reg = NULL_RTX;
628 for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
630 switch (GET_CODE (insn))
633 /* Look for expected value notes. */
634 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_EXPECTED_VALUE)
636 ev = NOTE_EXPECTED_VALUE (insn);
637 ev_reg = XEXP (ev, 0);
643 /* Never propagate across labels. */
648 /* Look for simple conditional branches. If we haven't got an
649 expected value yet, no point going further. */
650 if (GET_CODE (insn) != JUMP_INSN || ev == NULL_RTX
651 || ! any_condjump_p (insn))
656 /* Look for insns that clobber the EV register. */
657 if (ev && reg_set_p (ev_reg, insn))
662 /* Collect the branch condition, hopefully relative to EV_REG. */
663 /* ??? At present we'll miss things like
664 (expected_value (eq r70 0))
666 (set r80 (lt r70 r71))
667 (set pc (if_then_else (ne r80 0) ...))
668 as canonicalize_condition will render this to us as
670 Could use cselib to try and reduce this further. */
671 cond = XEXP (SET_SRC (pc_set (insn)), 0);
672 cond = canonicalize_condition (insn, cond, 0, NULL, ev_reg);
673 if (! cond || XEXP (cond, 0) != ev_reg
674 || GET_CODE (XEXP (cond, 1)) != CONST_INT)
677 /* Substitute and simplify. Given that the expression we're
678 building involves two constants, we should wind up with either
680 cond = gen_rtx_fmt_ee (GET_CODE (cond), VOIDmode,
681 XEXP (ev, 1), XEXP (cond, 1));
682 cond = simplify_rtx (cond);
684 /* Turn the condition into a scaled branch probability. */
685 if (cond != const_true_rtx && cond != const0_rtx)
687 predict_insn_def (insn, PRED_BUILTIN_EXPECT,
688 cond == const_true_rtx ? TAKEN : NOT_TAKEN);
692 /* Check whether this is the last basic block of function. Commonly tehre
693 is one extra common cleanup block. */
695 last_basic_block_p (bb)
698 if (bb == EXIT_BLOCK_PTR)
701 return (bb->next_bb == EXIT_BLOCK_PTR
702 || (bb->next_bb->next_bb == EXIT_BLOCK_PTR
703 && bb->succ && !bb->succ->succ_next
704 && bb->succ->dest->next_bb == EXIT_BLOCK_PTR));
707 /* Sets branch probabilities according to PREDiction and FLAGS. HEADS[bb->index]
708 should be index of basic block in that we need to alter branch predictions
709 (i.e. the first of our dominators such that we do not post-dominate it)
710 (but we fill this information on demand, so -1 may be there in case this
711 was not needed yet). */
714 process_note_prediction (bb, heads, dominators, post_dominators, pred, flags)
718 sbitmap *post_dominators;
726 taken = flags & IS_TAKEN;
728 if (heads[bb->index] < 0)
730 /* This is first time we need this field in heads array; so
731 find first dominator that we do not post-dominate (we are
732 using already known members of heads array). */
734 int next_ai = dominators[bb->index];
737 while (heads[next_ai] < 0)
739 if (!TEST_BIT (post_dominators[next_ai], bb->index))
743 next_ai = dominators[next_ai];
745 if (!TEST_BIT (post_dominators[next_ai], bb->index))
748 head = heads[next_ai];
749 while (next_ai != bb->index)
753 heads[next_ai] = head;
756 y = heads[bb->index];
758 /* Now find the edge that leads to our branch and aply the prediction. */
760 if (y == n_basic_blocks)
762 for (e = BASIC_BLOCK (y)->succ; e; e = e->succ_next)
763 if (e->dest->index >= 0
764 && TEST_BIT (post_dominators[e->dest->index], bb->index))
765 predict_edge_def (e, pred, taken);
768 /* Gathers NOTE_INSN_PREDICTIONs in given basic block and turns them
769 into branch probabilities. For description of heads array, see
770 process_note_prediction. */
773 process_note_predictions (bb, heads, dominators, post_dominators)
777 sbitmap *post_dominators;
782 /* Additionaly, we check here for blocks with no successors. */
783 int contained_noreturn_call = 0;
785 int noreturn_block = 1;
787 for (insn = bb->end; insn;
788 was_bb_head |= (insn == bb->head), insn = PREV_INSN (insn))
790 if (GET_CODE (insn) != NOTE)
796 /* Noreturn calls cause program to exit, therefore they are
797 always predicted as not taken. */
798 if (GET_CODE (insn) == CALL_INSN
799 && find_reg_note (insn, REG_NORETURN, NULL))
800 contained_noreturn_call = 1;
804 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_PREDICTION)
806 int alg = (int) NOTE_PREDICTION_ALG (insn);
807 /* Process single prediction note. */
808 process_note_prediction (bb,
812 alg, (int) NOTE_PREDICTION_FLAGS (insn));
816 for (e = bb->succ; e; e = e->succ_next)
817 if (!(e->flags & EDGE_FAKE))
819 if (contained_noreturn_call)
821 /* This block ended from other reasons than because of return.
