/* Branch prediction routines for the GNU compiler.
- Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc.
+ Copyright (C) 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
This file is part of GCC.
#include "config.h"
#include "system.h"
+#include "coretypes.h"
+#include "tm.h"
#include "tree.h"
#include "rtl.h"
#include "tm_p.h"
#include "recog.h"
#include "expr.h"
#include "predict.h"
-#include "profile.h"
-#include "real.h"
+#include "coverage.h"
+#include "sreal.h"
#include "params.h"
#include "target.h"
+#include "loop.h"
+#include "cfgloop.h"
-/* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE, 0.5,
- REAL_BB_FREQ_MAX. */
-static REAL_VALUE_TYPE real_zero, real_one, real_almost_one, real_br_prob_base,
- real_one_half, real_bb_freq_max;
+/* real constants: 0, 1, 1-1/REG_BR_PROB_BASE, REG_BR_PROB_BASE,
+ 1/REG_BR_PROB_BASE, 0.5, BB_FREQ_MAX. */
+static sreal real_zero, real_one, real_almost_one, real_br_prob_base,
+ real_inv_br_prob_base, real_one_half, real_bb_freq_max;
/* Random guesstimation given names. */
-#define PROB_NEVER (0)
#define PROB_VERY_UNLIKELY (REG_BR_PROB_BASE / 10 - 1)
-#define PROB_UNLIKELY (REG_BR_PROB_BASE * 4 / 10 - 1)
#define PROB_EVEN (REG_BR_PROB_BASE / 2)
-#define PROB_LIKELY (REG_BR_PROB_BASE - PROB_UNLIKELY)
#define PROB_VERY_LIKELY (REG_BR_PROB_BASE - PROB_VERY_UNLIKELY)
#define PROB_ALWAYS (REG_BR_PROB_BASE)
static void dump_prediction PARAMS ((enum br_predictor, int,
basic_block, int));
static void estimate_loops_at_level PARAMS ((struct loop *loop));
-static void propagate_freq PARAMS ((basic_block));
+static void propagate_freq PARAMS ((struct loop *));
static void estimate_bb_frequencies PARAMS ((struct loops *));
static void counts_to_freqs PARAMS ((void));
-static void process_note_predictions PARAMS ((basic_block, int *, int *,
- sbitmap *));
-static void process_note_prediction PARAMS ((basic_block, int *, int *,
- sbitmap *, int, int));
+static void process_note_predictions PARAMS ((basic_block, int *,
+ dominance_info,
+ dominance_info));
+static void process_note_prediction PARAMS ((basic_block, int *,
+ dominance_info,
+ dominance_info, int, int));
static bool last_basic_block_p PARAMS ((basic_block));
static void compute_function_frequency PARAMS ((void));
static void choose_function_section PARAMS ((void));
+static bool can_predict_insn_p PARAMS ((rtx));
/* Information we hold about each branch predictor.
Filled using information from predict.def. */
#undef DEF_PREDICTOR
/* Return true in case BB can be CPU intensive and should be optimized
- for maximal perofmrance. */
+ for maximal performance. */
bool
maybe_hot_bb_p (bb)
basic_block bb;
{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities
+ if (profile_info && flag_branch_probabilities
&& (bb->count
- < profile_info.max_counter_in_program
- / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
+ < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
return false;
if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
return false;
probably_cold_bb_p (bb)
basic_block bb;
{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities
+ if (profile_info && flag_branch_probabilities
&& (bb->count
- < profile_info.max_counter_in_program
- / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
+ < profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
return true;
if (bb->frequency < BB_FREQ_MAX / PARAM_VALUE (HOT_BB_FREQUENCY_FRACTION))
return true;
probably_never_executed_bb_p (bb)
basic_block bb;
{
- if (profile_info.count_profiles_merged
- && flag_branch_probabilities)
- return ((bb->count + profile_info.count_profiles_merged / 2)
- / profile_info.count_profiles_merged) == 0;
+ if (profile_info && flag_branch_probabilities)
+ return ((bb->count + profile_info->runs / 2) / profile_info->runs) == 0;
return false;
}
{
if (!any_condjump_p (insn))
abort ();
+ if (!flag_guess_branch_prob)
+ return;
REG_NOTES (insn)
= gen_rtx_EXPR_LIST (REG_BR_PRED,
predict_insn (last_insn, predictor, probability);
}
+/* Return true when we can store prediction on insn INSN.
