/* Branch prediction routines for the GNU compiler.
- Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc.
+ Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008, 2009
+ 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
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
+<http://www.gnu.org/licenses/>. */
/* References:
#include "recog.h"
#include "expr.h"
#include "predict.h"
-#include "profile.h"
+#include "coverage.h"
#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. */
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)
+/* Random guesstimation given names.
+ PROV_VERY_UNLIKELY should be small enough so basic block predicted
+ by it gets bellow HOT_BB_FREQUENCY_FRANCTION. */
+#define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 2000 - 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 PARAMS ((basic_block,
- enum br_predictor));
-static void combine_predictions_for_insn PARAMS ((rtx, basic_block));
-static void dump_prediction PARAMS ((enum br_predictor, int,
- basic_block, int));
-static void estimate_loops_at_level PARAMS ((struct loop *loop));
-static void propagate_freq PARAMS ((struct loop *));
-static void estimate_bb_frequencies PARAMS ((struct loops *));
-static void counts_to_freqs PARAMS ((void));
-static void process_note_predictions PARAMS ((basic_block, int *,
- dominance_info,
- dominance_info));
-static void process_note_prediction PARAMS ((basic_block, int *,
- dominance_info,
- dominance_info, int, int));
-static bool last_basic_block_p PARAMS ((basic_block));
-static void compute_function_frequency PARAMS ((void));
-static void choose_function_section PARAMS ((void));
-static bool can_predict_insn_p PARAMS ((rtx));
+static void combine_predictions_for_insn (rtx, basic_block);
+static void dump_prediction (FILE *, enum br_predictor, int, basic_block, int);
+static void predict_paths_leading_to (basic_block, enum br_predictor, enum prediction);
+static void choose_function_section (void);
+static bool can_predict_insn_p (const_rtx);
/* Information we hold about each branch predictor.
Filled using information from predict.def. */
};
#undef DEF_PREDICTOR
+/* Return TRUE if frequency FREQ is considered to be hot. */
+
+static inline bool
+maybe_hot_frequency_p (int freq)
+{
+ 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 (profile_status == PROFILE_ABSENT)
+ return true;
+ if (freq < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
+ return false;
+ return true;
+}
+
+/* Return TRUE if frequency FREQ is considered to be hot. */
+
+static inline bool
+maybe_hot_count_p (gcov_type count)
+{
+ if (profile_status != PROFILE_READ)
+ return true;
+ /* Code executed at most once is not hot. */
+ if (profile_info->runs >= count)
+ return false;
+ return (count
+ > profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION));
+}
+
/* Return true in case BB can be CPU intensive and should be optimized
for maximal performance. */
bool
-maybe_hot_bb_p (bb)
- basic_block bb;
-{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities
- && (bb->count
- < profile_info.max_counter_in_program
- / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
+maybe_hot_bb_p (const_basic_block bb)
+{
+ if (profile_status == PROFILE_READ)
+ return maybe_hot_count_p (bb->count);
+ return maybe_hot_frequency_p (bb->frequency);
+}
+
+/* Return true if the call can be hot. */
+
+bool
+cgraph_maybe_hot_edge_p (struct cgraph_edge *edge)
+{
+ if (profile_info && flag_branch_probabilities
+ && (edge->count
+ <= profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
return false;
- if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
+ if (lookup_attribute ("cold", DECL_ATTRIBUTES (edge->callee->decl))
+ || lookup_attribute ("cold", DECL_ATTRIBUTES (edge->caller->decl)))
+ return false;
+ if (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl)))
+ return true;
+ if (flag_guess_branch_prob
+ && edge->frequency <= (CGRAPH_FREQ_BASE
+ / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)))
return false;
return true;
}
-/* Return true in case BB is cold and should be optimized for size. */
+/* Return true in case BB can be CPU intensive and should be optimized
+ for maximal performance. */
+
+bool
+maybe_hot_edge_p (edge e)
+{
+ if (profile_status == PROFILE_READ)
+ return maybe_hot_count_p (e->count);
+ return maybe_hot_frequency_p (EDGE_FREQUENCY (e));
+}
+/* Return true in case BB is probably never executed. */
bool
-probably_cold_bb_p (bb)
- basic_block bb;
-{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities
- && (bb->count
- < profile_info.max_counter_in_program
- / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
+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;
- if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
+ return false;
+}
+
+/* Return true when current function should always be optimized for size. */
+
