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;
65 return (ef1->reg != ef2->reg);
68 static INLINE unsigned int exec_freq_hash(reg_exec_freq *e) {
69 return HASH_PTR(e->reg);
72 static INLINE void set_region_exec_freq(void *reg, double freq) {
76 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
79 double get_region_exec_freq(void *reg) {
80 reg_exec_freq ef, *found;
82 assert(exec_freq_set);
84 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
86 /* Not found if information is invalid. */
93 /* Returns the number of times the block is executed. */
94 double get_Block_exec_freq(ir_node *b) {
95 return get_region_exec_freq((void *)b);
98 double get_irn_exec_freq(ir_node *n) {
99 if (!is_Block(n)) n = get_nodes_block(n);
100 return get_Block_exec_freq(n);
104 /*------------------------------------------------------------------*/
105 /* A algorithm that precomputes whether Conds lead to an exception.
106 * Computes a field for all Projs from Conds that says the following:
107 * - The Proj projs from a normal dual Cond with probability 50:50
108 * - This Proj of the Cond leads to an exception, i.e., a raise node.
109 * It is taken with exception probability.
110 * - The Proj of the Cond avoids an exception. It is taken with
111 * 1 - exception probability. */
112 /*------------------------------------------------------------------*/
119 Cond_prob_avoid_exception,
120 Cond_prob_exception_taken,
121 Cond_prob_was_exception_taken,
124 static int just_passed_a_Raise = 0;
125 static ir_node *Cond_list = NULL;
127 /* We do not use an extra set, as Projs are not yet in the existing one. */
128 void set_ProjX_probability(ir_node *n, Cond_prob prob) {
132 set_insert(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
135 Cond_prob get_ProjX_probability(ir_node *n) {
136 reg_exec_freq ef, *found;
139 found = set_find(exec_freq_set, &ef, sizeof(ef), exec_freq_hash(&ef));
142 return (Cond_prob)found->prob;
144 return Cond_prob_none;
147 /* A walker that only visits the nodes we want to see. */
150 my_irg_walk_2_both(ir_node *node, irg_walk_func *pre, irg_walk_func *post, void * env) {
152 set_irn_visited(node, current_ir_graph->visited);
156 if (node->op != op_Block) {
158 if (node->op == op_Proj)
159 pred = get_irn_n(node, 0);
161 pred = get_irn_n(node, -1);
162 if (pred->visited < current_ir_graph->visited)
163 my_irg_walk_2_both(pred, pre, post, env);
167 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
168 ir_node *pred = get_irn_n(node, i);
169 if (pred->visited < current_ir_graph->visited)
170 my_irg_walk_2_both(pred, pre, post, env);
174 if (node->op == op_End) {
175 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
176 ir_node *pred = get_irn_n(node, i);
177 if ((pred->op == op_Block) && (pred->visited < current_ir_graph->visited))
178 my_irg_walk_2_both(pred, pre, post, env);
184 static void my_irg_walk_current_graph(irg_walk_func *pre, irg_walk_func *post, void *env) {
185 inc_irg_visited(current_ir_graph);
186 my_irg_walk_2_both(get_irg_end(current_ir_graph), pre, post, env);
190 static void walk_pre(ir_node *n, void *env) {
192 if (get_irn_op(n) == op_Raise)
193 just_passed_a_Raise = 1;
195 if ( (get_irn_op(n) == op_Proj)
196 && (get_irn_op(get_Proj_pred(n)) == op_Cond)
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;
215 if (get_irn_op(n) == op_Cond) {
216 set_irn_link(n, Cond_list);
221 static void walk_post(ir_node *n, void *env) {
223 if (get_irn_op(n) == op_Raise)
224 just_passed_a_Raise = 0;
226 if ( (get_irn_op(n) == op_Proj)
227 && (get_irn_op(get_Proj_pred(n)) == op_Cond)
228 && ((get_ProjX_probability(n) == Cond_prob_exception_taken) ||
229 (get_ProjX_probability(n) == Cond_prob_was_exception_taken) )) {
230 just_passed_a_Raise = 1;
234 /** Precompute which Conds test for an exception.
