blob: 68b111356804e2de6ce60aa08b64d0a9b1fc3fe7 [file] [log] [blame]
/* Branch prediction routines for the GNU compiler.
Copyright (C) 2000-2021 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 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
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 COPYING3. If not see
<http://www.gnu.org/licenses/>. */
/* References:
[1] "Branch Prediction for Free"
Ball and Larus; PLDI '93.
[2] "Static Branch Frequency and Program Profile Analysis"
Wu and Larus; MICRO-27.
[3] "Corpus-based Static Branch Prediction"
Calder, Grunwald, Lindsay, Martin, Mozer, and Zorn; PLDI '95. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "cfghooks.h"
#include "tree-pass.h"
#include "ssa.h"
#include "memmodel.h"
#include "emit-rtl.h"
#include "cgraph.h"
#include "coverage.h"
#include "diagnostic-core.h"
#include "gimple-predict.h"
#include "fold-const.h"
#include "calls.h"
#include "cfganal.h"
#include "profile.h"
#include "sreal.h"
#include "cfgloop.h"
#include "gimple-iterator.h"
#include "tree-cfg.h"
#include "tree-ssa-loop-niter.h"
#include "tree-ssa-loop.h"
#include "tree-scalar-evolution.h"
#include "ipa-utils.h"
#include "gimple-pretty-print.h"
#include "selftest.h"
#include "cfgrtl.h"
#include "stringpool.h"
#include "attribs.h"
/* Enum with reasons why a predictor is ignored. */
enum predictor_reason
{
REASON_NONE,
REASON_IGNORED,
REASON_SINGLE_EDGE_DUPLICATE,
REASON_EDGE_PAIR_DUPLICATE
};
/* String messages for the aforementioned enum. */
static const char *reason_messages[] = {"", " (ignored)",
" (single edge duplicate)", " (edge pair duplicate)"};
static void combine_predictions_for_insn (rtx_insn *, basic_block);
static void dump_prediction (FILE *, enum br_predictor, int, basic_block,
enum predictor_reason, edge);
static void predict_paths_leading_to (basic_block, enum br_predictor,
enum prediction,
class loop *in_loop = NULL);
static void predict_paths_leading_to_edge (edge, enum br_predictor,
enum prediction,
class loop *in_loop = NULL);
static bool can_predict_insn_p (const rtx_insn *);
static HOST_WIDE_INT get_predictor_value (br_predictor, HOST_WIDE_INT);
static void determine_unlikely_bbs ();
/* Information we hold about each branch predictor.
Filled using information from predict.def. */
struct predictor_info
{
const char *const name; /* Name used in the debugging dumps. */
const int hitrate; /* Expected hitrate used by
predict_insn_def call. */
const int flags;
};
/* Use given predictor without Dempster-Shaffer theory if it matches
using first_match heuristics. */
#define PRED_FLAG_FIRST_MATCH 1
/* Recompute hitrate in percent to our representation. */
#define HITRATE(VAL) ((int) ((VAL) * REG_BR_PROB_BASE + 50) / 100)
#define DEF_PREDICTOR(ENUM, NAME, HITRATE, FLAGS) {NAME, HITRATE, FLAGS},
static const struct predictor_info predictor_info[]= {
#include "predict.def"
/* Upper bound on predictors. */
{NULL, 0, 0}
};
#undef DEF_PREDICTOR
static gcov_type min_count = -1;
/* Determine the threshold for hot BB counts. */
gcov_type
get_hot_bb_threshold ()
{
if (min_count == -1)
{
const int hot_frac = param_hot_bb_count_fraction;
const gcov_type min_hot_count
= hot_frac
? profile_info->sum_max / hot_frac
: (gcov_type)profile_count::max_count;
set_hot_bb_threshold (min_hot_count);
if (dump_file)
fprintf (dump_file, "Setting hotness threshold to %" PRId64 ".\n",
min_hot_count);
}
return min_count;
}
/* Set the threshold for hot BB counts. */
void
set_hot_bb_threshold (gcov_type min)
{
min_count = min;
}
/* Return TRUE if COUNT is considered to be hot in function FUN. */
bool
maybe_hot_count_p (struct function *fun, profile_count count)
{
if (!count.initialized_p ())
return true;
if (count.ipa () == profile_count::zero ())
return false;
if (!count.ipa_p ())
{
struct cgraph_node *node = cgraph_node::get (fun->decl);
if (!profile_info || profile_status_for_fn (fun) != PROFILE_READ)
{
if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
return false;
if (node->frequency == NODE_FREQUENCY_HOT)
return true;
}
if (profile_status_for_fn (fun) == PROFILE_ABSENT)
return true;
if (node->frequency == NODE_FREQUENCY_EXECUTED_ONCE
&& count < (ENTRY_BLOCK_PTR_FOR_FN (fun)->count.apply_scale (2, 3)))
return false;
if (count.apply_scale (param_hot_bb_frequency_fraction, 1)
< ENTRY_BLOCK_PTR_FOR_FN (fun)->count)
return false;
return true;
}
/* Code executed at most once is not hot. */
if (count <= MAX (profile_info ? profile_info->runs : 1, 1))
return false;
return (count >= get_hot_bb_threshold ());
}
/* Return true if basic block BB of function FUN can be CPU intensive
and should thus be optimized for maximum performance. */
bool
maybe_hot_bb_p (struct function *fun, const_basic_block bb)
{
gcc_checking_assert (fun);
return maybe_hot_count_p (fun, bb->count);
}
/* Return true if edge E can be CPU intensive and should thus be optimized
for maximum performance. */
bool
maybe_hot_edge_p (edge e)
{
return maybe_hot_count_p (cfun, e->count ());
}
/* Return true if COUNT is considered to be never executed in function FUN
or if function FUN is considered so in the static profile. */
static bool
probably_never_executed (struct function *fun, profile_count count)
{
gcc_checking_assert (fun);
if (count.ipa () == profile_count::zero ())
return true;
/* Do not trust adjusted counts. This will make us to drop int cold section
code with low execution count as a result of inlining. These low counts
are not safe even with read profile and may lead us to dropping
code which actually gets executed into cold section of binary that is not
desirable. */
if (count.precise_p () && profile_status_for_fn (fun) == PROFILE_READ)
{
const int unlikely_frac = param_unlikely_bb_count_fraction;
if (count.apply_scale (unlikely_frac, 1) >= profile_info->runs)
return false;
return true;
}
if ((!profile_info || profile_status_for_fn (fun) != PROFILE_READ)
&& (cgraph_node::get (fun->decl)->frequency
== NODE_FREQUENCY_UNLIKELY_EXECUTED))
return true;
return false;
}
/* Return true if basic block BB of function FUN is probably never executed. */
bool
probably_never_executed_bb_p (struct function *fun, const_basic_block bb)
{
return probably_never_executed (fun, bb->count);
}
/* Return true if edge E is unlikely executed for obvious reasons. */
static bool
unlikely_executed_edge_p (edge e)
{
return (e->src->count == profile_count::zero ()
|| e->probability == profile_probability::never ())
|| (e->flags & (EDGE_EH | EDGE_FAKE));
}
/* Return true if edge E of function FUN is probably never executed. */
bool
probably_never_executed_edge_p (struct function *fun, edge e)
{
if (unlikely_executed_edge_p (e))
return true;
return probably_never_executed (fun, e->count ());
}
/* Return true if function FUN should always be optimized for size. */
optimize_size_level
