* Copyright: (c) 2006 Universität Karlsruhe
* Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
*/
-
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
-//#define USE_GSL
+#undef USE_GSL
#include <stdio.h>
#include <string.h>
+#include <limits.h>
#include <math.h>
#ifdef USE_GSL
#define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
+#define MAX_INT_FREQ 1000000
+
typedef struct _freq_t {
const ir_node *irn;
double freq;
size_t idx;
} walkerdata_t;
-struct _exec_freq_t {
+struct ir_exec_freq {
set *set;
hook_entry_t hook;
+ double max;
double min_non_zero;
+ double m, b;
unsigned infeasible : 1;
};
static int
cmp_freq(const void *a, const void *b, size_t size)
{
- const freq_t *p = a;
- const freq_t *q = b;
+ const freq_t *p = a;
+ const freq_t *q = b;
- return !(p->irn == q->irn);
+ return !(p->irn == q->irn);
}
static freq_t *
set_find_freq(set * set, const ir_node * irn)
{
- freq_t query;
+ freq_t query;
- query.irn = irn;
- return set_find(set, &query, sizeof(query), HASH_PTR(irn));
+ query.irn = irn;
+ return set_find(set, &query, sizeof(query), HASH_PTR(irn));
}
static freq_t *
set_insert_freq(set * set, const ir_node * irn)
{
- freq_t query;
+ freq_t query;
- query.irn = irn;
- query.freq = 0.0;
- return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
+ query.irn = irn;
+ query.freq = 0.0;
+ return set_insert(set, &query, sizeof(query), HASH_PTR(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;
assert(is_Block(irn));
freq = set_find_freq(freqs, irn);
assert(freq);
+
+ assert(freq->freq >= 0);
return freq->freq;
}
}
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);
- return (unsigned long) (f / ef->min_non_zero);
+ double f = get_block_execfreq(ef, bb);
+ int res = (int) (f > ef->min_non_zero ? ef->m * f + ef->b : 1.0);
+
+ // printf("%20.6f %10d\n", f, res);
+ 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)
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 *
+ir_exec_freq *create_execfreq(ir_graph *irg)
+{
+ ir_exec_freq *execfreq = xmalloc(sizeof(execfreq[0]));
+ memset(execfreq, 0, sizeof(execfreq[0]));
+ execfreq->set = new_set(cmp_freq, 32);
+
+ memset(&execfreq->hook, 0, sizeof(execfreq->hook));
+ execfreq->hook.context = execfreq;
+ execfreq->hook.hook._hook_node_info = exec_freq_node_info;
+ register_hook(hook_node_info, &execfreq->hook);
+
+ return execfreq;
+}
+
+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;
+}
+
+ir_exec_freq *
compute_execfreq(ir_graph * irg, double loop_weight)
{
- size_t size;
- double *matrix;
- double *rhs;
- int i;
- freq_t *freq;
- walkerdata_t wd;
- exec_freq_t *ef;
+ size_t size;
+ double *matrix;
+ double *rhs;
+ int i;
+ freq_t *freq;
+ walkerdata_t wd;
+ ir_exec_freq *ef;
set *freqs;
#ifdef USE_GSL
- gsl_vector *x;
+ gsl_vector *x;
#else
- double *x;
+ double *x;
#endif
ef = xmalloc(sizeof(ef[0]));
memset(ef, 0, sizeof(ef[0]));
ef->min_non_zero = 1e50; /* initialize with a reasonable large number. */
- freqs = ef->set = new_set(cmp_freq, 32);
+ freqs = ef->set = new_set(cmp_freq, 32);
- construct_cf_backedges(irg);
+ construct_cf_backedges(irg);
- wd.idx = 0;
- wd.set = freqs;
+ wd.idx = 0;
+ wd.set = freqs;
- irg_block_walk_graph(irg, block_walker, NULL, &wd);
+ irg_block_walk_graph(irg, block_walker, NULL, &wd);
- size = set_count(freqs);
- matrix = xmalloc(size*size*sizeof(*matrix));
- memset(matrix, 0, size*size*sizeof(*matrix));
- rhs = xmalloc(size*sizeof(*rhs));
- memset(rhs, 0, size*sizeof(*rhs));
+ size = set_count(freqs);
+ matrix = xmalloc(size*size*sizeof(*matrix));
+ memset(matrix, 0, size*size*sizeof(*matrix));
+ rhs = xmalloc(size*sizeof(*rhs));
+ memset(rhs, 0, size*sizeof(*rhs));
- set_foreach(freqs, freq) {
- ir_node *bb = (ir_node *)freq->irn;
- size_t idx = (int)get_irn_link(bb);
+ set_foreach(freqs, freq) {
+ ir_node *bb = (ir_node *)freq->irn;
+ size_t idx = (int)get_irn_link(bb);
- matrix[idx * (size + 1)] = -1.0;
+ matrix[idx * (size + 1)] = -1.0;
- if (bb == get_irg_start_block(irg)) {
- rhs[(int)get_irn_link(bb)] = -1.0;
- continue;
- }
+ if (bb == get_irg_start_block(irg)) {
+ rhs[(int)get_irn_link(bb)] = -1.0;
+ continue;
+ }
- for(i = get_Block_n_cfgpreds(bb) - 1; i >= 0; --i) {
- ir_node *pred = get_Block_cfgpred_block(bb, i);
- size_t pred_idx = (int)get_irn_link(pred);
+ for(i = get_Block_n_cfgpreds(bb) - 1; i >= 0; --i) {
+ ir_node *pred = get_Block_cfgpred_block(bb, i);
+ size_t pred_idx = (int)get_irn_link(pred);
-// matrix[pred_idx + idx*size] += 1.0/(double)get_Block_n_cfg_outs(pred);
- matrix[pred_idx + idx * size] += get_cf_probability(bb, i, loop_weight);
- }
- }
+ // matrix[pred_idx + idx*size] += 1.0/(double)get_Block_n_cfg_outs(pred);
+ matrix[pred_idx + idx * size] += get_cf_probability(bb, i, loop_weight);
+ }
+ }
- x = solve_lgs(matrix, rhs, size);
- if(x == NULL) {
+ x = solve_lgs(matrix, rhs, size);
+ if(x == NULL) {
ef->infeasible = 1;
return ef;
- }
+ }
+
+ 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 minimum non-zero execution frequency. */
- if(freq->freq > 0.0)
- ef->min_non_zero = MIN(ef->min_non_zero, 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);
+ }
+
+ /* 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 *fs = malloc(set_count(freqs) * sizeof(fs[0]));
+ int i, j, n = 0;
+
+ 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;
+
+ for(j = i + 1; j < n; ++j) {
+ double diff = fabs(fs[i] - fs[j]);
+
+ 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;
+
+ /*
+ * 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);
+ 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);
- return ef;
+ 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);
projects / libfirm / blobdiff