/*
- * 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) 1995-2008 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.
*/
-#ifdef HAVE_CONFIG_H
-#include "config.h"
-#endif
-#undef USE_GSL
+/**
+ * @file
+ * @brief Compute an estimate of basic block executions.
+ * @author Adam M. Szalkowski
+ * @date 28.05.2006
+ */
+#include "config.h"
#include <stdio.h>
#include <string.h>
#include <limits.h>
#include <math.h>
-#ifdef USE_GSL
-#include <gsl/gsl_linalg.h>
-#include <gsl/gsl_vector.h>
-#else
-#include "gaussjordan.h"
-#endif
-
-#include "execfreq.h"
+#include "gaussseidel.h"
-#include "firm_common_t.h"
#include "set.h"
#include "hashptr.h"
+#include "debug.h"
+#include "statev.h"
+#include "dfs_t.h"
+#include "absgraph.h"
#include "irprog_t.h"
#include "irgraph_t.h"
#include "irloop.h"
#include "irgwalk.h"
#include "iredges.h"
+#include "irouts.h"
#include "irprintf.h"
+#include "util.h"
#include "irhooks.h"
#include "execfreq.h"
-#define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
+/* enable to also solve the equations with Gauss-Jordan */
+#undef COMPARE_AGAINST_GAUSSJORDAN
+
+#ifdef COMPARE_AGAINST_GAUSSJORDAN
+#include "gaussjordan.h"
+#endif
+
+
+#define EPSILON 1e-5
+#define UNDEF(x) (fabs(x) < EPSILON)
+#define SEIDEL_TOLERANCE 1e-7
#define MAX_INT_FREQ 1000000
-typedef struct _freq_t {
+#define set_foreach(s,type,i) for ((i)=(type)set_first((s)); (i); (i)=(type)set_next((s)))
+
+typedef struct freq_t {
const ir_node *irn;
+ int idx;
double freq;
} freq_t;
-
-typedef struct _walkerdata_t {
- set *set;
- size_t idx;
-} walkerdata_t;
-
struct ir_exec_freq {
- set *set;
+ set *freqs;
hook_entry_t hook;
double max;
double min_non_zero;
unsigned infeasible : 1;
};
-static int
-cmp_freq(const void *a, const void *b, size_t size)
+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 = (const freq_t*) a;
+ const freq_t *q = (const freq_t*) b;
+ (void) size;
return !(p->irn == q->irn);
}
-static freq_t *
-set_find_freq(set * set, const ir_node * irn)
+static freq_t *set_find_freq(set *freqs, const ir_node *irn)
{
- freq_t query;
-
+ freq_t query;
query.irn = irn;
- return set_find(set, &query, sizeof(query), HASH_PTR(irn));
+ return (freq_t*) set_find(freqs, &query, sizeof(query), hash_ptr(irn));
}
-static freq_t *
-set_insert_freq(set * set, const ir_node * irn)
+static freq_t *set_insert_freq(set *freqs, const ir_node *irn)
{
freq_t query;
query.irn = irn;
query.freq = 0.0;
- return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
+ query.idx = -1;
+ return (freq_t*) set_insert(freqs, &query, sizeof(query), hash_ptr(irn));
}
-double
-get_block_execfreq(const ir_exec_freq *ef, const ir_node * irn)
+double get_block_execfreq(const ir_exec_freq *ef, const ir_node *irn)
{
- if(!ef->infeasible) {
- set *freqs = ef->set;
+ if (!ef->infeasible) {
+ set *freqs = ef->freqs;
freq_t *freq;
assert(is_Block(irn));
freq = set_find_freq(freqs, irn);
{
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 EPSILON 0.0001
-#define UNDEF(x) !(x > EPSILON)
-
-static void
-block_walker(ir_node * bb, void * data)
-{
- walkerdata_t *wd = data;
-
- set_insert_freq(wd->set, bb);
- set_irn_link(bb, (void*)wd->idx++);
-}
-
-#ifdef USE_GSL
-static gsl_vector *
-solve_lgs(double * a_data, double * b_data, size_t size)
+static double *solve_lgs(gs_matrix_t *mat, double *x, int size)
{
- gsl_matrix_view m
- = gsl_matrix_view_array (a_data, size, size);
-
- gsl_vector_view b
- = gsl_vector_view_array (b_data, size);
-
- gsl_vector *x = gsl_vector_alloc (size);
-
- int s;
-
- gsl_permutation * p = gsl_permutation_alloc (size);
+ double init = 1.0 / size;
+ double dev;
+ int i, iter = 0;
+
+ /* better convergence. */
+ for (i = 0; i < size; ++i)
+ x[i] = init;
+
+ stat_ev_dbl("execfreq_matrix_size", size);
+ stat_ev_tim_push();
+ do {
+ ++iter;
+ dev = gs_matrix_gauss_seidel(mat, x, size);
+ } while (fabs(dev) > SEIDEL_TOLERANCE);
+ stat_ev_tim_pop("execfreq_seidel_time");
+ stat_ev_dbl("execfreq_seidel_iter", iter);
+
+#ifdef COMPARE_AGAINST_GAUSSJORDAN
+ {
+ double *nw = XMALLOCN(double, size * size);
+ double *nx = XMALLOCNZ(double, size);
- gsl_linalg_LU_decomp (&m.matrix, p, &s);
+ gs_matrix_export(mat, nw, size);
- gsl_linalg_LU_solve (&m.matrix, p, &b.vector, x);
+ stat_ev_tim_push();
+ firm_gaussjordansolve(nw, nx, size);
+ stat_ev_tim_pop("execfreq_jordan_time");
- gsl_permutation_free (p);
+ xfree(nw);
+ xfree(nx);
+ }
+#endif
- return x;
+ return x;
}
-#else
-static double *
-solve_lgs(double * A, double * b, size_t size)
-{
- if(firm_gaussjordansolve(A,b,size) == 0) {
- return b;
- } else {
- return NULL;
- }
-}
-#endif
-static double
-get_cf_probability(ir_node *bb, int pos, double loop_weight)
+/*
+ * Determine probability that predecessor pos takes this cf edge.
