2 * Copyright (C) 1995-2008 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 Goetz Lindenmaier
29 #include "execution_frequency.h"
37 #include "irgraph_t.h"
42 #include "interval_analysis.h"
44 void set_irp_exec_freq_state(exec_freq_state s);
46 /*------------------------------------------------------------------*/
47 /* A hashmap mapping the frequency to block and loop nodes. Block
48 * and loop nodes are regions. */
49 /*------------------------------------------------------------------*/
57 /* We use this set for all nodes in all irgraphs. */
58 static set *exec_freq_set = NULL;
60 static int exec_freq_cmp(const void *e1, const void *e2, size_t size)
62 reg_exec_freq *ef1 = (reg_exec_freq *)e1;
63 reg_exec_freq *ef2 = (reg_exec_freq *)e2;
66 return (ef1->reg != ef2->reg);
69 static inline unsigned int exec_freq_hash(reg_exec_freq *e)
71 return HASH_PTR(e->reg);
74 static inline void set_region_exec_freq(void *reg, double freq)
79 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
82 double get_region_exec_freq(void *reg)
84 reg_exec_freq ef, *found;
86 assert(exec_freq_set);
88 found = (reg_exec_freq*) set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
90 /* Not found if information is invalid. */
97 /* Returns the number of times the block is executed. */
98 double get_Block_exec_freq(ir_node *b)
100 return get_region_exec_freq((void *)b);
103 double get_irn_exec_freq(ir_node *n)
105 if (!is_Block(n)) n = get_nodes_block(n);
106 return get_Block_exec_freq(n);
110 /*------------------------------------------------------------------*/
111 /* A algorithm that precomputes whether Conds lead to an exception.
112 * Computes a field for all Projs from Conds that says the following:
113 * - The Proj projs from a normal dual Cond with probability 50:50
114 * - This Proj of the Cond leads to an exception, i.e., a raise node.
115 * It is taken with exception probability.
116 * - The Proj of the Cond avoids an exception. It is taken with
117 * 1 - exception probability. */
118 /*------------------------------------------------------------------*/
125 Cond_prob_avoid_exception,
126 Cond_prob_exception_taken,
127 Cond_prob_was_exception_taken,
130 static int just_passed_a_Raise = 0;
131 static ir_node *Cond_list = NULL;
133 /* We do not use an extra set, as Projs are not yet in the existing one. */
134 static void set_ProjX_probability(ir_node *n, Cond_prob prob)
139 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
142 static Cond_prob get_ProjX_probability(ir_node *n)
144 reg_exec_freq ef, *found;
147 found = (reg_exec_freq*) set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
150 return (Cond_prob)found->prob;
152 return Cond_prob_none;
155 /* A walker that only visits the nodes we want to see. */
157 static void my_irg_walk_2_both(ir_node *node, irg_walk_func *pre, irg_walk_func *post, void * env)
160 set_irn_visited(node, current_ir_graph->visited);
164 if (!is_Block(node)) {
167 pred = get_irn_n(node, 0);
169 pred = get_irn_n(node, -1);
170 if (pred->visited < current_ir_graph->visited)
171 my_irg_walk_2_both(pred, pre, post, env);
175 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
176 ir_node *pred = get_irn_n(node, i);
177 if (pred->visited < current_ir_graph->visited)
178 my_irg_walk_2_both(pred, pre, post, env);
183 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
184 ir_node *pred = get_irn_n(node, i);
185 if ((pred->op == op_Block) && (pred->visited < current_ir_graph->visited))
186 my_irg_walk_2_both(pred, pre, post, env);
192 static void my_irg_walk_current_graph(irg_walk_func *pre, irg_walk_func *post, void *env)
194 inc_irg_visited(current_ir_graph);
195 my_irg_walk_2_both(get_irg_end(current_ir_graph), pre, post, env);
199 static void walk_pre(ir_node *n, void *env)
203 just_passed_a_Raise = 1;
205 if (get_irn_op(n) == op_Proj &&
206 is_Cond(get_Proj_pred(n)) &&
207 just_passed_a_Raise) {
209 ir_node *c = get_Proj_pred(n);
211 /* If we already visited the other Proj, and it also leads to a Raise,
212 we are in the middle of something. Continue searching. */
213 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
214 other_proj = get_irn_out(c, 0);
215 if (other_proj == n) other_proj = get_irn_out(c, 1);
216 if (get_ProjX_probability(other_proj) == Cond_prob_exception_taken) {
217 set_ProjX_probability(other_proj, Cond_prob_was_exception_taken);
218 /* Keep searching for the Proj, so keep just_passed_a_Raise. */
220 set_ProjX_probability(n, Cond_prob_exception_taken);
221 just_passed_a_Raise = 0;
226 set_irn_link(n, Cond_list);
231 static void walk_post(ir_node *n, void *env)
235 just_passed_a_Raise = 0;
237 if (get_irn_op(n) == op_Proj &&
238 is_Cond(get_Proj_pred(n)) && (
239 get_ProjX_probability(n) == Cond_prob_exception_taken ||
240 get_ProjX_probability(n) == Cond_prob_was_exception_taken
242 just_passed_a_Raise = 1;
246 /** Precompute which Conds test for an exception.