optimize_function_for_size_p (struct function *fun)
{
if (!fun || !fun->decl)
return optimize_size ? OPTIMIZE_SIZE_MAX : OPTIMIZE_SIZE_NO;
cgraph_node *n = cgraph_node::get (fun->decl);
if (n)
return n->optimize_for_size_p ();
return OPTIMIZE_SIZE_NO;
}
/* Return true if function FUN should always be optimized for speed. */
bool
optimize_function_for_speed_p (struct function *fun)
{
return !optimize_function_for_size_p (fun);
}
/* Return the optimization type that should be used for function FUN. */
optimization_type
function_optimization_type (struct function *fun)
{
return (optimize_function_for_speed_p (fun)
? OPTIMIZE_FOR_SPEED
: OPTIMIZE_FOR_SIZE);
}
/* Return TRUE if basic block BB should be optimized for size. */
optimize_size_level
optimize_bb_for_size_p (const_basic_block bb)
{
enum optimize_size_level ret = optimize_function_for_size_p (cfun);
if (bb && ret < OPTIMIZE_SIZE_MAX && bb->count == profile_count::zero ())
ret = OPTIMIZE_SIZE_MAX;
if (bb && ret < OPTIMIZE_SIZE_BALANCED && !maybe_hot_bb_p (cfun, bb))
ret = OPTIMIZE_SIZE_BALANCED;
return ret;
}
/* Return TRUE if basic block BB should be optimized for speed. */
bool
optimize_bb_for_speed_p (const_basic_block bb)
{
return !optimize_bb_for_size_p (bb);
}
/* Return the optimization type that should be used for basic block BB. */
optimization_type
bb_optimization_type (const_basic_block bb)
{
return (optimize_bb_for_speed_p (bb)
? OPTIMIZE_FOR_SPEED
: OPTIMIZE_FOR_SIZE);
}
/* Return TRUE if edge E should be optimized for size. */
optimize_size_level
optimize_edge_for_size_p (edge e)
{
enum optimize_size_level ret = optimize_function_for_size_p (cfun);
if (ret < OPTIMIZE_SIZE_MAX && unlikely_executed_edge_p (e))
ret = OPTIMIZE_SIZE_MAX;
if (ret < OPTIMIZE_SIZE_BALANCED && !maybe_hot_edge_p (e))
ret = OPTIMIZE_SIZE_BALANCED;
return ret;
}
/* Return TRUE if edge E should be optimized for speed. */
bool
optimize_edge_for_speed_p (edge e)
{
return !optimize_edge_for_size_p (e);
}
/* Return TRUE if the current function is optimized for size. */
optimize_size_level
optimize_insn_for_size_p (void)
{
enum optimize_size_level ret = optimize_function_for_size_p (cfun);
if (ret < OPTIMIZE_SIZE_BALANCED && !crtl->maybe_hot_insn_p)
ret = OPTIMIZE_SIZE_BALANCED;
return ret;
}
/* Return TRUE if the current function is optimized for speed. */
bool
optimize_insn_for_speed_p (void)
{
return !optimize_insn_for_size_p ();
}
/* Return TRUE if LOOP should be optimized for size. */
optimize_size_level
optimize_loop_for_size_p (class loop *loop)
{
return optimize_bb_for_size_p (loop->header);
}
/* Return TRUE if LOOP should be optimized for speed. */
bool
optimize_loop_for_speed_p (class loop *loop)
{
return optimize_bb_for_speed_p (loop->header);
}
/* Return TRUE if nest rooted at LOOP should be optimized for speed. */
bool
optimize_loop_nest_for_speed_p (class loop *loop)
{
class 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 if nest rooted at LOOP should be optimized for size. */
optimize_size_level
optimize_loop_nest_for_size_p (class loop *loop)
{
enum optimize_size_level ret = optimize_loop_for_size_p (loop);
class loop *l = loop;
l = loop->inner;
while (l && l != loop)
{
if (ret == OPTIMIZE_SIZE_NO)
break;
ret = MIN (optimize_loop_for_size_p (l), ret);
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 ret;
}
/* Return true if edge E is likely to be well predictable by branch
predictor. */
bool
predictable_edge_p (edge e)
{
if (!e->probability.initialized_p ())
return false;
if ((e->probability.to_reg_br_prob_base ()
<= param_predictable_branch_outcome * REG_BR_PROB_BASE / 100)
|| (REG_BR_PROB_BASE - e->probability.to_reg_br_prob_base ()
<= 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 (cfun, 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. */
bool
rtl_predicted_by_p (const_basic_block bb, enum br_predictor predictor)
{
rtx note;
if (!INSN_P (BB_END (bb)))
return false;
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;
}
/* Structure representing predictions in tree level. */
struct edge_prediction {
struct edge_prediction *ep_next;
edge ep_edge;
enum br_predictor ep_predictor;
int ep_probability;
};
/* This map contains for a basic block the list of predictions for the
outgoing edges. */
static hash_map<const_basic_block, edge_prediction *> *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)
{
struct edge_prediction *i;
edge_prediction **preds = bb_predictions->get (bb);
if (!preds)
return false;
for (i = *preds; i; i = i->ep_next)
if (i->ep_predictor == predictor)
return true;
return false;
}
/* Return true if the one of outgoing edges is already predicted by
PREDICTOR for edge E predicted as TAKEN. */
bool
edge_predicted_by_p (edge e, enum br_predictor predictor, bool taken)
{
struct edge_prediction *i;
basic_block bb = e->src;
edge_prediction **preds = bb_predictions->get (bb);
if (!preds)
return false;
int probability = predictor_info[(int) predictor].hitrate;
if (taken != TAKEN)
probability = REG_BR_PROB_BASE - probability;
for (i = *preds; i; i = i->ep_next)
if (i->ep_predictor == predictor
&& i->ep_edge == e
&& i->ep_probability == probability)
return true;
return false;
}
/* Same predicate as above, working on edges. */
bool
edge_probability_reliable_p (const_edge e)
{
return e->probability.probably_reliable_p ();
}
/* 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 profile_probability::from_reg_br_prob_note
(XINT (note, 0)).probably_reliable_p ();
}
static void
predict_insn (rtx_insn *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 (rtx_insn *insn, enum br_predictor predictor,
enum prediction taken)
{
int probability = predictor_info[(int) predictor].hitrate;
gcc_assert (probability != PROB_UNINITIALIZED);
if (taken != TAKEN)
probability = REG_BR_PROB_BASE - probability;
predict_insn (insn, predictor, probability);
}
/* Predict edge E with given probability if possible. */
void
rtl_predict_edge (edge e, enum br_predictor predictor, int probability)
{
rtx_insn *last_insn;
last_insn = BB_END (e->src);
/* We can store the branch prediction information only about
conditional jumps. */
if (!any_condjump_p (last_insn))
return;
/* We always store probability of branching. */
if (e->flags & EDGE_FALLTHRU)
probability = REG_BR_PROB_BASE - probability;
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)
{
if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun)
&& EDGE_COUNT (e->src->succs) > 1
&& flag_guess_branch_prob
&& optimize)
{
struct edge_prediction *i = XNEW (struct edge_prediction);
edge_prediction *&preds = bb_predictions->get_or_insert (e->src);
i->ep_next = preds;
preds = i;
i->ep_probability = probability;
i->ep_predictor = predictor;
i->ep_edge = e;
}
}
/* Filter edge predictions PREDS by a function FILTER: if FILTER return false
the prediction is removed.