+ */
+static double get_cf_probability(ir_node *bb, int pos, double 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);
+ double sum = 0.0;
+ double cur = 1.0;
+ double inv_loop_weight = 1./loop_weight;
+ const ir_node *pred = get_Block_cfgpred_block(bb, pos);
+ const ir_loop *pred_loop;
+ int pred_depth;
const ir_edge_t *edge;
+ const ir_loop *loop;
+ int depth;
+ int d;
+
+ if (is_Bad(pred))
+ return 0;
- cur = get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
+ loop = get_irn_loop(bb);
+ depth = get_loop_depth(loop);
+ pred_loop = get_irn_loop(pred);
+ pred_depth = get_loop_depth(pred_loop);
+
+ for (d = depth; d < pred_depth; ++d) {
+ cur *= inv_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;
+ const ir_node *succ = get_edge_src_irn(edge);
+ const ir_loop *succ_loop = get_irn_loop(succ);
+ int succ_depth = get_loop_depth(succ_loop);
+
+ double fac = 1.0;
+ for (d = succ_depth; d < pred_depth; ++d) {
+ fac *= inv_loop_weight;
+ }
+ sum += fac;
}
return cur/sum;
static void exec_freq_node_info(void *ctx, FILE *f, const ir_node *irn)
{
- if(is_Block(irn)) {
- ir_exec_freq *ef = ctx;
- fprintf(f, "execution frequency: %g/%lu\n", get_block_execfreq(ef, irn), get_block_execfreq_ulong(ef, irn));
- }
+ ir_exec_freq *ef = (ir_exec_freq*) ctx;
+ if (!is_Block(irn))
+ return;
+
+ fprintf(f, "execution frequency: %g/%lu\n", get_block_execfreq(ef, irn), get_block_execfreq_ulong(ef, irn));
}
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);
+ ir_exec_freq *execfreq = XMALLOCZ(ir_exec_freq);
+ execfreq->freqs = new_set(cmp_freq, 32);
memset(&execfreq->hook, 0, sizeof(execfreq->hook));
+
+ // set reasonable values to convert double execfreq to ulong execfreq
+ execfreq->m = 1.0;
+
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(ir_exec_freq *execfreq, const ir_node *block, double freq)
{
- freq_t *f = set_insert_freq(execfreq->set, block);
+ freq_t *f = set_insert_freq(execfreq->freqs, block);
f->freq = freq;
}
-ir_exec_freq *
-compute_execfreq(ir_graph * irg, double loop_weight)
+static void collect_blocks(ir_node *bl, void *data)
+{
+ set *freqs = (set*) data;
+ set_insert_freq(freqs, bl);
+}
+
+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;
- ir_exec_freq *ef;
+ gs_matrix_t *mat;
+ int size;
+ int n_keepalives;
+ int idx;
+ freq_t *freq, *s, *e;
+ ir_exec_freq *ef;
+ ir_node *end = get_irg_end(irg);
set *freqs;
-#ifdef USE_GSL
- gsl_vector *x;
-#else
+ dfs_t *dfs;
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);
+ double norm;
+
+ /*
+ * compute a DFS.
+ * using a toposort on the CFG (without back edges) will propagate
+ * the values better for the gauss/seidel iteration.
+ * => they can "flow" from start to end.
+ */
+ dfs = dfs_new(&absgraph_irg_cfg_succ, irg);
+ ef = XMALLOCZ(ir_exec_freq);
+ ef->min_non_zero = HUGE_VAL; /* initialize with a reasonable large number. */
+ freqs = ef->freqs = new_set(cmp_freq, dfs_get_n_nodes(dfs));
+
+ /*
+ * Populate the exec freq set.
+ * The DFS cannot be used alone, since the CFG might not be connected
+ * due to unreachable code.
