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"
36 #include "irgraph_t.h"
41 #include "interval_analysis.h"
43 void set_irp_exec_freq_state(exec_freq_state s);
45 /*------------------------------------------------------------------*/
46 /* A hashmap mapping the frequency to block and loop nodes. Block
47 * and loop nodes are regions. */
48 /*------------------------------------------------------------------*/
56 /* We use this set for all nodes in all irgraphs. */
57 static set *exec_freq_set = NULL;
59 static int exec_freq_cmp(const void *e1, const void *e2, size_t size) {
60 reg_exec_freq *ef1 = (reg_exec_freq *)e1;
61 reg_exec_freq *ef2 = (reg_exec_freq *)e2;
64 return (ef1->reg != ef2->reg);
67 static inline unsigned int exec_freq_hash(reg_exec_freq *e) {
68 return HASH_PTR(e->reg);
71 static inline void set_region_exec_freq(void *reg, double freq) {
75 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
78 double get_region_exec_freq(void *reg) {
79 reg_exec_freq ef, *found;
81 assert(exec_freq_set);
83 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
85 /* Not found if information is invalid. */
92 /* Returns the number of times the block is executed. */
93 double get_Block_exec_freq(ir_node *b) {
94 return get_region_exec_freq((void *)b);
97 double get_irn_exec_freq(ir_node *n) {
98 if (!is_Block(n)) n = get_nodes_block(n);
99 return get_Block_exec_freq(n);
103 /*------------------------------------------------------------------*/
104 /* A algorithm that precomputes whether Conds lead to an exception.
105 * Computes a field for all Projs from Conds that says the following:
106 * - The Proj projs from a normal dual Cond with probability 50:50
107 * - This Proj of the Cond leads to an exception, i.e., a raise node.
108 * It is taken with exception probability.
109 * - The Proj of the Cond avoids an exception. It is taken with
110 * 1 - exception probability. */
111 /*------------------------------------------------------------------*/
118 Cond_prob_avoid_exception,
119 Cond_prob_exception_taken,
120 Cond_prob_was_exception_taken,
123 static int just_passed_a_Raise = 0;
124 static ir_node *Cond_list = NULL;
126 /* We do not use an extra set, as Projs are not yet in the existing one. */
127 void set_ProjX_probability(ir_node *n, Cond_prob prob) {
131 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
134 Cond_prob get_ProjX_probability(ir_node *n) {
135 reg_exec_freq ef, *found;
138 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
141 return (Cond_prob)found->prob;
143 return Cond_prob_none;
146 /* A walker that only visits the nodes we want to see. */
149 my_irg_walk_2_both(ir_node *node, irg_walk_func *pre, irg_walk_func *post, void * env) {
151 set_irn_visited(node, current_ir_graph->visited);
155 if (node->op != op_Block) {
157 if (node->op == op_Proj)
158 pred = get_irn_n(node, 0);
160 pred = get_irn_n(node, -1);
161 if (pred->visited < current_ir_graph->visited)
162 my_irg_walk_2_both(pred, pre, post, env);
166 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
167 ir_node *pred = get_irn_n(node, i);
168 if (pred->visited < current_ir_graph->visited)
169 my_irg_walk_2_both(pred, pre, post, env);
173 if (node->op == op_End) {
174 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
175 ir_node *pred = get_irn_n(node, i);
176 if ((pred->op == op_Block) && (pred->visited < current_ir_graph->visited))
177 my_irg_walk_2_both(pred, pre, post, env);
183 static void my_irg_walk_current_graph(irg_walk_func *pre, irg_walk_func *post, void *env) {
184 inc_irg_visited(current_ir_graph);
185 my_irg_walk_2_both(get_irg_end(current_ir_graph), pre, post, env);
189 static void walk_pre(ir_node *n, void *env)
193 just_passed_a_Raise = 1;
195 if (get_irn_op(n) == op_Proj &&
196 is_Cond(get_Proj_pred(n)) &&
197 just_passed_a_Raise) {
199 ir_node *c = get_Proj_pred(n);
201 /* If we already visited the other Proj, and it also leads to a Raise,
202 we are in the middle of something. Continue searching. */
203 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
204 other_proj = get_irn_out(c, 0);
205 if (other_proj == n) other_proj = get_irn_out(c, 1);
206 if (get_ProjX_probability(other_proj) == Cond_prob_exception_taken) {
207 set_ProjX_probability(other_proj, Cond_prob_was_exception_taken);
208 /* Keep searching for the Proj, so keep just_passed_a_Raise. */
210 set_ProjX_probability(n, Cond_prob_exception_taken);
211 just_passed_a_Raise = 0;
216 set_irn_link(n, Cond_list);
221 static void walk_post(ir_node *n, void *env)
225 just_passed_a_Raise = 0;
227 if (get_irn_op(n) == op_Proj &&
228 is_Cond(get_Proj_pred(n)) && (
229 get_ProjX_probability(n) == Cond_prob_exception_taken ||
230 get_ProjX_probability(n) == Cond_prob_was_exception_taken
232 just_passed_a_Raise = 1;
236 /** Precompute which Conds test for an exception.