DATA are passed to the filter function. */
static void
filter_predictions (edge_prediction **preds,
bool (*filter) (edge_prediction *, void *), void *data)
{
if (!bb_predictions)
return;
if (preds)
{
struct edge_prediction **prediction = preds;
struct edge_prediction *next;
while (*prediction)
{
if ((*filter) (*prediction, data))
prediction = &((*prediction)->ep_next);
else
{
next = (*prediction)->ep_next;
free (*prediction);
*prediction = next;
}
}
}
}
/* Filter function predicate that returns true for a edge predicate P
if its edge is equal to DATA. */
static bool
not_equal_edge_p (edge_prediction *p, void *data)
{
return p->ep_edge != (edge)data;
}
/* Remove all predictions on given basic block that are attached
to edge E. */
void
remove_predictions_associated_with_edge (edge e)
{
if (!bb_predictions)
return;
edge_prediction **preds = bb_predictions->get (e->src);
filter_predictions (preds, not_equal_edge_p, e);
}
/* Clears the list of predictions stored for BB. */
static void
clear_bb_predictions (basic_block bb)
{
edge_prediction **preds = bb_predictions->get (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 (const rtx_insn *insn)
{
return (JUMP_P (insn)
&& any_condjump_p (insn)
&& EDGE_COUNT (BLOCK_FOR_INSN (insn)->succs) >= 2);
}
/* Predict edge E by given predictor if possible. */
void
predict_edge_def (edge e, enum br_predictor predictor,
enum prediction taken)
{
int probability = predictor_info[(int) predictor].hitrate;
if (taken != TAKEN)
probability = REG_BR_PROB_BASE - probability;
predict_edge (e, predictor, probability);
}
/* Invert all branch predictions or probability notes in the INSN. This needs
to be done each time we invert the condition used by the jump. */
void
invert_br_probabilities (rtx insn)
{
rtx note;
for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
if (REG_NOTE_KIND (note) == REG_BR_PROB)
XINT (note, 0) = profile_probability::from_reg_br_prob_note
(XINT (note, 0)).invert ().to_reg_br_prob_note ();
else if (REG_NOTE_KIND (note) == REG_BR_PRED)
XEXP (XEXP (note, 0), 1)
= GEN_INT (REG_BR_PROB_BASE - INTVAL (XEXP (XEXP (note, 0), 1)));
}
/* Dump information about the branch prediction to the output file. */
static void
dump_prediction (FILE *file, enum br_predictor predictor, int probability,
basic_block bb, enum predictor_reason reason = REASON_NONE,
edge ep_edge = NULL)
{
edge e = ep_edge;
edge_iterator ei;
if (!file)
return;
if (e == NULL)
FOR_EACH_EDGE (e, ei, bb->succs)
if (! (e->flags & EDGE_FALLTHRU))
break;
char edge_info_str[128];
if (ep_edge)
sprintf (edge_info_str, " of edge %d->%d", ep_edge->src->index,
ep_edge->dest->index);
else
edge_info_str[0] = '\0';
fprintf (file, " %s heuristics%s%s: %.2f%%",
predictor_info[predictor].name,
edge_info_str, reason_messages[reason],
probability * 100.0 / REG_BR_PROB_BASE);
if (bb->count.initialized_p ())
{
fprintf (file, " exec ");
bb->count.dump (file);
if (e)
{
fprintf (file, " hit ");
e->count ().dump (file);
fprintf (file, " (%.1f%%)", e->count ().to_gcov_type() * 100.0
/ bb->count.to_gcov_type ());
}
}
fprintf (file, "\n");
/* Print output that be easily read by analyze_brprob.py script. We are
interested only in counts that are read from GCDA files. */
if (dump_file && (dump_flags & TDF_DETAILS)
&& bb->count.precise_p ()
&& reason == REASON_NONE)
{
fprintf (file, ";;heuristics;%s;%" PRId64 ";%" PRId64 ";%.1f;\n",
predictor_info[predictor].name,
bb->count.to_gcov_type (), e->count ().to_gcov_type (),
probability * 100.0 / REG_BR_PROB_BASE);
}
}
/* Return true if STMT is known to be unlikely executed. */
static bool
unlikely_executed_stmt_p (gimple *stmt)
{
if (!is_gimple_call (stmt))
return false;
/* NORETURN attribute alone is not strong enough: exit() may be quite
likely executed once during program run. */
if (gimple_call_fntype (stmt)
&& lookup_attribute ("cold",
TYPE_ATTRIBUTES (gimple_call_fntype (stmt)))
&& !lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl)))
return true;
tree decl = gimple_call_fndecl (stmt);
if (!decl)
return false;
if (lookup_attribute ("cold", DECL_ATTRIBUTES (decl))
&& !lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl)))
return true;
cgraph_node *n = cgraph_node::get (decl);
if (!n)
return false;
availability avail;
n = n->ultimate_alias_target (&avail);
if (avail < AVAIL_AVAILABLE)
return false;
if (!n->analyzed
|| n->decl == current_function_decl)
return false;
return n->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED;
}
/* Return true if BB is unlikely executed. */
static bool
unlikely_executed_bb_p (basic_block bb)
{
if (bb->count == profile_count::zero ())
return true;
if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) || bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
return false;
for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
!gsi_end_p (gsi); gsi_next (&gsi))
{
if (unlikely_executed_stmt_p (gsi_stmt (gsi)))
return true;
if (stmt_can_terminate_bb_p (gsi_stmt (gsi)))
return false;
}
return false;
}
/* We cannot predict the probabilities of outgoing edges of bb. Set them
evenly and hope for the best. If UNLIKELY_EDGES is not null, distribute
even probability for all edges not mentioned in the set. These edges
are given PROB_VERY_UNLIKELY probability. Similarly for LIKELY_EDGES,
if we have exactly one likely edge, make the other edges predicted
as not probable. */
static void
set_even_probabilities (basic_block bb,
hash_set<edge> *unlikely_edges = NULL,
hash_set<edge_prediction *> *likely_edges = NULL)
{
unsigned nedges = 0, unlikely_count = 0;
edge e = NULL;
edge_iterator ei;
profile_probability all = profile_probability::always ();
FOR_EACH_EDGE (e, ei, bb->succs)
if (e->probability.initialized_p ())
all -= e->probability;
else if (!unlikely_executed_edge_p (e))
{
nedges++;
if (unlikely_edges != NULL && unlikely_edges->contains (e))
{
all -= profile_probability::very_unlikely ();
unlikely_count++;
}
}
/* Make the distribution even if all edges are unlikely. */
unsigned likely_count = likely_edges ? likely_edges->elements () : 0;
if (unlikely_count == nedges)
{
unlikely_edges = NULL;
unlikely_count = 0;
}
/* If we have one likely edge, then use its probability and distribute
remaining probabilities as even. */
if (likely_count == 1)
{
FOR_EACH_EDGE (e, ei, bb->succs)
if (e->probability.initialized_p ())
;
else if (!unlikely_executed_edge_p (e))
{
edge_prediction *prediction = *likely_edges->begin ();
int p = prediction->ep_probability;
profile_probability prob
= profile_probability::from_reg_br_prob_base (p);
if (prediction->ep_edge == e)
e->probability = prob;
else if (unlikely_edges != NULL && unlikely_edges->contains (e))
e->probability = profile_probability::very_unlikely ();
else
{
profile_probability remainder = prob.invert ();
remainder -= profile_probability::very_unlikely ()
.apply_scale (unlikely_count, 1);
int count = nedges - unlikely_count - 1;
gcc_assert (count >= 0);
e->probability = remainder.apply_scale (1, count);
}
}
else
e->probability = profile_probability::never ();
}
else
{
/* Make all unlikely edges unlikely and the rest will have even
probability. */
unsigned scale = nedges - unlikely_count;
FOR_EACH_EDGE (e, ei, bb->succs)
if (e->probability.initialized_p ())
;
else if (!unlikely_executed_edge_p (e))
{
if (unlikely_edges != NULL && unlikely_edges->contains (e))
e->probability = profile_probability::very_unlikely ();
else
e->probability = all.apply_scale (1, scale);
}
else
e->probability = profile_probability::never ();
}
}
/* Add REG_BR_PROB note to JUMP with PROB. */
void
add_reg_br_prob_note (rtx_insn *jump, profile_probability prob)
{
gcc_checking_assert (JUMP_P (jump) && !find_reg_note (jump, REG_BR_PROB, 0));
add_int_reg_note (jump, REG_BR_PROB, prob.to_reg_br_prob_note ());
}
/* 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 (rtx_insn *insn, basic_block bb)
{
rtx prob_note;
rtx *pnote;
rtx note;
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;
if (!can_predict_insn_p (insn))
{
set_even_probabilities (bb);
return;
}
prob_note = find_reg_note (insn, REG_BR_PROB, 0);
pnote = &REG_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. */
for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
if (REG_NOTE_KIND (note) == REG_BR_PRED)
{
enum br_predictor predictor = ((enum br_predictor)
INTVAL (XEXP (XEXP (note, 0), 0)));
int probability = INTVAL (XEXP (XEXP (note, 0), 1));
found = true;
if (best_predictor > predictor
&& predictor_info[predictor].flags & PRED_FLAG_FIRST_MATCH)
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 (best_predictor != END_PREDICTORS)
first_match = true;
if (!found)
dump_prediction (dump_file, PRED_NO_PREDICTION,
combined_probability, bb);
else
{
if (!first_match)
dump_prediction (dump_file, PRED_DS_THEORY, combined_probability,
bb, !first_match ? REASON_NONE : REASON_IGNORED);
else
dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability,
bb, first_match ? REASON_NONE : REASON_IGNORED);
}
if (first_match)
