/*
- * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @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
-
-#undef USE_GSL
#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 "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 "irnodehashmap.h"
-#include "execfreq.h"
+#include "execfreq_t.h"
-#define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
+#define EPSILON 1e-5
+#define UNDEF(x) (fabs(x) < EPSILON)
+#define SEIDEL_TOLERANCE 1e-7
#define MAX_INT_FREQ 1000000
-typedef struct _freq_t {
- const ir_node *irn;
- double freq;
-} freq_t;
-
-
-typedef struct _walkerdata_t {
- set *set;
- size_t idx;
-} walkerdata_t;
-
-struct ir_exec_freq {
- set *set;
- hook_entry_t hook;
- double max;
- double min_non_zero;
- double m, b;
- unsigned infeasible : 1;
-};
+static hook_entry_t hook;
-static int
-cmp_freq(const void *a, const void *b, size_t size)
+double get_block_execfreq(const ir_node *block)
{
- const freq_t *p = a;
- const freq_t *q = b;
- (void) size;
-
- return !(p->irn == q->irn);
+ return block->attr.block.execfreq;
}
-static freq_t *
-set_find_freq(set * set, const ir_node * irn)
+void set_block_execfreq(ir_node *block, double newfreq)
{
- freq_t query;
-
- query.irn = irn;
- return set_find(set, &query, sizeof(query), HASH_PTR(irn));
+ block->attr.block.execfreq = newfreq;
}
-static freq_t *
-set_insert_freq(set * set, const ir_node * irn)
+static void exec_freq_node_info(void *ctx, FILE *f, const ir_node *irn)
{
- freq_t query;
-
- query.irn = irn;
- query.freq = 0.0;
- return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
+ (void)ctx;
+ if (!is_Block(irn))
+ return;
+ double freq = get_block_execfreq(irn);
+ if (freq != 0.0)
+ fprintf(f, "execution frequency: %g\n", freq);
}
-double
-get_block_execfreq(const ir_exec_freq *ef, const ir_node * irn)
+void init_execfreq(void)
{
- if(!ef->infeasible) {
- set *freqs = ef->set;
- freq_t *freq;
- assert(is_Block(irn));
- freq = set_find_freq(freqs, irn);
- assert(freq);
-
- assert(freq->freq >= 0);
- return freq->freq;
- }
-
- return 1.0;
+ memset(&hook, 0, sizeof(hook));
+ hook.hook._hook_node_info = exec_freq_node_info;
+ register_hook(hook_node_info, &hook);
}
-unsigned long
-get_block_execfreq_ulong(const ir_exec_freq *ef, const ir_node *bb)
+void exit_execfreq(void)
{
- 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;
+ unregister_hook(hook_node_info, &hook);
}
-#define EPSILON 0.0001
-#define UNDEF(x) !(x > EPSILON)
-static void
-block_walker(ir_node * bb, void * data)
+static double *solve_lgs(gs_matrix_t *mat, double *x, int size)
{
- walkerdata_t *wd = data;
-
- set_insert_freq(wd->set, bb);
- set_irn_link(bb, (void*)wd->idx++);
+ 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);
+
+ return x;
}
-#ifdef USE_GSL
-static gsl_vector *
-solve_lgs(double * a_data, double * b_data, size_t size)
+/*
+ * Determine probability that predecessor pos takes this cf edge.
