X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=gcc%2Fpredict.c;h=dbdfce4efdd3a50a9443f4429474bf49a5e48598;hb=f7a00bdb09cb3716b05e56bcacfaab14f81102be;hp=5f7daea03d38bfe7d9ee3489ece77a0dd9d0ec03;hpb=eb940e633107cb92459fb7212039877cd5a8f53e;p=pf3gnuchains%2Fgcc-fork.git diff --git a/gcc/predict.c b/gcc/predict.c index 5f7daea03d3..dbdfce4efdd 100644 --- a/gcc/predict.c +++ b/gcc/predict.c @@ -1,11 +1,12 @@ /* Branch prediction routines for the GNU compiler. - Copyright (C) 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc. + Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2007 + Free Software Foundation, Inc. This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free -Software Foundation; either version 2, or (at your option) any later +Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY @@ -14,9 +15,8 @@ FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License -along with GCC; see the file COPYING. If not, write to the Free -Software Foundation, 59 Temple Place - Suite 330, Boston, MA -02111-1307, USA. */ +along with GCC; see the file COPYING3. If not see +. */ /* References: @@ -51,8 +51,15 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #include "sreal.h" #include "params.h" #include "target.h" -#include "loop.h" #include "cfgloop.h" +#include "tree-flow.h" +#include "ggc.h" +#include "tree-dump.h" +#include "tree-pass.h" +#include "timevar.h" +#include "tree-scalar-evolution.h" +#include "cfgloop.h" +#include "pointer-set.h" /* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE, 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */ @@ -60,24 +67,17 @@ static sreal real_zero, real_one, real_almost_one, real_br_prob_base, real_inv_br_prob_base, real_one_half, real_bb_freq_max; /* Random guesstimation given names. */ -#define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 10 - 1) +#define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 100 - 1) #define PROB_EVEN (REG_BR_PROB_BASE / 2) #define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY) #define PROB_ALWAYS (REG_BR_PROB_BASE) -static bool predicted_by_p (basic_block, enum br_predictor); static void combine_predictions_for_insn (rtx, basic_block); -static void dump_prediction (enum br_predictor, int, basic_block, int); -static void estimate_loops_at_level (struct loop *loop); -static void propagate_freq (struct loop *); -static void estimate_bb_frequencies (struct loops *); -static void counts_to_freqs (void); -static void process_note_predictions (basic_block, int *); -static void process_note_prediction (basic_block, int *, int, int); -static bool last_basic_block_p (basic_block); +static void dump_prediction (FILE *, enum br_predictor, int, basic_block, int); +static void predict_paths_leading_to (basic_block, int *, enum br_predictor, enum prediction); static void compute_function_frequency (void); static void choose_function_section (void); -static bool can_predict_insn_p (rtx); +static bool can_predict_insn_p (const_rtx); /* Information we hold about each branch predictor. Filled using information from predict.def. */ @@ -111,12 +111,19 @@ static const struct predictor_info predictor_info[]= { for maximal performance. */ bool -maybe_hot_bb_p (basic_block bb) +maybe_hot_bb_p (const_basic_block bb) { if (profile_info && flag_branch_probabilities && (bb->count < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION))) return false; + if (!profile_info || !flag_branch_probabilities) + { + if (cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED) + return false; + if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT) + return true; + } if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)) return false; return true; @@ -125,12 +132,15 @@ maybe_hot_bb_p (basic_block bb) /* Return true in case BB is cold and should be optimized for size. */ bool -probably_cold_bb_p (basic_block bb) +probably_cold_bb_p (const_basic_block bb) { if (profile_info && flag_branch_probabilities && (bb->count < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION))) return true; + if ((!profile_info || !flag_branch_probabilities) + && cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED) + return true; if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)) return true; return false; @@ -138,18 +148,21 @@ probably_cold_bb_p (basic_block bb) /* Return true in case BB is probably never executed. */ bool -probably_never_executed_bb_p (basic_block bb) +probably_never_executed_bb_p (const_basic_block bb) { if (profile_info && flag_branch_probabilities) return ((bb->count + profile_info->runs / 2) / profile_info->runs) == 0; + if ((!profile_info || !flag_branch_probabilities) + && cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED) + return true; return false; } /* Return true if the one of outgoing edges is already predicted by PREDICTOR. */ -static bool -predicted_by_p (basic_block bb, enum br_predictor predictor) +bool +rtl_predicted_by_p (basic_block bb, enum br_predictor predictor) { rtx note; if (!INSN_P (BB_END (bb))) @@ -161,11 +174,72 @@ predicted_by_p (basic_block bb, enum br_predictor predictor) return false; } -void +/* This map contains for a basic block the list of predictions for the + outgoing edges. */ + +static struct pointer_map_t *bb_predictions; + +/* Return true if the one of outgoing edges is already predicted by + PREDICTOR. */ + +bool +tree_predicted_by_p (basic_block bb, enum br_predictor predictor) +{ + struct edge_prediction *i; + void **preds = pointer_map_contains (bb_predictions, bb); + + if (!preds) + return false; + + for (i = *preds; i; i = i->ep_next) + if (i->ep_predictor == predictor) + return true; + return false; +} + +/* Return true when the probability of edge is reliable. + + The profile guessing code is good at predicting branch outcome (ie. + taken/not taken), that is predicted right slightly over 75% of time. + It is however notoriously poor on predicting the probability itself. + In general the profile appear a lot flatter (with probabilities closer + to 50%) than the reality so it is bad idea to use it to drive optimization + such as those disabling dynamic branch prediction for well predictable + branches. + + There are two exceptions - edges leading to noreturn edges and edges + predicted by number of iterations heuristics are predicted well. This macro + should be able to distinguish those, but at the moment it simply check for + noreturn heuristic that is only one giving probability over 99% or bellow + 1%. In future we might want to propagate reliability information across the + CFG if we find this information useful on multiple places. */ +static bool +probability_reliable_p (int prob) +{ + return (profile_status == PROFILE_READ + || (profile_status == PROFILE_GUESSED + && (prob <= HITRATE (1) || prob >= HITRATE (99)))); +} + +/* Same predicate as above, working on edges. */ +bool +edge_probability_reliable_p (const_edge e) +{ + return probability_reliable_p (e->probability); +} + +/* Same predicate as edge_probability_reliable_p, working on notes. */ +bool +br_prob_note_reliable_p (const_rtx note) +{ + gcc_assert (REG_NOTE_KIND (note) == REG_BR_PROB); + return probability_reliable_p (INTVAL (XEXP (note, 0))); +} + +static void predict_insn (rtx insn, enum br_predictor predictor, int probability) { - if (!any_condjump_p (insn)) - abort (); + gcc_assert (any_condjump_p (insn)); if (!flag_guess_branch_prob) return; @@ -194,7 +268,7 @@ predict_insn_def (rtx insn, enum br_predictor predictor, /* Predict edge E with given probability if possible. */ void -predict_edge (edge e, enum br_predictor predictor, int probability) +rtl_predict_edge (edge e, enum br_predictor predictor, int probability) { rtx last_insn; last_insn = BB_END (e->src); @@ -211,15 +285,84 @@ predict_edge (edge e, enum br_predictor predictor, int probability) predict_insn (last_insn, predictor, probability); } +/* Predict