+ At the moment we represent predictions only on conditional
+ jumps, not at computed jump or other complicated cases. */
+static bool
+can_predict_insn_p (insn)
+ rtx insn;
+{
+ return (GET_CODE (insn) == JUMP_INSN
+ && any_condjump_p (insn)
+ && BLOCK_FOR_INSN (insn)->succ->succ_next);
+}
+
/* Predict edge E by given predictor if possible. */
void
estimate_probability (loops_info)
struct loops *loops_info;
{
- sbitmap *dominators, *post_dominators;
- int i;
+ dominance_info dominators, post_dominators;
+ basic_block bb;
+ unsigned i;
- dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
- post_dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
- calculate_dominance_info (NULL, dominators, CDI_DOMINATORS);
- calculate_dominance_info (NULL, post_dominators, CDI_POST_DOMINATORS);
+ connect_infinite_loops_to_exit ();
+ dominators = calculate_dominance_info (CDI_DOMINATORS);
+ post_dominators = calculate_dominance_info (CDI_POST_DOMINATORS);
/* Try to predict out blocks in a loop that are not part of a
natural loop. */
- for (i = 0; i < loops_info->num; i++)
+ for (i = 1; i < loops_info->num; i++)
{
- int j;
+ basic_block bb, *bbs;
+ unsigned j;
int exits;
- struct loop *loop = &loops_info->array[i];
+ struct loop *loop = loops_info->parray[i];
+ struct loop_desc desc;
+ unsigned HOST_WIDE_INT niter;
flow_loop_scan (loops_info, loop, LOOP_EXIT_EDGES);
exits = loop->num_exits;
- for (j = loop->first->index; j <= loop->last->index; ++j)
- if (TEST_BIT (loop->nodes, j))
- {
- int header_found = 0;
- edge e;
+ if (simple_loop_p (loops_info, loop, &desc)
+ && desc.const_iter)
+ {
+ int prob;
+ niter = desc.niter + 1;
+ if (niter == 0) /* We might overflow here. */
+ niter = desc.niter;
+
+ prob = (REG_BR_PROB_BASE
+ - (REG_BR_PROB_BASE + niter /2) / niter);
+ /* Branch prediction algorithm gives 0 frequency for everything
+ after the end of loop for loop having 0 probability to finish. */
+ if (prob == REG_BR_PROB_BASE)
+ prob = REG_BR_PROB_BASE - 1;
+ predict_edge (desc.in_edge, PRED_LOOP_ITERATIONS,
+ prob);
+ }
+
+ bbs = get_loop_body (loop);
+ for (j = 0; j < loop->num_nodes; j++)
+ {
+ int header_found = 0;
+ edge e;
+
+ 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 (BASIC_BLOCK (j), PRED_CONTINUE))
+ if (!can_predict_insn_p (bb->end)
+ || predicted_by_p (bb, PRED_CONTINUE))
continue;
- /* Loop branch heuristics - predict an edge back to a
- loop's head as taken. */
- for (e = BASIC_BLOCK(j)->succ; e; e = e->succ_next)
- if (e->dest == loop->header
- && e->src == loop->latch)
- {
- header_found = 1;
- predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN);
- }
+ /* Loop branch heuristics - predict an edge back to a
+ loop's head as taken. */
+ for (e = bb->succ; e; e = e->succ_next)
+ if (e->dest == loop->header
+ && e->src == loop->latch)
+ {
+ header_found = 1;
+ predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN);
+ }
- /* Loop exit heuristics - predict an edge exiting the loop if the
- conditinal has no loop header successors as not taken. */
- if (!header_found)
- for (e = BASIC_BLOCK(j)->succ; e; e = e->succ_next)
- if (e->dest->index < 0
- || !TEST_BIT (loop->nodes, e->dest->index))
- predict_edge
- (e, PRED_LOOP_EXIT,
- (REG_BR_PROB_BASE
- - predictor_info [(int) PRED_LOOP_EXIT].hitrate)
- / exits);
- }
+ /* Loop exit heuristics - predict an edge exiting the loop if the
