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 Goetz Lindenmaier
31 #include "execution_frequency.h"
33 #include "firm_common_t.h"
39 #include "irgraph_t.h"
44 #include "interval_analysis.h"
46 void set_irp_exec_freq_state(exec_freq_state s);
48 /*------------------------------------------------------------------*/
49 /* A hashmap mapping the frequency to block and loop nodes. Block
50 * and loop nodes are regions. */
51 /*------------------------------------------------------------------*/
59 /* We use this set for all nodes in all irgraphs. */
60 static set *exec_freq_set = NULL;
62 static int exec_freq_cmp(const void *e1, const void *e2, size_t size) {
63 reg_exec_freq *ef1 = (reg_exec_freq *)e1;
64 reg_exec_freq *ef2 = (reg_exec_freq *)e2;
67 return (ef1->reg != ef2->reg);
70 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) {
78 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
81 double get_region_exec_freq(void *reg) {
82 reg_exec_freq ef, *found;
84 assert(exec_freq_set);
86 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
88 /* Not found if information is invalid. */
95 /* Returns the number of times the block is executed. */
96 double get_Block_exec_freq(ir_node *b) {
97 return get_region_exec_freq((void *)b);
100 double get_irn_exec_freq(ir_node *n) {
101 if (!is_Block(n)) n = get_nodes_block(n);
102 return get_Block_exec_freq(n);
106 /*------------------------------------------------------------------*/
107 /* A algorithm that precomputes whether Conds lead to an exception.
108 * Computes a field for all Projs from Conds that says the following:
109 * - The Proj projs from a normal dual Cond with probability 50:50
110 * - This Proj of the Cond leads to an exception, i.e., a raise node.
111 * It is taken with exception probability.
112 * - The Proj of the Cond avoids an exception. It is taken with
113 * 1 - exception probability. */
114 /*------------------------------------------------------------------*/
121 Cond_prob_avoid_exception,
122 Cond_prob_exception_taken,
123 Cond_prob_was_exception_taken,
126 static int just_passed_a_Raise = 0;
127 static ir_node *Cond_list = NULL;
129 /* We do not use an extra set, as Projs are not yet in the existing one. */
130 void set_ProjX_probability(ir_node *n, Cond_prob prob) {
134 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
137 Cond_prob get_ProjX_probability(ir_node *n) {
138 reg_exec_freq ef, *found;
141 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
144 return (Cond_prob)found->prob;
146 return Cond_prob_none;
149 /* A walker that only visits the nodes we want to see. */
152 my_irg_walk_2_both(ir_node *node, irg_walk_func *pre, irg_walk_func *post, void * env) {
154 set_irn_visited(node, current_ir_graph->visited);
158 if (node->op != op_Block) {
160 if (node->op == op_Proj)
161 pred = get_irn_n(node, 0);
163 pred = get_irn_n(node, -1);
164 if (pred->visited < current_ir_graph->visited)
165 my_irg_walk_2_both(pred, pre, post, env);
169 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
170 ir_node *pred = get_irn_n(node, i);
171 if (pred->visited < current_ir_graph->visited)
172 my_irg_walk_2_both(pred, pre, post, env);
176 if (node->op == op_End) {
177 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
178 ir_node *pred = get_irn_n(node, i);
179 if ((pred->op == op_Block) && (pred->visited < current_ir_graph->visited))
180 my_irg_walk_2_both(pred, pre, post, env);
186 static void my_irg_walk_current_graph(irg_walk_func *pre, irg_walk_func *post, void *env) {
187 inc_irg_visited(current_ir_graph);
188 my_irg_walk_2_both(get_irg_end(current_ir_graph), pre, post, env);
192 static void walk_pre(ir_node *n, void *env)
195 if (get_irn_op(n) == op_Raise)
196 just_passed_a_Raise = 1;
198 if ( (get_irn_op(n) == op_Proj)
199 && (get_irn_op(get_Proj_pred(n)) == op_Cond)
200 && (just_passed_a_Raise)) {
202 ir_node *c = get_Proj_pred(n);
204 /* If we already visited the other Proj, and it also leads to a Raise,
205 we are in the middle of something. Continue searching. */
206 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
207 other_proj = get_irn_out(c, 0);
208 if (other_proj == n) other_proj = get_irn_out(c, 1);
209 if (get_ProjX_probability(other_proj) == Cond_prob_exception_taken) {
210 set_ProjX_probability(other_proj, Cond_prob_was_exception_taken);
211 /* Keep searching for the Proj, so keep just_passed_a_Raise. */
213 set_ProjX_probability(n, Cond_prob_exception_taken);
214 just_passed_a_Raise = 0;
218 if (get_irn_op(n) == op_Cond) {
219 set_irn_link(n, Cond_list);
224 static void walk_post(ir_node *n, void *env)
227 if (get_irn_op(n) == op_Raise)
228 just_passed_a_Raise = 0;
230 if ( (get_irn_op(n) == op_Proj)
231 && (get_irn_op(get_Proj_pred(n)) == op_Cond)
232 && ((get_ProjX_probability(n) == Cond_prob_exception_taken) ||
233 (get_ProjX_probability(n) == Cond_prob_was_exception_taken) )) {
234 just_passed_a_Raise = 1;
238 /** Precompute which Conds test for an exception.
