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 Copy node statistics.
23 * @author Daniel Grund
37 #include "iredges_t.h"
39 #include "irnodeset.h"
41 #include "bechordal_t.h"
43 #include "becopyopt_t.h"
44 #include "becopystat.h"
47 #include "beintlive_t.h"
49 #define DEBUG_LVL SET_LEVEL_1
50 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
53 #define MAX_CLS_SIZE 20
54 #define MAX_CLS_PHIS 20
57 * For an explanation of these values see the code of copystat_dump_pretty
60 /* FROM HERE: PROBLEM CHARACTERIZATION */
66 I_PHI_CNT, /* number of phi nodes */
67 I_PHI_ARG_CNT, /* number of arguments of phis */
68 I_PHI_ARG_SELF, /* number of arguments of phis being the phi itself */
69 I_PHI_ARG_CONST, /* number of arguments of phis being consts */
70 I_PHI_ARG_PRED, /* ... being defined in a cf-pred */
71 I_PHI_ARG_GLOB, /* ... being defined elsewhere */
73 I_PHI_ARITY_E = I_PHI_ARITY_S+MAX_ARITY,
76 I_CPY_CNT, /* number of copynodes */
79 I_CLS_CNT, /* number of phi classes */
80 I_CLS_IF_FREE, /* number of pc having no interference */
81 I_CLS_IF_MAX, /* number of possible interferences in all classes */
82 I_CLS_IF_CNT, /* number of actual interferences in all classes */
84 I_CLS_SIZE_E = I_CLS_SIZE_S+MAX_CLS_SIZE,
86 I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
88 /* FROM HERE: RESULT VLAUES */
89 /* all of them are external set */
92 I_HEUR_TIME, /* solving time in milli seconds */
93 I_ILP_TIME, /* solving time in milli seconds */
96 I_ILP_ITER, /* number of simplex iterations */
98 /* copy instructions */
99 I_COPIES_MAX, /* max possible costs of copies*/
100 I_COPIES_INIT, /* number of copies in initial allocation */
101 I_COPIES_HEUR, /* number of copies after heuristic */
102 I_COPIES_5SEC, /* number of copies after ilp with max n sec */
103 I_COPIES_30SEC, /* number of copies after ilp with max n sec */
104 I_COPIES_OPT, /* number of copies after ilp */
105 I_COPIES_IF, /* number of copies inevitable due to root-arg-interf */
111 * Holds current values. Values are added till next copystat_reset
113 int curr_vals[ASIZE];
115 static ir_nodeset_t *all_phi_nodes;
116 static ir_nodeset_t *all_copy_nodes;
117 static ir_graph *last_irg;
119 void be_init_copystat(void) {
120 FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
122 all_phi_nodes = ir_nodeset_new(64);
123 all_copy_nodes = ir_nodeset_new(64);
124 memset(curr_vals, 0, sizeof(curr_vals));
126 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copystat);
128 void be_quit_copystat(void) {
129 ir_nodeset_del(all_phi_nodes);
130 ir_nodeset_del(all_copy_nodes);
132 BE_REGISTER_MODULE_DESTRUCTOR(be_quit_copystat);
134 void copystat_reset(void) {
140 * Collect general data
142 static void irg_stat_walker(ir_node *node, void *env) {
143 arch_env_t *arch_env = env;
144 curr_vals[I_ALL_NODES]++; /* count all nodes */
146 if (is_Block(node)) /* count all blocks */
147 curr_vals[I_BLOCKS]++;
149 if (is_Reg_Phi(node)) /* collect phis */
150 ir_nodeset_insert(all_phi_nodes, node);
152 if (is_Perm_Proj(arch_env, node))
153 ir_nodeset_insert(all_copy_nodes, node);
155 /* TODO: Add 2-Addr-Code nodes */
158 static void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
159 irg_walk_graph(irg, irg_stat_walker, NULL, arch_env);
164 * @return 1 if the block at pos @p pos removed a critical edge