+ conditional has no loop header successors as not taken. */
+ if (!header_found)
+ for (e = bb->succ; e; e = e->succ_next)
+ if (e->dest->index < 0
+ || !flow_bb_inside_loop_p (loop, e->dest))
+ predict_edge
+ (e, PRED_LOOP_EXIT,
+ (REG_BR_PROB_BASE
+ - predictor_info [(int) PRED_LOOP_EXIT].hitrate)
+ / exits);
+ }
}
/* Attempt to predict conditional jumps using a number of heuristics. */
- for (i = 0; i < n_basic_blocks; i++)
+ FOR_EACH_BB (bb)
{
- basic_block bb = BASIC_BLOCK (i);
rtx last_insn = bb->end;
rtx cond, earliest;
edge e;
- if (GET_CODE (last_insn) != JUMP_INSN || ! any_condjump_p (last_insn))
+ if (! can_predict_insn_p (last_insn))
continue;
for (e = bb->succ; e; e = e->succ_next)
/* 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
- && TEST_BIT (dominators[e->dest->index], e->src->index)
- && !TEST_BIT (post_dominators[e->src->index], e->dest->index))
+ && dominated_by_p (dominators, e->dest, e->src)
+ && !dominated_by_p (post_dominators, e->src, e->dest))
{
rtx insn;
case EQ:
case UNEQ:
/* Floating point comparisons appears to behave in a very
- inpredictable way because of special role of = tests in
+ 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 usefull to predict about them. */
+ nothing useful to predict about them. */
else if (XEXP (cond, 1) == const0_rtx
|| XEXP (cond, 0) == const0_rtx)
;
case NE:
case LTGT:
/* Floating point comparisons appears to behave in a very
- inpredictable way because of special role of = tests in
+ 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 usefull to predict about them. */
+ nothing useful to predict about them. */
else if (XEXP (cond, 1) == const0_rtx
|| XEXP (cond, 0) == const0_rtx)
;
}
/* Attach the combined probability to each conditional jump. */
- for (i = 0; i < n_basic_blocks; i++)
- if (GET_CODE (BLOCK_END (i)) == JUMP_INSN
- && any_condjump_p (BLOCK_END (i))
- && BASIC_BLOCK (i)->succ->succ_next != NULL)
- combine_predictions_for_insn (BLOCK_END (i), BASIC_BLOCK (i));
+ FOR_EACH_BB (bb)
+ if (GET_CODE (bb->end) == JUMP_INSN
+ && any_condjump_p (bb->end)
+ && bb->succ->succ_next != NULL)
+ combine_predictions_for_insn (bb->end, bb);
- sbitmap_vector_free (post_dominators);
- sbitmap_vector_free (dominators);
+ free_dominance_info (post_dominators);
+ free_dominance_info (dominators);
+ remove_fake_edges ();
estimate_bb_frequencies (loops_info);
}
\f
process_note_prediction (bb, heads, dominators, post_dominators, pred, flags)
basic_block bb;
int *heads;
- int *dominators;
- sbitmap *post_dominators;
+ dominance_info dominators;
+ dominance_info post_dominators;
int pred;
int flags;
{
/* This is first time we need this field in heads array; so
find first dominator that we do not post-dominate (we are
using already known members of heads array). */
- int ai = bb->index;
- int next_ai = dominators[bb->index];
+ basic_block ai = bb;
+ basic_block next_ai = get_immediate_dominator (dominators, bb);
int head;
- while (heads[next_ai] < 0)
+ while (heads[next_ai->index] < 0)
{
- if (!TEST_BIT (post_dominators[next_ai], bb->index))
+ if (!dominated_by_p (post_dominators, next_ai, bb))
break;
- heads[next_ai] = ai;
+ heads[next_ai->index] = ai->index;
ai = next_ai;
- next_ai = dominators[next_ai];
+ next_ai = get_immediate_dominator (dominators, next_ai);
}
- if (!TEST_BIT (post_dominators[next_ai], bb->index))
- head = next_ai;
+ if (!dominated_by_p (post_dominators, next_ai, bb))