+bool
+optimize_function_for_size_p (struct function *fun)
+{
+ return (optimize_size
+ || (fun && (fun->function_frequency
+ == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)));
+}
+
+/* Return true when current function should always be optimized for speed. */
+
+bool
+optimize_function_for_speed_p (struct function *fun)
+{
+ return !optimize_function_for_size_p (fun);
+}
+
+/* Return TRUE when BB should be optimized for size. */
+
+bool
+optimize_bb_for_size_p (const_basic_block bb)
+{
+ return optimize_function_for_size_p (cfun) || !maybe_hot_bb_p (bb);
+}
+
+/* Return TRUE when BB should be optimized for speed. */
+
+bool
+optimize_bb_for_speed_p (const_basic_block bb)
+{
+ return !optimize_bb_for_size_p (bb);
+}
+
+/* Return TRUE when BB should be optimized for size. */
+
+bool
+optimize_edge_for_size_p (edge e)
+{
+ return optimize_function_for_size_p (cfun) || !maybe_hot_edge_p (e);
+}
+
+/* Return TRUE when BB should be optimized for speed. */
+
+bool
+optimize_edge_for_speed_p (edge e)
+{
+ return !optimize_edge_for_size_p (e);
+}
+
+/* Return TRUE when BB should be optimized for size. */
+
+bool
+optimize_insn_for_size_p (void)
+{
+ return optimize_function_for_size_p (cfun) || !crtl->maybe_hot_insn_p;
+}
+
+/* Return TRUE when BB should be optimized for speed. */
+
+bool
+optimize_insn_for_speed_p (void)
+{
+ return !optimize_insn_for_size_p ();
+}
+
+/* Return TRUE when LOOP should be optimized for size. */
+
+bool
+optimize_loop_for_size_p (struct loop *loop)
+{
+ return optimize_bb_for_size_p (loop->header);
+}
+
+/* Return TRUE when LOOP should be optimized for speed. */
+
+bool
+optimize_loop_for_speed_p (struct loop *loop)
+{
+ return optimize_bb_for_speed_p (loop->header);
+}
+
+/* Return TRUE when LOOP nest should be optimized for speed. */
+
+bool
+optimize_loop_nest_for_speed_p (struct loop *loop)
+{
+ struct loop *l = loop;
+ if (optimize_loop_for_speed_p (loop))
return true;
+ l = loop->inner;
+ while (l && l != loop)
+ {
+ if (optimize_loop_for_speed_p (l))
+ return true;
+ if (l->inner)
+ l = l->inner;
+ else if (l->next)
+ l = l->next;
+ else
+ {
+ while (l != loop && !l->next)
+ l = loop_outer (l);
+ if (l != loop)
+ l = l->next;
+ }
+ }
return false;
}
-/* Return true in case BB is probably never executed. */
+/* Return TRUE when LOOP nest should be optimized for size. */
+
bool
-probably_never_executed_bb_p (bb)
- basic_block bb;
+optimize_loop_nest_for_size_p (struct loop *loop)
{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities)
- return ((bb->count + profile_info.count_profiles_merged / 2)
- / profile_info.count_profiles_merged) == 0;
+ return !optimize_loop_nest_for_speed_p (loop);
+}
+
+/* Return true when edge E is likely to be well predictable by branch
+ predictor. */
+
+bool
+predictable_edge_p (edge e)
+{
+ if (profile_status == PROFILE_ABSENT)
+ return false;
+ if ((e->probability
+ <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME) * REG_BR_PROB_BASE / 100)
+ || (REG_BR_PROB_BASE - e->probability
+ <= PARAM_VALUE (PARAM_PREDICTABLE_BRANCH_OUTCOME) * REG_BR_PROB_BASE / 100))
+ return true;
return false;
}
+
+/* Set RTL expansion for BB profile. */
+
+void
+rtl_profile_for_bb (basic_block bb)
+{
+ crtl->maybe_hot_insn_p = maybe_hot_bb_p (bb);
+}
+
+/* Set RTL expansion for edge profile. */
+
+void
+rtl_profile_for_edge (edge e)
+{
+ crtl->maybe_hot_insn_p = maybe_hot_edge_p (e);
+}
+
+/* Set RTL expansion to default mode (i.e. when profile info is not known). */
+void
+default_rtl_profile (void)
+{
+ crtl->maybe_hot_insn_p = true;
+}
+
/* Return true if the one of outgoing edges is already predicted by
PREDICTOR. */
-static bool
-predicted_by_p (bb, predictor)
- basic_block bb;
- enum br_predictor predictor;
+bool
+rtl_predicted_by_p (const_basic_block bb, enum br_predictor predictor)
{
rtx note;
- if (!INSN_P (bb->end))
+ if (!INSN_P (BB_END (bb)))
return false;
- for (note = REG_NOTES (bb->end); note; note = XEXP (note, 1))
+ for (note = REG_NOTES (BB_END (bb)); note; note = XEXP (note, 1))
if (REG_NOTE_KIND (note) == REG_BR_PRED
&& INTVAL (XEXP (XEXP (note, 0), 0)) == (int)predictor)
return true;
return false;
}
-void
-predict_insn (insn, predictor, probability)
- rtx insn;
- int probability;
- enum br_predictor predictor;
+/* 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
+gimple_predicted_by_p (const_basic_block bb, enum br_predictor predictor)
{
- if (!any_condjump_p (insn))
- abort ();
+ struct edge_prediction *i;
+ void **preds = pointer_map_contains (bb_predictions, bb);
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_BR_PRED,
- gen_rtx_CONCAT (VOIDmode,
- GEN_INT ((int) predictor),
- GEN_INT ((int) probability)),
- REG_NOTES (insn));
+ if (!preds)
+ return false;
+
+ for (i = (struct edge_prediction *) *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)
+{
+ gcc_assert (any_condjump_p (insn));