edge E with the given PROBABILITY. */ +void +tree_predict_edge (edge e, enum br_predictor predictor, int probability) +{ + gcc_assert (profile_status != PROFILE_GUESSED); + if ((e->src != ENTRY_BLOCK_PTR && EDGE_COUNT (e->src->succs) > 1) + && flag_guess_branch_prob && optimize) + { + struct edge_prediction *i = XNEW (struct edge_prediction); + void **preds = pointer_map_insert (bb_predictions, e->src); + + i->ep_next = *preds; + *preds = i; + i->ep_probability = probability; + i->ep_predictor = predictor; + i->ep_edge = e; + } +} + +/* Remove all predictions on given basic block that are attached + to edge E. */ +void +remove_predictions_associated_with_edge (edge e) +{ + void **preds; + + if (!bb_predictions) + return; + + preds = pointer_map_contains (bb_predictions, e->src); + + if (preds) + { + struct edge_prediction **prediction = (struct edge_prediction **) preds; + struct edge_prediction *next; + + while (*prediction) + { + if ((*prediction)->ep_edge == e) + { + next = (*prediction)->ep_next; + free (*prediction); + *prediction = next; + } + else + prediction = &((*prediction)->ep_next); + } + } +} + +/* Clears the list of predictions stored for BB. */ + +static void +clear_bb_predictions (basic_block bb) +{ + void **preds = pointer_map_contains (bb_predictions, bb); + struct edge_prediction *pred, *next; + + if (!preds) + return; + + for (pred = *preds; pred; pred = next) + { + next = pred->ep_next; + free (pred); + } + *preds = NULL; +} + /* Return true when we can store prediction on insn INSN. At the moment we represent predictions only on conditional jumps, not at computed jump or other complicated cases. */ static bool -can_predict_insn_p (rtx insn) +can_predict_insn_p (const_rtx insn) { - return (GET_CODE (insn) == JUMP_INSN + return (JUMP_P (insn) && any_condjump_p (insn) - && BLOCK_FOR_INSN (insn)->succ->succ_next); + && EDGE_COUNT (BLOCK_FOR_INSN (insn)->succs) >= 2); } /* Predict edge E by given predictor if possible. */ @@ -255,34 +398,55 @@ invert_br_probabilities (rtx insn) /* Dump information about the branch prediction to the output file. */ static void -dump_prediction (enum br_predictor predictor, int probability, +dump_prediction (FILE *file, enum br_predictor predictor, int probability, basic_block bb, int used) { - edge e = bb->succ; + edge e; + edge_iterator ei; - if (!rtl_dump_file) + if (!file) return; - while (e && (e->flags & EDGE_FALLTHRU)) - e = e->succ_next; + FOR_EACH_EDGE (e, ei, bb->succs) + if (! (e->flags & EDGE_FALLTHRU)) + break; - fprintf (rtl_dump_file, " %s heuristics%s: %.1f%%", + fprintf (file, " %s heuristics%s: %.1f%%", predictor_info[predictor].name, used ? "" : " (ignored)", probability * 100.0 / REG_BR_PROB_BASE); if (bb->count) { - fprintf (rtl_dump_file, " exec "); - fprintf (rtl_dump_file, HOST_WIDEST_INT_PRINT_DEC, bb->count); + fprintf (file, " exec "); + fprintf (file, HOST_WIDEST_INT_PRINT_DEC, bb->count); if (e) { - fprintf (rtl_dump_file, " hit "); - fprintf (rtl_dump_file, HOST_WIDEST_INT_PRINT_DEC, e->count); - fprintf (rtl_dump_file, " (%.1f%%)", e->count * 100.0 / bb->count); + fprintf (file, " hit "); + fprintf (file, HOST_WIDEST_INT_PRINT_DEC, e->count); + fprintf (file, " (%.1f%%)", e->count * 100.0 / bb->count); } } - fprintf (rtl_dump_file, "\n"); + fprintf (file, "\n"); +} + +/* We can not predict the probabilities of outgoing edges of bb. Set them + evenly and hope for the best. */ +static void +set_even_probabilities (basic_block bb) +{ + int nedges = 0; + edge e; + edge_iterator ei; + + FOR_EACH_EDGE (e, ei, bb->succs) + if (!(e->flags & (EDGE_EH | EDGE_FAKE))) + nedges ++; + FOR_EACH_EDGE (e, ei, bb->succs) + if (!(e->flags & (EDGE_EH | EDGE_FAKE))) + e->probability = (REG_BR_PROB_BASE + nedges / 2) / nedges; + else + e->probability = 0; } /* Combine all REG_BR_PRED notes into single probability and attach REG_BR_PROB @@ -291,8 +455,8 @@ dump_prediction (enum br_predictor predictor, int probability, static void combine_predictions_for_insn (rtx insn, basic_block bb) { - rtx prob_note = find_reg_note (insn, REG_BR_PROB, 0); - rtx *pnote = ®_NOTES (insn); + rtx prob_note; + rtx *pnote; rtx note; int best_probability = PROB_EVEN; int best_predictor = END_PREDICTORS; @@ -301,13 +465,20 @@ combine_predictions_for_insn (rtx insn, basic_block bb) bool first_match = false; bool found = false; - if (rtl_dump_file) - fprintf (rtl_dump_file, "Predictions for insn %i bb %i\n", INSN_UID (insn), + if (!can_predict_insn_p (insn)) + { + set_even_probabilities (bb); + return; + } + + prob_note = find_reg_note (insn, REG_BR_PROB, 0); + pnote = ®_NOTES (insn); + if (dump_file) + fprintf (dump_file, "Predictions for insn %i bb %i\n", INSN_UID (insn), bb->index); /* We implement "first match" heuristics and use probability guessed - by predictor with smallest index. In the future we will use better - probability combination techniques. */ + by predictor with smallest index. */ for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) if (REG_NOTE_KIND (note) == REG_BR_PRED) { @@ -339,16 +510,19 @@ combine_predictions_for_insn (rtx insn, basic_block bb) first_match = true; if (!found) - dump_prediction (PRED_NO_PREDICTION, combined_probability, bb, true); + dump_prediction (dump_file, PRED_NO_PREDICTION, + combined_probability, bb, true); else { - dump_prediction (PRED_DS_THEORY, combined_probability, bb, !first_match); - dump_prediction (PRED_FIRST_MATCH, best_probability, bb, first_match); + dump_prediction (dump_file, PRED_DS_THEORY, combined_probability, + bb, !first_match); + dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability, + bb, first_match); } if (first_match) combined_probability = best_probability; - dump_prediction (PRED_COMBINED, combined_probability, bb, true); + dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb, true); while (*pnote) { @@ -357,7 +531,7 @@ combine_predictions_for_insn (rtx insn, basic_block bb) int predictor = INTVAL (XEXP (XEXP (*pnote, 0), 0)); int probability = INTVAL (XEXP (XEXP (*pnote, 0), 1)); - dump_prediction (predictor, probability, bb, + dump_prediction (dump_file, predictor, probability, bb, !first_match || best_predictor == predictor); *pnote = XEXP (*pnote, 1); } @@ -373,67 +547,220 @@ combine_predictions_for_insn (rtx insn, basic_block bb) /* Save the prediction into CFG in case we are seeing non-degenerated conditional jump. */ - if (bb->succ->succ_next) + if (!single_succ_p (bb)) { BRANCH_EDGE (bb)->probability = combined_probability; FALLTHRU_EDGE (bb)->probability = REG_BR_PROB_BASE - combined_probability; } } + else if (!single_succ_p (bb)) + { + int prob = INTVAL (XEXP (prob_note, 0)); + + BRANCH_EDGE (bb)->probability = prob; + FALLTHRU_EDGE (bb)->probability = REG_BR_PROB_BASE - prob; + } + else + single_succ_edge (bb)->probability = REG_BR_PROB_BASE; } -/* Statically estimate the probability that a branch will be taken. - ??? In the next revision there will be a number of other predictors added - from the above references. Further, each heuristic will be factored out - into its own function for clarity (and to facilitate the combination of - predictions). */ +/* Combine predictions into single probability and store them into CFG. + Remove now useless prediction entries. */ -void -estimate_probability (struct loops *loops_info) +static void +combine_predictions_for_bb (basic_block bb) { - basic_block bb; - unsigned i; + int best_probability = PROB_EVEN; + int best_predictor = END_PREDICTORS; + int combined_probability = REG_BR_PROB_BASE / 2; + int d; + bool first_match = false; + bool found = false; + struct edge_prediction *pred; + int nedges = 0; + edge e, first = NULL, second = NULL; + edge_iterator ei; + void **preds; + + FOR_EACH_EDGE (e, ei, bb->succs) + if (!