/* Branch prediction routines for the GNU compiler.
- Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008
+ Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008, 2009
Free Software Foundation, Inc.
This file is part of GCC.
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.
+/* 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)
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 compute_function_frequency (void);
static void choose_function_section (void);
static bool can_predict_insn_p (const_rtx);
#undef DEF_PREDICTOR
/* Return TRUE if frequency FREQ is considered to be hot. */
-static bool
+
+static inline bool
maybe_hot_frequency_p (int freq)
{
+ struct cgraph_node *node = cgraph_node (current_function_decl);
if (!profile_info || !flag_branch_probabilities)
{
- if (cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
+ if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
return false;
- if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT)
+ if (node->frequency == NODE_FREQUENCY_HOT)
return true;
}
if (profile_status == PROFILE_ABSENT)
return true;
+ if (node->frequency == NODE_FREQUENCY_EXECUTED_ONCE
+ && freq <= (ENTRY_BLOCK_PTR->frequency * 2 / 3))
+ return false;
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 (const_basic_block bb)
{
- if (profile_info && flag_branch_probabilities
- && (bb->count
- < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
- return false;
+ if (profile_status == PROFILE_READ)
+ return maybe_hot_count_p (bb->count);
return maybe_hot_frequency_p (bb->frequency);
}
&& (edge->count
<= profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
return false;
- if (lookup_attribute ("cold", DECL_ATTRIBUTES (edge->callee->decl))
- || lookup_attribute ("cold", DECL_ATTRIBUTES (edge->caller->decl)))
+ if (edge->caller->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED
+ || edge->callee->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
return false;
- if (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl)))
+ if (optimize_size)
+ return false;
+ if (edge->caller->frequency == NODE_FREQUENCY_HOT)
return true;
+ if (edge->caller->frequency == NODE_FREQUENCY_EXECUTED_ONCE
+ && edge->frequency < CGRAPH_FREQ_BASE * 3 / 2)
+ return false;
if (flag_guess_branch_prob
- && edge->frequency < (CGRAPH_FREQ_MAX
- / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)))
+ && edge->frequency <= (CGRAPH_FREQ_BASE
+ / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION)))
return false;
return true;
}
bool
maybe_hot_edge_p (edge e)
{
- if (profile_info && flag_branch_probabilities
- && (e->count
- < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
- return false;
+ 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 cold and should be optimized for size. */
-
-bool
-probably_cold_bb_p (const_basic_block bb)
-{
- if (profile_info && flag_branch_probabilities
- && (bb->count
- < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
- return true;
- if ((!profile_info || !flag_branch_probabilities)
- && cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
- return true;
- if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
- return true;
- return false;
-}
-
/* Return true in case BB is probably never executed. */
bool
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)
+ && cgraph_node (current_function_decl)->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
return true;
return false;
}
optimize_function_for_size_p (struct function *fun)
{
return (optimize_size
- || fun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED);
+ || (fun && fun->decl
+ && (cgraph_node (fun->decl)->frequency
+ == NODE_FREQUENCY_UNLIKELY_EXECUTED)));
}
/* Return true when current function should always be optimized for speed. */
if (!preds)
return false;
-
+
for (i = (struct edge_prediction *) *preds; i; i = i->ep_next)
if (i->ep_predictor == predictor)
return true;
}
/* 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.
remove_predictions_associated_with_edge (edge e)
{
void **preds;
-
+
if (!bb_predictions)
return;
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;
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;
{
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 (dump_file, predictor, probability, bb,
combine_predictions_for_bb (basic_block bb)
{
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;
first = e;
}
- /* When there is no successor or only one choice, prediction is easy.
+ /* 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
by predictor with smallest index. */
for (pred = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
{
- int predictor = pred->ep_predictor;
+ 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)
- best_probability = probability, 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)
{
for (pred = (struct edge_prediction *) *preds; pred; pred = pred->ep_next)
{
- int predictor = pred->ep_predictor;
+ enum br_predictor predictor = pred->ep_predictor;
int probability = pred->ep_probability;
if (pred->ep_edge != EDGE_SUCC (bb, 0))
loop_iterator li;
struct loop *loop;
- scev_initialize ();
-
/* Try to predict out blocks in a loop that are not part of a
natural loop. */
FOR_EACH_LOOP (li, loop, 0)
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);
predict_edge (e, PRED_LOOP_EXIT, probability);
}
}
-
+
/* Free basic blocks from get_loop_body. */
free (bbs);
}
-
- scev_finalize ();
}
/* Attempt to predict probabilities of BB outgoing edges using local
return NULL;
}
-/* Return constant EXPR will likely have at execution time, NULL if unknown.