+ head = next_ai->index;
else
- head = heads[next_ai];
- while (next_ai != bb->index)
+ head = heads[next_ai->index];
+ while (next_ai != bb)
{
next_ai = ai;
- ai = heads[ai];
- heads[next_ai] = head;
+ if (heads[ai->index] == ENTRY_BLOCK)
+ ai = ENTRY_BLOCK_PTR;
+ else
+ ai = BASIC_BLOCK (heads[ai->index]);
+ heads[next_ai->index] = head;
}
}
y = heads[bb->index];
/* Now find the edge that leads to our branch and aply the prediction. */
- if (y == n_basic_blocks)
+ if (y == last_basic_block || !can_predict_insn_p (BASIC_BLOCK (y)->end))
return;
for (e = BASIC_BLOCK (y)->succ; e; e = e->succ_next)
if (e->dest->index >= 0
- && TEST_BIT (post_dominators[e->dest->index], bb->index))
+ && dominated_by_p (post_dominators, e->dest, bb))
predict_edge_def (e, pred, taken);
}
process_note_predictions (bb, heads, dominators, post_dominators)
basic_block bb;
int *heads;
- int *dominators;
- sbitmap *post_dominators;
+ dominance_info dominators;
+ dominance_info post_dominators;
{
rtx insn;
edge e;
- /* Additionaly, we check here for blocks with no successors. */
+ /* Additionally, we check here for blocks with no successors. */
int contained_noreturn_call = 0;
int was_bb_head = 0;
int noreturn_block = 1;
void
note_prediction_to_br_prob ()
{
- int i;
- sbitmap *post_dominators;
- int *dominators, *heads;
+ basic_block bb;
+ dominance_info post_dominators, dominators;
+ int *heads;
/* To enable handling of noreturn blocks. */
add_noreturn_fake_exit_edges ();
connect_infinite_loops_to_exit ();
- dominators = xmalloc (sizeof (int) * n_basic_blocks);
- memset (dominators, -1, sizeof (int) * n_basic_blocks);
- post_dominators = sbitmap_vector_alloc (n_basic_blocks, n_basic_blocks);
- calculate_dominance_info (NULL, post_dominators, CDI_POST_DOMINATORS);
- calculate_dominance_info (dominators, NULL, CDI_DOMINATORS);
+ post_dominators = calculate_dominance_info (CDI_POST_DOMINATORS);
+ dominators = calculate_dominance_info (CDI_DOMINATORS);
- heads = xmalloc (sizeof (int) * n_basic_blocks);
- memset (heads, -1, sizeof (int) * n_basic_blocks);
- heads[ENTRY_BLOCK_PTR->next_bb->index] = n_basic_blocks;
+ heads = xmalloc (sizeof (int) * last_basic_block);
+ memset (heads, -1, sizeof (int) * last_basic_block);
+ heads[ENTRY_BLOCK_PTR->next_bb->index] = last_basic_block;
/* Process all prediction notes. */
- for (i = 0; i < n_basic_blocks; ++i)
- {
- basic_block bb = BASIC_BLOCK (i);
- process_note_predictions (bb, heads, dominators, post_dominators);
- }
+ FOR_EACH_BB (bb)
+ process_note_predictions (bb, heads, dominators, post_dominators);
- sbitmap_vector_free (post_dominators);
- free (dominators);
+ free_dominance_info (post_dominators);
+ free_dominance_info (dominators);
free (heads);
remove_fake_edges ();
typedef struct block_info_def
{
/* Estimated frequency of execution of basic_block. */
- REAL_VALUE_TYPE frequency;
+ sreal frequency;
/* To keep queue of basic blocks to process. */
basic_block next;
/* In case edge is an loopback edge, the probability edge will be reached
in case header is. Estimated number of iterations of the loop can be
then computed as 1 / (1 - back_edge_prob). */
- REAL_VALUE_TYPE back_edge_prob;
+ sreal back_edge_prob;
/* True if the edge is an loopback edge in the natural loop. */
int back_edge:1;
} *edge_info;
#define EDGE_INFO(E) ((edge_info) (E)->aux)
/* Helper function for estimate_bb_frequencies.