+ if (!flag_guess_branch_prob)
+ return;
+
+ add_reg_note (insn, REG_BR_PRED,
+ gen_rtx_CONCAT (VOIDmode,
+ GEN_INT ((int) predictor),
+ GEN_INT ((int) probability)));
}
/* Predict insn by given predictor. */
void
-predict_insn_def (insn, predictor, taken)
- rtx insn;
- enum br_predictor predictor;
- enum prediction taken;
+predict_insn_def (rtx insn, enum br_predictor predictor,
+ enum prediction taken)
{
int probability = predictor_info[(int) predictor].hitrate;
/* Predict edge E with given probability if possible. */
void
-predict_edge (e, predictor, probability)
- edge e;
- int probability;
- enum br_predictor predictor;
+rtl_predict_edge (edge e, enum br_predictor predictor, int probability)
{
rtx last_insn;
- last_insn = e->src->end;
+ last_insn = BB_END (e->src);
/* We can store the branch prediction information only about
conditional jumps. */
predict_insn (last_insn, predictor, probability);
}
+/* Predict edge E with the given PROBABILITY. */
+void
+gimple_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 = (struct edge_prediction *) *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 = (struct edge_prediction *) *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 (insn)
- 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. */
void
-predict_edge_def (e, predictor, taken)
- edge e;
- enum br_predictor predictor;
- enum prediction taken;
+predict_edge_def (edge e, enum br_predictor predictor,
+ enum prediction taken)
{
int probability = predictor_info[(int) predictor].hitrate;
to be done each time we invert the condition used by the jump. */
void
-invert_br_probabilities (insn)
- rtx insn;
+invert_br_probabilities (rtx insn)
{
rtx note;
/* Dump information about the branch prediction to the output file. */
static void
-dump_prediction (predictor, probability, bb, used)
- enum br_predictor predictor;
- int probability;
- basic_block bb;
- int used;
+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
note if not already present. Remove now useless REG_BR_PRED notes. */
static void
-combine_predictions_for_insn (insn, bb)
- rtx insn;
- basic_block bb;
+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;
+ enum br_predictor best_predictor = END_PREDICTORS;
int combined_probability = REG_BR_PROB_BASE / 2;
int d;
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)
{
- int predictor = INTVAL (XEXP (XEXP (note, 0), 0));
+ enum br_predictor predictor = ((enum br_predictor)
+ INTVAL (XEXP (XEXP (note, 0), 0)));
int probability = INTVAL (XEXP (XEXP (note, 0), 1));
found = true;
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)
{
if (REG_NOTE_KIND (*pnote) == REG_BR_PRED)
{
- int predictor = INTVAL (XEXP (XEXP (*pnote, 0), 0));
+ enum br_predictor predictor = ((enum br_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);
}
if (!prob_note)
{
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_BR_PROB,
- GEN_INT (combined_probability), REG_NOTES (insn));
+ add_reg_note (insn, REG_BR_PROB, GEN_INT (combined_probability));
/* 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 (loops_info)
- struct loops *loops_info;
+static void
+combine_predictions_for_bb (basic_block bb)
{
- dominance_info dominators, post_dominators;
- basic_block bb;
- unsigned i;
+ int best_probability = PROB_EVEN;
+ enum br_predictor 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 ();
- dominators = calculate_dominance_info (CDI_DOMINATORS);
- post_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 = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
+ {
+ enum br_predictor predictor = pred->ep_predictor;
+ int probability = pred->ep_probability;
+
+ if (pred->ep_edge != first)
+ probability = REG_BR_PROB_BASE - probability;
+
+ found = true;
+ /* First match heuristics would be widly confused if we predicted
+ both directions. */
+ if (best_predictor > predictor)
+ {
+ struct edge_prediction *pred2;
+ int prob = probability;
+
+ for (pred2 = (struct edge_prediction *) *preds; pred2; pred2 = pred2->ep_next)
+ if (pred2 != pred && pred2->ep_predictor == pred->ep_predictor)
+ {
+ int probability2 = pred->ep_probability;
+
+ if (pred2->ep_edge != first)
+ probability2 = REG_BR_PROB_BASE - probability2;
+
+ if ((probability < REG_BR_PROB_BASE / 2) !=
+ (probability2 < REG_BR_PROB_BASE / 2))
+ break;
+
+ /* If the same predictor later gave better result, go for it! */
+ if ((probability >= REG_BR_PROB_BASE / 2 && (probability2 > probability))
+ || (probability <= REG_BR_PROB_BASE / 2 && (probability2 < probability)))
+ prob = probability2;
+ }
+ if (!pred2)
+ best_probability = prob, 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 = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
+ {
+ enum br_predictor 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;
/* 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 (loops_info, loop, LOOP_EXIT_EDGES);
