more debug info

[libfirm] / ir / ana / execfreq.c

diff --git a/ir/ana/execfreq.c b/ir/ana/execfreq.c
index ff9ef3e..41909a6 100644 (file)
--- a/ir/ana/execfreq.c
+++ b/ir/ana/execfreq.c
@@ -1,5 +1,5 @@
 /*
- * 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.
  *
@@ -112,6 +112,7 @@ set_insert_freq(set * set, const ir_node * irn)
 
        query.irn = irn;
        query.freq = 0.0;
+       query.idx  = -1;
        return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
 }
 
@@ -162,10 +163,9 @@ 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();
@@ -212,8 +212,7 @@ 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));
@@ -231,6 +230,12 @@ void set_execfreq(ir_exec_freq *execfreq, const ir_node *block, double freq)
        f->freq = freq;
 }
 
+static void collect_blocks(ir_node *bl, void *data)
+{
+       set *freqs = data;
+       set_insert_freq(freqs, bl);
+}
+
 ir_exec_freq *
 compute_execfreq(ir_graph * irg, double loop_weight)
 {
@@ -251,21 +256,23 @@ 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, 32);
+       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);
-       /* 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); */
+       edges_assure(irg);
 
        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);
@@ -275,15 +282,17 @@ compute_execfreq(ir_graph * irg, double loop_weight)
                freq = set_insert_freq(freqs, bb);
                freq->idx = idx;
 
-               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);
                        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));
                }
+               gs_matrix_set(mat, idx, idx, -1.0);
        }
 
+       dfs_free(dfs);
+
        /*
         * Add a loop from end to start.
         * The problem is then an eigenvalue problem:
@@ -291,7 +300,8 @@ compute_execfreq(ir_graph * irg, double loop_weight)
         */
        s = set_find_freq(freqs, get_irg_start_block(irg));
        e = set_find_freq(freqs, get_irg_end_block(irg));
-       gs_matrix_set(mat, s->idx, e->idx, 1.0);
+       if (e->idx >= 0)
+               gs_matrix_set(mat, s->idx, e->idx, 1.0);
 
        /* solve the system and delete the matrix */
        solve_lgs(mat, x, size);
@@ -301,16 +311,14 @@ compute_execfreq(ir_graph * irg, double loop_weight)
         * compute the normalization factor.
         * 1.0 / exec freq of start block.
         */
-       assert(x[s->idx] > 0.0);
-       norm = 1.0 / x[s->idx];
+       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;
 
-               /* freq->freq = UNDEF(x[idx]) ? EPSILON : x[idx]; */
-               /* TODO: Do we need the check for zero? */
-               freq->freq = x[idx] * norm;
+               /* 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);