167 static INLINE int was_edge_critical(const ir_node *bl, int pos) {
168 const ir_edge_t *edge;
169 const ir_node *bl_at_pos, *bl_before;
170 assert(is_Block(bl));
172 /* Does bl have several predecessors ?*/
173 if (get_irn_arity(bl) <= 1)
176 /* Does the pred have exactly one predecessor */
177 bl_at_pos = get_irn_n(bl, pos);
178 if (get_irn_arity(bl_at_pos) != 1)
181 /* Does the pred of the pred have several successors */
182 bl_before = get_irn_n(bl_at_pos, 0);
183 edge = get_block_succ_first(bl_before);
184 return get_block_succ_next(bl_before, edge) ? 1 : 0;
188 * Collect phi node data
190 static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi)
197 /* count all phi phis */
198 curr_vals[I_PHI_CNT]++;
201 arity = get_irn_arity(phi);
202 curr_vals[I_PHI_ARG_CNT] += arity;
203 if (arity > MAX_ARITY)
204 curr_vals[I_PHI_ARITY_E]++;
206 curr_vals[I_PHI_ARITY_S + arity]++;
208 phi_bl = get_nodes_block(phi);
209 /* type of argument {self, const, pred, glob} */
210 for (i = 0; i < arity; i++) {
211 ir_node *block_of_arg, *block_ith_pred;
212 ir_node *arg = get_irn_n(phi, i);
215 curr_vals[I_PHI_ARG_SELF]++;
219 if (iro_Const == get_irn_opcode(arg)) {
220 curr_vals[I_PHI_ARG_CONST]++;
224 /* get the pred block skipping blocks on critical edges */
225 block_ith_pred = get_Block_cfgpred_block(phi_bl, i);
226 if (was_edge_critical(phi_bl, i))
227 block_ith_pred = get_Block_cfgpred_block(block_ith_pred, 0);
229 block_of_arg = get_nodes_block(arg);
230 if (block_of_arg == block_ith_pred) {
231 curr_vals[I_PHI_ARG_PRED]++;
235 curr_vals[I_PHI_ARG_GLOB]++;
240 * Collect register-constrained node data
242 static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
243 curr_vals[I_CPY_CNT]++;
244 curr_vals[I_COPIES_MAX]++;
245 if (values_interfere(chordal_env->birg, root, get_Perm_src(root))) {
246 curr_vals[I_COPIES_IF]++;
247 assert(0 && "A Perm pair (in/out) should never interfere!");
252 * Collect phi class data
254 static void stat_phi_class(be_chordal_env_t *chordal_env, ir_node **pc) {
255 int i, o, size, if_free, phis;
257 /* phi class count */
258 curr_vals[I_CLS_CNT]++;
262 if (size > MAX_CLS_SIZE)
263 curr_vals[I_CLS_SIZE_E]++;
265 curr_vals[I_CLS_SIZE_S + size]++;
267 /* determine number of phis on this class */
268 for (phis = i = 0; i < size; ++i)
272 if (phis > MAX_CLS_PHIS)
273 curr_vals[I_CLS_PHIS_E]++;
275 curr_vals[I_CLS_PHIS_S + phis]++;
277 /* determine interference of phi class members */
278 curr_vals[I_CLS_IF_MAX] += size * (size - 1) / 2;
279 for (if_free = 1, i = 0; i < size - 1; ++i)
280 for (o = i + 1; o < size; ++o)
281 if (values_interfere(chordal_env->birg, pc[i], pc[o])) {
283 curr_vals[I_CLS_IF_CNT]++;
286 /* Does this phi class have an inner interference? */
287 curr_vals[I_CLS_IF_FREE] += if_free;
290 static void copystat_collect_cls(be_chordal_env_t *cenv) {
291 ir_graph *irg = cenv->irg;
292 arch_env_t *aenv = cenv->birg->main_env->arch_env;
294 phi_classes_t *pc_obj;
295 pset *all_phi_classes;
296 ir_nodeset_iterator_t iter;
299 copystat_collect_irg(irg, aenv);
301 /* compute the Phi classes of the collected Phis */
302 pc_obj = phi_class_new_from_set(cenv->irg, all_phi_nodes, 0);
303 all_phi_classes = get_all_phi_classes(pc_obj);
305 foreach_ir_nodeset(all_phi_nodes, n, iter) {
306 if (arch_get_irn_reg_class(n, -1) == cenv->cls)
307 stat_phi_node(cenv, n);