822 If it is because of noreturn call, this should certainly not
823 be taken. Otherwise it is probably some error recovery. */
824 process_note_prediction (bb,
827 post_dominators, PRED_NORETURN, NOT_TAKEN);
831 /* Gathers NOTE_INSN_PREDICTIONs and turns them into
832 branch probabilities. */
835 note_prediction_to_br_prob ()
838 sbitmap *post_dominators;
839 int *dominators, *heads;
841 /* To enable handling of noreturn blocks. */
842 add_noreturn_fake_exit_edges ();
843 connect_infinite_loops_to_exit ();
845 dominators = xmalloc (sizeof (int) * n_basic_blocks);
846 memset (dominators, -1, sizeof (int) * n_basic_blocks);
847 post_dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
848 calculate_dominance_info (NULL, post_dominators, CDI_POST_DOMINATORS);
849 calculate_dominance_info (dominators, NULL, CDI_DOMINATORS);
851 heads = xmalloc (sizeof (int) * n_basic_blocks);
852 memset (heads, -1, sizeof (int) * n_basic_blocks);
853 heads[ENTRY_BLOCK_PTR->next_bb->index] = n_basic_blocks;
855 /* Process all prediction notes. */
857 for (i = 0; i < n_basic_blocks; ++i)
859 basic_block bb = BASIC_BLOCK (i);
860 process_note_predictions (bb, heads, dominators, post_dominators);
863 sbitmap_vector_free (post_dominators);
867 remove_fake_edges ();
870 /* This is used to carry information about basic blocks. It is
871 attached to the AUX field of the standard CFG block. */
873 typedef struct block_info_def
875 /* Estimated frequency of execution of basic_block. */
876 REAL_VALUE_TYPE frequency;
878 /* To keep queue of basic blocks to process. */
881 /* True if block needs to be visited in prop_freqency. */
884 /* Number of predecessors we need to visit first. */
888 /* Similar information for edges. */
889 typedef struct edge_info_def
891 /* In case edge is an loopback edge, the probability edge will be reached
892 in case header is. Estimated number of iterations of the loop can be
893 then computed as 1 / (1 - back_edge_prob). */
894 REAL_VALUE_TYPE back_edge_prob;
895 /* True if the edge is an loopback edge in the natural loop. */
899 #define BLOCK_INFO(B) ((block_info) (B)->aux)
900 #define EDGE_INFO(E) ((edge_info) (E)->aux)
902 /* Helper function for estimate_bb_frequencies.