- exits = loop->num_exits;
+ exits = get_loop_exit_edges (loop);
+ n_exits = VEC_length (edge, exits);
- if (simple_loop_p (loops_info, loop, &desc)
- && desc.const_iter)
+ for (j = 0; VEC_iterate (edge, exits, j, ex); j++)
{
- niter = desc.niter + 1;
- if (niter == 0) /* We might overflow here. */
- niter = desc.niter;
-
- predict_edge (desc.in_edge, PRED_LOOP_ITERATIONS,
- REG_BR_PROB_BASE
- - (REG_BR_PROB_BASE + niter /2)
- / niter);
+ 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];
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)
- || 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 (bbs);
}
+}
- /* Attempt to predict conditional jumps using a number of heuristics. */
- FOR_EACH_BB (bb)
+/* 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;
- 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;
+ }
+}
+
+/* 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);
+}
+\f
+static tree expr_expected_value (tree, bitmap);
+
+/* Helper function for expr_expected_value. */
+
+static tree
+expr_expected_value_1 (tree type, tree op0, enum tree_code code, tree op1, bitmap visited)
+{
+ gimple def;
+
+ if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
+ {
+ if (TREE_CONSTANT (op0))
+ return op0;
+
+ if (code != SSA_NAME)
+ return NULL_TREE;
+
+ def = SSA_NAME_DEF_STMT (op0);
- for (e = bb->succ; e; e = e->succ_next)
+ /* If we were already here, break the infinite cycle. */
+ if (bitmap_bit_p (visited, SSA_NAME_VERSION (op0)))
+ return NULL;
+ bitmap_set_bit (visited, SSA_NAME_VERSION (op0));
+
+ if (gimple_code (def) == GIMPLE_PHI)
{
- /* 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);
-
- /* 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 (dominators, e->dest, e->src)
- && !dominated_by_p (post_dominators, e->src, e->dest))
+ /* All the arguments of the PHI node must have the same constant
+ length. */
+ int i, n = gimple_phi_num_args (def);
+ tree val = NULL, new_val;
+
+ for (i = 0; i < n; 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 = e->dest->head; insn != NEXT_INSN (e->dest->end);
- 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 (is_gimple_assign (def))
+ {
+ if (gimple_assign_lhs (def) != op0)
+ return NULL;
+
+ return expr_expected_value_1 (TREE_TYPE (gimple_assign_lhs (def)),
+ gimple_assign_rhs1 (def),
+ gimple_assign_rhs_code (def),
+ gimple_assign_rhs2 (def),
+ visited);
}
- cond = get_condition (last_insn, &earliest);
- if (! cond)
- continue;
-
- /* 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)))))
+ if (is_gimple_call (def))
{
- 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);
+ tree decl = gimple_call_fndecl (def);
+ if (!decl)
+ return NULL;
+ if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL
+ && DECL_FUNCTION_CODE (decl) == BUILT_IN_EXPECT)
+ {
+ tree val;
+
+ if (gimple_call_num_args (def) != 2)
+ return NULL;
+ val = gimple_call_arg (def, 0);
+ if (TREE_CONSTANT (val))
+ return val;
+ return gimple_call_arg (def, 1);
+ }
}
- else
- /* 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;
- }
+ return NULL;
}
- /* Attach the combined probability to each conditional jump. */
- FOR_EACH_BB (bb)
- if (GET_CODE (bb->end) == JUMP_INSN
- && any_condjump_p (bb->end)
- && bb->succ->succ_next != NULL)
- combine_predictions_for_insn (bb->end, bb);
+ if (get_gimple_rhs_class (code) == GIMPLE_BINARY_RHS)
+ {
+ tree res;
+ op0 = expr_expected_value (op0, visited);
+ if (!op0)
+ return NULL;
+ op1 = expr_expected_value (op1, visited);
+ if (!op1)
+ return NULL;
+ res = fold_build2 (code, type, op0, op1);
+ if (TREE_CONSTANT (res))
+ return res;
+ return NULL;
+ }
+ if (get_gimple_rhs_class (code) == GIMPLE_UNARY_RHS)
+ {
+ tree res;
+ op0 = expr_expected_value (op0, visited);
+ if (!op0)
+ return NULL;
+ res = fold_build1 (code, type, op0);
+ if (TREE_CONSTANT (res))
+ return res;
+ return NULL;
+ }
+ return NULL;
+}
+
+/* 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)
+{
+ enum tree_code code;
+ tree op0, op1;
- free_dominance_info (post_dominators);
- free_dominance_info (dominators);
+ if (TREE_CONSTANT (expr))
+ return expr;
- remove_fake_edges ();
- estimate_bb_frequencies (loops_info);
+ extract_ops_from_tree (expr, &code, &op0, &op1);
+ return expr_expected_value_1 (TREE_TYPE (expr),
+ op0, code, op1, visited);
}
-\f
-/* __builtin_expect dropped tokens into the insn stream describing expected
- values of registers. Generate branch probabilities based off these
- values. */
-void
-expected_value_to_br_prob ()
+\f
+/* Get rid of all builtin_expect calls and GIMPLE_PREDICT statements
+ we no longer need. */
+static unsigned int
+strip_predict_hints (void)
{
- rtx insn, cond, ev = NULL_RTX, ev_reg = NULL_RTX;
+ basic_block bb;
+ gimple ass_stmt;
+ tree var;
- for (insn = get_insns (); insn ; insn = NEXT_INSN (insn))
+ FOR_EACH_BB (bb)
{
- switch (GET_CODE (insn))
+ gimple_stmt_iterator bi;