310 foreach_ir_nodeset(all_copy_nodes, n, iter) {
311 if (arch_get_irn_reg_class(n, -1) == cenv->cls)
312 stat_copy_node(cenv, n);
315 foreach_pset(all_phi_classes, pc) {
316 ir_node *member = pc[0];
317 if (arch_get_irn_reg_class(member, -1) == cenv->cls)
318 stat_phi_class(cenv, pc);
321 /* free the phi class object */
322 phi_class_free(pc_obj);
325 void copystat_add_max_costs(int costs) {
326 curr_vals[I_COPIES_MAX] += costs;
328 void copystat_add_inevit_costs(int costs) {
329 curr_vals[I_COPIES_IF] += costs;
331 void copystat_add_init_costs(int costs) {
332 curr_vals[I_COPIES_INIT] += costs;
334 void copystat_add_heur_costs(int costs) {
335 curr_vals[I_COPIES_HEUR] += costs;
337 void copystat_add_ilp_5_sec_costs(int costs) {
338 curr_vals[I_COPIES_5SEC] += costs;
340 void copystat_add_ilp_30_sec_costs(int costs) {
341 curr_vals[I_COPIES_30SEC] += costs;
343 void copystat_add_opt_costs(int costs) {
344 curr_vals[I_COPIES_OPT] += costs;
346 void copystat_add_heur_time(int time) {
347 curr_vals[I_HEUR_TIME] += time;
352 void copystat_add_ilp_time(int time) {
353 curr_vals[I_ILP_TIME] += time;
355 void copystat_add_ilp_vars(int vars) {
356 curr_vals[I_ILP_VARS] += vars;
358 void copystat_add_ilp_csts(int csts) {
359 curr_vals[I_ILP_CSTR] += csts;
361 void copystat_add_ilp_iter(int iters) {
362 curr_vals[I_ILP_ITER] += iters;
365 #endif /* WITH_ILP */
367 void copystat_dump(ir_graph *irg) {
372 snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
373 buf[sizeof(buf) - 1] = '\0';
374 out = be_ffopen(buf, "stat", "wt");
376 fprintf(out, "%d\n", ASIZE);
377 for (i = 0; i < ASIZE; i++) {
379 if (i >= I_PHI_ARITY_S && i <= I_PHI_ARITY_E)
380 fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_PHI_CNT]);
381 else if (i >= I_CLS_SIZE_S && i <= I_CLS_SIZE_E)
382 fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_CLS_CNT]);
385 fprintf(out, "%i\n", curr_vals[i]);
391 void copystat_dump_pretty(ir_graph *irg) {
396 snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
397 buf[sizeof(buf) - 1] = '\0';
398 out = be_ffopen(buf, "pstat", "wt");
400 fprintf(out, "Nodes %4d\n", curr_vals[I_ALL_NODES]);
401 fprintf(out, "Blocks %4d\n", curr_vals[I_BLOCKS]);
402 fprintf(out, "CopyIrn %4d\n", curr_vals[I_CPY_CNT]);
404 fprintf(out, "\nPhis %4d\n", curr_vals[I_PHI_CNT]);
405 fprintf(out, "... argument types\n");
406 fprintf(out, " Total %4d\n", curr_vals[I_PHI_ARG_CNT]);
407 fprintf(out, " Self %4d\n", curr_vals[I_PHI_ARG_SELF]);
408 fprintf(out, " Constants %4d\n", curr_vals[I_PHI_ARG_CONST]);
409 fprintf(out, " CF-Pred %4d\n", curr_vals[I_PHI_ARG_PRED]);
410 fprintf(out, " Others %4d\n", curr_vals[I_PHI_ARG_GLOB]);
411 fprintf(out, "... arities\n");
412 for (i = I_PHI_ARITY_S; i<=I_PHI_ARITY_E; i++)
413 fprintf(out, " %2i %4d\n", i-I_PHI_ARITY_S, curr_vals[i]);
415 fprintf(out, "\nPhi classes %4d\n", curr_vals[I_CLS_CNT]);
416 fprintf(out, " compl. free %4d\n", curr_vals[I_CLS_IF_FREE]);
417 fprintf(out, " inner intf. %4d / %4d\n", curr_vals[I_CLS_IF_CNT], curr_vals[I_CLS_IF_MAX]);
418 fprintf(out, "... sizes\n");
419 for (i = I_CLS_SIZE_S; i<=I_CLS_SIZE_E; i++)
420 fprintf(out, " %2i %4d\n", i-I_CLS_SIZE_S, curr_vals[i]);
421 fprintf(out, "... contained phis\n");
422 for (i = I_CLS_PHIS_S; i<=I_CLS_PHIS_E; i++)
423 fprintf(out, " %2i %4d\n", i-I_CLS_PHIS_S, curr_vals[i]);