combined_probability = best_probability;
dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb);
while (*pnote)
{
if (REG_NOTE_KIND (*pnote) == REG_BR_PRED)
{
enum br_predictor predictor = ((enum br_predictor)
INTVAL (XEXP (XEXP (*pnote, 0), 0)));
int probability = INTVAL (XEXP (XEXP (*pnote, 0), 1));
dump_prediction (dump_file, predictor, probability, bb,
(!first_match || best_predictor == predictor)
? REASON_NONE : REASON_IGNORED);
*pnote = XEXP (*pnote, 1);
}
else
pnote = &XEXP (*pnote, 1);
}
if (!prob_note)
{
profile_probability p
= profile_probability::from_reg_br_prob_base (combined_probability);
add_reg_br_prob_note (insn, p);
/* Save the prediction into CFG in case we are seeing non-degenerated
conditional jump. */
if (!single_succ_p (bb))
{
BRANCH_EDGE (bb)->probability = p;
FALLTHRU_EDGE (bb)->probability
= BRANCH_EDGE (bb)->probability.invert ();
}
}
else if (!single_succ_p (bb))
{
profile_probability prob = profile_probability::from_reg_br_prob_note
(XINT (prob_note, 0));
BRANCH_EDGE (bb)->probability = prob;
FALLTHRU_EDGE (bb)->probability = prob.invert ();
}
else
single_succ_edge (bb)->probability = profile_probability::always ();
}
/* Edge prediction hash traits. */
struct predictor_hash: pointer_hash <edge_prediction>
{
static inline hashval_t hash (const edge_prediction *);
static inline bool equal (const edge_prediction *, const edge_prediction *);
};
/* Calculate hash value of an edge prediction P based on predictor and
normalized probability. */
inline hashval_t
predictor_hash::hash (const edge_prediction *p)
{
inchash::hash hstate;
hstate.add_int (p->ep_predictor);
int prob = p->ep_probability;
if (prob > REG_BR_PROB_BASE / 2)
prob = REG_BR_PROB_BASE - prob;
hstate.add_int (prob);
return hstate.end ();
}
/* Return true whether edge predictions P1 and P2 use the same predictor and
have equal (or opposed probability). */
inline bool
predictor_hash::equal (const edge_prediction *p1, const edge_prediction *p2)
{
return (p1->ep_predictor == p2->ep_predictor
&& (p1->ep_probability == p2->ep_probability
|| p1->ep_probability == REG_BR_PROB_BASE - p2->ep_probability));
}
struct predictor_hash_traits: predictor_hash,
typed_noop_remove <edge_prediction *> {};
/* Return true if edge prediction P is not in DATA hash set. */
static bool
not_removed_prediction_p (edge_prediction *p, void *data)
{
hash_set<edge_prediction *> *remove = (hash_set<edge_prediction *> *) data;
return !remove->contains (p);
}
/* Prune predictions for a basic block BB. Currently we do following
clean-up steps:
1) remove duplicate prediction that is guessed with the same probability
(different than 1/2) to both edge
2) remove duplicates for a prediction that belongs with the same probability
to a single edge
*/
static void
prune_predictions_for_bb (basic_block bb)
{
edge_prediction **preds = bb_predictions->get (bb);
if (preds)
{
hash_table <predictor_hash_traits> s (13);
hash_set <edge_prediction *> remove;
/* Step 1: identify predictors that should be removed. */
for (edge_prediction *pred = *preds; pred; pred = pred->ep_next)
{
edge_prediction *existing = s.find (pred);
if (existing)
{
if (pred->ep_edge == existing->ep_edge
&& pred->ep_probability == existing->ep_probability)
{
/* Remove a duplicate predictor. */
dump_prediction (dump_file, pred->ep_predictor,
pred->ep_probability, bb,
REASON_SINGLE_EDGE_DUPLICATE, pred->ep_edge);
remove.add (pred);
}
else if (pred->ep_edge != existing->ep_edge
&& pred->ep_probability == existing->ep_probability
&& pred->ep_probability != REG_BR_PROB_BASE / 2)
{
/* Remove both predictors as they predict the same
for both edges. */
dump_prediction (dump_file, existing->ep_predictor,
pred->ep_probability, bb,
REASON_EDGE_PAIR_DUPLICATE,
existing->ep_edge);
dump_prediction (dump_file, pred->ep_predictor,
pred->ep_probability, bb,
REASON_EDGE_PAIR_DUPLICATE,
pred->ep_edge);
remove.add (existing);
remove.add (pred);
}
}
edge_prediction **slot2 = s.find_slot (pred, INSERT);
*slot2 = pred;
}
/* Step 2: Remove predictors. */
filter_predictions (preds, not_removed_prediction_p, &remove);
}
}
/* Combine predictions into single probability and store them into CFG.
Remove now useless prediction entries.
If DRY_RUN is set, only produce dumps and do not modify profile. */
static void
combine_predictions_for_bb (basic_block bb, bool dry_run)
{
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;
int nzero = 0;
int nunknown = 0;
FOR_EACH_EDGE (e, ei, bb->succs)
{
if (!unlikely_executed_edge_p (e))
{
nedges ++;
if (first && !second)
second = e;
if (!first)
first = e;
}
else if (!e->probability.initialized_p ())
e->probability = profile_probability::never ();
if (!e->probability.initialized_p ())
nunknown++;
else if (e->probability == profile_probability::never ())
nzero++;
}
/* When there is no successor or only one choice, prediction is easy.
When we have a basic block with more than 2 successors, the situation
is more complicated as DS theory cannot be used literally.
More precisely, let's assume we predicted edge e1 with probability p1,
thus: m1({b1}) = p1. As we're going to combine more than 2 edges, we
need to find probability of e.g. m1({b2}), which we don't know.
The only approximation is to equally distribute 1-p1 to all edges
different from b1.
According to numbers we've got from SPEC2006 benchark, there's only
one interesting reliable predictor (noreturn call), which can be
handled with a bit easier approach. */
if (nedges != 2)
{
hash_set<edge> unlikely_edges (4);
hash_set<edge_prediction *> likely_edges (4);
/* Identify all edges that have a probability close to very unlikely.
Doing the approach for very unlikely doesn't worth for doing as
there's no such probability in SPEC2006 benchmark. */
edge_prediction **preds = bb_predictions->get (bb);
if (preds)
for (pred = *preds; pred; pred = pred->ep_next)
{
if (pred->ep_probability <= PROB_VERY_UNLIKELY
|| pred->ep_predictor == PRED_COLD_LABEL)
unlikely_edges.add (pred->ep_edge);
else if (pred->ep_probability >= PROB_VERY_LIKELY
|| pred->ep_predictor == PRED_BUILTIN_EXPECT
|| pred->ep_predictor == PRED_HOT_LABEL)
likely_edges.add (pred);
}
/* It can happen that an edge is both in likely_edges and unlikely_edges.
Clear both sets in that situation. */
for (hash_set<edge_prediction *>::iterator it = likely_edges.begin ();
it != likely_edges.end (); ++it)
if (unlikely_edges.contains ((*it)->ep_edge))
{
likely_edges.empty ();
unlikely_edges.empty ();
break;
}
if (!dry_run)
set_even_probabilities (bb, &unlikely_edges, &likely_edges);
clear_bb_predictions (bb);
if (dump_file)
{
fprintf (dump_file, "Predictions for bb %i\n", bb->index);
if (unlikely_edges.is_empty ())
fprintf (dump_file,
"%i edges in bb %i predicted to even probabilities\n",
nedges, bb->index);
else
{
fprintf (dump_file,
"%i edges in bb %i predicted with some unlikely edges\n",
nedges, bb->index);
FOR_EACH_EDGE (e, ei, bb->succs)
if (!unlikely_executed_edge_p (e))
dump_prediction (dump_file, PRED_COMBINED,
e->probability.to_reg_br_prob_base (), bb, REASON_NONE, e);
}
}
return;
}
if (dump_file)
fprintf (dump_file, "Predictions for bb %i\n", bb->index);
prune_predictions_for_bb (bb);
edge_prediction **preds = bb_predictions->get (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)
{
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
&& predictor_info[predictor].flags & PRED_FLAG_FIRST_MATCH)
{
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 = pred2->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 (best_predictor != END_PREDICTORS)
first_match = true;
if (!found)
dump_prediction (dump_file, PRED_NO_PREDICTION, combined_probability, bb);
else
{
if (!first_match)
dump_prediction (dump_file, PRED_DS_THEORY, combined_probability, bb,
!first_match ? REASON_NONE : REASON_IGNORED);
else
dump_prediction (dump_file, PRED_FIRST_MATCH, best_probability, bb,
first_match ? REASON_NONE : REASON_IGNORED);
}
if (first_match)
combined_probability = best_probability;
dump_prediction (dump_file, PRED_COMBINED, combined_probability, bb);
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;
dump_prediction (dump_file, predictor, probability, bb,
(!first_match || best_predictor == predictor)
? REASON_NONE : REASON_IGNORED, pred->ep_edge);
}
}
clear_bb_predictions (bb);
/* If we have only one successor which is unknown, we can compute missing
probability. */
if (nunknown == 1)
{
profile_probability prob = profile_probability::always ();
edge missing = NULL;
FOR_EACH_EDGE (e, ei, bb->succs)
if (e->probability.initialized_p ())
prob -= e->probability;
else if (missing == NULL)
missing = e;
else
gcc_unreachable ();
missing->probability = prob;
}
/* If nothing is unknown, we have nothing to update. */
else if (!nunknown && nzero != (int)EDGE_COUNT (bb->succs))
;
else if (!dry_run)
{
first->probability
= profile_probability::from_reg_br_prob_base (combined_probability);
second->probability = first->probability.invert ();
}
}
/* Check if T1 and T2 satisfy the IV_COMPARE condition.