236 * Operates on current_ir_graph. */
237 void precompute_cond_evaluation(void) {
240 compute_irg_outs(current_ir_graph);
242 just_passed_a_Raise = 0;
244 my_irg_walk_current_graph(walk_pre, walk_post, NULL);
246 for (c = Cond_list; c; c = get_irn_link(c)) {
249 assert(get_irn_n_outs(c) == 2 && "encountered a switch cond");
250 p0 = get_irn_out(c, 0);
251 p1 = get_irn_out(c, 1);
253 /* both are exceptions */
254 if ((get_ProjX_probability(p0) == Cond_prob_exception_taken) &&
255 (get_ProjX_probability(p1) == Cond_prob_exception_taken) ) {
256 assert(0 && "I tried to avoid these!");
258 set_ProjX_probability(p0, Cond_prob_normal);
259 set_ProjX_probability(p1, Cond_prob_normal);
262 /* p0 is exception */
263 else if (get_ProjX_probability(p0) == Cond_prob_exception_taken) {
264 set_ProjX_probability(p1, Cond_prob_avoid_exception);
267 /* p1 is exception */
268 else if (get_ProjX_probability(p1) == Cond_prob_exception_taken) {
269 set_ProjX_probability(p0, Cond_prob_avoid_exception);
272 /* none is exception */
274 set_ProjX_probability(p0, Cond_prob_normal);
275 set_ProjX_probability(p1, Cond_prob_normal);
280 int is_fragile_Proj(ir_node *n) {
281 return is_Proj(n) && (get_ProjX_probability(n) == Cond_prob_exception_taken);
284 /*------------------------------------------------------------------*/
285 /* The algorithm to compute the execution frequencies.
287 * Walk the control flow loop tree which we consider the interval
288 * tree. Compute the execution for the lowest loop, add inner loops
289 * to worklist. Consider the inner loops as simple nodes. Check that
290 * there is only one loop header in each loop. */
291 /*------------------------------------------------------------------*/
293 static double exception_prob = 0.001;
295 static INLINE int is_loop_head(ir_node *cond) {
299 /** Weight a single region in edge.
301 * Given all outs of the predecessor region, we can compute the weight of
302 * this single edge. */
303 static INLINE double get_weighted_region_exec_freq(void *reg, int pos) {
304 void *pred_reg = get_region_in(reg, pos);
305 double res, full_freq = get_region_exec_freq (pred_reg);
306 int n_outs = get_region_n_outs (pred_reg);
307 int n_exc_outs = get_region_n_exc_outs(pred_reg);
310 if (is_ir_node(reg)) {
311 cfop = get_Block_cfgpred((ir_node *)reg, pos);
312 if (is_Proj(cfop) && (get_irn_op(get_Proj_pred(cfop)) != op_Cond))
313 cfop = skip_Proj(cfop);
315 assert(is_ir_loop(reg));
316 cfop = get_loop_cfop(reg, pos);
319 if (is_fragile_op(cfop) || is_fragile_Proj(cfop)) {
320 res = full_freq * exception_prob;
323 /* Equally distribute the weight after exceptions to the left over outs. */
324 res = (full_freq *(1 - exception_prob * n_exc_outs)) / (n_outs - n_exc_outs);
330 static INLINE void compute_region_freqency(void *reg, double head_weight) {
331 int i, n_ins = get_region_n_ins(reg);
334 //printf("head weight %lf: ", head_weight); DDMR(reg);
336 for (i = 0; i < n_ins; ++i) {
337 void *pred_reg = get_region_in(reg, i);
339 my_freq += get_weighted_region_exec_freq(reg, i);
343 if (my_freq == 0.0) {
344 /* All preds are from outer loop. We are a head or so. */
345 my_freq = head_weight;
347 set_region_exec_freq(reg, my_freq);
350 static void check_proper_head(ir_loop *l, void *reg) {
351 int i, n_ins = get_region_n_ins(reg);
352 for (i = 0; i < n_ins; ++i) {
353 assert(!get_region_in(reg, i));
357 /* Compute the ex freq for current_ir_graph */
358 static void compute_frequency(int default_loop_weight) {
359 ir_loop *outermost_l = get_irg_loop(current_ir_graph);
360 pdeq *block_worklist = new_pdeq1(outermost_l);
362 /* Outermost start is considered a loop head. We will soon multiply
363 by default_loop_weight. */
364 set_region_exec_freq(outermost_l, 1.0/default_loop_weight);
366 while (!pdeq_empty(block_worklist)) {
367 ir_loop *l = (ir_loop *)pdeq_getl(block_worklist);
368 int i, n_elems = get_loop_n_elements(l);
370 /* The header is initialized with the frequency of the full loop times the iteration weight. */
371 check_proper_head(l, get_loop_element(l, 0).son);
373 for (i = 0; i < n_elems; ++i) {
374 loop_element e = get_loop_element(l, i);
375 if (is_ir_loop(e.son)) pdeq_putr(block_worklist, e.son);
376 compute_region_freqency(e.son, default_loop_weight * get_region_exec_freq(l));
379 del_pdeq(block_worklist);
382 /* Compute the execution frequency for all blocks in the given
385 * irg: The graph to be analyzed.