903 Propagate the frequencies for loops headed by HEAD. */
906 propagate_freq (head)
909 basic_block bb = head;
910 basic_block last = bb;
915 /* For each basic block we need to visit count number of his predecessors
916 we need to visit first. */
917 for (n = 0; n < n_basic_blocks; n++)
919 basic_block bb = BASIC_BLOCK (n);
920 if (BLOCK_INFO (bb)->tovisit)
924 for (e = bb->pred; e; e = e->pred_next)
925 if (BLOCK_INFO (e->src)->tovisit && !(e->flags & EDGE_DFS_BACK))
927 else if (BLOCK_INFO (e->src)->tovisit
928 && rtl_dump_file && !EDGE_INFO (e)->back_edge)
929 fprintf (rtl_dump_file,
930 "Irreducible region hit, ignoring edge to %i->%i\n",
931 e->src->index, bb->index);
932 BLOCK_INFO (bb)->npredecessors = count;
936 memcpy (&BLOCK_INFO (head)->frequency, &real_one, sizeof (real_one));
937 for (; bb; bb = nextbb)
939 REAL_VALUE_TYPE cyclic_probability, frequency;
941 memcpy (&cyclic_probability, &real_zero, sizeof (real_zero));
942 memcpy (&frequency, &real_zero, sizeof (real_zero));
944 nextbb = BLOCK_INFO (bb)->next;
945 BLOCK_INFO (bb)->next = NULL;
947 /* Compute frequency of basic block. */
950 #ifdef ENABLE_CHECKING
951 for (e = bb->pred; e; e = e->pred_next)
952 if (BLOCK_INFO (e->src)->tovisit && !(e->flags & EDGE_DFS_BACK))
956 for (e = bb->pred; e; e = e->pred_next)
957 if (EDGE_INFO (e)->back_edge)
959 REAL_ARITHMETIC (cyclic_probability, PLUS_EXPR,
961 EDGE_INFO (e)->back_edge_prob);
963 else if (!(e->flags & EDGE_DFS_BACK))
967 /* frequency += (e->probability
968 * BLOCK_INFO (e->src)->frequency /
969 REG_BR_PROB_BASE); */
971 REAL_VALUE_FROM_INT (tmp, e->probability, 0,
972 TYPE_MODE (double_type_node));
973 REAL_ARITHMETIC (tmp, MULT_EXPR, tmp,
974 BLOCK_INFO (e->src)->frequency);
975 REAL_ARITHMETIC (tmp, RDIV_EXPR, tmp, real_br_prob_base);
976 REAL_ARITHMETIC (frequency, PLUS_EXPR, frequency, tmp);
979 if (REAL_VALUES_LESS (real_almost_one, cyclic_probability))
980 memcpy (&cyclic_probability, &real_almost_one, sizeof (real_zero));
982 /* BLOCK_INFO (bb)->frequency = frequency / (1 - cyclic_probability)
985 REAL_ARITHMETIC (cyclic_probability, MINUS_EXPR, real_one,
987 REAL_ARITHMETIC (BLOCK_INFO (bb)->frequency,
988 RDIV_EXPR, frequency, cyclic_probability);
991 BLOCK_INFO (bb)->tovisit = 0;
993 /* Compute back edge frequencies. */
994 for (e = bb->succ; e; e = e->succ_next)
999 /* EDGE_INFO (e)->back_edge_prob
1000 = ((e->probability * BLOCK_INFO (bb)->frequency)
1001 / REG_BR_PROB_BASE); */
1002 REAL_VALUE_FROM_INT (tmp, e->probability, 0,
1003 TYPE_MODE (double_type_node));
1004 REAL_ARITHMETIC (tmp, MULT_EXPR, tmp,
1005 BLOCK_INFO (bb)->frequency);
1006 REAL_ARITHMETIC (EDGE_INFO (e)->back_edge_prob,
1007 RDIV_EXPR, tmp, real_br_prob_base);
1011 /* Propagate to successor blocks. */
1012 for (e = bb->succ; e; e = e->succ_next)
1013 if (!(e->flags & EDGE_DFS_BACK)
1014 && BLOCK_INFO (e->dest)->npredecessors)
1016 BLOCK_INFO (e->dest)->npredecessors--;
1017 if (!BLOCK_INFO (e->dest)->npredecessors)
1022 BLOCK_INFO (last)->next = e->dest;
1030 /* Estimate probabilities of loopback edges in loops at same nest level. */
1033 estimate_loops_at_level (first_loop)
1034 struct loop *first_loop;
1036 struct loop *l, *loop = first_loop;
1038 for (loop = first_loop; loop; loop = loop->next)
1043 estimate_loops_at_level (loop->inner);
1045 /* Find current loop back edge and mark it. */