(e->flags & (EDGE_EH | EDGE_FAKE))) + { + nedges ++; + if (first && !second) + second = e; + if (!first) + first = e; + } - connect_infinite_loops_to_exit (); - calculate_dominance_info (CDI_DOMINATORS); - calculate_dominance_info (CDI_POST_DOMINATORS); + /* When there is no successor or only one choice, prediction is easy. + + We are lazy for now and predict only basic blocks with two outgoing + edges. It is possible to predict generic case too, but we have to + ignore first match heuristics and do more involved combining. Implement + this later. */ + if (nedges != 2) + { + if (!bb->count) + set_even_probabilities (bb); + clear_bb_predictions (bb); + if (dump_file) + fprintf (dump_file, "%i edges in bb %i predicted to even probabilities\n", + nedges, bb->index); + return; + } + + if (dump_file) + fprintf (dump_file, "Predictions for bb %i\n", bb->index); + + preds = pointer_map_contains (bb_predictions, bb); + if (preds) + { + /* We implement "first match" heuristics and use probability guessed + by predictor with smallest index. */ + for (pred = *preds; pred; pred = pred->ep_next) + { + int predictor = pred->ep_predictor; + int probability = pred->ep_probability; + + if (pred->ep_edge != first) + probability = REG_BR_PROB_BASE - probability; + + found = true; + if (best_predictor > predictor) + best_probability = probability, best_predictor = predictor; + + d = (combined_probability * probability + + (REG_BR_PROB_BASE - combined_probability) + * (REG_BR_PROB_BASE - probability)); + + /* Use FP math to avoid overflows of 32bit integers. */ + if (d == 0) + /* If one probability is 0% and one 100%, avoid division by zero. */ + combined_probability = REG_BR_PROB_BASE / 2; + else + combined_probability = (((double) combined_probability) + * probability + * REG_BR_PROB_BASE / d + 0.5); + } + } + + /* Decide which heuristic to use. In case we didn't match anything, + use no_prediction heuristic, in case we did match, use either + first match or Dempster-Shaffer theory depending on the flags. */ + + if (predictor_info [best_predictor].flags & PRED_FLAG_FIRST_MATCH) + first_match = true; + + if (!found) + dump_prediction (dump_file, PRED_NO_PREDICTION, combined_probability, bb, true); + else + { + dump_prediction (dump_file, PRED_DS_THEORY, combined_probability, bb, + !first_match); + dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability, bb, + first_match); + } + + if (first_match) + combined_probability = best_probability; + dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb, true); + + if (preds) + { + for (pred = *preds; pred; pred = pred->ep_next) + { + int predictor = pred->ep_predictor; + int probability = pred->ep_probability; + + if (pred->ep_edge != EDGE_SUCC (bb, 0)) + probability = REG_BR_PROB_BASE - probability; + dump_prediction (dump_file, predictor, probability, bb, + !first_match || best_predictor == predictor); + } + } + clear_bb_predictions (bb); + + if (!bb->count) + { + first->probability = combined_probability; + second->probability = REG_BR_PROB_BASE - combined_probability; + } +} + +/* Predict edge probabilities by exploiting loop structure. */ + +static void +predict_loops (void) +{ + loop_iterator li; + struct loop *loop; + + scev_initialize (); /* Try to predict out blocks in a loop that are not part of a natural loop. */ - for (i = 1; i < loops_info->num; i++) + FOR_EACH_LOOP (li, loop, 0) { basic_block bb, *bbs; - unsigned j; - int exits; - struct loop *loop = loops_info->parray[i]; - struct loop_desc desc; - unsigned HOST_WIDE_INT niter; + unsigned j, n_exits; + VEC (edge, heap) *exits; + struct tree_niter_desc niter_desc; + edge ex; - flow_loop_scan (loop, LOOP_EXIT_EDGES); - exits = loop->num_exits; + exits = get_loop_exit_edges (loop); + n_exits = VEC_length (edge, exits); - if (simple_loop_p (loop, &desc) && desc.const_iter) + for (j = 0; VEC_iterate (edge, exits, j, ex); j++) { - int prob; - niter = desc.niter + 1; - if (niter == 0) /* We might overflow here. */ - niter = desc.niter; - - prob = (REG_BR_PROB_BASE - - (REG_BR_PROB_BASE + niter /2) / niter); - /* Branch prediction algorithm gives 0 frequency for everything - after the end of loop for loop having 0 probability to finish. */ - if (prob == REG_BR_PROB_BASE) - prob = REG_BR_PROB_BASE - 1; - predict_edge (desc.in_edge, PRED_LOOP_ITERATIONS, - prob); + tree niter = NULL; + HOST_WIDE_INT nitercst; + int max = PARAM_VALUE (PARAM_MAX_PREDICTED_ITERATIONS); + int probability; + enum br_predictor predictor; + + if (number_of_iterations_exit (loop, ex, &niter_desc, false)) + niter = niter_desc.niter; + if (!niter || TREE_CODE (niter_desc.niter) != INTEGER_CST) + niter = loop_niter_by_eval (loop, ex); + + if (TREE_CODE (niter) == INTEGER_CST) + { + if (host_integerp (niter, 1) + && compare_tree_int (niter, max-1) == -1) + nitercst = tree_low_cst (niter, 1) + 1; + else + nitercst = max; + predictor = PRED_LOOP_ITERATIONS; + } + /* If we have just one exit and we can derive some information about + the number of iterations of the loop from the statements inside + the loop, use it to predict this exit. */ + else if (n_exits == 1) + { + nitercst = estimated_loop_iterations_int (loop, false); + if (nitercst < 0) + continue; + if (nitercst > max) + nitercst = max; + + predictor = PRED_LOOP_ITERATIONS_GUESSED; + } + else + continue; + + probability = ((REG_BR_PROB_BASE + nitercst / 2) / nitercst); + predict_edge (ex, predictor, probability); } + VEC_free (edge, heap, exits); bbs = get_loop_body (loop); + for (j = 0; j < loop->num_nodes; j++) { int header_found = 0; edge e; + edge_iterator ei; bb = bbs[j]; @@ -441,275 +768,707 @@ estimate_probability (struct loops *loops_info) statements construct loops via "non-loop" constructs in the source language and are better to be handled separately. */ - if (!can_predict_insn_p (BB_END (bb)) - || predicted_by_p (bb, PRED_CONTINUE)) + if (predicted_by_p (bb, PRED_CONTINUE)) continue; /* Loop branch heuristics - predict an edge back to a loop's head as taken. */ - for (e = bb->succ; e; e = e->succ_next) - if (e->dest == loop->header - && e->src == loop->latch) - { - header_found = 1; - predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); - } + if (bb == loop->latch) + { + e = find_edge (loop->latch, loop->header); + if (e) + { + header_found = 1; + predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); + } + } /* Loop exit heuristics - predict an edge exiting the loop if the conditional has no loop header successors as not taken. */ - if (!header_found) - for (e = bb->succ; e; e = e->succ_next) - if (e->dest->index < 0 - || !flow_bb_inside_loop_p (loop, e->dest)) - predict_edge - (e, PRED_LOOP_EXIT, - (REG_BR_PROB_BASE - - predictor_info [(int) PRED_LOOP_EXIT].hitrate) - / exits); + if (!header_found + /* If we already used more reliable loop exit predictors, do not + bother with PRED_LOOP_EXIT. */ + && !predicted_by_p (bb, PRED_LOOP_ITERATIONS_GUESSED) + && !predicted_by_p (bb, PRED_LOOP_ITERATIONS)) + { + /* For loop with many exits we don't want to predict all exits + with the pretty large probability, because if all exits are + considered in row, the loop would be predicted to iterate + almost never. The code to divide probability by number of + exits is very rough. It should compute the number of exits + taken in each patch through function (not the overall number + of exits that might be a lot higher for loops with wide switch + statements in them) and compute n-th square root. + + We limit the minimal probability by 2% to avoid + EDGE_PROBABILITY_RELIABLE from trusting the branch prediction + as this was causing regression in perl benchmark containing such + a wide loop. */ + + int probability = ((REG_BR_PROB_BASE + - predictor_info [(int) PRED_LOOP_EXIT].hitrate) + / n_exits); + if (probability < HITRATE (2)) + probability = HITRATE (2); + FOR_EACH_EDGE (e, ei, bb->succs) + if (e->dest->index < NUM_FIXED_BLOCKS + || !flow_bb_inside_loop_p (loop, e->dest)) + predict_edge (e, PRED_LOOP_EXIT, probability); + } } - /* Free basic