386 * default_loop_weight: The number of executions of a loop.
388 void compute_execution_frequency(ir_graph *irg, int default_loop_weight, double exception_probability) {
389 ir_graph *rem = current_ir_graph;
390 current_ir_graph = irg;
391 exception_prob = exception_probability;
392 if (!exec_freq_set) exec_freq_set = new_set(exec_freq_cmp, 256);
394 precompute_cond_evaluation();
395 construct_intervals(current_ir_graph);
396 compute_frequency(default_loop_weight);
398 set_irg_exec_freq_state(irg, exec_freq_consistent);
399 if (get_irp_exec_freq_state() == exec_freq_none)
400 set_irp_exec_freq_state(exec_freq_inconsistent);
403 dump_loop_tree (current_ir_graph, "-execfreq");
404 dump_ir_block_graph(current_ir_graph, "-execfreq");
405 dump_interval_graph(current_ir_graph, "-execfreq");
408 current_ir_graph = rem;
412 void compute_execution_frequencies(int default_loop_weight, double exception_probability) {
413 int i, n_irgs = get_irp_n_irgs();
415 for (i = 0; i < n_irgs; ++i) {
416 compute_execution_frequency(get_irp_irg(i), default_loop_weight, exception_probability);
418 set_irp_exec_freq_state(exec_freq_consistent);
421 /** free occupied memory, reset */
422 void free_execution_frequency(void) {
423 int i, n_irgs = get_irp_n_irgs();
425 del_set(exec_freq_set);
427 for (i = 0; i < n_irgs; ++i)
428 set_irg_exec_freq_state(get_irp_irg(i), exec_freq_none);
429 set_irp_exec_freq_state(exec_freq_none);
432 exec_freq_state get_irg_exec_freq_state(ir_graph *irg) {
433 return irg->execfreq_state;
435 void set_irg_exec_freq_state(ir_graph *irg, exec_freq_state s) {
436 if ((get_irp_exec_freq_state() == exec_freq_consistent && s != exec_freq_consistent) ||
437 (get_irp_exec_freq_state() == exec_freq_none && s != exec_freq_none))
438 irp->execfreq_state = exec_freq_inconsistent;
439 irg->execfreq_state = s;
442 /* Sets irg and irp exec freq state to inconsistent if it is set to consistent. */
443 void set_irg_exec_freq_state_inconsistent(ir_graph *irg) {
444 if (get_irg_exec_freq_state(irg) == exec_freq_consistent)
445 set_irg_exec_freq_state(irg, exec_freq_inconsistent);
448 void set_irp_exec_freq_state(exec_freq_state s) {
449 irp->execfreq_state = s;
452 exec_freq_state get_irp_exec_freq_state(void) {
453 return irp->execfreq_state;
456 /* Sets irp and all irg exec freq states to inconsistent if it is set to consistent. */
457 void set_irp_exec_freq_state_inconsistent(void) {
458 if (get_irp_exec_freq_state() != exec_freq_none) {
459 int i, n_irgs = get_irp_n_irgs();
460 set_irp_exec_freq_state(exec_freq_inconsistent);
461 for (i = 0; i < n_irgs; ++i) {
462 ir_graph *irg = get_irp_irg(i);
463 if (get_irg_exec_freq_state(irg) != exec_freq_none)
464 irg->execfreq_state = exec_freq_inconsistent;