X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=sidebyside;f=ir%2Fana%2Fexecfreq.c;h=2e7a136a6333e3e0fe06890502e29a4d997fb32b;hb=2af4a97900b435e35d7c20350604e3863aff2b4c;hp=308a11d869b1f177362825f775153615197c442f;hpb=e07c0dc0ce91e9d3d97963472b06569b8a73af86;p=libfirm diff --git a/ir/ana/execfreq.c b/ir/ana/execfreq.c index 308a11d86..2e7a136a6 100644 --- a/ir/ana/execfreq.c +++ b/ir/ana/execfreq.c @@ -24,9 +24,7 @@ * @date 28.05.2006 * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include #include @@ -35,7 +33,6 @@ #include "gaussseidel.h" -#include "firm_common_t.h" #include "set.h" #include "hashptr.h" #include "debug.h" @@ -49,6 +46,7 @@ #include "irloop.h" #include "irgwalk.h" #include "iredges.h" +#include "irouts.h" #include "irprintf.h" #include "irtools.h" #include "irhooks.h" @@ -163,10 +161,9 @@ solve_lgs(gs_matrix_t *mat, double *x, int size) #ifdef COMPARE_AGAINST_GAUSSJORDAN { - double *nw = xmalloc(size * size * sizeof(*nw)); - double *nx = xmalloc(size * sizeof(*nx)); + double *nw = XMALLOCN(double, size * size); + double *nx = XMALLOCNZ(double, size); - memset(nx, 0, size * sizeof(*nx)); gs_matrix_export(mat, nw, size); stat_ev_tim_push(); @@ -181,23 +178,45 @@ solve_lgs(gs_matrix_t *mat, double *x, int size) return x; } +/* + * 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; @@ -213,11 +232,14 @@ static void exec_freq_node_info(void *ctx, FILE *f, const ir_node *irn) ir_exec_freq *create_execfreq(ir_graph *irg) { - ir_exec_freq *execfreq = xmalloc(sizeof(execfreq[0])); - memset(execfreq, 0, sizeof(execfreq[0])); + ir_exec_freq *execfreq = XMALLOCZ(ir_exec_freq); execfreq->set = 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); @@ -243,9 +265,11 @@ compute_execfreq(ir_graph * irg, double loop_weight) { 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; @@ -258,8 +282,7 @@ compute_execfreq(ir_graph * irg, double loop_weight) * => 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 = 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)); @@ -275,7 +298,7 @@ compute_execfreq(ir_graph * irg, double loop_weight) size = dfs_get_n_nodes(dfs); mat = gs_new_matrix(size, size); - x = xmalloc(size*sizeof(*x)); + 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); @@ -285,27 +308,47 @@ compute_execfreq(ir_graph * irg, double loop_weight) 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 a loop from end to start. + * 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);