240 * Operates on current_ir_graph. */
241 void precompute_cond_evaluation(void) {
244 compute_irg_outs(current_ir_graph);
246 just_passed_a_Raise = 0;
248 my_irg_walk_current_graph(walk_pre, walk_post, NULL);
250 for (c = Cond_list; c; c = get_irn_link(c)) {
253 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
254 p0 = get_irn_out(c, 0);
255 p1 = get_irn_out(c, 1);
257 /* both are exceptions */
258 if ((get_ProjX_probability(p0) == Cond_prob_exception_taken) &&
259 (get_ProjX_probability(p1) == Cond_prob_exception_taken) ) {
260 assert(0 && "I tried to avoid these!");
262 set_ProjX_probability(p0, Cond_prob_normal);
263 set_ProjX_probability(p1, Cond_prob_normal);
266 /* p0 is exception */
267 else if (get_ProjX_probability(p0) == Cond_prob_exception_taken) {
268 set_ProjX_probability(p1, Cond_prob_avoid_exception);
271 /* p1 is exception */
272 else if (get_ProjX_probability(p1) == Cond_prob_exception_taken) {
273 set_ProjX_probability(p0, Cond_prob_avoid_exception);
276 /* none is exception */
278 set_ProjX_probability(p0, Cond_prob_normal);
279 set_ProjX_probability(p1, Cond_prob_normal);
284 int is_fragile_Proj(ir_node *n) {
285 return is_Proj(n) && (get_ProjX_probability(n) == Cond_prob_exception_taken);
288 /*------------------------------------------------------------------*/
289 /* The algorithm to compute the execution frequencies.
291 * Walk the control flow loop tree which we consider the interval
292 * tree. Compute the execution for the lowest loop, add inner loops
293 * to worklist. Consider the inner loops as simple nodes. Check that
294 * there is only one loop header in each loop. */
295 /*------------------------------------------------------------------*/
297 static double exception_prob = 0.001;
299 static INLINE int is_loop_head(ir_node *cond)
305 /** Weight a single region in edge.
307 * Given all outs of the predecessor region, we can compute the weight of
308 * this single edge. */
309 static INLINE double get_weighted_region_exec_freq(void *reg, int pos) {
310 void *pred_reg = get_region_in(reg, pos);
311 double res, full_freq = get_region_exec_freq (pred_reg);
312 int n_outs = get_region_n_outs (pred_reg);
313 int n_exc_outs = get_region_n_exc_outs(pred_reg);
316 if (is_ir_node(reg)) {
317 cfop = get_Block_cfgpred((ir_node *)reg, pos);
318 if (is_Proj(cfop) && (get_irn_op(get_Proj_pred(cfop)) != op_Cond))
319 cfop = skip_Proj(cfop);
321 assert(is_ir_loop(reg));
322 cfop = get_loop_cfop(reg, pos);
325 if (is_fragile_op(cfop) || is_fragile_Proj(cfop)) {
326 res = full_freq * exception_prob;
329 /* Equally distribute the weight after exceptions to the left over outs. */
330 res = (full_freq *(1 - exception_prob * n_exc_outs)) / (n_outs - n_exc_outs);
336 static INLINE void compute_region_freqency(void *reg, double head_weight) {
337 int i, n_ins = get_region_n_ins(reg);
340 //printf("head weight %lf: ", head_weight); DDMR(reg);
342 for (i = 0; i < n_ins; ++i) {
343 void *pred_reg = get_region_in(reg, i);
345 my_freq += get_weighted_region_exec_freq(reg, i);
349 if (my_freq == 0.0) {
350 /* All preds are from outer loop. We are a head or so. */
351 my_freq = head_weight;
353 set_region_exec_freq(reg, my_freq);
356 static void check_proper_head(ir_loop *l, void *reg)
358 int i, n_ins = get_region_n_ins(reg);
360 for (i = 0; i < n_ins; ++i) {
361 assert(!get_region_in(reg, i));
365 /* Compute the ex freq for current_ir_graph */
366 static void compute_frequency(int default_loop_weight) {
367 ir_loop *outermost_l = get_irg_loop(current_ir_graph);
368 pdeq *block_worklist = new_pdeq1(outermost_l);
370 /* Outermost start is considered a loop head. We will soon multiply
371 by default_loop_weight. */
372 set_region_exec_freq(outermost_l, 1.0/default_loop_weight);
374 while (!pdeq_empty(block_worklist)) {
375 ir_loop *l = (ir_loop *)pdeq_getl(block_worklist);
376 int i, n_elems = get_loop_n_elements(l);
378 /* The header is initialized with the frequency of the full loop times the iteration weight. */
379 check_proper_head(l, get_loop_element(l, 0).son);
381 for (i = 0; i < n_elems; ++i) {
382 loop_element e = get_loop_element(l, i);
383 if (is_ir_loop(e.son)) pdeq_putr(block_worklist, e.son);
384 compute_region_freqency(e.son, default_loop_weight * get_region_exec_freq(l));
387 del_pdeq(block_worklist);
390 /* Compute the execution frequency for all blocks in the given
393 * irg: The graph to be analyzed.