Return the SSA_NAME if the condition satisfies, NULL otherwise.
T1 and T2 should be one of the following cases:
1. T1 is SSA_NAME, T2 is NULL
2. T1 is SSA_NAME, T2 is INTEGER_CST between [-4, 4]
3. T2 is SSA_NAME, T1 is INTEGER_CST between [-4, 4] */
static tree
strips_small_constant (tree t1, tree t2)
{
tree ret = NULL;
int value = 0;
if (!t1)
return NULL;
else if (TREE_CODE (t1) == SSA_NAME)
ret = t1;
else if (tree_fits_shwi_p (t1))
value = tree_to_shwi (t1);
else
return NULL;
if (!t2)
return ret;
else if (tree_fits_shwi_p (t2))
value = tree_to_shwi (t2);
else if (TREE_CODE (t2) == SSA_NAME)
{
if (ret)
return NULL;
else
ret = t2;
}
if (value <= 4 && value >= -4)
return ret;
else
return NULL;
}
/* Return the SSA_NAME in T or T's operands.
Return NULL if SSA_NAME cannot be found. */
static tree
get_base_value (tree t)
{
if (TREE_CODE (t) == SSA_NAME)
return t;
if (!BINARY_CLASS_P (t))
return NULL;
switch (TREE_OPERAND_LENGTH (t))
{
case 1:
return strips_small_constant (TREE_OPERAND (t, 0), NULL);
case 2:
return strips_small_constant (TREE_OPERAND (t, 0),
TREE_OPERAND (t, 1));
default:
return NULL;
}
}
/* Check the compare STMT in LOOP. If it compares an induction
variable to a loop invariant, return true, and save
LOOP_INVARIANT, COMPARE_CODE and LOOP_STEP.
Otherwise return false and set LOOP_INVAIANT to NULL. */
static bool
is_comparison_with_loop_invariant_p (gcond *stmt, class loop *loop,
tree *loop_invariant,
enum tree_code *compare_code,
tree *loop_step,
tree *loop_iv_base)
{
tree op0, op1, bound, base;
affine_iv iv0, iv1;
enum tree_code code;
tree step;
code = gimple_cond_code (stmt);
*loop_invariant = NULL;
switch (code)
{
case GT_EXPR:
case GE_EXPR:
case NE_EXPR:
case LT_EXPR:
case LE_EXPR:
case EQ_EXPR:
break;
default:
return false;
}
op0 = gimple_cond_lhs (stmt);
op1 = gimple_cond_rhs (stmt);
if ((TREE_CODE (op0) != SSA_NAME && TREE_CODE (op0) != INTEGER_CST)
|| (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op1) != INTEGER_CST))
return false;
if (!simple_iv (loop, loop_containing_stmt (stmt), op0, &iv0, true))
return false;
if (!simple_iv (loop, loop_containing_stmt (stmt), op1, &iv1, true))
return false;
if (TREE_CODE (iv0.step) != INTEGER_CST
|| TREE_CODE (iv1.step) != INTEGER_CST)
return false;
if ((integer_zerop (iv0.step) && integer_zerop (iv1.step))
|| (!integer_zerop (iv0.step) && !integer_zerop (iv1.step)))
return false;
if (integer_zerop (iv0.step))
{
if (code != NE_EXPR && code != EQ_EXPR)
code = invert_tree_comparison (code, false);
bound = iv0.base;
base = iv1.base;
if (tree_fits_shwi_p (iv1.step))
step = iv1.step;
else
return false;
}
else
{
bound = iv1.base;
base = iv0.base;
if (tree_fits_shwi_p (iv0.step))
step = iv0.step;
else
return false;
}
if (TREE_CODE (bound) != INTEGER_CST)
bound = get_base_value (bound);
if (!bound)
return false;
if (TREE_CODE (base) != INTEGER_CST)
base = get_base_value (base);
if (!base)
return false;
*loop_invariant = bound;
*compare_code = code;
*loop_step = step;
*loop_iv_base = base;
return true;
}
/* Compare two SSA_NAMEs: returns TRUE if T1 and T2 are value coherent. */
static bool
expr_coherent_p (tree t1, tree t2)
{
gimple *stmt;
tree ssa_name_1 = NULL;
tree ssa_name_2 = NULL;
gcc_assert (TREE_CODE (t1) == SSA_NAME || TREE_CODE (t1) == INTEGER_CST);
gcc_assert (TREE_CODE (t2) == SSA_NAME || TREE_CODE (t2) == INTEGER_CST);
if (t1 == t2)
return true;
if (TREE_CODE (t1) == INTEGER_CST && TREE_CODE (t2) == INTEGER_CST)
return true;
if (TREE_CODE (t1) == INTEGER_CST || TREE_CODE (t2) == INTEGER_CST)
return false;
/* Check to see if t1 is expressed/defined with t2. */
stmt = SSA_NAME_DEF_STMT (t1);
gcc_assert (stmt != NULL);
if (is_gimple_assign (stmt))
{
ssa_name_1 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
if (ssa_name_1 && ssa_name_1 == t2)
return true;
}
/* Check to see if t2 is expressed/defined with t1. */
stmt = SSA_NAME_DEF_STMT (t2);
gcc_assert (stmt != NULL);
if (is_gimple_assign (stmt))
{
ssa_name_2 = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE);
if (ssa_name_2 && ssa_name_2 == t1)
return true;
}
/* Compare if t1 and t2's def_stmts are identical. */
if (ssa_name_2 != NULL && ssa_name_1 == ssa_name_2)
return true;
else
return false;
}
/* Return true if E is predicted by one of loop heuristics. */
static bool
predicted_by_loop_heuristics_p (basic_block bb)
{
struct edge_prediction *i;
edge_prediction **preds = bb_predictions->get (bb);
if (!preds)
return false;
for (i = *preds; i; i = i->ep_next)
if (i->ep_predictor == PRED_LOOP_ITERATIONS_GUESSED
|| i->ep_predictor == PRED_LOOP_ITERATIONS_MAX
|| i->ep_predictor == PRED_LOOP_ITERATIONS
|| i->ep_predictor == PRED_LOOP_EXIT
|| i->ep_predictor == PRED_LOOP_EXIT_WITH_RECURSION
|| i->ep_predictor == PRED_LOOP_EXTRA_EXIT)
return true;
return false;
}
/* Predict branch probability of BB when BB contains a branch that compares
an induction variable in LOOP with LOOP_IV_BASE_VAR to LOOP_BOUND_VAR. The
loop exit is compared using LOOP_BOUND_CODE, with step of LOOP_BOUND_STEP.
E.g.