+ for (bi = gsi_start_bb (bb); !gsi_end_p (bi);)
{
- case NOTE:
- /* Look for expected value notes. */
- if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_EXPECTED_VALUE)
+ gimple stmt = gsi_stmt (bi);
+
+ if (gimple_code (stmt) == GIMPLE_PREDICT)
{
- ev = NOTE_EXPECTED_VALUE (insn);
- ev_reg = XEXP (ev, 0);
- delete_insn (insn);
+ gsi_remove (&bi, true);
+ continue;
}
- 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;
+ else if (gimple_code (stmt) == GIMPLE_CALL)
+ {
+ tree fndecl = gimple_call_fndecl (stmt);
- default:
- /* Look for insns that clobber the EV register. */
- if (ev && reg_set_p (ev_reg, insn))
- ev = NULL_RTX;
- continue;
- }
+ if (fndecl
+ && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
+ && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_EXPECT
+ && gimple_call_num_args (stmt) == 2)
+ {
+ var = gimple_call_lhs (stmt);
+ ass_stmt = gimple_build_assign (var, gimple_call_arg (stmt, 0));
- /* 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);
- 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);
+ gsi_replace (&bi, ass_stmt, true);
+ }
+ }
+ gsi_next (&bi);
+ }
}
+ return 0;
}
\f
-/* Check whether this is the last basic block of function. Commonly tehre
- is one extra common cleanup block. */
-static bool
-last_basic_block_p (bb)
- basic_block bb;
+/* Predict using opcode of the last statement in basic block. */
+static void
+tree_predict_by_opcode (basic_block bb)
{
- if (bb == EXIT_BLOCK_PTR)
- return false;
+ gimple stmt = last_stmt (bb);
+ edge then_edge;
+ tree op0, op1;
+ tree type;
+ tree val;
+ enum tree_code cmp;
+ bitmap visited;
+ edge_iterator ei;
+
+ if (!stmt || gimple_code (stmt) != GIMPLE_COND)
+ return;
+ FOR_EACH_EDGE (then_edge, ei, bb->succs)
+ if (then_edge->flags & EDGE_TRUE_VALUE)
+ break;
+ op0 = gimple_cond_lhs (stmt);
+ op1 = gimple_cond_rhs (stmt);
+ cmp = gimple_cond_code (stmt);
+ type = TREE_TYPE (op0);
+ visited = BITMAP_ALLOC (NULL);
+ val = expr_expected_value_1 (boolean_type_node, op0, cmp, op1, 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 (cmp == EQ_EXPR)
+ predict_edge_def (then_edge, PRED_TREE_POINTER, NOT_TAKEN);
+ else if (cmp == NE_EXPR)
+ predict_edge_def (then_edge, PRED_TREE_POINTER, TAKEN);
+ }
+ else
- 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));
+ /* 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 (cmp)
+ {
+ 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 (op1))
+ ;
+ 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 (op1))
+ ;
+ 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 (op1)
+ || integer_onep (op1)
+ || integer_all_onesp (op1)
+ || real_zerop (op1)
+ || real_onep (op1)
+ || real_minus_onep (op1))
+ predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, NOT_TAKEN);
+ break;
+
+ case GE_EXPR:
+ case GT_EXPR:
+ if (integer_zerop (op1)
+ || integer_onep (op1)
+ || integer_all_onesp (op1)
+ || real_zerop (op1)
+ || real_onep (op1)
+ || real_minus_onep (op1))
+ predict_edge_def (then_edge, PRED_TREE_OPCODE_POSITIVE, TAKEN);
+ break;
+
+ default:
+ break;
+ }
}
-/* 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 such that we do not post-dominate it)
- (but we fill this information on demand, so -1 may be there in case this
- was not needed yet). */
+/* Try to guess whether the value of return means error code. */
-static void
-process_note_prediction (bb, heads, dominators, post_dominators, pred, flags)
- basic_block bb;
- int *heads;
- dominance_info dominators;
- dominance_info post_dominators;
- int pred;
- int flags;
+static enum br_predictor
+return_prediction (tree val, enum prediction *prediction)
{
- edge e;
- int y;
- bool taken;
-
- taken = flags & IS_TAKEN;
-
- if (heads[bb->index] < 0)
+ /* VOID. */
+ if (!val)
+ return PRED_NO_PREDICTION;
+ /* Different heuristics for pointers and scalars. */
+ if (POINTER_TYPE_P (TREE_TYPE (val)))
{
- /* This is first time we need this field in heads array; so
- find first dominator that we do not post-dominate (we are
- using already known members of heads array). */
- basic_block ai = bb;
- basic_block next_ai = get_immediate_dominator (dominators, bb);
- int head;
-
- while (heads[next_ai->index] < 0)
+ /* NULL is usually not returned. */
+ if (integer_zerop (val))
{
- if (!dominated_by_p (post_dominators, next_ai, bb))
- break;
- heads[next_ai->index] = ai->index;
- ai = next_ai;
- next_ai = get_immediate_dominator (dominators, next_ai);
+ *prediction = NOT_TAKEN;
+ return PRED_NULL_RETURN;
}
- if (!dominated_by_p (post_dominators, next_ai, bb))
- head = next_ai->index;
- else
- head = heads[next_ai->index];
- while (next_ai != bb)
+ }
+ 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)
{
- next_ai = ai;
- if (heads[ai->index] == ENTRY_BLOCK)
- ai = ENTRY_BLOCK_PTR;
- else
- ai = BASIC_BLOCK (heads[ai->index]);
- heads[next_ai->index] = head;