394 * default_loop_weight: The number of executions of a loop.
396 void compute_execution_frequency(ir_graph *irg, int default_loop_weight, double exception_probability) {
397 ir_graph *rem = current_ir_graph;
398 current_ir_graph = irg;
399 exception_prob = exception_probability;
400 if (!exec_freq_set) exec_freq_set = new_set(exec_freq_cmp, 256);
402 precompute_cond_evaluation();
403 construct_intervals(current_ir_graph);
404 compute_frequency(default_loop_weight);
406 set_irg_exec_freq_state(irg, exec_freq_consistent);
407 if (get_irp_exec_freq_state() == exec_freq_none)
408 set_irp_exec_freq_state(exec_freq_inconsistent);
411 dump_loop_tree (current_ir_graph, "-execfreq");
412 dump_ir_block_graph(current_ir_graph, "-execfreq");
413 dump_interval_graph(current_ir_graph, "-execfreq");
416 current_ir_graph = rem;
420 void compute_execution_frequencies(int default_loop_weight, double exception_probability) {
421 int i, n_irgs = get_irp_n_irgs();
423 for (i = 0; i < n_irgs; ++i) {
424 compute_execution_frequency(get_irp_irg(i), default_loop_weight, exception_probability);
426 set_irp_exec_freq_state(exec_freq_consistent);
429 /** free occupied memory, reset */
430 void free_execution_frequency(void) {
431 int i, n_irgs = get_irp_n_irgs();
433 del_set(exec_freq_set);
435 for (i = 0; i < n_irgs; ++i)
436 set_irg_exec_freq_state(get_irp_irg(i), exec_freq_none);
437 set_irp_exec_freq_state(exec_freq_none);
440 exec_freq_state get_irg_exec_freq_state(ir_graph *irg) {
441 return irg->execfreq_state;
443 void set_irg_exec_freq_state(ir_graph *irg, exec_freq_state s) {
444 if ((get_irp_exec_freq_state() == exec_freq_consistent && s != exec_freq_consistent) ||
445 (get_irp_exec_freq_state() == exec_freq_none && s != exec_freq_none))
446 irp->execfreq_state = exec_freq_inconsistent;
447 irg->execfreq_state = s;
450 /* Sets irg and irp exec freq state to inconsistent if it is set to consistent. */
451 void set_irg_exec_freq_state_inconsistent(ir_graph *irg) {
452 if (get_irg_exec_freq_state(irg) == exec_freq_consistent)
453 set_irg_exec_freq_state(irg, exec_freq_inconsistent);
456 void set_irp_exec_freq_state(exec_freq_state s) {
457 irp->execfreq_state = s;
460 exec_freq_state get_irp_exec_freq_state(void) {
461 return irp->execfreq_state;
464 /* Sets irp and all irg exec freq states to inconsistent if it is set to consistent. */
465 void set_irp_exec_freq_state_inconsistent(void) {
466 if (get_irp_exec_freq_state() != exec_freq_none) {
467 int i, n_irgs = get_irp_n_irgs();
468 set_irp_exec_freq_state(exec_freq_inconsistent);
469 for (i = 0; i < n_irgs; ++i) {
470 ir_graph *irg = get_irp_irg(i);
471 if (get_irg_exec_freq_state(irg) != exec_freq_none)
472 irg->execfreq_state = exec_freq_inconsistent;