248 * Operates on current_ir_graph. */
249 static void precompute_cond_evaluation(void)
253 compute_irg_outs(current_ir_graph);
255 just_passed_a_Raise = 0;
257 my_irg_walk_current_graph(walk_pre, walk_post, NULL);
259 for (c = Cond_list; c; c = (ir_node*)get_irn_link(c)) {
262 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
263 p0 = get_irn_out(c, 0);
264 p1 = get_irn_out(c, 1);
266 /* both are exceptions */
267 if ((get_ProjX_probability(p0) == Cond_prob_exception_taken) &&
268 (get_ProjX_probability(p1) == Cond_prob_exception_taken) ) {
269 panic("I tried to avoid these!");
272 set_ProjX_probability(p0, Cond_prob_normal);
273 set_ProjX_probability(p1, Cond_prob_normal);
277 /* p0 is exception */
278 else if (get_ProjX_probability(p0) == Cond_prob_exception_taken) {
279 set_ProjX_probability(p1, Cond_prob_avoid_exception);
282 /* p1 is exception */
283 else if (get_ProjX_probability(p1) == Cond_prob_exception_taken) {
284 set_ProjX_probability(p0, Cond_prob_avoid_exception);
287 /* none is exception */
289 set_ProjX_probability(p0, Cond_prob_normal);
290 set_ProjX_probability(p1, Cond_prob_normal);
295 int is_fragile_Proj(ir_node *n)
297 return is_Proj(n) && (get_ProjX_probability(n) == Cond_prob_exception_taken);
300 /*------------------------------------------------------------------*/
301 /* The algorithm to compute the execution frequencies.
303 * Walk the control flow loop tree which we consider the interval
304 * tree. Compute the execution for the lowest loop, add inner loops
305 * to worklist. Consider the inner loops as simple nodes. Check that
306 * there is only one loop header in each loop. */
307 /*------------------------------------------------------------------*/
309 static double exception_prob = 0.001;
311 static inline int is_loop_head(ir_node *cond)
317 /** Weight a single region in edge.
319 * Given all outs of the predecessor region, we can compute the weight of
320 * this single edge. */
321 static inline double get_weighted_region_exec_freq(void *reg, int pos)
323 void *pred_reg = get_region_in(reg, pos);
324 double res, full_freq = get_region_exec_freq (pred_reg);
325 int n_outs = get_region_n_outs (pred_reg);
326 int n_exc_outs = get_region_n_exc_outs(pred_reg);
329 if (is_ir_node(reg)) {
330 cfop = get_Block_cfgpred((ir_node *)reg, pos);
331 if (is_Proj(cfop) && !is_Cond(get_Proj_pred(cfop)))
332 cfop = skip_Proj(cfop);
334 assert(is_ir_loop(reg));
335 cfop = (ir_node*)get_loop_cfop(reg, pos);
338 if (is_fragile_op(cfop) || is_fragile_Proj(cfop)) {
339 res = full_freq * exception_prob;
342 /* Equally distribute the weight after exceptions to the left over outs. */
343 res = (full_freq *(1 - exception_prob * n_exc_outs)) / (n_outs - n_exc_outs);
349 static inline void compute_region_freqency(void *reg, double head_weight)
351 int i, n_ins = get_region_n_ins(reg);
354 for (i = 0; i < n_ins; ++i) {
355 void *pred_reg = get_region_in(reg, i);
357 my_freq += get_weighted_region_exec_freq(reg, i);
361 if (my_freq == 0.0) {
362 /* All preds are from outer loop. We are a head or so. */
363 my_freq = head_weight;
365 set_region_exec_freq(reg, my_freq);
368 static void check_proper_head(ir_loop *l, void *reg)
370 int i, n_ins = get_region_n_ins(reg);
372 for (i = 0; i < n_ins; ++i) {
373 assert(!get_region_in(reg, i));
377 /* Compute the ex freq for current_ir_graph */
378 static void compute_frequency(int default_loop_weight)
380 ir_loop *outermost_l = get_irg_loop(current_ir_graph);
381 pdeq *block_worklist = new_pdeq1(outermost_l);
383 /* Outermost start is considered a loop head. We will soon multiply
384 by default_loop_weight. */
385 set_region_exec_freq(outermost_l, 1.0/default_loop_weight);
387 while (!pdeq_empty(block_worklist)) {
388 ir_loop *l = (ir_loop *)pdeq_getl(block_worklist);
389 size_t i, n_elems = get_loop_n_elements(l);
391 /* The header is initialized with the frequency of the full loop times the iteration weight. */
392 check_proper_head(l, get_loop_element(l, 0).son);
394 for (i = 0; i < n_elems; ++i) {
395 loop_element e = get_loop_element(l, i);
396 if (is_ir_loop(e.son)) pdeq_putr(block_worklist, e.son);
397 compute_region_freqency(e.son, default_loop_weight * get_region_exec_freq(l));
400 del_pdeq(block_worklist);
403 /* Compute the execution frequency for all blocks in the given
406 * irg: The graph to be analyzed.