blocks from get_loop_body. */ + /* Free basic blocks from get_loop_body. */ free (bbs); } - /* Attempt to predict conditional jumps using a number of heuristics. */ - FOR_EACH_BB (bb) + scev_finalize (); +} + +/* Attempt to predict probabilities of BB outgoing edges using local + properties. */ +static void +bb_estimate_probability_locally (basic_block bb) +{ + rtx last_insn = BB_END (bb); + rtx cond; + + if (! can_predict_insn_p (last_insn)) + return; + cond = get_condition (last_insn, NULL, false, false); + if (! cond) + return; + + /* Try "pointer heuristic." + A comparison ptr == 0 is predicted as false. + Similarly, a comparison ptr1 == ptr2 is predicted as false. */ + if (COMPARISON_P (cond) + && ((REG_P (XEXP (cond, 0)) && REG_POINTER (XEXP (cond, 0))) + || (REG_P (XEXP (cond, 1)) && REG_POINTER (XEXP (cond, 1))))) { - rtx last_insn = BB_END (bb); - rtx cond, earliest; - edge e; + if (GET_CODE (cond) == EQ) + predict_insn_def (last_insn, PRED_POINTER, NOT_TAKEN); + else if (GET_CODE (cond) == NE) + predict_insn_def (last_insn, PRED_POINTER, TAKEN); + } + else - if (! can_predict_insn_p (last_insn)) - continue; + /* Try "opcode heuristic." + EQ tests are usually false and NE tests are usually true. Also, + most quantities are positive, so we can make the appropriate guesses + about signed comparisons against zero. */ + switch (GET_CODE (cond)) + { + case CONST_INT: + /* Unconditional branch. */ + predict_insn_def (last_insn, PRED_UNCONDITIONAL, + cond == const0_rtx ? NOT_TAKEN : TAKEN); + break; + + case EQ: + case UNEQ: + /* Floating point comparisons appears to behave in a very + unpredictable way because of special role of = tests in + FP code. */ + if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) + ; + /* Comparisons with 0 are often used for booleans and there is + nothing useful to predict about them. */ + else if (XEXP (cond, 1) == const0_rtx + || XEXP (cond, 0) == const0_rtx) + ; + else + predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, NOT_TAKEN); + break; + + case NE: + case LTGT: + /* Floating point comparisons appears to behave in a very + unpredictable way because of special role of = tests in + FP code. */ + if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) + ; + /* Comparisons with 0 are often used for booleans and there is + nothing useful to predict about them. */ + else if (XEXP (cond, 1) == const0_rtx + || XEXP (cond, 0) == const0_rtx) + ; + else + predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, TAKEN); + break; + + case ORDERED: + predict_insn_def (last_insn, PRED_FPOPCODE, TAKEN); + break; + + case UNORDERED: + predict_insn_def (last_insn, PRED_FPOPCODE, NOT_TAKEN); + break; + + case LE: + case LT: + if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx + || XEXP (cond, 1) == constm1_rtx) + predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, NOT_TAKEN); + break; + + case GE: + case GT: + if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx + || XEXP (cond, 1) == constm1_rtx) + predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, TAKEN); + break; + + default: + break; + } +} - for (e = bb->succ; e; e = e->succ_next) +/* Set edge->probability for each successor edge of BB. */ +void +guess_outgoing_edge_probabilities (basic_block bb) +{ + bb_estimate_probability_locally (bb); + combine_predictions_for_insn (BB_END (bb), bb); +} + +/* Return constant EXPR will likely have at execution time, NULL if unknown. + The function is used by builtin_expect branch predictor so the evidence + must come from this construct and additional possible constant folding. + + We may want to implement more involved value guess (such as value range + propagation based prediction), but such tricks shall go to new + implementation. */ + +static tree +expr_expected_value (tree expr, bitmap visited) +{ + if (TREE_CONSTANT (expr)) + return expr; + else if (TREE_CODE (expr) == SSA_NAME) + { + tree def = SSA_NAME_DEF_STMT (expr); + + /* If we were already here, break the infinite cycle. */ + if (bitmap_bit_p (visited, SSA_NAME_VERSION (expr))) + return NULL; + bitmap_set_bit (visited, SSA_NAME_VERSION (expr)); + + if (TREE_CODE (def) == PHI_NODE) { - /* Predict early returns to be probable, as we've already taken - care for error returns and other are often used for fast paths - trought function. */ - if ((e->dest == EXIT_BLOCK_PTR - || (e->dest->succ && !e->dest->succ->succ_next - && e->dest->succ->dest == EXIT_BLOCK_PTR)) - && !predicted_by_p (bb, PRED_NULL_RETURN) - && !predicted_by_p (bb, PRED_CONST_RETURN) - && !predicted_by_p (bb, PRED_NEGATIVE_RETURN) - && !last_basic_block_p (e->dest)) - predict_edge_def (e, PRED_EARLY_RETURN, TAKEN); + /* All the arguments of the PHI node must have the same constant + length. */ + int i; + tree val = NULL, new_val; - /* Look for block we are guarding (ie we dominate it, - but it doesn't postdominate us). */ - if (e->dest != EXIT_BLOCK_PTR && e->dest != bb - && dominated_by_p (CDI_DOMINATORS, e->dest, e->src) - && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e->dest)) + for (i = 0; i < PHI_NUM_ARGS (def); i++) { - rtx insn; - - /* The call heuristic claims that a guarded function call - is improbable. This is because such calls are often used - to signal exceptional situations such as printing error - messages. */ - for (insn = BB_HEAD (e->dest); insn != NEXT_INSN (BB_END (e->dest)); - insn = NEXT_INSN (insn)) - if (GET_CODE (insn) == CALL_INSN - /* Constant and pure calls are hardly used to signalize - something exceptional. */ - && ! CONST_OR_PURE_CALL_P (insn)) - { - predict_edge_def (e, PRED_CALL, NOT_TAKEN); - break; - } + tree arg = PHI_ARG_DEF (def, i); + + /* If this PHI has itself as an argument, we cannot + determine the string length of this argument. However, + if we can find an expected constant value for the other + PHI args then we can still be sure that this is + likely a constant. So be optimistic and just + continue with the next argument. */ + if (arg == PHI_RESULT (def)) + continue; + + new_val = expr_expected_value (arg, visited); + if (!new_val) + return NULL; + if (!val) + val = new_val; + else if (!operand_equal_p (val, new_val, false)) + return NULL; + } + return val; + } + if (TREE_CODE (def) != GIMPLE_MODIFY_STMT + || GIMPLE_STMT_OPERAND (def, 0) != expr) + return NULL; + return expr_expected_value (GIMPLE_STMT_OPERAND (def, 1), visited); + } + else if (TREE_CODE (expr) == CALL_EXPR) + { + tree decl = get_callee_fndecl (expr); + if (!decl) + return NULL; + if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL + && DECL_FUNCTION_CODE (decl) == BUILT_IN_EXPECT) + { + tree val; + + if (call_expr_nargs (expr) != 2) + return NULL; + val = CALL_EXPR_ARG (expr, 0); + if (TREE_CONSTANT (val)) + return val; + return CALL_EXPR_ARG (expr, 1); + } + } + if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr)) + { + tree op0, op1, res; + op0 = expr_expected_value (TREE_OPERAND (expr, 0), visited); + if (!op0) + return NULL; + op1 = expr_expected_value (TREE_OPERAND (expr, 1), visited); + if (!op1) + return NULL; + res = fold_build2 (TREE_CODE (expr), TREE_TYPE (expr), op0, op1); + if (TREE_CONSTANT (res)) + return res; + return NULL; + } + if (UNARY_CLASS_P (expr)) + { + tree op0, res; + op0 = expr_expected_value (TREE_OPERAND (expr, 0), visited); + if (!op0) + return NULL; + res = fold_build1 (TREE_CODE (expr), TREE_TYPE (expr), op0); + if (TREE_CONSTANT (res)) + return res; + return NULL; + } + return NULL; +} + +/* Get rid of all builtin_expect calls we no longer need. */ +static void +strip_builtin_expect (void) +{ + basic_block bb; + FOR_EACH_BB (bb) + { + block_stmt_iterator bi; + for (bi = bsi_start (bb); !bsi_end_p (bi); bsi_next (&bi)) + { + tree stmt = bsi_stmt (bi); + tree fndecl; + tree call; + + if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT + && (call = GIMPLE_STMT_OPERAND (stmt, 1)) + && TREE_CODE (call) == CALL_EXPR + && (fndecl = get_callee_fndecl (call)) + && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL + && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_EXPECT + && call_expr_nargs (call) == 2) + { + GIMPLE_STMT_OPERAND (stmt, 1) = CALL_EXPR_ARG (call, 0); + update_stmt (stmt); } } + } +} + +/* Predict using opcode of the last statement in basic block. */ +static void +tree_predict_by_opcode (basic_block bb) +{ + tree stmt = last_stmt (bb); + edge then_edge; + tree cond; + tree op0; + tree type; + tree val; + bitmap visited; + edge_iterator ei; + + if (!stmt || TREE_CODE (stmt) != COND_EXPR) + return; + FOR_EACH_EDGE (then_edge, ei, bb->succs) + if (then_edge->flags & EDGE_TRUE_VALUE) + break; + cond = TREE_OPERAND (stmt, 0); + if (!COMPARISON_CLASS_P (cond)) + return; + op0 = TREE_OPERAND (cond, 0); + type = TREE_TYPE (op0); + visited = BITMAP_ALLOC (NULL); + val = expr_expected_value (cond, visited); + BITMAP_FREE (visited); + if (val) + { + if (integer_zerop (val)) + predict_edge_def (then_edge, PRED_BUILTIN_EXPECT, NOT_TAKEN); + else + predict_edge_def (then_edge, PRED_BUILTIN_EXPECT, TAKEN); + return; + } + /* Try "pointer heuristic." + A comparison ptr == 0 is predicted as false. + Similarly, a comparison ptr1 == ptr2 is predicted as false. */ + if (POINTER_TYPE_P (type)) + { + if (TREE_CODE (cond) == EQ_EXPR) + predict_edge_def (then_edge, PRED_TREE_POINTER, NOT_TAKEN); + else if (TREE_CODE (cond) == NE_EXPR) + predict_edge_def (then_edge, PRED_TREE_POINTER, TAKEN); + } + else - cond = get_condition (last_insn, &earliest, false); - if (! cond) - continue; + /* Try "opcode heuristic." + EQ tests are usually false and NE tests are usually true. Also, + most quantities are positive, so we can make the appropriate guesses + about signed comparisons against zero. */ + switch (TREE_CODE (cond)) + { + case EQ_EXPR: + case UNEQ_EXPR: + /* Floating point comparisons appears to behave in a very + unpredictable way because of special role of = tests in + FP code. */ + if (FLOAT_TYPE_P (type)) + ; + /* Comparisons with 0 are often used for booleans and there is + nothing useful to predict about them. */ + else if (integer_zerop (op0) + || integer_zerop (TREE_OPERAND (cond, 1))) + ; + else + predict_edge_def (then_edge, PRED_TREE_OPCODE_NONEQUAL, NOT_TAKEN); + break; + + case NE_EXPR: + case LTGT_EXPR: + /* Floating point comparisons appears to behave in a very + unpredictable way because of special role of = tests in + FP code. */ + if (FLOAT_TYPE_P (type)) + ; + /* Comparisons with 0 are often used for booleans and there is + nothing useful to predict about them. */ + else if (integer_zerop (op0) + || integer_zerop (TREE_OPERAND (cond, 1))) + ; + else + predict_edge_def (then_edge, PRED_TREE_OPCODE_NONEQUAL, TAKEN); + break; + + case ORDERED_EXPR: + predict_edge_def (then_edge, PRED_TREE_FPOPCODE, TAKEN); + break; + + case UNORDERED_EXPR: + predict_edge_def (then_edge, PRED_TREE_FPOPCODE, NOT_TAKEN); + break; + + case LE_EXPR: + case LT_EXPR: + if (integer_zerop (TREE_OPERAND (cond, 1)) + || integer_onep (TREE_OPERAND (cond, 1)) + || integer_all_onesp (TREE_OPERAND (cond, 1)) + || real_zerop (TREE_OPERAND (cond, 1)) + || real_onep (TREE_OPERAND (cond, 1)) + || real_minus_onep (TREE_OPERAND (cond, 1))) + predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, NOT_TAKEN); + break; + + case GE_EXPR: + case GT_EXPR: + if (integer_zerop (TREE_OPERAND (cond, 1)) + || integer_onep (TREE_OPERAND (cond, 1)) + || integer_all_onesp (TREE_OPERAND (cond, 1)) + || real_zerop (TREE_OPERAND (cond, 1)) + || real_onep (TREE_OPERAND (cond, 1)) + || real_minus_onep (TREE_OPERAND (cond, 1))) + predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, TAKEN); + break; + + default: + break; + } +} - /* Try "pointer heuristic." - A comparison ptr == 0 is predicted as false. - Similarly, a comparison ptr1 == ptr2 is predicted as false. */ - if (GET_RTX_CLASS (GET_CODE (cond)) == '<' - && ((REG_P (XEXP (cond, 0)) && REG_POINTER (XEXP (cond, 0))) - || (REG_P (XEXP (cond, 1)) && REG_POINTER (XEXP (cond, 1))))) +/* Try to guess whether the value of return means error code. */ +static enum br_predictor +return_prediction (tree val, enum prediction *prediction) +{ + /* VOID. */ + if (!val) + return PRED_NO_PREDICTION; + /* Different heuristics for pointers and scalars. */ + if (POINTER_TYPE_P (TREE_TYPE (val))) + { + /* NULL is usually not returned. */ + if (integer_zerop (val)) { - if (GET_CODE (cond) == EQ) - predict_insn_def (last_insn, PRED_POINTER, NOT_TAKEN); - else if (GET_CODE (cond) == NE) - predict_insn_def (last_insn, PRED_POINTER, TAKEN); + *prediction = NOT_TAKEN; + return PRED_NULL_RETURN; } - else + } + else if (INTEGRAL_TYPE_P (TREE_TYPE (val))) + { + /* Negative return values are often used to indicate + errors. */ + if (TREE_CODE (val) == INTEGER_CST + && tree_int_cst_sgn (val) < 0) + { + *prediction = NOT_TAKEN; + return PRED_NEGATIVE_RETURN; + } + /* Constant return values seems to be commonly taken. + Zero/one often represent booleans so exclude them from the + heuristics. */ + if (TREE_CONSTANT (val) + && (!integer_zerop (val) && !integer_onep (val))) + { + *prediction = TAKEN; + return PRED_CONST_RETURN; + } + } + return PRED_NO_PREDICTION; +} - /* Try "opcode heuristic." - EQ tests are usually false and NE tests are usually true. Also, - most quantities are positive, so we can make the appropriate guesses - about signed comparisons against zero. */ - switch (GET_CODE (cond)) - { - case CONST_INT: - /* Unconditional branch. */ - predict_insn_def (last_insn, PRED_UNCONDITIONAL, - cond == const0_rtx ? NOT_TAKEN : TAKEN); - break; +/* Find the basic block with return expression and look up for possible + return value trying to apply RETURN_PREDICTION heuristics. */ +static void +apply_return_prediction (int *heads) +{ + tree return_stmt = NULL; + tree return_val; + edge e; + tree phi; + int phi_num_args, i; + enum br_predictor pred; + enum prediction direction; + edge_iterator ei; - case EQ: - case UNEQ: - /* Floating point comparisons appears to behave in a very - unpredictable way because of special role of = tests in - FP code. */ - if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) - ; - /* Comparisons with 0 are often used for booleans and there is - nothing useful to predict about them. */ - else if (XEXP (cond, 1) == const0_rtx - || XEXP (cond, 0) == const0_rtx) - ; - else - predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, NOT_TAKEN); - break; + FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) + { + return_stmt = last_stmt (e->src); + if (return_stmt + && TREE_CODE (return_stmt) == RETURN_EXPR) + break; + } + if (!e) + return; + return_val = TREE_OPERAND (return_stmt, 0); + if (!return_val) + return; + if (TREE_CODE (return_val) == GIMPLE_MODIFY_STMT) + return_val = GIMPLE_STMT_OPERAND (return_val, 1); + if (TREE_CODE (return_val) != SSA_NAME + || !SSA_NAME_DEF_STMT (return_val) + || TREE_CODE (SSA_NAME_DEF_STMT (return_val)) != PHI_NODE) + return; + for (phi = SSA_NAME_DEF_STMT (return_val); phi; phi = PHI_CHAIN (phi)) + if (PHI_RESULT (phi) == return_val) + break; + if (!phi) + return; + phi_num_args = PHI_NUM_ARGS (phi); + pred = return_prediction (PHI_ARG_DEF (phi, 0), &direction); + + /* Avoid the degenerate case where all return values form the function + belongs to same category (ie they are all positive constants) + so we can hardly say something about them. */ + for (i = 1; i < phi_num_args; i++) + if (pred != return_prediction (PHI_ARG_DEF (phi, i), &direction)) + break; + if (i != phi_num_args) + for (i = 0; i < phi_num_args; i++) + { + pred = return_prediction (PHI_ARG_DEF (phi, i), &direction); + if (pred != PRED_NO_PREDICTION) + predict_paths_leading_to (PHI_ARG_EDGE (phi, i)->src, heads, pred, + direction); + } +} - case NE: - case LTGT: - /* Floating point comparisons