2 * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Compute an estimate of basic block executions.
23 * @author Adam M. Szalkowski
39 #include <gsl/gsl_linalg.h>
40 #include <gsl/gsl_vector.h>
42 #include "gaussjordan.h"
45 #include "firm_common_t.h"
51 #include "irgraph_t.h"
61 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
63 #define MAX_INT_FREQ 1000000
65 typedef struct _freq_t {
71 typedef struct _walkerdata_t {
82 unsigned infeasible : 1;
86 cmp_freq(const void *a, const void *b, size_t size)
91 return !(p->irn == q->irn);
95 set_find_freq(set * set, const ir_node * irn)
100 return set_find(set, &query, sizeof(query), HASH_PTR(irn));
104 set_insert_freq(set * set, const ir_node * irn)
110 return set_insert(set, &query, sizeof(query), HASH_PTR(irn));
114 get_block_execfreq(const ir_exec_freq *ef, const ir_node * irn)
116 if(!ef->infeasible) {
117 set *freqs = ef->set;
119 assert(is_Block(irn));
120 freq = set_find_freq(freqs, irn);
123 assert(freq->freq >= 0);
131 get_block_execfreq_ulong(const ir_exec_freq *ef, const ir_node *bb)
133 double f = get_block_execfreq(ef, bb);
134 int res = (int) (f > ef->min_non_zero ? ef->m * f + ef->b : 1.0);
136 // printf("%20.6f %10d\n", f, res);
140 #define EPSILON 0.0001
141 #define UNDEF(x) !(x > EPSILON)
144 block_walker(ir_node * bb, void * data)
146 walkerdata_t *wd = data;
148 set_insert_freq(wd->set, bb);
149 set_irn_link(bb, (void*)wd->idx++);
154 solve_lgs(double * a_data, double * b_data, size_t size)
157 = gsl_matrix_view_array (a_data, size, size);
160 = gsl_vector_view_array (b_data, size);
162 gsl_vector *x = gsl_vector_alloc (size);
166 gsl_permutation * p = gsl_permutation_alloc (size);
168 gsl_linalg_LU_decomp (&m.matrix, p, &s);
170 gsl_linalg_LU_solve (&m.matrix, p, &b.vector, x);
172 gsl_permutation_free (p);
178 solve_lgs(double * A, double * b, size_t size)
180 if(firm_gaussjordansolve(A,b,size) == 0) {
189 get_cf_probability(ir_node *bb, int pos, double loop_weight)
193 const ir_node *pred = get_Block_cfgpred_block(bb, pos);
194 const ir_loop *pred_loop = get_irn_loop(pred);
195 int pred_depth = get_loop_depth(pred_loop);
196 const ir_edge_t *edge;
198 cur = get_loop_depth(get_irn_loop(bb)) < get_loop_depth(get_irn_loop(pred)) ? 1.0 : loop_weight;
200 foreach_block_succ(pred, edge) {
201 const ir_node *block = get_edge_src_irn(edge);
202 const ir_loop *loop = get_irn_loop(block);
203 int depth = get_loop_depth(loop);
204 sum += depth < pred_depth ? 1.0 : loop_weight;
210 static void exec_freq_node_info(void *ctx, FILE *f, const ir_node *irn)
213 ir_exec_freq *ef = ctx;
214 fprintf(f, "execution frequency: %g/%lu\n", get_block_execfreq(ef, irn), get_block_execfreq_ulong(ef, irn));
218 ir_exec_freq *create_execfreq(ir_graph *irg)
220 ir_exec_freq *execfreq = xmalloc(sizeof(execfreq[0]));
221 memset(execfreq, 0, sizeof(execfreq[0]));
222 execfreq->set = new_set(cmp_freq, 32);
224 memset(&execfreq->hook, 0, sizeof(execfreq->hook));
225 execfreq->hook.context = execfreq;
226 execfreq->hook.hook._hook_node_info = exec_freq_node_info;
227 register_hook(hook_node_info, &execfreq->hook);
232 void set_execfreq(ir_exec_freq *execfreq, const ir_node *block, double freq)
234 freq_t *f = set_insert_freq(execfreq->set, block);