+ *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;
}
}
- y = heads[bb->index];
+ return PRED_NO_PREDICTION;
+}
- /* Now find the edge that leads to our branch and aply the prediction. */
+/* 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 (void)
+{
+ gimple return_stmt = NULL;
+ tree return_val;
+ edge e;
+ gimple phi;
+ int phi_num_args, i;
+ enum br_predictor pred;
+ enum prediction direction;
+ edge_iterator ei;
- if (y == last_basic_block || !can_predict_insn_p (BASIC_BLOCK (y)->end))
+ FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
+ {
+ return_stmt = last_stmt (e->src);
+ if (return_stmt
+ && gimple_code (return_stmt) == GIMPLE_RETURN)
+ break;
+ }
+ if (!e)
return;
- for (e = BASIC_BLOCK (y)->succ; e; e = e->succ_next)
- if (e->dest->index >= 0
- && dominated_by_p (post_dominators, e->dest, bb))
- predict_edge_def (e, pred, taken);
+ return_val = gimple_return_retval (return_stmt);
+ if (!return_val)
+ return;
+ if (TREE_CODE (return_val) != SSA_NAME
+ || !SSA_NAME_DEF_STMT (return_val)
+ || gimple_code (SSA_NAME_DEF_STMT (return_val)) != GIMPLE_PHI)
+ return;
+ phi = SSA_NAME_DEF_STMT (return_val);
+ phi_num_args = gimple_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 (gimple_phi_arg_edge (phi, i)->src, pred,
+ direction);
+ }
}
-/* Gathers NOTE_INSN_PREDICTIONs in given basic block and turns them
- into branch probabilities. For description of heads array, see
- process_note_prediction. */
+/* 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. */
static void
-process_note_predictions (bb, heads, dominators, post_dominators)
- basic_block bb;
- int *heads;
- dominance_info dominators;
- dominance_info post_dominators;
+tree_bb_level_predictions (void)
{
- rtx insn;
+ basic_block bb;
+ bool has_return_edges = false;
edge e;
+ edge_iterator ei;
- /* Additionally, we check here for blocks with no successors. */
- int contained_noreturn_call = 0;
- int was_bb_head = 0;
- int noreturn_block = 1;
+ FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
+ if (!(e->flags & (EDGE_ABNORMAL | EDGE_FAKE | EDGE_EH)))
+ {
+ has_return_edges = true;
+ break;
+ }
- for (insn = bb->end; insn;
- was_bb_head |= (insn == bb->head), insn = PREV_INSN (insn))
+ apply_return_prediction ();
+
+ FOR_EACH_BB (bb)
{
- if (GET_CODE (insn) != NOTE)
+ gimple_stmt_iterator gsi;
+
+ for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
- if (was_bb_head)
- break;
- else
+ gimple stmt = gsi_stmt (gsi);
+ tree decl;
+
+ if (is_gimple_call (stmt))
{
- /* 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 ((gimple_call_flags (stmt) & ECF_NORETURN)
+ && has_return_edges)
+ predict_paths_leading_to (bb, PRED_NORETURN,
+ NOT_TAKEN);
+ decl = gimple_call_fndecl (stmt);
+ if (decl
+ && lookup_attribute ("cold",
+ DECL_ATTRIBUTES (decl)))
+ predict_paths_leading_to (bb, PRED_COLD_FUNCTION,
+ NOT_TAKEN);
+ }
+ else if (gimple_code (stmt) == GIMPLE_PREDICT)
+ {
+ predict_paths_leading_to (bb, gimple_predict_predictor (stmt),
+ gimple_predict_outcome (stmt));
+ /* Keep GIMPLE_PREDICT around so early inlining will propagate
+ hints to callers. */
}
- }
- if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_PREDICTION)
- {
- int alg = (int) NOTE_PREDICTION_ALG (insn);
- /* Process single prediction note. */
- process_note_prediction (bb,
- heads,
- dominators,
- post_dominators,
- 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)
+}
+
+#ifdef ENABLE_CHECKING
+
+/* Callback for pointer_map_traverse, asserts that the pointer map is
+ empty. */
+
+static bool
+assert_is_empty (const void *key ATTRIBUTE_UNUSED, void **value,
+ void *data ATTRIBUTE_UNUSED)
+{
+ gcc_assert (!*value);
+ return false;
+}
+#endif
+
+/* Predict branch probabilities and estimate profile for basic block BB. */
+
+static void
+tree_estimate_probability_bb (basic_block bb)
+{
+ edge e;
+ edge_iterator ei;
+ gimple last;
+
+ FOR_EACH_EDGE (e, ei, bb->succs)
{
- /* 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,
- dominators,
- post_dominators, PRED_NORETURN, NOT_TAKEN);
+ /* 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
+ && (last = last_stmt (e->dest)) != NULL
+ && gimple_code (last) == GIMPLE_RETURN)
+ {
+ 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);
+ }
+
+ /* 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))
+ {
+ gimple_stmt_iterator bi;
+
+ /* 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 = gsi_start_bb (e->dest); !gsi_end_p (bi);
+ gsi_next (&bi))
+ {
+ gimple stmt = gsi_stmt (bi);
+ if (is_gimple_call (stmt)
+ /* Constant and pure calls are hardly used to signalize
+ something exceptional. */
+ && gimple_has_side_effects (stmt))
+ {
+ predict_edge_def (e, PRED_CALL, NOT_TAKEN);
+ break;
+ }
+ }
+ }
}
+ tree_predict_by_opcode (bb);
}
-/* Gathers NOTE_INSN_PREDICTIONs and turns them into
- branch probabilities. */
+/* Predict branch probabilities and estimate profile of the tree CFG.