1046 for (e = loop->latch->succ; e->dest != loop->header; e = e->succ_next)
1049 EDGE_INFO (e)->back_edge = 1;
1051 /* In case the loop header is shared, ensure that it is the last
1052 one sharing the same header, so we avoid redundant work. */
1055 for (l = loop->next; l; l = l->next)
1056 if (l->header == loop->header)
1063 /* Now merge all nodes of all loops with given header as not visited. */
1064 for (l = loop->shared ? first_loop : loop; l != loop->next; l = l->next)
1065 if (loop->header == l->header)
1066 EXECUTE_IF_SET_IN_SBITMAP (l->nodes, 0, n,
1067 BLOCK_INFO (BASIC_BLOCK (n))->tovisit = 1
1070 propagate_freq (loop->header);
1074 /* Convert counts measured by profile driven feedback to frequencies. */
1079 HOST_WIDEST_INT count_max = 1;
1082 for (i = 0; i < n_basic_blocks; i++)
1083 count_max = MAX (BASIC_BLOCK (i)->count, count_max);
1085 for (i = -2; i < n_basic_blocks; i++)
1090 bb = ENTRY_BLOCK_PTR;
1092 bb = EXIT_BLOCK_PTR;
1094 bb = BASIC_BLOCK (i);
1096 bb->frequency = (bb->count * BB_FREQ_MAX + count_max / 2) / count_max;
1100 /* Return true if function is likely to be expensive, so there is no point to
1101 optimize performance of prologue, epilogue or do inlining at the expense
1102 of code size growth. THRESHOLD is the limit of number of isntructions
1103 function can execute at average to be still considered not expensive. */
1106 expensive_function_p (threshold)
1109 unsigned int sum = 0;
1113 /* We can not compute accurately for large thresholds due to scaled
1115 if (threshold > BB_FREQ_MAX)
1118 /* Frequencies are out of range. This either means that function contains
1119 internal loop executing more than BB_FREQ_MAX times or profile feedback
1120 is available and function has not been executed at all. */
1121 if (ENTRY_BLOCK_PTR->frequency == 0)
1124 /* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
1125 limit = ENTRY_BLOCK_PTR->frequency * threshold;
1126 for (i = 0; i < n_basic_blocks; i++)
1128 basic_block bb = BASIC_BLOCK (i);
1131 for (insn = bb->head; insn != NEXT_INSN (bb->end);
1132 insn = NEXT_INSN (insn))
1133 if (active_insn_p (insn))
1135 sum += bb->frequency;
1144 /* Estimate basic blocks frequency by given branch probabilities. */
1147 estimate_bb_frequencies (loops)
1148 struct loops *loops;
1151 REAL_VALUE_TYPE freq_max;
1152 enum machine_mode double_mode = TYPE_MODE (double_type_node);
1154 if (flag_branch_probabilities)
1158 REAL_VALUE_FROM_INT (real_zero, 0, 0, double_mode);
1159 REAL_VALUE_FROM_INT (real_one, 1, 0, double_mode);
1160 REAL_VALUE_FROM_INT (real_br_prob_base, REG_BR_PROB_BASE, 0, double_mode);
1161 REAL_VALUE_FROM_INT (real_bb_freq_max, BB_FREQ_MAX, 0, double_mode);
1162 REAL_VALUE_FROM_INT (real_one_half, 2, 0, double_mode);
1164 REAL_ARITHMETIC (real_one_half, RDIV_EXPR, real_one, real_one_half);
1166 REAL_ARITHMETIC (real_almost_one, RDIV_EXPR, real_one, real_br_prob_base);
1167 REAL_ARITHMETIC (real_almost_one, MINUS_EXPR, real_one, real_almost_one);
1169 mark_dfs_back_edges ();
1170 /* Fill in the probability values in flowgraph based on the REG_BR_PROB
1172 for (i = 0; i < n_basic_blocks; i++)
1174 rtx last_insn = BLOCK_END (i);
1176 if (GET_CODE (last_insn) != JUMP_INSN || !any_condjump_p (last_insn)
1177 /* Avoid handling of conditional jumps jumping to fallthru edge. */
1178 || BASIC_BLOCK (i)->succ->succ_next == NULL)
1180 /* We can predict only conditional jumps at the moment.