+ */
+ irg_block_walk_graph(irg, collect_blocks, NULL, freqs);
construct_cf_backedges(irg);
edges_assure(irg);
- wd.idx = 0;
- wd.set = freqs;
+ size = dfs_get_n_nodes(dfs);
+ mat = gs_new_matrix(size, size);
+ x = XMALLOCN(double, size);
- irg_block_walk_graph(irg, block_walker, NULL, &wd);
+ for (idx = dfs_get_n_nodes(dfs) - 1; idx >= 0; --idx) {
+ ir_node *bb = (ir_node *) dfs_get_post_num_node(dfs, size - idx - 1);
+ int i;
- 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));
+ freq = set_insert_freq(freqs, bb);
+ freq->idx = idx;
- set_foreach(freqs, freq) {
- ir_node *bb = (ir_node *)freq->irn;
- size_t idx = (int)get_irn_link(bb);
+ /* Sum of (execution frequency of predecessor * probability of cf edge) ... */
+ for (i = get_Block_n_cfgpreds(bb) - 1; i >= 0; --i) {
+ ir_node *pred = get_Block_cfgpred_block(bb, i);
+ int pred_idx = size - dfs_get_post_num(dfs, pred) - 1;
- matrix[idx * (size + 1)] = -1.0;
-
- if (bb == get_irg_start_block(irg)) {
- rhs[(int)get_irn_link(bb)] = -1.0;
- continue;
+ gs_matrix_set(mat, idx, pred_idx, get_cf_probability(bb, i, loop_weight));
}
+ /* ... equals my execution frequency */
+ gs_matrix_set(mat, idx, idx, -1.0);
+ }
- 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);
+ dfs_free(dfs);
+
+ /*
+ * Add an edge from end to start.
+ * The problem is then an eigenvalue problem:
+ * Solve A*x = 1*x => (A-I)x = 0
+ */
+ s = set_find_freq(freqs, get_irg_start_block(irg));
+
+ e = set_find_freq(freqs, get_irg_end_block(irg));
+ if (e->idx >= 0)
+ gs_matrix_set(mat, s->idx, e->idx, 1.0);
+
+ /*
+ * Also add an edge for each kept block to start.
+ *
+ * This avoid strange results for e.g. an irg containing a exit()-call
+ * which block has no cfg successor.
+ */
+ n_keepalives = get_End_n_keepalives(end);
+ for (idx = n_keepalives - 1; idx >= 0; --idx) {
+ ir_node *keep = get_End_keepalive(end, idx);
+
+ if (is_Block(keep) && get_Block_n_cfg_outs(keep) == 0) {
+ freq_t *k = set_find_freq(freqs, keep);
+ if (k->idx >= 0)
+ gs_matrix_set(mat, s->idx, k->idx, 1.0);
}
}
- x = solve_lgs(matrix, rhs, size);
- if(x == NULL) {
- ef->infeasible = 1;
- return ef;
- }
+ /* solve the system and delete the matrix */
+ solve_lgs(mat, x, size);
+ gs_delete_matrix(mat);
- ef->max = 0.0;
+ /*
+ * compute the normalization factor.
+ * 1.0 / exec freq of start block.
+ */
+ norm = x[s->idx] != 0.0 ? 1.0 / x[s->idx] : 1.0;
- set_foreach(freqs, freq) {
- const ir_node *bb = freq->irn;
- size_t idx = PTR_TO_INT(get_irn_link(bb));
+ ef->max = 0.0;
+ set_foreach(freqs, freq_t*, freq) {
+ idx = freq->idx;
-#ifdef USE_GSL
- freq->freq = UNDEF(gsl_vector_get(x, idx)) ? EPSILON : gsl_vector_get(x, idx);
-#else
- freq->freq = UNDEF(x[idx]) ? EPSILON : x[idx];
-#endif
+ /* take abs because it sometimes can be -0 in case of endless loops */
+ freq->freq = fabs(x[idx]) * norm;
/* get the maximum exec freq */
ef->max = MAX(ef->max, freq->freq);
/* Get the minimum non-zero execution frequency. */
- if(freq->freq > 0.0)
+ 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 */
+ /* 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 l1 = 1.0;
double h1 = MAX_INT_FREQ;
- double *fs = malloc(set_count(freqs) * sizeof(fs[0]));
+ double *fs = (double*) malloc(set_count(freqs) * sizeof(fs[0]));
int i, j, n = 0;
- set_foreach(freqs, freq)
+ set_foreach(freqs, freq_t*, 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)
+ for (i = 0; i < n; ++i) {
+ if (fs[i] <= 0.0)
continue;
- for(j = i + 1; j < n; ++j) {
+ for (j = i + 1; j < n; ++j) {
double diff = fabs(fs[i] - fs[j]);
- if(!UNDEF(diff))
+ if (!UNDEF(diff))
smallest_diff = MIN(diff, 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) {
+ 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);
-#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);
+ xfree(x);
+
return ef;
}
-void
-free_execfreq(ir_exec_freq *ef)
+void free_execfreq(ir_exec_freq *ef)
{
- del_set(ef->set);
+ del_set(ef->freqs);
unregister_hook(hook_node_info, &ef->hook);
free(ef);
}
-
-#undef ELEM
projects / libfirm / blobdiff