/*
- * Project: libFIRM
- * File name: ir/ana/execfreq.c
- * Purpose: Compute an estimate of basic block executions.
- * Author: Adam M. Szalkowski
- * Modified by:
- * Created: 28.05.2006
- * CVS-ID: $Id$
- * Copyright: (c) 2006 Universität Karlsruhe
- * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
+ * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
*/
+/**
+ * @file
+ * @brief Compute an estimate of basic block executions.
+ * @author Adam M. Szalkowski
+ * @date 28.05.2006
+ * @version $Id$
+ */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "gaussjordan.h"
#endif
-#include "execfreq.h"
-
#include "firm_common_t.h"
#include "set.h"
#include "hashptr.h"
+#include "debug.h"
#include "irprog_t.h"
#include "irgraph_t.h"
#include "irnode_t.h"
#include "irloop.h"
#include "irgwalk.h"
-#include "irouts.h"
+#include "iredges.h"
#include "irprintf.h"
+#include "irtools.h"
#include "irhooks.h"
#include "execfreq.h"
size_t idx;
} walkerdata_t;
-struct _exec_freq_t {
+struct ir_exec_freq {
set *set;
hook_entry_t hook;
double max;
{
const freq_t *p = a;
const freq_t *q = b;
+ (void) size;
return !(p->irn == q->irn);
}
}
double
-get_block_execfreq(const exec_freq_t *ef, const ir_node * irn)
+get_block_execfreq(const ir_exec_freq *ef, const ir_node * irn)
{
if(!ef->infeasible) {
set *freqs = ef->set;
}
unsigned long
-get_block_execfreq_ulong(const exec_freq_t *ef, const ir_node *bb)
+get_block_execfreq_ulong(const ir_exec_freq *ef, const ir_node *bb)
{
double f = get_block_execfreq(ef, bb);
int res = (int) (f > ef->min_non_zero ? ef->m * f + ef->b : 1.0);
return res;
}
-#define ZERO(x) (fabs(x) < 0.0001)
+#define EPSILON 0.0001
+#define UNDEF(x) !(x > EPSILON)
static void
block_walker(ir_node * bb, void * data)
return NULL;
}
}
-#endif
+#endif /* USE_GSL */
static double
get_cf_probability(ir_node *bb, int pos, double loop_weight)
{
- double sum = 0.0;
- double cur = 0.0;
- int i;
- ir_node *pred = get_Block_cfgpred_block(bb, pos);
-
- cur = get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
-
- for(i = get_Block_n_cfg_outs(pred) - 1; i >= 0; --i) {
- ir_node *succ = get_Block_cfg_out(pred, i);
-
- sum += get_loop_depth(get_irn_loop(succ)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
- }
+ double sum = 0.0;
+ double cur = 0.0;
+ const ir_node *pred = get_Block_cfgpred_block(bb, pos);
+ const ir_loop *pred_loop = get_irn_loop(pred);
+ int pred_depth = get_loop_depth(pred_loop);
+ const ir_edge_t *edge;
+
+ cur = get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
+
+ foreach_block_succ(pred, edge) {
+ const ir_node *block = get_edge_src_irn(edge);
+ const ir_loop *loop = get_irn_loop(block);
+ int depth = get_loop_depth(loop);
+ sum += depth < pred_depth ? 1.0 : loop_weight;
+ }
- return cur/sum;
+ return cur/sum;
}
static void exec_freq_node_info(void *ctx, FILE *f, const ir_node *irn)
{
if(is_Block(irn)) {
- exec_freq_t *ef = ctx;
+ ir_exec_freq *ef = ctx;
fprintf(f, "execution frequency: %g/%lu\n", get_block_execfreq(ef, irn), get_block_execfreq_ulong(ef, irn));
}
}
-exec_freq_t *create_execfreq(ir_graph *irg)
+ir_exec_freq *create_execfreq(ir_graph *irg)
{
- exec_freq_t *execfreq = xmalloc(sizeof(execfreq[0]));
+ ir_exec_freq *execfreq = xmalloc(sizeof(execfreq[0]));
memset(execfreq, 0, sizeof(execfreq[0]));
execfreq->set = new_set(cmp_freq, 32);
execfreq->hook.context = execfreq;
execfreq->hook.hook._hook_node_info = exec_freq_node_info;
register_hook(hook_node_info, &execfreq->hook);
+ (void) irg;
return execfreq;
}
-void set_execfreq(exec_freq_t *execfreq, const ir_node *block, double freq)
+void set_execfreq(ir_exec_freq *execfreq, const ir_node *block, double freq)
{
freq_t *f = set_insert_freq(execfreq->set, block);
f->freq = freq;
}
-exec_freq_t *
+ir_exec_freq *
compute_execfreq(ir_graph * irg, double loop_weight)
{
size_t size;
int i;
freq_t *freq;
walkerdata_t wd;
- exec_freq_t *ef;
+ ir_exec_freq *ef;
set *freqs;
#ifdef USE_GSL
gsl_vector *x;
freqs = ef->set = new_set(cmp_freq, 32);
construct_cf_backedges(irg);
+ /* TODO: edges are corrupt for EDGE_KIND_BLOCK after the local optimize
+ graph phase merges blocks in the x86 backend */
+ edges_deactivate(irg);
+ edges_activate(irg);
+ /* edges_assure(irg); */
wd.idx = 0;
wd.set = freqs;
}
x = solve_lgs(matrix, rhs, size);