for (int i = 0; i < bound; i++) {
if (i < bound - 2)
computation_1();
else
computation_2();
}
In this loop, we will predict the branch inside the loop to be taken. */
static void
predict_iv_comparison (class loop *loop, basic_block bb,
tree loop_bound_var,
tree loop_iv_base_var,
enum tree_code loop_bound_code,
int loop_bound_step)
{
gimple *stmt;
tree compare_var, compare_base;
enum tree_code compare_code;
tree compare_step_var;
edge then_edge;
edge_iterator ei;
if (predicted_by_loop_heuristics_p (bb))
return;
stmt = last_stmt (bb);
if (!stmt || gimple_code (stmt) != GIMPLE_COND)
return;
if (!is_comparison_with_loop_invariant_p (as_a <gcond *> (stmt),
loop, &compare_var,
&compare_code,
&compare_step_var,
&compare_base))
return;
/* Find the taken edge. */
FOR_EACH_EDGE (then_edge, ei, bb->succs)
if (then_edge->flags & EDGE_TRUE_VALUE)
break;
/* When comparing an IV to a loop invariant, NE is more likely to be
taken while EQ is more likely to be not-taken. */
if (compare_code == NE_EXPR)
{
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
return;
}
else if (compare_code == EQ_EXPR)
{
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
return;
}
if (!expr_coherent_p (loop_iv_base_var, compare_base))
return;
/* If loop bound, base and compare bound are all constants, we can
calculate the probability directly. */
if (tree_fits_shwi_p (loop_bound_var)
&& tree_fits_shwi_p (compare_var)
&& tree_fits_shwi_p (compare_base))
{
int probability;
wi::overflow_type overflow;
bool overall_overflow = false;
widest_int compare_count, tem;
/* (loop_bound - base) / compare_step */
tem = wi::sub (wi::to_widest (loop_bound_var),
wi::to_widest (compare_base), SIGNED, &overflow);
overall_overflow |= overflow;
widest_int loop_count = wi::div_trunc (tem,
wi::to_widest (compare_step_var),
SIGNED, &overflow);
overall_overflow |= overflow;
if (!wi::neg_p (wi::to_widest (compare_step_var))
^ (compare_code == LT_EXPR || compare_code == LE_EXPR))
{
/* (loop_bound - compare_bound) / compare_step */
tem = wi::sub (wi::to_widest (loop_bound_var),
wi::to_widest (compare_var), SIGNED, &overflow);
overall_overflow |= overflow;
compare_count = wi::div_trunc (tem, wi::to_widest (compare_step_var),
SIGNED, &overflow);
overall_overflow |= overflow;
}
else
{
/* (compare_bound - base) / compare_step */
tem = wi::sub (wi::to_widest (compare_var),
wi::to_widest (compare_base), SIGNED, &overflow);
overall_overflow |= overflow;
compare_count = wi::div_trunc (tem, wi::to_widest (compare_step_var),
SIGNED, &overflow);
overall_overflow |= overflow;
}
if (compare_code == LE_EXPR || compare_code == GE_EXPR)
++compare_count;
if (loop_bound_code == LE_EXPR || loop_bound_code == GE_EXPR)
++loop_count;
if (wi::neg_p (compare_count))
compare_count = 0;
if (wi::neg_p (loop_count))
loop_count = 0;
if (loop_count == 0)
probability = 0;
else if (wi::cmps (compare_count, loop_count) == 1)
probability = REG_BR_PROB_BASE;
else
{
tem = compare_count * REG_BR_PROB_BASE;
tem = wi::udiv_trunc (tem, loop_count);
probability = tem.to_uhwi ();
}
/* FIXME: The branch prediction seems broken. It has only 20% hitrate. */
if (!overall_overflow)
predict_edge (then_edge, PRED_LOOP_IV_COMPARE, probability);
return;
}
if (expr_coherent_p (loop_bound_var, compare_var))
{
if ((loop_bound_code == LT_EXPR || loop_bound_code == LE_EXPR)
&& (compare_code == LT_EXPR || compare_code == LE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else if ((loop_bound_code == GT_EXPR || loop_bound_code == GE_EXPR)
&& (compare_code == GT_EXPR || compare_code == GE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else if (loop_bound_code == NE_EXPR)
{
/* If the loop backedge condition is "(i != bound)", we do
the comparison based on the step of IV:
* step < 0 : backedge condition is like (i > bound)
* step > 0 : backedge condition is like (i < bound) */
gcc_assert (loop_bound_step != 0);
if (loop_bound_step > 0
&& (compare_code == LT_EXPR
|| compare_code == LE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else if (loop_bound_step < 0
&& (compare_code == GT_EXPR
|| compare_code == GE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
}
else
/* The branch is predicted not-taken if loop_bound_code is
opposite with compare_code. */
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
}
else if (expr_coherent_p (loop_iv_base_var, compare_var))
{
/* For cases like:
for (i = s; i < h; i++)
if (i > s + 2) ....
The branch should be predicted taken. */
if (loop_bound_step > 0
&& (compare_code == GT_EXPR || compare_code == GE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else if (loop_bound_step < 0
&& (compare_code == LT_EXPR || compare_code == LE_EXPR))
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, TAKEN);
else
predict_edge_def (then_edge, PRED_LOOP_IV_COMPARE_GUESS, NOT_TAKEN);
}
}
/* Predict for extra loop exits that will lead to EXIT_EDGE. The extra loop
exits are resulted from short-circuit conditions that will generate an
if_tmp. E.g.:
if (foo() || global > 10)
break;
This will be translated into:
BB3:
loop header...
BB4:
if foo() goto BB6 else goto BB5
BB5:
if global > 10 goto BB6 else goto BB7
BB6:
goto BB7
BB7:
iftmp = (PHI 0(BB5), 1(BB6))
if iftmp == 1 goto BB8 else goto BB3
BB8:
outside of the loop...
The edge BB7->BB8 is loop exit because BB8 is outside of the loop.
From the dataflow, we can infer that BB4->BB6 and BB5->BB6 are also loop
exits. This function takes BB7->BB8 as input, and finds out the extra loop
exits to predict them using PRED_LOOP_EXTRA_EXIT. */
static void
predict_extra_loop_exits (edge exit_edge)
{
unsigned i;
bool check_value_one;
gimple *lhs_def_stmt;
gphi *phi_stmt;
tree cmp_rhs, cmp_lhs;
gimple *last;
gcond *cmp_stmt;
last = last_stmt (exit_edge->src);
if (!last)
return;
cmp_stmt = dyn_cast <gcond *> (last);
if (!cmp_stmt)
return;
cmp_rhs = gimple_cond_rhs (cmp_stmt);
cmp_lhs = gimple_cond_lhs (cmp_stmt);
if (!TREE_CONSTANT (cmp_rhs)
|| !(integer_zerop (cmp_rhs) || integer_onep (cmp_rhs)))
return;
if (TREE_CODE (cmp_lhs) != SSA_NAME)
return;
/* If check_value_one is true, only the phi_args with value '1' will lead
to loop exit. Otherwise, only the phi_args with value '0' will lead to
loop exit. */
check_value_one = (((integer_onep (cmp_rhs))
^ (gimple_cond_code (cmp_stmt) == EQ_EXPR))
^ ((exit_edge->flags & EDGE_TRUE_VALUE) != 0));
lhs_def_stmt = SSA_NAME_DEF_STMT (cmp_lhs);
if (!lhs_def_stmt)
return;
phi_stmt = dyn_cast <gphi *> (lhs_def_stmt);
if (!phi_stmt)
return;
for (i = 0; i < gimple_phi_num_args (phi_stmt); i++)
{
edge e1;
edge_iterator ei;
tree val = gimple_phi_arg_def (phi_stmt, i);
edge e = gimple_phi_arg_edge (phi_stmt, i);
if (!TREE_CONSTANT (val) || !(integer_zerop (val) || integer_onep (val)))
continue;
if ((check_value_one ^ integer_onep (val)) == 1)
continue;
if (EDGE_COUNT (e->src->succs) != 1)
{
predict_paths_leading_to_edge (e, PRED_LOOP_EXTRA_EXIT, NOT_TAKEN);
continue;
}
FOR_EACH_EDGE (e1, ei, e->src->preds)
predict_paths_leading_to_edge (e1, PRED_LOOP_EXTRA_EXIT, NOT_TAKEN);
}
}
/* Predict edge probabilities by exploiting loop structure. */
static void
predict_loops (void)
{
class loop *loop;
basic_block bb;
hash_set <class loop *> with_recursion(10);
FOR_EACH_BB_FN (bb, cfun)
{
gimple_stmt_iterator gsi;
tree decl;
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
if (is_gimple_call (gsi_stmt (gsi))
&& (decl = gimple_call_fndecl (gsi_stmt (gsi))) != NULL
&& recursive_call_p (current_function_decl, decl))
{
loop = bb->loop_father;
while (loop && !with_recursion.add (loop))
loop = loop_outer (loop);
}
}
/* Try to predict out blocks in a loop that are not part of a
natural loop. */
for (auto loop : loops_list (cfun, LI_FROM_INNERMOST))
{
basic_block bb, *bbs;
unsigned j, n_exits = 0;
class tree_niter_desc niter_desc;
