X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fana%2Fexecfreq.c;h=7950e5c5eec2f2eada56640a438d7e84b9ff709c;hb=e97a86a5e90141c2cb4ccdc34195cab035627c3f;hp=308a11d869b1f177362825f775153615197c442f;hpb=e07c0dc0ce91e9d3d97963472b06569b8a73af86;p=libfirm diff --git a/ir/ana/execfreq.c b/ir/ana/execfreq.c index 308a11d86..7950e5c5e 100644 --- a/ir/ana/execfreq.c +++ b/ir/ana/execfreq.c @@ -1,20 +1,6 @@ /* - * 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. + * Copyright (C) 2012 University of Karlsruhe. */ /** @@ -22,11 +8,8 @@ * @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 #include @@ -35,11 +18,10 @@ #include "gaussseidel.h" -#include "firm_common_t.h" #include "set.h" #include "hashptr.h" #include "debug.h" -#include "statev.h" +#include "statev_t.h" #include "dfs_t.h" #include "absgraph.h" @@ -49,100 +31,56 @@ #include "irloop.h" #include "irgwalk.h" #include "iredges.h" +#include "irouts.h" #include "irprintf.h" -#include "irtools.h" +#include "util.h" #include "irhooks.h" +#include "irnodehashmap.h" -#include "execfreq.h" +#include "execfreq_t.h" -/* 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 EPSILON 1e-5 #define UNDEF(x) (fabs(x) < EPSILON) #define SEIDEL_TOLERANCE 1e-7 #define MAX_INT_FREQ 1000000 -#define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s))) - -typedef struct _freq_t { - const ir_node *irn; - int idx; - double freq; -} 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; - (void) size; +static hook_entry_t hook; - return !(p->irn == q->irn); +double get_block_execfreq(const ir_node *block) +{ + 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; - query.idx = -1; - 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); - return res; + unregister_hook(hook_node_info, &hook); } -static double * -solve_lgs(gs_matrix_t *mat, double *x, int size) + +static double *solve_lgs(gs_matrix_t *mat, double *x, int size) { double init = 1.0 / size; double dev; @@ -157,243 +95,218 @@ solve_lgs(gs_matrix_t *mat, double *x, int size) do { ++iter; dev = gs_matrix_gauss_seidel(mat, x, size); - } while(fabs(dev) > SEIDEL_TOLERANCE); + } 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 = xmalloc(size * size * sizeof(*nw)); - double *nx = xmalloc(size * sizeof(*nx)); - - memset(nx, 0, size * sizeof(*nx)); - 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 - return x; } -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(const ir_node *bb, int pos, double loop_weight) { - double sum = 0.0; - double cur = 0.0; + 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 = 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; + 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; + } 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) +static double *freqs; +static double min_non_zero; +static double max_freq; + +static void collect_freqs(ir_node *node, void *data) { - 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)); - } + (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); } -ir_exec_freq *create_execfreq(ir_graph *irg) +void ir_calculate_execfreq_int_factors(ir_execfreq_int_factors *factors, + ir_graph *irg) { - ir_exec_freq *execfreq = xmalloc(sizeof(execfreq[0])); - memset(execfreq, 0, sizeof(execfreq[0])); - execfreq->set = new_set(cmp_freq, 32); + /* 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); - 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; + /* + * 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; - return execfreq; -} + for (size_t j = i + 1; j < n_freqs; ++j) { + double diff = fabs(freqs[i] - freqs[j]); -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; + if (!UNDEF(diff)) + smallest_diff = MIN(diff, smallest_diff); + } + } + + 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; + } + + DEL_ARR_F(freqs); } -static void collect_blocks(ir_node *bl, void *data) +int get_block_execfreq_int(const ir_execfreq_int_factors *factors, + const ir_node *block) { - set *freqs = data; - set_insert_freq(freqs, bl); + 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) { - gs_matrix_t *mat; - int size; - int idx; - freq_t *freq, *s, *e; - ir_exec_freq *ef; - set *freqs; - dfs_t *dfs; - double *x; - double norm; + double loop_weight = 10.0; - /* - * compute a DFS. + 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 = dfs_new(&absgraph_irg_cfg_succ, irg); - ef = xmalloc(sizeof(ef[0])); - memset(ef, 0, sizeof(ef[0])); - 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); + dfs_t *dfs = dfs_new(&absgraph_irg_cfg_succ, irg); - size = dfs_get_n_nodes(dfs); - mat = gs_new_matrix(size, size); - x = xmalloc(size*sizeof(*x)); + int size = dfs_get_n_nodes(dfs); + gs_matrix_t *mat = gs_new_matrix(size, 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; + ir_node *const start_block = get_irg_start_block(irg); + ir_node *const end_block = get_irg_end_block(irg); - freq = set_insert_freq(freqs, bb); - freq->idx = idx; + for (int idx = size - 1; idx >= 0; --idx) { + const ir_node *bb = (ir_node*)dfs_get_post_num_node(dfs, size-idx-1); - 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)); + /* 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); } + /* ... 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 + */ + if (bb == end_block) { + int const s_idx = size - dfs_get_post_num(dfs, start_block) - 1; + gs_matrix_set(mat, s_idx, idx, 1.0); + } + } /* - * Add a loop from end to start. - * The problem is then an eigenvalue problem: - * Solve A*x = 1*x => (A-I)x = 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. */ - 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); + 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; + + 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); + } /* 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. + /* 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) */ - norm = x[s->idx] != 0.0 ? 1.0 / x[s->idx] : 1.0; + int start_idx = size - dfs_get_post_num(dfs, start_block) - 1; + double start_freq = x[start_idx]; + double norm = start_freq != 0.0 ? 1.0 / start_freq : 1.0; - ef->max = 0.0; - set_foreach(freqs, freq) { - int idx = freq->idx; + 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 */ - 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); + double freq = fabs(x[idx]) * norm; + set_block_execfreq(bb, freq); } - 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); + dfs_free(dfs); xfree(x); - - return ef; -} - -void -free_execfreq(ir_exec_freq *ef) -{ - del_set(ef->set); - unregister_hook(hook_node_info, &ef->hook); - free(ef); }