appears to behave in a very - unpredictable way because of special role of = tests in - FP code. */ - if (FLOAT_MODE_P (GET_MODE (XEXP (cond, 0)))) - ; - /* Comparisons with 0 are often used for booleans and there is - nothing useful to predict about them. */ - else if (XEXP (cond, 1) == const0_rtx - || XEXP (cond, 0) == const0_rtx) - ; - else - predict_insn_def (last_insn, PRED_OPCODE_NONEQUAL, TAKEN); - break; +/* Look for basic block that contains unlikely to happen events + (such as noreturn calls) and mark all paths leading to execution + of this basic blocks as unlikely. */ - case ORDERED: - predict_insn_def (last_insn, PRED_FPOPCODE, TAKEN); - break; +static void +tree_bb_level_predictions (void) +{ + basic_block bb; + int *heads; - case UNORDERED: - predict_insn_def (last_insn, PRED_FPOPCODE, NOT_TAKEN); - break; + heads = XCNEWVEC (int, last_basic_block); + heads[ENTRY_BLOCK_PTR->next_bb->index] = last_basic_block; - case LE: - case LT: - if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx - || XEXP (cond, 1) == constm1_rtx) - predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, NOT_TAKEN); - break; + apply_return_prediction (heads); - case GE: - case GT: - if (XEXP (cond, 1) == const0_rtx || XEXP (cond, 1) == const1_rtx - || XEXP (cond, 1) == constm1_rtx) - predict_insn_def (last_insn, PRED_OPCODE_POSITIVE, TAKEN); - break; + FOR_EACH_BB (bb) + { + block_stmt_iterator bsi = bsi_last (bb); - default: - break; - } + for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) + { + tree stmt = bsi_stmt (bsi); + tree decl; + switch (TREE_CODE (stmt)) + { + case GIMPLE_MODIFY_STMT: + if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)) == CALL_EXPR) + { + stmt = GIMPLE_STMT_OPERAND (stmt, 1); + goto call_expr; + } + break; + case CALL_EXPR: +call_expr:; + if (call_expr_flags (stmt) & ECF_NORETURN) + predict_paths_leading_to (bb, heads, PRED_NORETURN, + NOT_TAKEN); + decl = get_callee_fndecl (stmt); + if (decl + && lookup_attribute ("cold", + DECL_ATTRIBUTES (decl))) + predict_paths_leading_to (bb, heads, PRED_COLD_FUNCTION, + NOT_TAKEN); + break; + default: + break; + } + } } - /* Attach the combined probability to each conditional jump. */ - FOR_EACH_BB (bb) - if (GET_CODE (BB_END (bb)) == JUMP_INSN - && any_condjump_p (BB_END (bb)) - && bb->succ->succ_next != NULL) - combine_predictions_for_insn (BB_END (bb), bb); + free (heads); +} - free_dominance_info (CDI_POST_DOMINATORS); +#ifdef ENABLE_CHECKING + +/* Callback for pointer_map_traverse, asserts that the pointer map is + empty. */ - remove_fake_edges (); - estimate_bb_frequencies (loops_info); +static bool +assert_is_empty (const void *key ATTRIBUTE_UNUSED, void **value, + void *data ATTRIBUTE_UNUSED) +{ + gcc_assert (!*value); + return false; } - -/* __builtin_expect dropped tokens into the insn stream describing expected - values of registers. Generate branch probabilities based off these - values. */ +#endif -void -expected_value_to_br_prob (void) +/* Predict branch probabilities and estimate profile of the tree CFG. */ +static unsigned int +tree_estimate_probability (void) { - rtx insn, cond, ev = NULL_RTX, ev_reg = NULL_RTX; + basic_block bb; + + loop_optimizer_init (0); + if (dump_file && (dump_flags & TDF_DETAILS)) + flow_loops_dump (dump_file, NULL, 0); - for (insn = get_insns (); insn ; insn = NEXT_INSN (insn)) + add_noreturn_fake_exit_edges (); + connect_infinite_loops_to_exit (); + /* We use loop_niter_by_eval, which requires that the loops have + preheaders. */ + create_preheaders (CP_SIMPLE_PREHEADERS); + calculate_dominance_info (CDI_POST_DOMINATORS); + + bb_predictions = pointer_map_create (); + tree_bb_level_predictions (); + + mark_irreducible_loops (); + record_loop_exits (); + if (number_of_loops () > 1) + predict_loops (); + + FOR_EACH_BB (bb) { - switch (GET_CODE (insn)) + edge e; + edge_iterator ei; + + FOR_EACH_EDGE (e, ei, bb->succs) { - case NOTE: - /* Look for expected value notes. */ - if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_EXPECTED_VALUE) + /* Predict early returns to be probable, as we've already taken + care for error returns and other cases are often used for + fast paths through function. + + Since we've already removed the return statements, we are + looking for CFG like: + + if (conditional) + { + .. + goto return_block + } + some other blocks + return_block: + return_stmt. */ + if (e->dest != bb->next_bb + && e->dest != EXIT_BLOCK_PTR + && single_succ_p (e->dest) + && single_succ_edge (e->dest)->dest == EXIT_BLOCK_PTR + && TREE_CODE (last_stmt (e->dest)) == RETURN_EXPR) { - ev = NOTE_EXPECTED_VALUE (insn); - ev_reg = XEXP (ev, 0); - delete_insn (insn); + edge e1; + edge_iterator ei1; + + if (single_succ_p (bb)) + { + FOR_EACH_EDGE (e1, ei1, bb->preds) + if (!predicted_by_p (e1->src, PRED_NULL_RETURN) + && !predicted_by_p (e1->src, PRED_CONST_RETURN) + && !predicted_by_p (e1->src, PRED_NEGATIVE_RETURN)) + predict_edge_def (e1, PRED_TREE_EARLY_RETURN, NOT_TAKEN); + } + else + if (!predicted_by_p (e->src, PRED_NULL_RETURN) + && !predicted_by_p (e->src, PRED_CONST_RETURN) + && !predicted_by_p (e->src, PRED_NEGATIVE_RETURN)) + predict_edge_def (e, PRED_TREE_EARLY_RETURN, NOT_TAKEN); } - continue; - - case CODE_LABEL: - /* Never propagate across labels. */ - ev = NULL_RTX; - continue; - - case JUMP_INSN: - /* Look for simple conditional branches. If we haven't got an - expected value yet, no point going further. */ - if (GET_CODE (insn) != JUMP_INSN || ev == NULL_RTX - || ! any_condjump_p (insn)) - continue; - break; - default: - /* Look for insns that clobber the EV register. */ - if (ev && reg_set_p (ev_reg, insn)) - ev = NULL_RTX; - continue; - } + /* Look for block we are guarding (ie we dominate it, + but it doesn't postdominate us). */ + if (e->dest != EXIT_BLOCK_PTR && e->dest != bb + && dominated_by_p (CDI_DOMINATORS, e->dest, e->src) + && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e->dest)) + { + block_stmt_iterator bi; - /* Collect the branch condition, hopefully relative to EV_REG. */ - /* ??? At present we'll miss things like - (expected_value (eq r70 0)) - (set r71 -1) - (set r80 (lt r70 r71)) - (set pc (if_then_else (ne r80 0) ...)) - as canonicalize_condition will render this to us as - (lt r70, r71) - Could use cselib to try and reduce this further. */ - cond = XEXP (SET_SRC (pc_set (insn)), 0); - cond = canonicalize_condition (insn, cond, 0, NULL, ev_reg, false); - if (! cond || XEXP (cond, 0) != ev_reg - || GET_CODE (XEXP (cond, 1)) != CONST_INT) - continue; - - /* Substitute and simplify. Given that the expression we're - building involves two constants, we should wind up with either - true or false. */ - cond = gen_rtx_fmt_ee (GET_CODE (cond), VOIDmode, - XEXP (ev, 1), XEXP (cond, 1)); - cond = simplify_rtx (cond); - - /* Turn the condition into a scaled branch probability. */ - if (cond != const_true_rtx && cond != const0_rtx) - abort (); - predict_insn_def (insn, PRED_BUILTIN_EXPECT, - cond == const_true_rtx ? TAKEN : NOT_TAKEN); + /* The call heuristic claims that a guarded function call + is improbable. This is because such calls are often used + to signal exceptional situations such as printing error + messages. */ + for (bi = bsi_start (e->dest); !bsi_end_p (bi); + bsi_next (&bi)) + { + tree stmt = bsi_stmt (bi); + if ((TREE_CODE (stmt) == CALL_EXPR + || (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT + && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)) + == CALL_EXPR)) + /* Constant and pure calls are hardly used to signalize + something exceptional. */ + && TREE_SIDE_EFFECTS (stmt)) + { + predict_edge_def (e, PRED_CALL, NOT_TAKEN); + break; + } + } + } + } + tree_predict_by_opcode (bb); } -} - -/* Check whether this is the last basic block of function. Commonly - there is one extra common cleanup block. */ -static bool -last_basic_block_p (basic_block bb) -{ - if (bb == EXIT_BLOCK_PTR) - return false; + FOR_EACH_BB (bb) + combine_predictions_for_bb (bb); - return (bb->next_bb == EXIT_BLOCK_PTR - || (bb->next_bb->next_bb == EXIT_BLOCK_PTR - && bb->succ && !bb->succ->succ_next - && bb->succ->dest->next_bb == EXIT_BLOCK_PTR)); -} +#ifdef ENABLE_CHECKING + pointer_map_traverse (bb_predictions, assert_is_empty, NULL); +#endif + pointer_map_destroy (bb_predictions); + bb_predictions = NULL; + strip_builtin_expect (); + estimate_bb_frequencies (); + free_dominance_info (CDI_POST_DOMINATORS); + remove_fake_exit_edges (); + loop_optimizer_finalize (); + if (dump_file && (dump_flags & TDF_DETAILS)) + dump_tree_cfg (dump_file, dump_flags); + if (profile_status == PROFILE_ABSENT) + profile_status = PROFILE_GUESSED; + return 0; +} + /* Sets branch probabilities according to PREDiction and FLAGS. HEADS[bb->index] should be index of basic block in that we need to alter branch predictions (i.e. the first of our dominators @@ -717,15 +1476,14 @@ last_basic_block_p (basic_block bb) on demand, so -1 may be there in case this was not needed yet). */ static void -process_note_prediction (basic_block bb, int *heads, int pred, int flags) +predict_paths_leading_to (basic_block bb, int *heads, enum br_predictor pred, + enum prediction taken) { edge e; + edge_iterator ei; int y; - bool taken; - - taken = flags & IS_TAKEN; - if (heads[bb->index] < 0) + if (heads[bb->index] == ENTRY_BLOCK) { /* This is first time we need this field in heads array; so find first dominator that we do not post-dominate (we are @@ -734,7 +1492,7 @@ process_note_prediction (basic_block bb, int *heads, int pred, int flags) basic_block next_ai = get_immediate_dominator (CDI_DOMINATORS, bb); int head; - while (heads[next_ai->index] < 0) + while (heads[next_ai->index] == ENTRY_BLOCK) { if (!dominated_by_p (CDI_POST_DOMINATORS, next_ai, bb)) break; @@ -749,10 +1507,7 @@ process_note_prediction (basic_block bb, int *heads, int pred, int flags) while (next_ai != bb) { next_ai = ai; - if (heads[ai->index] == ENTRY_BLOCK) - ai = ENTRY_BLOCK_PTR; - else - ai = BASIC_BLOCK (heads[ai->index]); + ai = BASIC_BLOCK (heads[ai->index]); heads[next_ai->index] = head; } } @@ -760,99 +1515,13 @@ process_note_prediction (basic_block bb, int *heads, int pred, int flags) /* Now find the edge that leads to our branch and aply the prediction. */ - if (y == last_basic_block || !can_predict_insn_p (BB_END (BASIC_BLOCK (y)))) + if (y == last_basic_block) return; - for (e = BASIC_BLOCK (y)->succ; e; e = e->succ_next) - if (e->dest->index >= 0 + FOR_EACH_EDGE (e, ei, BASIC_BLOCK (y)->succs) + if (e->dest->index >= NUM_FIXED_BLOCKS && dominated_by_p (CDI_POST_DOMINATORS, e->dest, bb)) predict_edge_def (e, pred, taken); } - -/* Gathers NOTE_INSN_PREDICTIONs in given basic block and turns them - into branch probabilities. For description of heads array, see - process_note_prediction. */ - -static void -process_note_predictions (basic_block bb, int *heads) -{ - rtx insn; - edge e; - - /* Additionally, we check here for blocks with no successors. */ - int contained_noreturn_call = 0; - int was_bb_head = 0; - int noreturn_block = 1; - - for (insn = BB_END (bb); insn; - was_bb_head |= (insn == BB_HEAD (bb)), insn = PREV_INSN (insn)) - { - if (GET_CODE (insn) != NOTE) - { - if (was_bb_head) - break; - else - { - /* Noreturn calls cause program to exit, therefore they are - always predicted as not taken. */ - if (GET_CODE (insn) == CALL_INSN - && find_reg_note (insn, REG_NORETURN, NULL)) - contained_noreturn_call = 1; - continue; - } - } - if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_PREDICTION) - { - int alg = (int) NOTE_PREDICTION_ALG (insn); - /* Process single prediction note. */ - process_note_prediction (bb, - heads, - alg, (int) NOTE_PREDICTION_FLAGS (insn)); - delete_insn (insn); - } - } - for (e = bb->succ; e; e = e->succ_next) - if (!(e->flags & EDGE_FAKE)) - noreturn_block = 0; - if (contained_noreturn_call) - { - /* This block ended from other reasons than because of return. - If it is because of noreturn call, this should certainly not - be taken. Otherwise it is probably some error recovery. */ - process_note_prediction (bb, heads, PRED_NORETURN, NOT_TAKEN); - } -} - -/* Gathers NOTE_INSN_PREDICTIONs and turns them into - branch probabilities. */ - -void -note_prediction_to_br_prob (void) -{ - basic_block bb; - int *heads; - - /* To enable handling of noreturn blocks. */ - add_noreturn_fake_exit_edges (); - connect_infinite_loops_to_exit (); - - calculate_dominance_info (CDI_POST_DOMINATORS); - calculate_dominance_info (CDI_DOMINATORS); - - heads = xmalloc (sizeof (int) * last_basic_block); - memset (heads, -1, sizeof (int) * last_basic_block); - heads[ENTRY_BLOCK_PTR->next_bb->index] = last_basic_block; - - /* Process all prediction notes. */ - - FOR_EACH_BB (bb) - process_note_predictions (bb, heads); - - free_dominance_info (CDI_POST_DOMINATORS); - free_dominance_info (CDI_DOMINATORS); - free (heads); - - remove_fake_edges (); -} /* This is used to carry information about basic blocks. It is attached to the AUX field of the standard CFG block. */ @@ -865,9 +1534,6 @@ typedef struct block_info_def /* To keep queue of basic blocks to process. */ basic_block next; - /* True if block needs to be visited in propagate_freq. */ - unsigned int tovisit:1; - /* Number of predecessors we need to visit first. */ int npredecessors; } *block_info; @@ -875,11 +1541,11 @@ typedef struct block_info_def /* Similar information for edges. */ typedef struct edge_info_def { - /* In case edge is an loopback edge, the probability edge will be reached + /* In case edge is a loopback edge, the probability edge will be reached in case header is. Estimated number of iterations of the loop can be then computed as 1 / (1 - back_edge_prob). */ sreal back_edge_prob; - /* True if the edge is an loopback edge in the natural loop. */ + /* True if the edge is a loopback edge in the natural loop. */ unsigned int back_edge:1; } *edge_info; @@ -887,41 +1553,50 @@ typedef struct edge_info_def #define EDGE_INFO(E) ((edge_info) (E)->aux) /* Helper function for estimate_bb_frequencies. - Propagate the frequencies for LOOP. */ + Propagate the frequencies in blocks marked in + TOVISIT, starting in HEAD. */ static void -propagate_freq (struct loop *loop) +propagate_freq (basic_block head, bitmap tovisit) { - basic_block head = loop->header; basic_block bb; basic_block last; + unsigned i; edge e; basic_block nextbb; + bitmap_iterator bi; /* For each basic block we need to visit count number of his predecessors we need to visit first. */ - FOR_EACH_BB (bb) + EXECUTE_IF_SET_IN_BITMAP (tovisit, 0, i, bi) { - if (BLOCK_INFO (bb)->tovisit) + edge_iterator ei; + int count = 0; + + /* The outermost "loop" includes the exit block, which we can not + look up via BASIC_BLOCK. Detect this and use EXIT_BLOCK_PTR + directly. Do the same for the entry block. */ + bb = BASIC_BLOCK (i); + + FOR_EACH_EDGE (e, ei, bb->preds) { - int count = 0; - - for (e = bb->pred; e; e = e->pred_next) - if (BLOCK_INFO (e->src)->tovisit && !(e->flags & EDGE_DFS_BACK)) - count++; - else if (BLOCK_INFO (e->src)->tovisit - && rtl_dump_file && !EDGE_INFO (e)->back_edge) - fprintf (rtl_dump_file, - "Irreducible region hit, ignoring edge to %i->%i\n", - e->src->index, bb->index); - BLOCK_INFO (bb)->npredecessors = count; + bool visit = bitmap_bit_p (tovisit, e->src->index); + + if (visit && !(e->flags & EDGE_DFS_BACK)) + count++; + else if (visit && dump_file && !EDGE_INFO (e)->back_edge) + fprintf (dump_file, + "Irreducible region hit, ignoring edge to %i->%i\n", + e->src->index, bb->index); } + BLOCK_INFO (bb)->npredecessors = count; } memcpy (&BLOCK_INFO (head)->frequency, &real_one, sizeof (real_one)); last = head; for (bb = head; bb; bb = nextbb) { + edge_iterator ei; sreal cyclic_probability, frequency; memcpy (&cyclic_probability, &real_zero, sizeof (real_zero)); @@ -934,12 +1609,12 @@ propagate_freq (struct loop *loop) if (bb != head) { #ifdef ENABLE_CHECKING - for (e = bb->pred; e; e = e->pred_next) - if (BLOCK_INFO (e->src)->tovisit && !