1181 Expect each edge to be equally probable.
1182 ?? In the future we want to make abnormal edges improbable. */
1186 for (e = BASIC_BLOCK (i)->succ; e; e = e->succ_next)
1189 if (e->probability != 0)
1193 for (e = BASIC_BLOCK (i)->succ; e; e = e->succ_next)
1194 e->probability = (REG_BR_PROB_BASE + nedges / 2) / nedges;
1198 ENTRY_BLOCK_PTR->succ->probability = REG_BR_PROB_BASE;
1200 /* Set up block info for each basic block. */
1201 alloc_aux_for_blocks (sizeof (struct block_info_def));
1202 alloc_aux_for_edges (sizeof (struct edge_info_def));
1203 for (i = -2; i < n_basic_blocks; i++)
1209 bb = ENTRY_BLOCK_PTR;
1211 bb = EXIT_BLOCK_PTR;
1213 bb = BASIC_BLOCK (i);
1215 BLOCK_INFO (bb)->tovisit = 0;
1216 for (e = bb->succ; e; e = e->succ_next)
1219 REAL_VALUE_FROM_INT (EDGE_INFO (e)->back_edge_prob,
1220 e->probability, 0, double_mode);
1221 REAL_ARITHMETIC (EDGE_INFO (e)->back_edge_prob,
1222 RDIV_EXPR, EDGE_INFO (e)->back_edge_prob,
1227 /* First compute probabilities locally for each loop from innermost
1228 to outermost to examine probabilities for back edges. */
1229 estimate_loops_at_level (loops->tree_root);
1231 /* Now fake loop around whole function to finalize probabilities. */
1232 for (i = 0; i < n_basic_blocks; i++)
1233 BLOCK_INFO (BASIC_BLOCK (i))->tovisit = 1;
1235 BLOCK_INFO (ENTRY_BLOCK_PTR)->tovisit = 1;
1236 BLOCK_INFO (EXIT_BLOCK_PTR)->tovisit = 1;
1237 propagate_freq (ENTRY_BLOCK_PTR);
1239 memcpy (&freq_max, &real_zero, sizeof (real_zero));
1240 for (i = 0; i < n_basic_blocks; i++)
1241 if (REAL_VALUES_LESS
1242 (freq_max, BLOCK_INFO (BASIC_BLOCK (i))->frequency))
1243 memcpy (&freq_max, &BLOCK_INFO (BASIC_BLOCK (i))->frequency,
1246 for (i = -2; i < n_basic_blocks; i++)
1249 REAL_VALUE_TYPE tmp;
1252 bb = ENTRY_BLOCK_PTR;
1254 bb = EXIT_BLOCK_PTR;
1256 bb = BASIC_BLOCK (i);
1258 REAL_ARITHMETIC (tmp, MULT_EXPR, BLOCK_INFO (bb)->frequency,
1260 REAL_ARITHMETIC (tmp, RDIV_EXPR, tmp, freq_max);
1261 REAL_ARITHMETIC (tmp, PLUS_EXPR, tmp, real_one_half);
1262 bb->frequency = REAL_VALUE_UNSIGNED_FIX (tmp);
1265 free_aux_for_blocks ();
1266 free_aux_for_edges ();
1268 compute_function_frequency ();
1269 if (flag_reorder_functions)
1270 choose_function_section ();
1273 /* Decide whether function is hot, cold or unlikely executed. */
1275 compute_function_frequency ()
1278 if (!profile_info.count_profiles_merged
1279 || !flag_branch_probabilities)
1281 cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED;
1282 for (i = 0; i < n_basic_blocks; i++)
1284 basic_block bb = BASIC_BLOCK (i);
1285 if (maybe_hot_bb_p (bb))
1287 cfun->function_frequency = FUNCTION_FREQUENCY_HOT;
1290 if (!probably_never_executed_bb_p (bb))
1291 cfun->function_frequency = FUNCTION_FREQUENCY_NORMAL;
1295 /* Choose appropriate section for the function. */
1297 choose_function_section ()
1299 if (DECL_SECTION_NAME (current_function_decl)
1300 || !targetm.have_named_sections)
1302 if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT)
1303 DECL_SECTION_NAME (current_function_decl) =
1304 build_string (strlen (HOT_TEXT_SECTION_NAME), HOT_TEXT_SECTION_NAME);
1305 if (cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
1306 DECL_SECTION_NAME (current_function_decl) =
1307 build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME),
1308 UNLIKELY_EXECUTED_TEXT_SECTION_NAME);