+ This function can be called from the loop optimizers to recompute
+ the profile information. */
void
-note_prediction_to_br_prob ()
+tree_estimate_probability (void)
{
basic_block bb;
- dominance_info post_dominators, dominators;
- int *heads;
- /* To enable handling of noreturn blocks. */
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);
- post_dominators = calculate_dominance_info (CDI_POST_DOMINATORS);
- dominators = calculate_dominance_info (CDI_DOMINATORS);
+ bb_predictions = pointer_map_create ();
+ tree_bb_level_predictions ();
+ record_loop_exits ();
- 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;
+ if (number_of_loops () > 1)
+ predict_loops ();
- /* Process all prediction notes. */
+ FOR_EACH_BB (bb)
+ tree_estimate_probability_bb (bb);
FOR_EACH_BB (bb)
- process_note_predictions (bb, heads, dominators, post_dominators);
+ combine_predictions_for_bb (bb);
+
+#ifdef ENABLE_CHECKING
+ pointer_map_traverse (bb_predictions, assert_is_empty, NULL);
+#endif
+ pointer_map_destroy (bb_predictions);
+ bb_predictions = NULL;
+
+ estimate_bb_frequencies ();
+ free_dominance_info (CDI_POST_DOMINATORS);
+ remove_fake_exit_edges ();
+}
- free_dominance_info (post_dominators);
- free_dominance_info (dominators);
- free (heads);
+/* Predict branch probabilities and estimate profile of the tree CFG.
+ This is the driver function for PASS_PROFILE. */
- remove_fake_edges ();
+static unsigned int
+tree_estimate_probability_driver (void)
+{
+ unsigned nb_loops;
+
+ loop_optimizer_init (0);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ flow_loops_dump (dump_file, NULL, 0);
+
+ mark_irreducible_loops ();
+
+ nb_loops = number_of_loops ();
+ if (nb_loops > 1)
+ scev_initialize ();
+
+ tree_estimate_probability ();
+
+ if (nb_loops > 1)
+ scev_finalize ();
+
+ loop_optimizer_finalize ();
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ gimple_dump_cfg (dump_file, dump_flags);
+ if (profile_status == PROFILE_ABSENT)
+ profile_status = PROFILE_GUESSED;
+ return 0;
+}
+\f
+/* Predict edges to successors of CUR whose sources are not postdominated by
+ BB by PRED and recurse to all postdominators. */
+
+static void
+predict_paths_for_bb (basic_block cur, basic_block bb,
+ enum br_predictor pred,
+ enum prediction taken)
+{
+ edge e;
+ edge_iterator ei;
+ basic_block son;
+
+ /* We are looking for all edges forming edge cut induced by
+ set of all blocks postdominated by BB. */
+ FOR_EACH_EDGE (e, ei, cur->preds)
+ if (e->src->index >= NUM_FIXED_BLOCKS
+ && !dominated_by_p (CDI_POST_DOMINATORS, e->src, bb))
+ {
+ gcc_assert (bb == cur || dominated_by_p (CDI_POST_DOMINATORS, cur, bb));
+ predict_edge_def (e, pred, taken);
+ }
+ for (son = first_dom_son (CDI_POST_DOMINATORS, cur);
+ son;
+ son = next_dom_son (CDI_POST_DOMINATORS, son))
+ predict_paths_for_bb (son, bb, pred, taken);
+}
+
+/* Sets branch probabilities according to PREDiction and
+ FLAGS. */
+
+static void
+predict_paths_leading_to (basic_block bb, enum br_predictor pred,
+ enum prediction taken)
+{
+ predict_paths_for_bb (bb, bb, pred, taken);
}
\f
/* This is used to carry information about basic blocks. It is
/* To keep queue of basic blocks to process. */
basic_block next;
- /* True if block needs to be visited in prop_freqency. */
- int tovisit:1;
-
/* Number of predecessors we need to visit first. */
int npredecessors;
} *block_info;
/* 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. */
- int back_edge:1;
+ /* True if the edge is a loopback edge in the natural loop. */
+ unsigned int back_edge:1;
} *edge_info;
#define BLOCK_INFO(B) ((block_info) (B)->aux)
#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 (loop)
- 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));
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,
sizeof (real_almost_one));
}
- /* BLOCK_INFO (bb)->frequency = frequency
+ /* BLOCK_INFO (bb)->frequency = frequency
/ (1 - cyclic_probability) */
sreal_sub (&cyclic_probability, &real_one, &cyclic_probability);
}
}
- BLOCK_INFO (bb)->tovisit = 0;
+ bitmap_clear_bit (tovisit, bb->index);
- /* Compute back edge frequencies. */
- for (e = bb->succ; e; e = e->succ_next)
- if (e->dest == head)
- {
- sreal tmp;
+ 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); */
+ /* 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);
- }
+ 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)
{
last = e->dest;
}
- }
+ }