- Propagate the frequencies for loops headed by HEAD. */
+ Propagate the frequencies for LOOP. */
static void
-propagate_freq (head)
- basic_block head;
+propagate_freq (loop)
+ struct loop *loop;
{
- basic_block bb = head;
- basic_block last = bb;
+ basic_block head = loop->header;
+ basic_block bb;
+ basic_block last;
edge e;
basic_block nextbb;
- int n;
/* For each basic block we need to visit count number of his predecessors
we need to visit first. */
- for (n = 0; n < n_basic_blocks; n++)
+ FOR_EACH_BB (bb)
{
- basic_block bb = BASIC_BLOCK (n);
if (BLOCK_INFO (bb)->tovisit)
{
int count = 0;
}
memcpy (&BLOCK_INFO (head)->frequency, &real_one, sizeof (real_one));
- for (; bb; bb = nextbb)
+ last = head;
+ for (bb = head; bb; bb = nextbb)
{
- REAL_VALUE_TYPE cyclic_probability, frequency;
+ sreal cyclic_probability, frequency;
memcpy (&cyclic_probability, &real_zero, sizeof (real_zero));
memcpy (&frequency, &real_zero, sizeof (real_zero));
for (e = bb->pred; e; e = e->pred_next)
if (EDGE_INFO (e)->back_edge)
{
- REAL_ARITHMETIC (cyclic_probability, PLUS_EXPR,
- cyclic_probability,
- EDGE_INFO (e)->back_edge_prob);
+ sreal_add (&cyclic_probability, &cyclic_probability,
+ &EDGE_INFO (e)->back_edge_prob);
}
else if (!(e->flags & EDGE_DFS_BACK))
{
- REAL_VALUE_TYPE tmp;
+ sreal tmp;
/* frequency += (e->probability
* BLOCK_INFO (e->src)->frequency /
REG_BR_PROB_BASE); */
- REAL_VALUE_FROM_INT (tmp, e->probability, 0,
- TYPE_MODE (double_type_node));
- REAL_ARITHMETIC (tmp, MULT_EXPR, tmp,
- BLOCK_INFO (e->src)->frequency);
- REAL_ARITHMETIC (tmp, RDIV_EXPR, tmp, real_br_prob_base);
- REAL_ARITHMETIC (frequency, PLUS_EXPR, frequency, tmp);
+ sreal_init (&tmp, e->probability, 0);
+ sreal_mul (&tmp, &tmp, &BLOCK_INFO (e->src)->frequency);
+ sreal_mul (&tmp, &tmp, &real_inv_br_prob_base);
+ sreal_add (&frequency, &frequency, &tmp);
}
- if (REAL_VALUES_LESS (real_almost_one, cyclic_probability))
- memcpy (&cyclic_probability, &real_almost_one, sizeof (real_zero));
+ if (sreal_compare (&cyclic_probability, &real_zero) == 0)
+ {
+ memcpy (&BLOCK_INFO (bb)->frequency, &frequency,
+ sizeof (frequency));
+ }
+ else
+ {
+ if (sreal_compare (&cyclic_probability, &real_almost_one) > 0)
+ {
+ memcpy (&cyclic_probability, &real_almost_one,
+ sizeof (real_almost_one));
+ }
- /* BLOCK_INFO (bb)->frequency = frequency / (1 - cyclic_probability)
- */
+ /* BLOCK_INFO (bb)->frequency = frequency
+ / (1 - cyclic_probability) */
- REAL_ARITHMETIC (cyclic_probability, MINUS_EXPR, real_one,
- cyclic_probability);
- REAL_ARITHMETIC (BLOCK_INFO (bb)->frequency,
- RDIV_EXPR, frequency, cyclic_probability);
+ sreal_sub (&cyclic_probability, &real_one, &cyclic_probability);
