X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fana%2Fexecfreq.c;h=83c9e4c5020d6cff6b77b7294f7525c08a79794f;hb=a1e9069afa4fa1e16e2d176bcd7905d6a1ed4677;hp=fa2960bbf0df506c89a6971aacb23793a48f8c3b;hpb=1ddb7bc5b6bb20093f85cc0e3b501583a0a46667;p=libfirm diff --git a/ir/ana/execfreq.c b/ir/ana/execfreq.c index fa2960bbf..83c9e4c50 100644 --- a/ir/ana/execfreq.c +++ b/ir/ana/execfreq.c @@ -1,249 +1,434 @@ /* - * 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 +/** + * @file + * @brief Compute an estimate of basic block executions. + * @author Adam M. Szalkowski + * @date 28.05.2006 + * @version $Id$ + */ #include "config.h" -#endif - -//#define USE_GSL #include #include +#include +#include -#ifdef USE_GSL -#include -#include -#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 "irnode_t.h" #include "irloop.h" #include "irgwalk.h" +#include "iredges.h" #include "irouts.h" #include "irprintf.h" +#include "irtools.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 -typedef struct _walkerdata_t { - set *set; - size_t idx; -} walkerdata_t; +#ifdef COMPARE_AGAINST_GAUSSJORDAN +#include "gaussjordan.h" +#endif -static int -cmp_freq(const void *a, const void *b, size_t size) -{ - const freq_t *p = a; - const freq_t *q = b; - return !(p->irn == q->irn); -} +#define EPSILON 1e-5 +#define UNDEF(x) (fabs(x) < EPSILON) +#define SEIDEL_TOLERANCE 1e-7 -static freq_t * -set_find_freq(set * set, const ir_node * irn) -{ - freq_t query; +#define MAX_INT_FREQ 1000000 - query.irn = irn; - return set_find(set, &query, sizeof(query), HASH_PTR(irn)); -} +#define set_foreach(s,i) for ((i)=set_first((s)); (i); (i)=set_next((s))) -static freq_t * -set_insert_freq(set * set, const ir_node * irn) -{ - freq_t query; +typedef struct freq_t { + const ir_node *irn; + int idx; + double freq; +} freq_t; - query.irn = irn; - query.freq = 0.0; - return set_insert(set, &query, sizeof(query), HASH_PTR(irn)); -} +struct ir_exec_freq { + set *set; + hook_entry_t hook; + double max; + double min_non_zero; + double m, b; + unsigned infeasible : 1; +}; -double -get_block_execfreq(set * freqs, const ir_node * irn) +static int cmp_freq(const void *a, const void *b, size_t size) { - assert(is_Block(irn)); + const freq_t *p = a; + const freq_t *q = b; + (void) size; - freq_t *freq = set_find_freq(freqs, irn); - assert(freq); - - return freq->freq; + return !(p->irn == q->irn); } -#define ZERO(x) (((x) > 0) ? ((x) < 0.0001) : ((x) > -0.0001)) - -static void -block_walker(ir_node * bb, void * data) +static freq_t *set_find_freq(set *set, const ir_node *irn) { - walkerdata_t *wd = data; + freq_t query; - set_insert_freq(wd->set, bb); - set_irn_link(bb, (void*)wd->idx++); + query.irn = irn; + return set_find(set, &query, sizeof(query), HASH_PTR(irn)); } -#ifdef USE_GSL -static gsl_vector * -solve_lgs(double * a_data, double * b_data, size_t size) +static freq_t *set_insert_freq(set *set, const ir_node *irn) { - 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); - - gsl_linalg_LU_decomp (&m.matrix, p, &s); - - gsl_linalg_LU_solve (&m.matrix, p, &b.vector, x); + freq_t query; - gsl_permutation_free (p); - - return x; + query.irn = irn; + query.freq = 0.0; + query.idx = -1; + return set_insert(set, &query, sizeof(query), HASH_PTR(irn)); } -#else -static double * -solve_lgs(double * A, double * b, size_t size) + +double get_block_execfreq(const ir_exec_freq *ef, const ir_node *irn) { - if(firm_gaussjordansolve(A,b,size) == 0) { - return b; - } else { - return NULL; - } + 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; } -#endif -static double -get_cf_probability(const ir_node * bb, int pos) +unsigned long +get_block_execfreq_ulong(const ir_exec_freq *ef, const ir_node *bb) { -#define LOOP_WEIGHT 9.0 - - double sum = 0.0; - double cur = 0.0; - int i, - n; - ir_node *pred = get_Block_cfgpred_block(bb, pos); - - if(get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred))) { - cur = 1.0; - } else { - cur = LOOP_WEIGHT; - } - - for(i = 0, n = get_Block_n_cfg_outs(pred); i < n; ++i) { - ir_node *succ = get_Block_cfg_out(pred, i); - - if(get_loop_depth(get_irn_loop(succ)) < get_loop_depth(get_irn_loop(pred))) { - sum += 1.0; - } else { - sum += LOOP_WEIGHT; - } - } - - return cur/sum; + double f = get_block_execfreq(ef, bb); + int res = (int) (f > ef->min_non_zero ? ef->m * f + ef->b : 1.0); + return res; } -set * -compute_execfreq(ir_graph * irg) +static double *solve_lgs(gs_matrix_t *mat, double *x, int size) { - set *freqs = new_set(cmp_freq, 32); - size_t size; - double *matrix; - double *rhs; - size_t i = 0; - freq_t *freq; - walkerdata_t wd; -#ifdef USE_GSL - gsl_vector *x; -#else - double *x; + 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); + + gs_matrix_export(mat, nw, size); + + stat_ev_tim_push(); + firm_gaussjordansolve(nw, nx, size); + stat_ev_tim_pop("execfreq_jordan_time"); + + xfree(nw); + xfree(nx); + } #endif - construct_cf_backedges(irg); - - wd.idx = 0; - wd.set = freqs; - - irg_block_walk_graph(irg, block_walker, NULL, &wd); - - size = set_count(freqs); - matrix = malloc(size*size*sizeof(*matrix)); - memset(matrix, 0, size*size*sizeof(*matrix)); - rhs = malloc(size*sizeof(*rhs)); - memset(rhs, 0, size*sizeof(*rhs)); - - set_foreach(freqs, freq) { - const ir_node *bb = freq->irn; - size_t idx = (int)get_irn_link(bb); + return x; +} - matrix[idx*(size+1)] = -1.0; +/* + * 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 = 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; + + 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 *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; +} - if(bb == get_irg_start_block(irg)) { - rhs[(int)get_irn_link(bb)] = -1.0; - continue; - } +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)); + } +} - for(i = 0; i < get_Block_n_cfgpreds(bb); ++i) { - ir_node *pred = get_Block_cfgpred_block(bb, i); - size_t pred_idx = (int)get_irn_link(pred); +ir_exec_freq *create_execfreq(ir_graph *irg) +{ + ir_exec_freq *execfreq = XMALLOCZ(ir_exec_freq); + execfreq->set = new_set(cmp_freq, 32); -// matrix[pred_idx + idx*size] += 1.0/(double)get_Block_n_cfg_outs(pred); - matrix[pred_idx + idx*size] += get_cf_probability(bb, i); - } - } + memset(&execfreq->hook, 0, sizeof(execfreq->hook)); - x = solve_lgs(matrix, rhs, size); - if(x == NULL) { - del_set(freqs); - return NULL; - } + // set reasonable values to convert double execfreq to ulong execfreq + execfreq->m = 1.0; - set_foreach(freqs, freq) { - const ir_node *bb = freq->irn; - size_t idx = PTR_TO_INT(get_irn_link(bb)); + execfreq->hook.context = execfreq; + execfreq->hook.hook._hook_node_info = exec_freq_node_info; + register_hook(hook_node_info, &execfreq->hook); + (void) irg; -#ifdef USE_GSL - freq->freq = ZERO(gsl_vector_get(x, idx)) ? 0.0 : gsl_vector_get(x, idx); -#else - freq->freq = ZERO(x[idx]) ? 0.0 : x[idx]; -#endif - ir_fprintf(stderr, "execfreq %+F: %f\n", bb, freq->freq); - } + return execfreq; +} -#ifdef USE_GSL - gsl_vector_free(x); -#endif - free(matrix); +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; +} - return freqs; +static void collect_blocks(ir_node *bl, void *data) +{ + set *freqs = data; + set_insert_freq(freqs, bl); } -void -free_execfreq(set * freqs) +ir_exec_freq *compute_execfreq(ir_graph *irg, double loop_weight) { - if(freqs) del_set(freqs); + 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; + dfs_t *dfs; + double *x; + 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->set = 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); + + size = dfs_get_n_nodes(dfs); + mat = gs_new_matrix(size, size); + x = XMALLOCN(double, size); + + 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); + freq_t *freq; + int i; + + freq = set_insert_freq(freqs, bb); + freq->idx = idx; + + /* 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; + + 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); + } + + 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); + } + } + + /* solve the system and delete the matrix */ + solve_lgs(mat, x, size); + gs_delete_matrix(mat); + + /* + * compute the normalization factor. + * 1.0 / exec freq of start block. + */ + norm = x[s->idx] != 0.0 ? 1.0 / x[s->idx] : 1.0; + + ef->max = 0.0; + set_foreach(freqs, freq) { + int idx = freq->idx; + + /* 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) + 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; + } + + free(fs); + } + + 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; } -#undef ELEM +void free_execfreq(ir_exec_freq *ef) +{ + del_set(ef->set); + unregister_hook(hook_node_info, &ef->hook); + free(ef); +}