}
}
/* Estimate probabilities of loopback edges in loops at same nest level. */
static void
-estimate_loops_at_level (first_loop)
- struct loop *first_loop;
+estimate_loops_at_level (struct loop *first_loop)
{
struct loop *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
-counts_to_freqs ()
+estimate_loops (void)
{
- gcov_type count_max = 1;
+ bitmap tovisit = BITMAP_ALLOC (NULL);
basic_block bb;
- FOR_EACH_BB (bb)
- count_max = MAX (bb->count, count_max);
+ /* 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, true_count_max = 0;
+ basic_block bb;
+
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
+ 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
function can execute at average to be still considered not expensive. */
bool
-expensive_function_p (threshold)
- int threshold;
+expensive_function_p (int threshold)
{
unsigned int sum = 0;
basic_block bb;
/* 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
{
rtx insn;
- for (insn = bb->head; insn != NEXT_INSN (bb->end);
+ for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb));
insn = NEXT_INSN (insn))
if (active_insn_p (insn))
{
/* Estimate basic blocks frequency by given branch probabilities. */
-static void
-estimate_bb_frequencies (loops)
- struct loops *loops;
+void
+estimate_bb_frequencies (void)
{
basic_block bb;
sreal freq_max;
- if (flag_branch_probabilities)
- counts_to_freqs ();
- else
+ if (profile_status != PROFILE_READ || !counts_to_freqs ())
{
- sreal_init (&real_zero, 0, 0);
- sreal_init (&real_one, 1, 0);
- sreal_init (&real_br_prob_base, REG_BR_PROB_BASE, 0);
- sreal_init (&real_bb_freq_max, BB_FREQ_MAX, 0);
- sreal_init (&real_one_half, 1, -1);
- sreal_div (&real_inv_br_prob_base, &real_one, &real_br_prob_base);
- sreal_sub (&real_almost_one, &real_one, &real_inv_br_prob_base);
+ static int real_values_initialized = 0;
+
+ if (!real_values_initialized)
+ {
+ real_values_initialized = 1;
+ sreal_init (&real_zero, 0, 0);
+ sreal_init (&real_one, 1, 0);
+ sreal_init (&real_br_prob_base, REG_BR_PROB_BASE, 0);
+ sreal_init (&real_bb_freq_max, BB_FREQ_MAX, 0);
+ sreal_init (&real_one_half, 1, -1);
+ sreal_div (&real_inv_br_prob_base, &real_one, &real_br_prob_base);
+ sreal_sub (&real_almost_one, &real_one, &real_inv_br_prob_base);
+ }
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;
-
- 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));
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,
/* 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)
}
/* Decide whether function is hot, cold or unlikely executed. */
-static void
-compute_function_frequency ()
+void
+compute_function_frequency (void)
{
basic_block bb;
- if (!profile_info.count_profiles_merged
- || !flag_branch_probabilities)
- return;
+ if (!profile_info || !flag_branch_probabilities)
+ {
+ 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)
{
/* Choose appropriate section for the function. */
static void
-choose_function_section ()
+choose_function_section (void)
{
if (DECL_SECTION_NAME (current_function_decl)
|| !targetm.have_named_sections
/* Theoretically we can split the gnu.linkonce text section too,
- but this requires more work as the frequency needs to match
+ but this requires more work as the frequency needs to match
for all generated objects so we need to merge the frequency
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);
build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME),
UNLIKELY_EXECUTED_TEXT_SECTION_NAME);
}
+
+static bool
+gate_estimate_probability (void)
+{
+ return flag_guess_branch_prob;
+}
+
+/* Build PREDICT_EXPR. */
+tree
+build_predict_expr (enum br_predictor predictor, enum prediction taken)
+{
+ tree t = build1 (PREDICT_EXPR, void_type_node,
+ build_int_cst (NULL, predictor));
+ SET_PREDICT_EXPR_OUTCOME (t, taken);
+ return t;
+}
+
+const char *
+predictor_name (enum br_predictor predictor)
+{
+ return predictor_info[predictor].name;
+}
+
+struct gimple_opt_pass pass_profile =
+{
+ {
+ GIMPLE_PASS,
+ "profile", /* name */
+ gate_estimate_probability, /* gate */
+ tree_estimate_probability_driver, /* 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 */
+ }
+};
+
+struct gimple_opt_pass pass_strip_predict_hints =
+{
+ {
+ GIMPLE_PASS,
+ "*strip_predict_hints", /* name */
+ NULL, /* gate */
+ strip_predict_hints, /* 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 */
+ }
+};
projects / gcc.git / blobdiff