+ sreal_div (&BLOCK_INFO (bb)->frequency,
+ &frequency, &cyclic_probability);
+ }
}
BLOCK_INFO (bb)->tovisit = 0;
for (e = bb->succ; e; e = e->succ_next)
if (e->dest == head)
{
- REAL_VALUE_TYPE tmp;
+ sreal tmp;
/* EDGE_INFO (e)->back_edge_prob
= ((e->probability * BLOCK_INFO (bb)->frequency)
/ REG_BR_PROB_BASE); */
- REAL_VALUE_FROM_INT (tmp, e->probability, 0,
- TYPE_MODE (double_type_node));
- REAL_ARITHMETIC (tmp, MULT_EXPR, tmp,
- BLOCK_INFO (bb)->frequency);
- REAL_ARITHMETIC (EDGE_INFO (e)->back_edge_prob,
- RDIV_EXPR, tmp, real_br_prob_base);
+ sreal_init (&tmp, e->probability, 0);
+ sreal_mul (&tmp, &tmp, &BLOCK_INFO (bb)->frequency);
+ sreal_mul (&EDGE_INFO (e)->back_edge_prob,
+ &tmp, &real_inv_br_prob_base);
}
/* Propagate to successor blocks. */
estimate_loops_at_level (first_loop)
struct loop *first_loop;
{
- struct loop *l, *loop = first_loop;
+ struct loop *loop;
for (loop = first_loop; loop; loop = loop->next)
{
- int n;
edge e;
+ basic_block *bbs;
+ unsigned i;
estimate_loops_at_level (loop->inner);
-
- /* Find current loop back edge and mark it. */
- for (e = loop->latch->succ; e->dest != loop->header; e = e->succ_next)
- ;
-
- EDGE_INFO (e)->back_edge = 1;
-
- /* In case the loop header is shared, ensure that it is the last
- one sharing the same header, so we avoid redundant work. */
- if (loop->shared)
+
+ if (loop->latch->succ) /* Do not do this for dummy function loop. */
{
- for (l = loop->next; l; l = l->next)
- if (l->header == loop->header)
- break;
-
- if (l)
- continue;
- }
-
- /* Now merge all nodes of all loops with given header as not visited. */
- for (l = loop->shared ? first_loop : loop; l != loop->next; l = l->next)
- if (loop->header == l->header)
- EXECUTE_IF_SET_IN_SBITMAP (l->nodes, 0, n,
- BLOCK_INFO (BASIC_BLOCK (n))->tovisit = 1
- );
-
- propagate_freq (loop->header);
+ /* Find current loop back edge and mark it. */
+ e = loop_latch_edge (loop);
+ EDGE_INFO (e)->back_edge = 1;
+ }
+
+ bbs = get_loop_body (loop);
+ for (i = 0; i < loop->num_nodes; i++)
+ BLOCK_INFO (bbs[i])->tovisit = 1;
+ free (bbs);
+ propagate_freq (loop);
}
}
static void
counts_to_freqs ()
{
- HOST_WIDEST_INT count_max = 1;
- int i;
+ gcov_type count_max = 1;
+ basic_block bb;
- for (i = 0; i < n_basic_blocks; i++)
- count_max = MAX (BASIC_BLOCK (i)->count, count_max);
+ FOR_EACH_BB (bb)
+ count_max = MAX (bb->count, count_max);
- for (i = -2; i < n_basic_blocks; i++)
- {
- basic_block bb;
-
- if (i == -2)
- bb = ENTRY_BLOCK_PTR;
- else if (i == -1)
- bb = EXIT_BLOCK_PTR;
- else
- bb = BASIC_BLOCK (i);
-
- bb->frequency = (bb->count * BB_FREQ_MAX + count_max / 2) / count_max;
- }
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