+ */
+static double get_cf_probability(const ir_node *bb, int pos, double loop_weight)
{
- 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;
+ 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_loop *loop;
+ int depth;
+ int d;
+
+ if (is_Bad(pred))
+ return 0;
+
+ 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;
+ }
- gsl_permutation * p = gsl_permutation_alloc (size);
+ foreach_block_succ(pred, edge) {
+ 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);
- gsl_linalg_LU_decomp (&m.matrix, p, &s);
+ double fac = 1.0;
+ for (d = succ_depth; d < pred_depth; ++d) {
+ fac *= inv_loop_weight;
+ }
+ sum += fac;
+ }
- gsl_linalg_LU_solve (&m.matrix, p, &b.vector, x);
+ return cur/sum;
+}
- gsl_permutation_free (p);
+static double *freqs;
+static double min_non_zero;
+static double max_freq;
- return x;
-}
-#else
-static double *
-solve_lgs(double * A, double * b, size_t size)
+static void collect_freqs(ir_node *node, void *data)
{
- if(firm_gaussjordansolve(A,b,size) == 0) {
- return b;
- } else {
- return NULL;
- }
+ (void) data;
+ double freq = get_block_execfreq(node);
+ if (freq > max_freq)
+ max_freq = freq;
+ if (freq > 0.0 && freq < min_non_zero)
+ min_non_zero = freq;
+ ARR_APP1(double, freqs, freq);
}
-#endif /* USE_GSL */
-static double
-get_cf_probability(ir_node *bb, int pos, double loop_weight)
+void ir_calculate_execfreq_int_factors(ir_execfreq_int_factors *factors,
+ ir_graph *irg)
{
- 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;
+ /* compute m and b of the transformation used to convert the doubles into
+ * scaled ints */
+ freqs = NEW_ARR_F(double, 0);
+ min_non_zero = HUGE_VAL;
+ max_freq = 0.0;
+ irg_block_walk_graph(irg, collect_freqs, NULL, NULL);
+
+ /*
+ * find the smallest difference of the execution frequencies
+ * we try to ressolve it with 1 integer.
+ */
+ size_t n_freqs = ARR_LEN(freqs);
+ double smallest_diff = 1.0;
+ for (size_t i = 0; i < n_freqs; ++i) {
+ if (freqs[i] <= 0.0)
+ continue;
- cur = get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
+ for (size_t j = i + 1; j < n_freqs; ++j) {
+ double diff = fabs(freqs[i] - freqs[j]);
- 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;
+ if (!UNDEF(diff))
+ smallest_diff = MIN(diff, smallest_diff);
+ }
}
- 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));
+ double l2 = min_non_zero;
+ double h2 = max_freq;
+ double l1 = 1.0;
+ double h1 = MAX_INT_FREQ;
+
+ /* according to that the slope of the translation function is
+ * 1.0 / smallest_diff */
+ factors->m = 1.0 / smallest_diff;
+
+ /* the abscissa is then given by */
+ factors->b = l1 - factors->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 (factors->m * h2 + factors->b > MAX_INT_FREQ) {
+ factors->m = (h1 - l1) / (h2 - l2);
+ factors->b = l1 - factors->m * l2;
}
-}
-
-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);
- (void) irg;
-
- return execfreq;
+ DEL_ARR_F(freqs);
}
-void set_execfreq(ir_exec_freq *execfreq, const ir_node *block, double freq)
+int get_block_execfreq_int(const ir_execfreq_int_factors *factors,
+ const ir_node *block)
{
- freq_t *f = set_insert_freq(execfreq->set, block);
- f->freq = freq;
+ double f = get_block_execfreq(block);
+ int res = (int) (f > factors->min_non_zero ? factors->m * f + factors->b : 1.0);
+ return res;
}
-ir_exec_freq *
-compute_execfreq(ir_graph * irg, double loop_weight)
+void ir_estimate_execfreq(ir_graph *irg)
{
- 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;
-#else
- 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);
-
- 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;
-
- 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));
-
- 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;
-
- if (bb == get_irg_start_block(irg)) {
- rhs[(int)get_irn_link(bb)] = -1.0;
- continue;
+ double loop_weight = 10.0;
+
+ assure_irg_properties(irg,
+ IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
+ | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO
+ | IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE);
+
+ /* 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_t *dfs = dfs_new(&absgraph_irg_cfg_succ, irg);
+
+ int size = dfs_get_n_nodes(dfs);
+ gs_matrix_t *mat = gs_new_matrix(size, size);
+
+ ir_node *end_block = get_irg_end_block(irg);
+
+ for (int idx = size - 1; idx >= 0; --idx) {
+ const ir_node *bb = (ir_node*)dfs_get_post_num_node(dfs, size-idx-1);
+
+ /* Sum of (execution frequency of predecessor * probability of cf edge) ... */
+ for (int i = get_Block_n_cfgpreds(bb) - 1; i >= 0; --i) {
+ const ir_node *pred = get_Block_cfgpred_block(bb, i);
+ int pred_idx = size - dfs_get_post_num(dfs, pred)-1;
+ double cf_probability = get_cf_probability(bb, i, loop_weight);
+ gs_matrix_set(mat, idx, pred_idx, cf_probability);
}
-
- 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);
+ /* ... equals my execution frequency */
+ gs_matrix_set(mat, idx, idx, -1.0);
+
+ /* Add an edge from end to start.