407 * default_loop_weight: The number of executions of a loop.
409 void compute_execution_frequency(ir_graph *irg, int default_loop_weight, double exception_probability)
411 ir_graph *rem = current_ir_graph;
412 current_ir_graph = irg;
413 exception_prob = exception_probability;
414 if (!exec_freq_set) exec_freq_set = new_set(exec_freq_cmp, 256);
416 precompute_cond_evaluation();
417 construct_intervals(current_ir_graph);
418 compute_frequency(default_loop_weight);
420 set_irg_exec_freq_state(irg, exec_freq_consistent);
421 if (get_irp_exec_freq_state() == exec_freq_none)
422 set_irp_exec_freq_state(exec_freq_inconsistent);
425 dump_loop_tree (current_ir_graph, "-execfreq");
426 dump_ir_block_graph(current_ir_graph, "-execfreq");
427 dump_interval_graph(current_ir_graph, "-execfreq");
430 current_ir_graph = rem;
434 void compute_execution_frequencies(int default_loop_weight, double exception_probability)
436 size_t i, n_irgs = get_irp_n_irgs();
438 for (i = 0; i < n_irgs; ++i) {
439 compute_execution_frequency(get_irp_irg(i), default_loop_weight, exception_probability);
441 set_irp_exec_freq_state(exec_freq_consistent);
444 /** free occupied memory, reset */
445 void free_execution_frequency(void)
447 size_t i, n_irgs = get_irp_n_irgs();
449 del_set(exec_freq_set);
451 for (i = 0; i < n_irgs; ++i)
452 set_irg_exec_freq_state(get_irp_irg(i), exec_freq_none);
453 set_irp_exec_freq_state(exec_freq_none);
456 exec_freq_state get_irg_exec_freq_state(ir_graph *irg)
458 return irg->execfreq_state;
460 void set_irg_exec_freq_state(ir_graph *irg, exec_freq_state s)
462 if ((get_irp_exec_freq_state() == exec_freq_consistent && s != exec_freq_consistent) ||
463 (get_irp_exec_freq_state() == exec_freq_none && s != exec_freq_none))
464 irp->execfreq_state = exec_freq_inconsistent;
465 irg->execfreq_state = s;
468 /* Sets irg and irp exec freq state to inconsistent if it is set to consistent. */
469 void set_irg_exec_freq_state_inconsistent(ir_graph *irg)
471 if (get_irg_exec_freq_state(irg) == exec_freq_consistent)
472 set_irg_exec_freq_state(irg, exec_freq_inconsistent);
475 void set_irp_exec_freq_state(exec_freq_state s)
477 irp->execfreq_state = s;
480 exec_freq_state get_irp_exec_freq_state(void)
482 return irp->execfreq_state;
485 /* Sets irp and all irg exec freq states to inconsistent if it is set to consistent. */
486 void set_irp_exec_freq_state_inconsistent(void)
488 if (get_irp_exec_freq_state() != exec_freq_none) {
489 size_t i, n_irgs = get_irp_n_irgs();
490 set_irp_exec_freq_state(exec_freq_inconsistent);
491 for (i = 0; i < n_irgs; ++i) {
492 ir_graph *irg = get_irp_irg(i);
493 if (get_irg_exec_freq_state(irg) != exec_freq_none)
494 irg->execfreq_state = exec_freq_inconsistent;