(e->flags & EDGE_DFS_BACK)) - abort (); + FOR_EACH_EDGE (e, ei, bb->preds) + gcc_assert (!bitmap_bit_p (tovisit, e->src->index) + || (e->flags & EDGE_DFS_BACK)); #endif - for (e = bb->pred; e; e = e->pred_next) + FOR_EACH_EDGE (e, ei, bb->preds) if (EDGE_INFO (e)->back_edge) { sreal_add (&cyclic_probability, &cyclic_probability, @@ -981,26 +1656,25 @@ propagate_freq (struct loop *loop) } } - BLOCK_INFO (bb)->tovisit = 0; - - /* Compute back edge frequencies. */ - for (e = bb->succ; e; e = e->succ_next) - if (e->dest == head) - { - sreal tmp; - - /* EDGE_INFO (e)->back_edge_prob - = ((e->probability * BLOCK_INFO (bb)->frequency) - / REG_BR_PROB_BASE); */ + bitmap_clear_bit (tovisit, bb->index); - sreal_init (&tmp, e->probability, 0); - sreal_mul (&tmp, &tmp, &BLOCK_INFO (bb)->frequency); - sreal_mul (&EDGE_INFO (e)->back_edge_prob, - &tmp, &real_inv_br_prob_base); - } + e = find_edge (bb, head); + if (e) + { + sreal tmp; + + /* EDGE_INFO (e)->back_edge_prob + = ((e->probability * BLOCK_INFO (bb)->frequency) + / REG_BR_PROB_BASE); */ + + sreal_init (&tmp, e->probability, 0); + sreal_mul (&tmp, &tmp, &BLOCK_INFO (bb)->frequency); + sreal_mul (&EDGE_INFO (e)->back_edge_prob, + &tmp, &real_inv_br_prob_base); + } /* Propagate to successor blocks. */ - for (e = bb->succ; e; e = e->succ_next) + FOR_EACH_EDGE (e, ei, bb->succs) if (!(e->flags & EDGE_DFS_BACK) && BLOCK_INFO (e->dest)->npredecessors) { @@ -1011,10 +1685,10 @@ propagate_freq (struct loop *loop) nextbb = e->dest; else BLOCK_INFO (last)->next = e->dest; - + last = e->dest; } - } + } } } @@ -1030,37 +1704,61 @@ estimate_loops_at_level (struct loop *first_loop) edge e; basic_block *bbs; unsigned i; + bitmap tovisit = BITMAP_ALLOC (NULL); estimate_loops_at_level (loop->inner); - if (loop->latch->succ) /* Do not do this for dummy function loop. */ - { - /* Find current loop back edge and mark it. */ - e = loop_latch_edge (loop); - EDGE_INFO (e)->back_edge = 1; - } + /* Find current loop back edge and mark it. */ + e = loop_latch_edge (loop); + EDGE_INFO (e)->back_edge = 1; bbs = get_loop_body (loop); for (i = 0; i < loop->num_nodes; i++) - BLOCK_INFO (bbs[i])->tovisit = 1; + bitmap_set_bit (tovisit, bbs[i]->index); free (bbs); - propagate_freq (loop); + propagate_freq (loop->header, tovisit); + BITMAP_FREE (tovisit); } } -/* Convert counts measured by profile driven feedback to frequencies. */ +/* Propagates frequencies through structure of loops. */ static void +estimate_loops (void) +{ + bitmap tovisit = BITMAP_ALLOC (NULL); + basic_block bb; + + /* Start by estimating the frequencies in the loops. */ + if (number_of_loops () > 1) + estimate_loops_at_level (current_loops->tree_root->inner); + + /* Now propagate the frequencies through all the blocks. */ + FOR_ALL_BB (bb) + { + bitmap_set_bit (tovisit, bb->index); + } + propagate_freq (ENTRY_BLOCK_PTR, tovisit); + BITMAP_FREE (tovisit); +} + +/* Convert counts measured by profile driven feedback to frequencies. + Return nonzero iff there was any nonzero execution count. */ + +int counts_to_freqs (void) { - gcov_type count_max = 1; + gcov_type count_max, true_count_max = 0; basic_block bb; FOR_EACH_BB (bb) - count_max = MAX (bb->count, count_max); + true_count_max = MAX (bb->count, true_count_max); + count_max = MAX (true_count_max, 1); FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb) bb->frequency = (bb->count * BB_FREQ_MAX + count_max / 2) / count_max; + + return true_count_max; } /* Return true if function is likely to be expensive, so there is no point to @@ -1077,8 +1775,7 @@ expensive_function_p (int threshold) /* We can not compute accurately for large thresholds due to scaled frequencies. */ - if (threshold > BB_FREQ_MAX) - abort (); + gcc_assert (threshold <= BB_FREQ_MAX); /* Frequencies are out of range. This either means that function contains internal loop executing more than BB_FREQ_MAX times or profile feedback @@ -1107,15 +1804,13 @@ expensive_function_p (int threshold) /* Estimate basic blocks frequency by given branch probabilities. */ -static void -estimate_bb_frequencies (struct loops *loops) +void +estimate_bb_frequencies (void) { basic_block bb; sreal freq_max; - if (flag_branch_probabilities) - counts_to_freqs (); - else + if (!flag_branch_probabilities || !counts_to_freqs ()) { static int real_values_initialized = 0; @@ -1132,33 +1827,8 @@ estimate_bb_frequencies (struct loops *loops) } mark_dfs_back_edges (); - /* Fill in the probability values in flowgraph based on the REG_BR_PROB - notes. */ - FOR_EACH_BB (bb) - { - rtx last_insn = BB_END (bb); - if (!can_predict_insn_p (last_insn)) - { - /* We can predict only conditional jumps at the moment. - Expect each edge to be equally probable. - ?? In the future we want to make abnormal edges improbable. */ - int nedges = 0; - edge e; - - for (e = bb->succ; e; e = e->succ_next) - { - nedges++; - if (e->probability != 0) - break; - } - if (!e) - for (e = bb->succ; e; e = e->succ_next) - e->probability = (REG_BR_PROB_BASE + nedges / 2) / nedges; - } - } - - ENTRY_BLOCK_PTR->succ->probability = REG_BR_PROB_BASE; + single_succ_edge (ENTRY_BLOCK_PTR)->probability = REG_BR_PROB_BASE; /* Set up block info for each basic block. */ alloc_aux_for_blocks (sizeof (struct block_info_def)); @@ -1166,9 +1836,9 @@ estimate_bb_frequencies (struct loops *loops) FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb) { edge e; + edge_iterator ei; - BLOCK_INFO (bb)->tovisit = 0; - for (e = bb->succ; e; e = e->succ_next) + FOR_EACH_EDGE (e, ei, bb->succs) { sreal_init (&EDGE_INFO (e)->back_edge_prob, e->probability, 0); sreal_mul (&EDGE_INFO (e)->back_edge_prob, @@ -1179,7 +1849,7 @@ estimate_bb_frequencies (struct loops *loops) /* First compute probabilities locally for each loop from innermost to outermost to examine probabilities for back edges. */ - estimate_loops_at_level (loops->tree_root); + estimate_loops (); memcpy (&freq_max, &real_zero, sizeof (real_zero)); FOR_EACH_BB (bb) @@ -1211,7 +1881,15 @@ compute_function_frequency (void) basic_block bb; if (!profile_info || !flag_branch_probabilities) - return; + { + if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl)) + != NULL) + cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED; + else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl)) + != NULL) + cfun->function_frequency = FUNCTION_FREQUENCY_HOT; + return; + } cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED; FOR_EACH_BB (bb) { @@ -1237,6 +1915,13 @@ choose_function_section (void) of all instances. For now just never set frequency for these. */ || DECL_ONE_ONLY (current_function_decl)) return; + + /* If we are doing the partitioning optimization, let the optimization + choose the correct section into which to put things. */ + + if (flag_reorder_blocks_and_partition) + return; + if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT) DECL_SECTION_NAME (current_function_decl) = build_string (strlen (HOT_TEXT_SECTION_NAME), HOT_TEXT_SECTION_NAME); @@ -1245,3 +1930,26 @@ choose_function_section (void) build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME), UNLIKELY_EXECUTED_TEXT_SECTION_NAME); } + +static bool +gate_estimate_probability (void) +{ + return flag_guess_branch_prob; +} + +struct tree_opt_pass pass_profile = +{ + "profile", /* name */ + gate_estimate_probability, /* gate */ + tree_estimate_probability, /* execute */ + NULL, /* sub */ + NULL, /* next */ + 0, /* static_pass_number */ + TV_BRANCH_PROB, /* tv_id */ + PROP_cfg, /* properties_required */ + 0, /* properties_provided */ + 0, /* properties_destroyed */ + 0, /* todo_flags_start */ + TODO_ggc_collect | TODO_verify_ssa, /* todo_flags_finish */ + 0 /* letter */ +};