edge ex;
class nb_iter_bound *nb_iter;
enum tree_code loop_bound_code = ERROR_MARK;
tree loop_bound_step = NULL;
tree loop_bound_var = NULL;
tree loop_iv_base = NULL;
gcond *stmt = NULL;
bool recursion = with_recursion.contains (loop);
auto_vec<edge> exits = get_loop_exit_edges (loop);
FOR_EACH_VEC_ELT (exits, j, ex)
if (!unlikely_executed_edge_p (ex) && !(ex->flags & EDGE_ABNORMAL_CALL))
n_exits ++;
if (!n_exits)
continue;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Predicting loop %i%s with %i exits.\n",
loop->num, recursion ? " (with recursion)":"", n_exits);
if (dump_file && (dump_flags & TDF_DETAILS)
&& max_loop_iterations_int (loop) >= 0)
{
fprintf (dump_file,
"Loop %d iterates at most %i times.\n", loop->num,
(int)max_loop_iterations_int (loop));
}
if (dump_file && (dump_flags & TDF_DETAILS)
&& likely_max_loop_iterations_int (loop) >= 0)
{
fprintf (dump_file, "Loop %d likely iterates at most %i times.\n",
loop->num, (int)likely_max_loop_iterations_int (loop));
}
FOR_EACH_VEC_ELT (exits, j, ex)
{
tree niter = NULL;
HOST_WIDE_INT nitercst;
int max = param_max_predicted_iterations;
int probability;
enum br_predictor predictor;
widest_int nit;
if (unlikely_executed_edge_p (ex)
|| (ex->flags & EDGE_ABNORMAL_CALL))
continue;
/* Loop heuristics do not expect exit conditional to be inside
inner loop. We predict from innermost to outermost loop. */
if (predicted_by_loop_heuristics_p (ex->src))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Skipping exit %i->%i because "
"it is already predicted.\n",
ex->src->index, ex->dest->index);
continue;
}
predict_extra_loop_exits (ex);
if (number_of_iterations_exit (loop, ex, &niter_desc, false, false))
niter = niter_desc.niter;
if (!niter || TREE_CODE (niter_desc.niter) != INTEGER_CST)
niter = loop_niter_by_eval (loop, ex);
if (dump_file && (dump_flags & TDF_DETAILS)
&& TREE_CODE (niter) == INTEGER_CST)
{
fprintf (dump_file, "Exit %i->%i %d iterates ",
ex->src->index, ex->dest->index,
loop->num);
print_generic_expr (dump_file, niter, TDF_SLIM);
fprintf (dump_file, " times.\n");
}
if (TREE_CODE (niter) == INTEGER_CST)
{
if (tree_fits_uhwi_p (niter)
&& max
&& compare_tree_int (niter, max - 1) == -1)
nitercst = tree_to_uhwi (niter) + 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
&& estimated_stmt_executions (loop, &nit))
{
if (wi::gtu_p (nit, max))
nitercst = max;
else
nitercst = nit.to_shwi ();
predictor = PRED_LOOP_ITERATIONS_GUESSED;
}
/* If we have likely upper bound, trust it for very small iteration
counts. Such loops would otherwise get mispredicted by standard
LOOP_EXIT heuristics. */
else if (n_exits == 1
&& likely_max_stmt_executions (loop, &nit)
&& wi::ltu_p (nit,
RDIV (REG_BR_PROB_BASE,
REG_BR_PROB_BASE
- predictor_info
[recursion
? PRED_LOOP_EXIT_WITH_RECURSION
: PRED_LOOP_EXIT].hitrate)))
{
nitercst = nit.to_shwi ();
predictor = PRED_LOOP_ITERATIONS_MAX;
}
else
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Nothing known about exit %i->%i.\n",
ex->src->index, ex->dest->index);
continue;
}
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Recording prediction to %i iterations by %s.\n",
(int)nitercst, predictor_info[predictor].name);
/* If the prediction for number of iterations is zero, do not
predict the exit edges. */
if (nitercst == 0)
continue;
probability = RDIV (REG_BR_PROB_BASE, nitercst);
predict_edge (ex, predictor, probability);
}
/* Find information about loop bound variables. */
for (nb_iter = loop->bounds; nb_iter;
nb_iter = nb_iter->next)
if (nb_iter->stmt
&& gimple_code (nb_iter->stmt) == GIMPLE_COND)
{
stmt = as_a <gcond *> (nb_iter->stmt);
break;
}
if (!stmt && last_stmt (loop->header)
&& gimple_code (last_stmt (loop->header)) == GIMPLE_COND)
stmt = as_a <gcond *> (last_stmt (loop->header));
if (stmt)
is_comparison_with_loop_invariant_p (stmt, loop,
&loop_bound_var,
&loop_bound_code,
&loop_bound_step,
&loop_iv_base);
bbs = get_loop_body (loop);
for (j = 0; j < loop->num_nodes; j++)
{
edge e;
edge_iterator ei;
bb = bbs[j];
/* Bypass loop heuristics on continue statement. These
statements construct loops via "non-loop" constructs
in the source language and are better to be handled
separately. */
if (predicted_by_p (bb, PRED_CONTINUE))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "BB %i predicted by continue.\n",
bb->index);
continue;
}
/* If we already used more reliable loop exit predictors, do not
bother with PRED_LOOP_EXIT. */
if (!predicted_by_loop_heuristics_p (bb))
{
/* 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
[recursion
? PRED_LOOP_EXIT_WITH_RECURSION
: 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))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"Predicting exit %i->%i with prob %i.\n",
e->src->index, e->dest->index, probability);
predict_edge (e,
recursion ? PRED_LOOP_EXIT_WITH_RECURSION
: PRED_LOOP_EXIT, probability);
}
}
if (loop_bound_var)
predict_iv_comparison (loop, bb, loop_bound_var, loop_iv_base,
loop_bound_code,
tree_to_shwi (loop_bound_step));
}
/* In the following code
for (loop1)
if (cond)
for (loop2)
body;
guess that cond is unlikely. */
if (loop_outer (loop)->num)
{
basic_block bb = NULL;
edge preheader_edge = loop_preheader_edge (loop);
if (single_pred_p (preheader_edge->src)
&& single_succ_p (preheader_edge->src))
preheader_edge = single_pred_edge (preheader_edge->src);
gimple *stmt = last_stmt (preheader_edge->src);
/* Pattern match fortran loop preheader:
_16 = BUILTIN_EXPECT (_15, 1, PRED_FORTRAN_LOOP_PREHEADER);
_17 = (logical(kind=4)) _16;
if (_17 != 0)
goto <bb 11>;
else
goto <bb 13>;
Loop guard branch prediction says nothing about duplicated loop
headers produced by fortran frontend and in this case we want
to predict paths leading to this preheader. */
if (stmt
&& gimple_code (stmt) == GIMPLE_COND
&& gimple_cond_code (stmt) == NE_EXPR
&& TREE_CODE (gimple_cond_lhs (stmt)) == SSA_NAME
&& integer_zerop (gimple_cond_rhs (stmt)))
{
gimple *call_stmt = SSA_NAME_DEF_STMT (gimple_cond_lhs (stmt));
if (gimple_code (call_stmt) == GIMPLE_ASSIGN
&& CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (call_stmt))
&& TREE_CODE (gimple_assign_rhs1 (call_stmt)) == SSA_NAME)
call_stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (call_stmt));
if (gimple_call_internal_p (call_stmt, IFN_BUILTIN_EXPECT)
&& TREE_CODE (gimple_call_arg (call_stmt, 2)) == INTEGER_CST
&& tree_fits_uhwi_p (gimple_call_arg (call_stmt, 2))
&& tree_to_uhwi (gimple_call_arg (call_stmt, 2))
== PRED_FORTRAN_LOOP_PREHEADER)
bb = preheader_edge->src;
}
if (!bb)
{
if (!dominated_by_p (CDI_DOMINATORS,
loop_outer (loop)->latch, loop->header))
predict_paths_leading_to_edge (loop_preheader_edge (loop),
recursion
? PRED_LOOP_GUARD_WITH_RECURSION
: PRED_LOOP_GUARD,
NOT_TAKEN,
loop_outer (loop));
}
else
{
if (!dominated_by_p (CDI_DOMINATORS,
loop_outer (loop)->latch, bb))
predict_paths_leading_to (bb,
recursion
? PRED_LOOP_GUARD_WITH_RECURSION
: PRED_LOOP_GUARD,
NOT_TAKEN,
loop_outer (loop));
}
}
/* Free basic blocks from get_loop_body. */
free (bbs);
}
}
/* Attempt to predict probabilities of BB outgoing edges using local
properties. */
static void
bb_estimate_probability_locally (basic_block bb)
{
rtx_insn *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)))))
{
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
/* 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);
}
static tree expr_expected_value (tree, bitmap, enum br_predictor *predictor,
HOST_WIDE_INT *probability);
/* Helper function for expr_expected_value. */
static tree
expr_expected_value_1 (tree type, tree op0, enum tree_code code,
tree op1, bitmap visited, enum br_predictor *predictor,
HOST_WIDE_INT *probability)
{
gimple *def;
/* Reset returned probability value. */
*probability = -1;