+ bb->frequency = (bb->count * BB_FREQ_MAX + count_max / 2) / count_max;
}
/* Return true if function is likely to be expensive, so there is no point to
optimize performance of prologue, epilogue or do inlining at the expense
- of code size growth. THRESHOLD is the limit of number of isntructions
+ of code size growth. THRESHOLD is the limit of number of instructions
function can execute at average to be still considered not expensive. */
bool
int threshold;
{
unsigned int sum = 0;
- int i;
+ basic_block bb;
unsigned int limit;
/* We can not compute accurately for large thresholds due to scaled
/* Maximally BB_FREQ_MAX^2 so overflow won't happen. */
limit = ENTRY_BLOCK_PTR->frequency * threshold;
- for (i = 0; i < n_basic_blocks; i++)
+ FOR_EACH_BB (bb)
{
- basic_block bb = BASIC_BLOCK (i);
rtx insn;
for (insn = bb->head; insn != NEXT_INSN (bb->end);
estimate_bb_frequencies (loops)
struct loops *loops;
{
- int i;
- REAL_VALUE_TYPE freq_max;
- enum machine_mode double_mode = TYPE_MODE (double_type_node);
+ basic_block bb;
+ sreal freq_max;
if (flag_branch_probabilities)
counts_to_freqs ();
else
{
- REAL_VALUE_FROM_INT (real_zero, 0, 0, double_mode);
- REAL_VALUE_FROM_INT (real_one, 1, 0, double_mode);
- REAL_VALUE_FROM_INT (real_br_prob_base, REG_BR_PROB_BASE, 0, double_mode);
- REAL_VALUE_FROM_INT (real_bb_freq_max, BB_FREQ_MAX, 0, double_mode);
- REAL_VALUE_FROM_INT (real_one_half, 2, 0, double_mode);
-
- REAL_ARITHMETIC (real_one_half, RDIV_EXPR, real_one, real_one_half);
-
- REAL_ARITHMETIC (real_almost_one, RDIV_EXPR, real_one, real_br_prob_base);
- REAL_ARITHMETIC (real_almost_one, MINUS_EXPR, real_one, real_almost_one);
+ static int real_values_initialized = 0;
+
+ if (!real_values_initialized)
+ {
+ real_values_initialized = 1;
+ sreal_init (&real_zero, 0, 0);
+ sreal_init (&real_one, 1, 0);
+ sreal_init (&real_br_prob_base, REG_BR_PROB_BASE, 0);
+ sreal_init (&real_bb_freq_max, BB_FREQ_MAX, 0);
+ sreal_init (&real_one_half, 1, -1);
+ sreal_div (&real_inv_br_prob_base, &real_one, &real_br_prob_base);
+ sreal_sub (&real_almost_one, &real_one, &real_inv_br_prob_base);
+ }
mark_dfs_back_edges ();
/* Fill in the probability values in flowgraph based on the REG_BR_PROB
notes. */
- for (i = 0; i < n_basic_blocks; i++)
+ FOR_EACH_BB (bb)
{
- rtx last_insn = BLOCK_END (i);
+ rtx last_insn = bb->end;
- if (GET_CODE (last_insn) != JUMP_INSN || !any_condjump_p (last_insn)
- /* Avoid handling of conditional jumps jumping to fallthru edge. */
- || BASIC_BLOCK (i)->succ->succ_next == NULL)
+ if (!can_predict_insn_p (last_insn))
{
/* We can predict only conditional jumps at the moment.
Expect each edge to be equally probable.