239 compute_execfreq(ir_graph * irg, double loop_weight)
255 ef = xmalloc(sizeof(ef[0]));
256 memset(ef, 0, sizeof(ef[0]));
257 ef->min_non_zero = 1e50; /* initialize with a reasonable large number. */
258 freqs = ef->set = new_set(cmp_freq, 32);
260 construct_cf_backedges(irg);
261 /* workaround for defect edges (happens after optimisations merges blocks)*/
262 edges_deactivate(irg);
264 /* edges_assure(irg); */
269 irg_block_walk_graph(irg, block_walker, NULL, &wd);
271 size = set_count(freqs);
272 matrix = xmalloc(size*size*sizeof(*matrix));
273 memset(matrix, 0, size*size*sizeof(*matrix));
274 rhs = xmalloc(size*sizeof(*rhs));
275 memset(rhs, 0, size*sizeof(*rhs));
277 set_foreach(freqs, freq) {
278 ir_node *bb = (ir_node *)freq->irn;
279 size_t idx = (int)get_irn_link(bb);
281 matrix[idx * (size + 1)] = -1.0;
283 if (bb == get_irg_start_block(irg)) {
284 rhs[(int)get_irn_link(bb)] = -1.0;
288 for(i = get_Block_n_cfgpreds(bb) - 1; i >= 0; --i) {
289 ir_node *pred = get_Block_cfgpred_block(bb, i);
290 size_t pred_idx = (int)get_irn_link(pred);
292 // matrix[pred_idx + idx*size] += 1.0/(double)get_Block_n_cfg_outs(pred);
293 matrix[pred_idx + idx * size] += get_cf_probability(bb, i, loop_weight);
297 x = solve_lgs(matrix, rhs, size);
299 DEBUG_ONLY(ir_fprintf(stderr, "Debug Warning: Couldn't estimate execution frequencies for %+F\n", irg));
304 set_foreach(freqs, freq) {
305 const ir_node *bb = freq->irn;
306 size_t idx = PTR_TO_INT(get_irn_link(bb));
309 freq->freq = UNDEF(gsl_vector_get(x, idx)) ? EPSILON : gsl_vector_get(x, idx);
311 freq->freq = UNDEF(x[idx]) ? EPSILON : x[idx];
314 /* get the maximum exec freq */
315 ef->max = MAX(ef->max, freq->freq);
317 /* Get the minimum non-zero execution frequency. */
319 ef->min_non_zero = MIN(ef->min_non_zero, freq->freq);
322 /* compute m and b of the transformation used to convert the doubles into scaled ints */
324 double smallest_diff = 1.0;
326 double l2 = ef->min_non_zero;
329 double h1 = MAX_INT_FREQ;
331 double *fs = malloc(set_count(freqs) * sizeof(fs[0]));
334 set_foreach(freqs, freq)
335 fs[n++] = freq->freq;
338 * find the smallest difference of the execution frequencies
339 * we try to ressolve it with 1 integer.
341 for(i = 0; i < n; ++i) {
345 for(j = i + 1; j < n; ++j) {
346 double diff = fabs(fs[i] - fs[j]);
349 smallest_diff = MIN(diff, smallest_diff);
353 /* according to that the slope of the translation function is 1.0 / smallest diff */
354 ef->m = 1.0 / smallest_diff;
356 /* the abscissa is then given by */
357 ef->b = l1 - ef->m * l2;
360 * if the slope is so high that the largest integer would be larger than MAX_INT_FREQ
361 * set the largest int freq to that upper limit and recompute the translation function
363 if(ef->m * h2 + ef->b > MAX_INT_FREQ) {
364 ef->m = (h1 - l1) / (h2 - l2);
365 ef->b = l1 - ef->m * l2;
368 // printf("smallest_diff: %g, l1: %f, h1: %f, l2: %f, h2: %f, m: %f, b: %f\n", smallest_diff, l1, h1, l2, h2, ef->m, ef->b);
375 memset(&ef->hook, 0, sizeof(ef->hook));
376 ef->hook.context = ef;
377 ef->hook.hook._hook_node_info = exec_freq_node_info;
378 register_hook(hook_node_info, &ef->hook);
388 free_execfreq(ir_exec_freq *ef)
391 unregister_hook(hook_node_info, &ef->hook);