- if(x == NULL) {
+ if (x == NULL) {
+ DEBUG_ONLY(ir_fprintf(stderr, "Debug Warning: Couldn't estimate execution frequencies for %+F\n", irg));
ef->infeasible = 1;
- return ef;
- }
+ } else {
+ ef->max = 0.0;
- ef->max = 0.0;
-
- set_foreach(freqs, freq) {
- const ir_node *bb = freq->irn;
- size_t idx = PTR_TO_INT(get_irn_link(bb));
+ set_foreach(freqs, freq) {
+ const ir_node *bb = freq->irn;
+ size_t idx = PTR_TO_INT(get_irn_link(bb));
#ifdef USE_GSL
- freq->freq = ZERO(gsl_vector_get(x, idx)) ? 0.0 : gsl_vector_get(x, idx);
+ freq->freq = UNDEF(gsl_vector_get(x, idx)) ? EPSILON : gsl_vector_get(x, idx);
#else
- freq->freq = ZERO(x[idx]) ? 0.0 : x[idx];
+ freq->freq = UNDEF(x[idx]) ? EPSILON : x[idx];
#endif
- /* get the maximum exec freq */
- ef->max = MAX(ef->max, freq->freq);
+ /* get the maximum exec freq */
+ ef->max = MAX(ef->max, freq->freq);
- /* Get the minimum non-zero execution frequency. */
- if(freq->freq > 0.0)
- ef->min_non_zero = MIN(ef->min_non_zero, freq->freq);
- }
+ /* Get the minimum non-zero execution frequency. */
+ if(freq->freq > 0.0)
+ ef->min_non_zero = MIN(ef->min_non_zero, freq->freq);
+ }
- /* compute m and b of the transformation used to convert the doubles into scaled ints */
- {
- double smallest_diff = 1.0;
+ /* compute m and b of the transformation used to convert the doubles into scaled ints */
+ {
+ double smallest_diff = 1.0;
- double l2 = ef->min_non_zero;
- double h2 = ef->max;
- double l1 = 1.0;
- double h1 = MAX_INT_FREQ;
+ double l2 = ef->min_non_zero;
+ double h2 = ef->max;
+ double l1 = 1.0;
+ double h1 = MAX_INT_FREQ;
- double *fs = malloc(set_count(freqs) * sizeof(fs[0]));
- int i, j, n = 0;
+ double *fs = malloc(set_count(freqs) * sizeof(fs[0]));
+ int i, j, n = 0;
- set_foreach(freqs, freq)
- fs[n++] = freq->freq;
+ set_foreach(freqs, freq)
+ fs[n++] = freq->freq;
- /*
- * find the smallest difference of the execution frequencies
- * we try to ressolve it with 1 integer.
- */
- for(i = 0; i < n; ++i) {
- if(fs[i] <= 0.0)
- continue;
+ /*
+ * find the smallest difference of the execution frequencies
+ * we try to ressolve it with 1 integer.
+ */
+ for(i = 0; i < n; ++i) {
+ if(fs[i] <= 0.0)
+ continue;
- for(j = i + 1; j < n; ++j) {
- double diff = fabs(fs[i] - fs[j]);
+ for(j = i + 1; j < n; ++j) {
+ double diff = fabs(fs[i] - fs[j]);
- if(!ZERO(diff))
- smallest_diff = MIN(diff, smallest_diff);
+ if(!UNDEF(diff))
+ smallest_diff = MIN(diff, smallest_diff);
+ }
}
- }
-
- /* according to that the slope of the translation function is 1.0 / smallest diff */
- ef->m = 1.0 / smallest_diff;
- /* the abscissa is then given by */
- ef->b = l1 - ef->m * l2;
+ /* according to that the slope of the translation function is 1.0 / smallest diff */
+ ef->m = 1.0 / smallest_diff;
- /*
- * if the slope is so high that the largest integer would be larger than MAX_INT_FREQ
- * set the largest int freq to that upper limit and recompute the translation function
- */
- if(ef->m * h2 + ef->b > MAX_INT_FREQ) {
- ef->m = (h1 - l1) / (h2 - l2);
+ /* the abscissa is then given by */
ef->b = l1 - ef->m * l2;
- }
- // printf("smallest_diff: %g, l1: %f, h1: %f, l2: %f, h2: %f, m: %f, b: %f\n", smallest_diff, l1, h1, l2, h2, ef->m, ef->b);
- free(fs);
- }
+ /*
+ * if the slope is so high that the largest integer would be larger than MAX_INT_FREQ
+ * set the largest int freq to that upper limit and recompute the translation function
+ */
+ if(ef->m * h2 + ef->b > MAX_INT_FREQ) {
+ ef->m = (h1 - l1) / (h2 - l2);
+ ef->b = l1 - ef->m * l2;
+ }
+
+ // printf("smallest_diff: %g, l1: %f, h1: %f, l2: %f, h2: %f, m: %f, b: %f\n", smallest_diff, l1, h1, l2, h2, ef->m, ef->b);
+ free(fs);
+ }
#ifdef USE_GSL
- gsl_vector_free(x);
+ gsl_vector_free(x);
#endif
- free(matrix);
+ memset(&ef->hook, 0, sizeof(ef->hook));
+ ef->hook.context = ef;
+ ef->hook.hook._hook_node_info = exec_freq_node_info;
+ register_hook(hook_node_info, &ef->hook);
+ }
- memset(&ef->hook, 0, sizeof(ef->hook));
- ef->hook.context = ef;
- ef->hook.hook._hook_node_info = exec_freq_node_info;
- register_hook(hook_node_info, &ef->hook);
+ free(matrix);
+ free(rhs);
return ef;
}
void
-free_execfreq(exec_freq_t *ef)
+free_execfreq(ir_exec_freq *ef)
{
del_set(ef->set);
unregister_hook(hook_node_info, &ef->hook);
free(ef);
}
-
-#undef ELEM