int nedges = 0;
edge e;
- for (e = BASIC_BLOCK (i)->succ; e; e = e->succ_next)
+ for (e = bb->succ; e; e = e->succ_next)
{
nedges++;
if (e->probability != 0)
break;
}
if (!e)
- for (e = BASIC_BLOCK (i)->succ; e; e = e->succ_next)
+ for (e = bb->succ; e; e = e->succ_next)
e->probability = (REG_BR_PROB_BASE + nedges / 2) / nedges;
}
}
/* Set up block info for each basic block. */
alloc_aux_for_blocks (sizeof (struct block_info_def));
alloc_aux_for_edges (sizeof (struct edge_info_def));
- for (i = -2; i < n_basic_blocks; i++)
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
{
edge e;
- basic_block bb;
-
- if (i == -2)
- bb = ENTRY_BLOCK_PTR;
- else if (i == -1)
- bb = EXIT_BLOCK_PTR;
- else
- bb = BASIC_BLOCK (i);
BLOCK_INFO (bb)->tovisit = 0;
for (e = bb->succ; e; e = e->succ_next)
{
-
- REAL_VALUE_FROM_INT (EDGE_INFO (e)->back_edge_prob,
- e->probability, 0, double_mode);
- REAL_ARITHMETIC (EDGE_INFO (e)->back_edge_prob,
- RDIV_EXPR, EDGE_INFO (e)->back_edge_prob,
- real_br_prob_base);
+ sreal_init (&EDGE_INFO (e)->back_edge_prob, e->probability, 0);
+ sreal_mul (&EDGE_INFO (e)->back_edge_prob,
+ &EDGE_INFO (e)->back_edge_prob,
+ &real_inv_br_prob_base);
}
}
to outermost to examine probabilities for back edges. */
estimate_loops_at_level (loops->tree_root);
- /* Now fake loop around whole function to finalize probabilities. */
- for (i = 0; i < n_basic_blocks; i++)
- BLOCK_INFO (BASIC_BLOCK (i))->tovisit = 1;
-
- BLOCK_INFO (ENTRY_BLOCK_PTR)->tovisit = 1;
- BLOCK_INFO (EXIT_BLOCK_PTR)->tovisit = 1;
- propagate_freq (ENTRY_BLOCK_PTR);
-
memcpy (&freq_max, &real_zero, sizeof (real_zero));
- for (i = 0; i < n_basic_blocks; i++)
- if (REAL_VALUES_LESS
- (freq_max, BLOCK_INFO (BASIC_BLOCK (i))->frequency))
- memcpy (&freq_max, &BLOCK_INFO (BASIC_BLOCK (i))->frequency,
- sizeof (freq_max));
+ FOR_EACH_BB (bb)
+ if (sreal_compare (&freq_max, &BLOCK_INFO (bb)->frequency) < 0)
+ memcpy (&freq_max, &BLOCK_INFO (bb)->frequency, sizeof (freq_max));
- for (i = -2; i < n_basic_blocks; i++)
+ sreal_div (&freq_max, &real_bb_freq_max, &freq_max);
+ FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, NULL, next_bb)
{
- basic_block bb;
- REAL_VALUE_TYPE tmp;
+ sreal tmp;
- if (i == -2)
- bb = ENTRY_BLOCK_PTR;
- else if (i == -1)
- bb = EXIT_BLOCK_PTR;
- else
- bb = BASIC_BLOCK (i);
-
- REAL_ARITHMETIC (tmp, MULT_EXPR, BLOCK_INFO (bb)->frequency,
- real_bb_freq_max);
- REAL_ARITHMETIC (tmp, RDIV_EXPR, tmp, freq_max);
- REAL_ARITHMETIC (tmp, PLUS_EXPR, tmp, real_one_half);
- bb->frequency = REAL_VALUE_UNSIGNED_FIX (tmp);
+ sreal_mul (&tmp, &BLOCK_INFO (bb)->frequency, &freq_max);
+ sreal_add (&tmp, &tmp, &real_one_half);
+ bb->frequency = sreal_to_int (&tmp);
}
free_aux_for_blocks ();
static void
compute_function_frequency ()
{
- int i;
- if (!profile_info.count_profiles_merged
- || !flag_branch_probabilities)
+ basic_block bb;
+
+ if (!profile_info || !flag_branch_probabilities)
return;
cfun->function_frequency = FUNCTION_FREQUENCY_UNLIKELY_EXECUTED;
- for (i = 0; i < n_basic_blocks; i++)
+ FOR_EACH_BB (bb)
{
- basic_block bb = BASIC_BLOCK (i);
if (maybe_hot_bb_p (bb))
{
cfun->function_frequency = FUNCTION_FREQUENCY_HOT;
choose_function_section ()
{
if (DECL_SECTION_NAME (current_function_decl)
- || !targetm.have_named_sections)
+ || !targetm.have_named_sections
+ /* Theoretically we can split the gnu.linkonce text section too,
+ but this requires more work as the frequency needs to match
+ for all generated objects so we need to merge the frequency
+ of all instances. For now just never set frequency for these. */
+ || DECL_ONE_ONLY (current_function_decl))
return;
if (cfun->function_frequency == FUNCTION_FREQUENCY_HOT)
DECL_SECTION_NAME (current_function_decl) =