238 * Operates on current_ir_graph. */
239 void precompute_cond_evaluation(void) {
242 compute_irg_outs(current_ir_graph);
244 just_passed_a_Raise = 0;
246 my_irg_walk_current_graph(walk_pre, walk_post, NULL);
248 for (c = Cond_list; c; c = get_irn_link(c)) {
251 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
252 p0 = get_irn_out(c, 0);
253 p1 = get_irn_out(c, 1);
255 /* both are exceptions */
256 if ((get_ProjX_probability(p0) == Cond_prob_exception_taken) &&
257 (get_ProjX_probability(p1) == Cond_prob_exception_taken) ) {
258 assert(0 && "I tried to avoid these!");
260 set_ProjX_probability(p0, Cond_prob_normal);
261 set_ProjX_probability(p1, Cond_prob_normal);
264 /* p0 is exception */
265 else if (get_ProjX_probability(p0) == Cond_prob_exception_taken) {
266 set_ProjX_probability(p1, Cond_prob_avoid_exception);
269 /* p1 is exception */
270 else if (get_ProjX_probability(p1) == Cond_prob_exception_taken) {
271 set_ProjX_probability(p0, Cond_prob_avoid_exception);
274 /* none is exception */
276 set_ProjX_probability(p0, Cond_prob_normal);
277 set_ProjX_probability(p1, Cond_prob_normal);
282 int is_fragile_Proj(ir_node *n) {
283 return is_Proj(n) && (get_ProjX_probability(n) == Cond_prob_exception_taken);
286 /*------------------------------------------------------------------*/
287 /* The algorithm to compute the execution frequencies.
289 * Walk the control flow loop tree which we consider the interval
290 * tree. Compute the execution for the lowest loop, add inner loops
291 * to worklist. Consider the inner loops as simple nodes. Check that
292 * there is only one loop header in each loop. */
293 /*------------------------------------------------------------------*/
295 static double exception_prob = 0.001;
297 static inline int is_loop_head(ir_node *cond)
303 /** Weight a single region in edge.
305 * Given all outs of the predecessor region, we can compute the weight of
306 * this single edge. */
307 static inline double get_weighted_region_exec_freq(void *reg, int pos) {
308 void *pred_reg = get_region_in(reg, pos);
309 double res, full_freq = get_region_exec_freq (pred_reg);
310 int n_outs = get_region_n_outs (pred_reg);
311 int n_exc_outs = get_region_n_exc_outs(pred_reg);
314 if (is_ir_node(reg)) {
315 cfop = get_Block_cfgpred((ir_node *)reg, pos);
316 if (is_Proj(cfop) && !is_Cond(get_Proj_pred(cfop)))
317 cfop = skip_Proj(cfop);
319 assert(is_ir_loop(reg));
320 cfop = get_loop_cfop(reg, pos);
323 if (is_fragile_op(cfop) || is_fragile_Proj(cfop)) {
324 res = full_freq * exception_prob;
327 /* Equally distribute the weight after exceptions to the left over outs. */
328 res = (full_freq *(1 - exception_prob * n_exc_outs)) / (n_outs - n_exc_outs);
334 static inline void compute_region_freqency(void *reg, double head_weight) {
335 int i, n_ins = get_region_n_ins(reg);
338 //printf("head weight %lf: ", head_weight); DDMR(reg);
340 for (i = 0; i < n_ins; ++i) {
341 void *pred_reg = get_region_in(reg, i);
343 my_freq += get_weighted_region_exec_freq(reg, i);
347 if (my_freq == 0.0) {
348 /* All preds are from outer loop. We are a head or so. */
349 my_freq = head_weight;
351 set_region_exec_freq(reg, my_freq);
354 static void check_proper_head(ir_loop *l, void *reg)
356 int i, n_ins = get_region_n_ins(reg);
358 for (i = 0; i < n_ins; ++i) {
359 assert(!get_region_in(reg, i));
363 /* Compute the ex freq for current_ir_graph */
364 static void compute_frequency(int default_loop_weight) {
365 ir_loop *outermost_l = get_irg_loop(current_ir_graph);
366 pdeq *block_worklist = new_pdeq1(outermost_l);
368 /* Outermost start is considered a loop head. We will soon multiply
369 by default_loop_weight. */
370 set_region_exec_freq(outermost_l, 1.0/default_loop_weight);
372 while (!pdeq_empty(block_worklist)) {
373 ir_loop *l = (ir_loop *)pdeq_getl(block_worklist);
374 int i, n_elems = get_loop_n_elements(l);
376 /* The header is initialized with the frequency of the full loop times the iteration weight. */
377 check_proper_head(l, get_loop_element(l, 0).son);
379 for (i = 0; i < n_elems; ++i) {
380 loop_element e = get_loop_element(l, i);
381 if (is_ir_loop(e.son)) pdeq_putr(block_worklist, e.son);
382 compute_region_freqency(e.son, default_loop_weight * get_region_exec_freq(l));
385 del_pdeq(block_worklist);
388 /* Compute the execution frequency for all blocks in the given
391 * irg: The graph to be analyzed.