+/* 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. */
tree_bb_level_predictions (void)
{
basic_block bb;
+ bool has_return_edges = false;
+ edge e;
+ edge_iterator ei;
+
+ FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
+ if (!(e->flags & (EDGE_ABNORMAL | EDGE_FAKE | EDGE_EH)))
+ {
+ has_return_edges = true;
+ break;
+ }
apply_return_prediction ();
if (is_gimple_call (stmt))
{
- if (gimple_call_flags (stmt) & ECF_NORETURN)
+ if ((gimple_call_flags (stmt) & ECF_NORETURN)
+ && has_return_edges)
predict_paths_leading_to (bb, PRED_NORETURN,
NOT_TAKEN);
decl = gimple_call_fndecl (stmt);
}
#endif
-/* Predict branch probabilities and estimate profile of the tree CFG. */
-static unsigned int
+/* 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)
+ {
+ /* 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);
+}
+
+/* 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
tree_estimate_probability (void)
{
basic_block bb;
- loop_optimizer_init (0);
- if (dump_file && (dump_flags & TDF_DETAILS))
- flow_loops_dump (dump_file, NULL, 0);
-
add_noreturn_fake_exit_edges ();
connect_infinite_loops_to_exit ();
/* We use loop_niter_by_eval, which requires that the loops have
bb_predictions = pointer_map_create ();
tree_bb_level_predictions ();
-
- mark_irreducible_loops ();
record_loop_exits ();
+
if (number_of_loops () > 1)
predict_loops ();
FOR_EACH_BB (bb)
- {
- edge e;
- edge_iterator ei;
+ tree_estimate_probability_bb (bb);
- FOR_EACH_EDGE (e, ei, bb->succs)
- {
- /* 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
- && gimple_code (last_stmt (e->dest)) == 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);
- }
FOR_EACH_BB (bb)
combine_predictions_for_bb (bb);
estimate_bb_frequencies ();
free_dominance_info (CDI_POST_DOMINATORS);
remove_fake_exit_edges ();
+}
+
+/* Predict branch probabilities and estimate profile of the tree CFG.
+ This is the driver function for PASS_PROFILE. */
+
+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 (e)
{
sreal tmp;
-
+
/* EDGE_INFO (e)->back_edge_prob
= ((e->probability * BLOCK_INFO (bb)->frequency)
/ REG_BR_PROB_BASE); */
-
+
sreal_init (&tmp, e->probability, 0);
sreal_mul (&tmp, &tmp, &BLOCK_INFO (bb)->frequency);
sreal_mul (&EDGE_INFO (e)->back_edge_prob,
nextbb = e->dest;
else
BLOCK_INFO (last)->next = e->dest;
-
+
last = e->dest;
}
}
gcov_type count_max, true_count_max = 0;
basic_block bb;
- FOR_EACH_BB (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);
basic_block bb;
sreal freq_max;
- if (!flag_branch_probabilities || !counts_to_freqs ())
+ if (profile_status != PROFILE_READ || !counts_to_freqs ())
{
static int real_values_initialized = 0;
}
/* Decide whether function is hot, cold or unlikely executed. */
-static void
+void
compute_function_frequency (void)
{
basic_block bb;
+ struct cgraph_node *node = cgraph_node (current_function_decl);
if (!profile_info || !flag_branch_probabilities)
{
+ int flags = flags_from_decl_or_type (current_function_decl);
if (lookup_attribute ("cold", DECL_ATTRIBUTES (current_function_decl))
!= NULL)
- cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED;
+ node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
else if (lookup_attribute ("hot", DECL_ATTRIBUTES (current_function_decl))
!= NULL)
- cfun->function_frequency = FUNCTION_FREQUENCY_HOT;
+ node->frequency = NODE_FREQUENCY_HOT;
+ else if (flags & ECF_NORETURN)
+ node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
+ else if (MAIN_NAME_P (DECL_NAME (current_function_decl)))
+ node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
+ else if (DECL_STATIC_CONSTRUCTOR (current_function_decl)
+ || DECL_STATIC_DESTRUCTOR (current_function_decl))
+ node->frequency = NODE_FREQUENCY_EXECUTED_ONCE;
return;
}
- cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED;
+ node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
FOR_EACH_BB (bb)
{
if (maybe_hot_bb_p (bb))
{
- cfun->function_frequency = FUNCTION_FREQUENCY_HOT;
+ node->frequency = NODE_FREQUENCY_HOT;
return;
}
if (!probably_never_executed_bb_p (bb))
- cfun->function_frequency = FUNCTION_FREQUENCY_NORMAL;
+ node->frequency = NODE_FREQUENCY_NORMAL;
}
}
static void
choose_function_section (void)
{
+ struct cgraph_node *node = cgraph_node (current_function_decl);
if (DECL_SECTION_NAME (current_function_decl)
|| !targetm.have_named_sections
/* Theoretically we can split the gnu.linkonce text section too,
if (flag_reorder_blocks_and_partition)
return;
- if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT)
+ if (node->frequency == NODE_FREQUENCY_HOT)
DECL_SECTION_NAME (current_function_decl) =
build_string (strlen (HOT_TEXT_SECTION_NAME), HOT_TEXT_SECTION_NAME);
- if (cfun->function_frequency == FUNCTION_FREQUENCY_UNLIKELY_EXECUTED)
+ if (node->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED)
DECL_SECTION_NAME (current_function_decl) =
build_string (strlen (UNLIKELY_EXECUTED_TEXT_SECTION_NAME),
UNLIKELY_EXECUTED_TEXT_SECTION_NAME);
{
tree t = build1 (PREDICT_EXPR, void_type_node,
build_int_cst (NULL, predictor));
- PREDICT_EXPR_OUTCOME (t) = taken;
+ SET_PREDICT_EXPR_OUTCOME (t, taken);
return t;
}
return predictor_info[predictor].name;
}
-struct gimple_opt_pass pass_profile =
+struct gimple_opt_pass pass_profile =
{
{
GIMPLE_PASS,
"profile", /* name */
gate_estimate_probability, /* gate */
- tree_estimate_probability, /* execute */
+ tree_estimate_probability_driver, /* execute */
NULL, /* sub */
NULL, /* next */
0, /* static_pass_number */
}
};
-struct gimple_opt_pass pass_strip_predict_hints =
+struct gimple_opt_pass pass_strip_predict_hints =
{
{
GIMPLE_PASS,
- "", /* name */
+ "*strip_predict_hints", /* name */
NULL, /* gate */
strip_predict_hints, /* execute */
NULL, /* sub */