*predictor = PRED_UNCONDITIONAL;
if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
{
if (TREE_CONSTANT (op0))
return op0;
if (code == IMAGPART_EXPR)
{
if (TREE_CODE (TREE_OPERAND (op0, 0)) == SSA_NAME)
{
def = SSA_NAME_DEF_STMT (TREE_OPERAND (op0, 0));
if (is_gimple_call (def)
&& gimple_call_internal_p (def)
&& (gimple_call_internal_fn (def)
== IFN_ATOMIC_COMPARE_EXCHANGE))
{
/* Assume that any given atomic operation has low contention,
and thus the compare-and-swap operation succeeds. */
*predictor = PRED_COMPARE_AND_SWAP;
return build_one_cst (TREE_TYPE (op0));
}
}
}
if (code != SSA_NAME)
return NULL_TREE;
def = SSA_NAME_DEF_STMT (op0);
/* If we were already here, break the infinite cycle. */
if (!bitmap_set_bit (visited, SSA_NAME_VERSION (op0)))
return NULL;
if (gimple_code (def) == GIMPLE_PHI)
{
/* 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++)
{
tree arg = PHI_ARG_DEF (def, i);
enum br_predictor predictor2;
/* 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;
HOST_WIDE_INT probability2;
new_val = expr_expected_value (arg, visited, &predictor2,
&probability2);
/* It is difficult to combine value predictors. Simply assume
that later predictor is weaker and take its prediction. */
if (*predictor < predictor2)
{
*predictor = predictor2;
*probability = probability2;
}
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, predictor, probability);
}
if (is_gimple_call (def))
{
tree decl = gimple_call_fndecl (def);
if (!decl)
{
if (gimple_call_internal_p (def)
&& gimple_call_internal_fn (def) == IFN_BUILTIN_EXPECT)
{
gcc_assert (gimple_call_num_args (def) == 3);
tree val = gimple_call_arg (def, 0);
if (TREE_CONSTANT (val))
return val;
tree val2 = gimple_call_arg (def, 2);
gcc_assert (TREE_CODE (val2) == INTEGER_CST
&& tree_fits_uhwi_p (val2)
&& tree_to_uhwi (val2) < END_PREDICTORS);
*predictor = (enum br_predictor) tree_to_uhwi (val2);
if (*predictor == PRED_BUILTIN_EXPECT)
*probability
= HITRATE (param_builtin_expect_probability);
return gimple_call_arg (def, 1);
}
return NULL;
}
if (DECL_IS_MALLOC (decl) || DECL_IS_OPERATOR_NEW_P (decl))
{
if (predictor)
*predictor = PRED_MALLOC_NONNULL;
return boolean_true_node;
}
if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
switch (DECL_FUNCTION_CODE (decl))
{
case 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;
*predictor = PRED_BUILTIN_EXPECT;
*probability
= HITRATE (param_builtin_expect_probability);
return gimple_call_arg (def, 1);
}
case BUILT_IN_EXPECT_WITH_PROBABILITY:
{
tree val;
if (gimple_call_num_args (def) != 3)
return NULL;
val = gimple_call_arg (def, 0);
if (TREE_CONSTANT (val))
return val;
/* Compute final probability as:
probability * REG_BR_PROB_BASE. */
tree prob = gimple_call_arg (def, 2);
tree t = TREE_TYPE (prob);
tree base = build_int_cst (integer_type_node,
REG_BR_PROB_BASE);
base = build_real_from_int_cst (t, base);
tree r = fold_build2_initializer_loc (UNKNOWN_LOCATION,
MULT_EXPR, t, prob, base);
if (TREE_CODE (r) != REAL_CST)
{
error_at (gimple_location (def),
"probability %qE must be "
"constant floating-point expression", prob);
return NULL;
}
HOST_WIDE_INT probi
= real_to_integer (TREE_REAL_CST_PTR (r));
if (probi >= 0 && probi <= REG_BR_PROB_BASE)
{
*predictor = PRED_BUILTIN_EXPECT_WITH_PROBABILITY;
*probability = probi;
}
else
error_at (gimple_location (def),
"probability %qE is outside "
"the range [0.0, 1.0]", prob);
return gimple_call_arg (def, 1);
}
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_N:
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1:
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2:
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4:
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8:
case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_N:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8:
case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16:
/* Assume that any given atomic operation has low contention,
and thus the compare-and-swap operation succeeds. */
*predictor = PRED_COMPARE_AND_SWAP;
return boolean_true_node;
case BUILT_IN_REALLOC:
if (predictor)
*predictor = PRED_MALLOC_NONNULL;
return boolean_true_node;
default:
break;
}
}
return NULL;
}
if (get_gimple_rhs_class (code) == GIMPLE_BINARY_RHS)
{
tree res;
enum br_predictor predictor2;
HOST_WIDE_INT probability2;
op0 = expr_expected_value (op0, visited, predictor, probability);
if (!op0)
return NULL;
op1 = expr_expected_value (op1, visited, &predictor2, &probability2);
if (!op1)
return NULL;
res = fold_build2 (code, type, op0, op1);
if (TREE_CODE (res) == INTEGER_CST
&& TREE_CODE (op0) == INTEGER_CST
&& TREE_CODE (op1) == INTEGER_CST)
{
/* Combine binary predictions. */
if (*probability != -1 || probability2 != -1)
{
HOST_WIDE_INT p1 = get_predictor_value (*predictor, *probability);
HOST_WIDE_INT p2 = get_predictor_value (predictor2, probability2);
*probability = RDIV (p1 * p2, REG_BR_PROB_BASE);
}
if (*predictor < predictor2)
*predictor = predictor2;
return res;
}
return NULL;
}
if (get_gimple_rhs_class (code) == GIMPLE_UNARY_RHS)
{
tree res;
op0 = expr_expected_value (op0, visited, predictor, probability);
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 br_predictor *predictor,
HOST_WIDE_INT *probability)
{
enum tree_code code;
tree op0, op1;
if (TREE_CONSTANT (expr))
{
*predictor = PRED_UNCONDITIONAL;
*probability = -1;
return expr;
}
extract_ops_from_tree (expr, &code, &op0, &op1);
return expr_expected_value_1 (TREE_TYPE (expr),
op0, code, op1, visited, predictor,
probability);
}
/* Return probability of a PREDICTOR. If the predictor has variable
probability return passed PROBABILITY. */
static HOST_WIDE_INT
get_predictor_value (br_predictor predictor, HOST_WIDE_INT probability)
{
switch (predictor)
{
case PRED_BUILTIN_EXPECT:
case PRED_BUILTIN_EXPECT_WITH_PROBABILITY:
gcc_assert (probability != -1);
return probability;
default:
gcc_assert (probability == -1);
return predictor_info[(int) predictor].hitrate;
}
}
/* Predict using opcode of the last statement in basic block. */
static void
tree_predict_by_opcode (basic_block bb)
{
gimple *stmt = last_stmt (bb);
edge then_edge;
tree op0, op1;
tree type;
tree val;
enum tree_code cmp;
edge_iterator ei;
enum br_predictor predictor;
HOST_WIDE_INT probability;
if (!stmt)
return;
if (gswitch *sw = dyn_cast <gswitch *> (stmt))
{
tree index = gimple_switch_index (sw);
tree val = expr_expected_value (index, auto_bitmap (),
&predictor, &probability);
if (val && TREE_CODE (val) == INTEGER_CST)
{
edge e = find_taken_edge_switch_expr (sw, val);
if (predictor == PRED_BUILTIN_EXPECT)
{
int percent = param_builtin_expect_probability;
gcc_assert (percent >= 0 && percent <= 100);
predict_edge (e, PRED_BUILTIN_EXPECT,
HITRATE (percent));
}
else
predict_edge_def (e, predictor, TAKEN);
}
}
if (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);
val = expr_expected_value_1 (boolean_type_node, op0, cmp, op1, auto_bitmap (),
&predictor, &probability);
if (val && TREE_CODE (val) == INTEGER_CST)
{
HOST_WIDE_INT prob = get_predictor_value (predictor, probability);
if (integer_zerop (val))
prob = REG_BR_PROB_BASE - prob;
predict_edge (then_edge, predictor, prob);
}
/* 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
/* 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;
}
}
/* Returns TRUE if the STMT is exit(0) like statement. */
static bool
is_exit_with_zero_arg (const gimple *stmt)
{
/* This is not exit, _exit or _Exit. */
if (!gimple_call_builtin_p (stmt, BUILT_IN_EXIT)
&& !gimple_call_builtin_p (stmt, BUILT_IN__EXIT)
&& !gimple_call_builtin_p (stmt, BUILT_IN__EXIT2))
return false;
/* Argument is an interger zero. */
return integer_zerop (gimple_call_arg (stmt, 0));
}
/* 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)))
{