392 * default_loop_weight: The number of executions of a loop.
394 void compute_execution_frequency(ir_graph *irg, int default_loop_weight, double exception_probability) {
395 ir_graph *rem = current_ir_graph;
396 current_ir_graph = irg;
397 exception_prob = exception_probability;
398 if (!exec_freq_set) exec_freq_set = new_set(exec_freq_cmp, 256);
400 precompute_cond_evaluation();
401 construct_intervals(current_ir_graph);
402 compute_frequency(default_loop_weight);
404 set_irg_exec_freq_state(irg, exec_freq_consistent);
405 if (get_irp_exec_freq_state() == exec_freq_none)
406 set_irp_exec_freq_state(exec_freq_inconsistent);
409 dump_loop_tree (current_ir_graph, "-execfreq");
410 dump_ir_block_graph(current_ir_graph, "-execfreq");
411 dump_interval_graph(current_ir_graph, "-execfreq");
414 current_ir_graph = rem;
418 void compute_execution_frequencies(int default_loop_weight, double exception_probability) {
419 int i, n_irgs = get_irp_n_irgs();
421 for (i = 0; i < n_irgs; ++i) {
422 compute_execution_frequency(get_irp_irg(i), default_loop_weight, exception_probability);
424 set_irp_exec_freq_state(exec_freq_consistent);
427 /** free occupied memory, reset */
428 void free_execution_frequency(void) {
429 int i, n_irgs = get_irp_n_irgs();
431 del_set(exec_freq_set);
433 for (i = 0; i < n_irgs; ++i)
434 set_irg_exec_freq_state(get_irp_irg(i), exec_freq_none);
435 set_irp_exec_freq_state(exec_freq_none);
438 exec_freq_state get_irg_exec_freq_state(ir_graph *irg) {
439 return irg->execfreq_state;
441 void set_irg_exec_freq_state(ir_graph *irg, exec_freq_state s) {
442 if ((get_irp_exec_freq_state() == exec_freq_consistent && s != exec_freq_consistent) ||
443 (get_irp_exec_freq_state() == exec_freq_none && s != exec_freq_none))
444 irp->execfreq_state = exec_freq_inconsistent;
445 irg->execfreq_state = s;
448 /* Sets irg and irp exec freq state to inconsistent if it is set to consistent. */
449 void set_irg_exec_freq_state_inconsistent(ir_graph *irg) {
450 if (get_irg_exec_freq_state(irg) == exec_freq_consistent)
451 set_irg_exec_freq_state(irg, exec_freq_inconsistent);
454 void set_irp_exec_freq_state(exec_freq_state s) {
455 irp->execfreq_state = s;
458 exec_freq_state get_irp_exec_freq_state(void) {
459 return irp->execfreq_state;
462 /* Sets irp and all irg exec freq states to inconsistent if it is set to consistent. */
463 void set_irp_exec_freq_state_inconsistent(void) {
464 if (get_irp_exec_freq_state() != exec_freq_none) {
465 int i, n_irgs = get_irp_n_irgs();
466 set_irp_exec_freq_state(exec_freq_inconsistent);
467 for (i = 0; i < n_irgs; ++i) {
468 ir_graph *irg = get_irp_irg(i);
469 if (get_irg_exec_freq_state(irg) != exec_freq_none)
470 irg->execfreq_state = exec_freq_inconsistent;