/*
- * 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.
*
* @date 28.05.2006
* @version $Id$
*/
-#ifdef HAVE_CONFIG_H
#include "config.h"
-#endif
#include <stdio.h>
#include <string.h>
#include "gaussseidel.h"
-#include "firm_common_t.h"
#include "set.h"
#include "hashptr.h"
#include "debug.h"
#include "irloop.h"
#include "irgwalk.h"
#include "iredges.h"
+#include "irouts.h"
#include "irprintf.h"
#include "irtools.h"
#include "irhooks.h"
query.irn = irn;
query.freq = 0.0;
+ query.idx = -1;
return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
}
#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();
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;
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);
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)
{
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;
* => 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);
freq = set_insert_freq(freqs, bb);
freq->idx = idx;
- gs_matrix_set(mat, idx, idx, -1.0);
+ /* 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));
- gs_matrix_set(mat, s->idx, e->idx, 1.0);
+ 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);
* 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);
projects / libfirm / blobdiff