+ * The problem is then an eigenvalue problem:
+ * Solve A*x = 1*x => (A-I)x = 0
+ */
+ if (bb == end_block) {
+ const ir_node *start_block = get_irg_start_block(irg);
+ int s_idx = size - dfs_get_post_num(dfs, start_block)-1;
+ gs_matrix_set(mat, s_idx, idx, 1.0);
}
}
- x = solve_lgs(matrix, rhs, size);
- if (x == NULL) {
- DEBUG_ONLY(ir_fprintf(stderr, "Debug Warning: Couldn't estimate execution frequencies for %+F\n", irg));
- ef->infeasible = 1;
- } else {
- ef->max = 0.0;
-
- 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 = UNDEF(gsl_vector_get(x, idx)) ? EPSILON : gsl_vector_get(x, idx);
-#else
- freq->freq = UNDEF(x[idx]) ? EPSILON : x[idx];
-#endif
-
- /* 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);
- }
+ /*
+ * 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.
+ */
+ ir_node *start_block = get_irg_start_block(irg);
+ int s_idx = size - dfs_get_post_num(dfs, start_block)-1;
+ const ir_node *end = get_irg_end(irg);
+ int n_keepalives = get_End_n_keepalives(end);
+ for (int idx = n_keepalives - 1; idx >= 0; --idx) {
+ ir_node *keep = get_End_keepalive(end, idx);
+ if (!is_Block(keep) || get_irn_n_edges_kind(keep, EDGE_KIND_BLOCK) > 0)
+ continue;
-#ifdef USE_GSL
- gsl_vector_free(x);
-#endif
- 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);
+ int k_idx = size-dfs_get_post_num(dfs, keep)-1;
+ if (k_idx > 0)
+ gs_matrix_set(mat, s_idx, k_idx, 1.0);
}
- free(matrix);
- free(rhs);
+ /* solve the system and delete the matrix */
+ double *x = XMALLOCN(double, size);
+ solve_lgs(mat, x, size);
+ gs_delete_matrix(mat);
+
+ /* compute the normalization factor.
+ * 1.0 / exec freq of start block.
+ * (note: start_idx is != 0 in strange cases involving endless loops,
+ * probably a misfeature/bug)
+ */
+ int start_idx = size-dfs_get_post_num(dfs, get_irg_start_block(irg))-1;
+ double start_freq = x[start_idx];
+ double norm = start_freq != 0.0 ? 1.0 / start_freq : 1.0;
+
+ for (int idx = size - 1; idx >= 0; --idx) {
+ ir_node *bb = (ir_node *) dfs_get_post_num_node(dfs, size - idx - 1);
+
+ /* take abs because it sometimes can be -0 in case of endless loops */
+ double freq = fabs(x[idx]) * norm;
+ set_block_execfreq(bb, freq);
+ }
- return ef;
-}
+ dfs_free(dfs);
-void
-free_execfreq(ir_exec_freq *ef)
-{
- del_set(ef->set);
- unregister_hook(hook_node_info, &ef->hook);
- free(ef);
+ xfree(x);
}