425 fprintf(out, "\nILP stat\n");
426 fprintf(out, " Time %8d\n", curr_vals[I_ILP_TIME]);
427 fprintf(out, " Iter %8d\n", curr_vals[I_ILP_ITER]);
429 fprintf(out, "\nCopy stat\n");
430 fprintf(out, " Max %4d\n", curr_vals[I_COPIES_MAX]);
431 fprintf(out, " Init %4d\n", curr_vals[I_COPIES_INIT]);
432 fprintf(out, " Heur %4d\n", curr_vals[I_COPIES_HEUR]);
433 fprintf(out, " Opt %4d\n", curr_vals[I_COPIES_OPT]);
434 fprintf(out, " Intf %4d\n", curr_vals[I_COPIES_IF]);
440 * Helpers for saving and restoring colors of nodes.
441 * Used to get dependable and comparable benchmark results.
443 typedef struct color_saver {
444 arch_env_t *arch_env;
445 be_chordal_env_t *chordal_env;
447 int flag; /* 0 save, 1 load */
450 static void save_load(ir_node *irn, void *env) {
451 color_save_t *saver = env;
452 if (saver->chordal_env->cls == arch_get_irn_reg_class(irn, -1)) {
453 if (saver->flag == 0) { /* save */
454 const arch_register_t *reg = arch_get_irn_register(irn);
455 pmap_insert(saver->saved_colors, irn, (void *) reg);
457 arch_register_t *reg = pmap_get(saver->saved_colors, irn);
458 arch_set_irn_register(saver->arch_env, irn, reg);
463 static void save_colors(color_save_t *color_saver) {
464 color_saver->flag = 0;
465 irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
469 static void load_colors(color_save_t *color_saver) {
470 color_saver->flag = 1;
471 irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
476 * Main compare routine
478 void co_compare_solvers(be_chordal_env_t *chordal_env) {
482 int costs_inevit, costs_init, costs_solved, lower_bound;
484 copystat_collect_cls(chordal_env);
486 co = new_copy_opt(chordal_env, co_get_costs_loop_depth);
487 co_build_ou_structure(co);
488 co_build_graph_structure(co);
489 DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
492 saver.arch_env = chordal_env->birg->main_env->arch_env;
493 saver.chordal_env = chordal_env;
494 saver.saved_colors = pmap_create();
498 costs_inevit = co_get_inevit_copy_costs(co);
499 lower_bound = co_get_lower_bound(co);
500 costs_init = co_get_copy_costs(co);
502 DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
503 DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
504 DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
506 copystat_add_inevit_costs(costs_inevit);
507 copystat_add_init_costs(costs_init);
508 copystat_add_max_costs(co_get_max_copy_costs(co));
510 /* heuristic 1 (Daniel Grund) */
511 timer = ir_timer_register("heur1", NULL);
512 ir_timer_reset_and_start(timer);
514 co_solve_heuristic(co);
516 ir_timer_stop(timer);
518 costs_solved = co_get_copy_costs(co);
519 DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
520 copystat_add_heur_time(ir_timer_elapsed_msec(timer));
521 copystat_add_heur_costs(costs_solved);
522 assert(lower_bound <= costs_solved);
524 /* heuristic 2 (Sebastian Hack) */
525 timer = ir_timer_register("heur2", NULL);
526 ir_timer_reset_and_start(timer);
528 co_solve_heuristic_new(co);
530 ir_timer_stop(timer);
532 costs_solved = co_get_copy_costs(co);
533 DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
534 copystat_add_heur_time(ir_timer_elapsed_msec(timer));
535 copystat_add_heur_costs(costs_solved);
536 assert(lower_bound <= costs_solved);
538 /* Park & Moon register coalescing (Kimon Hoffmann) */
539 timer = ir_timer_register("park", NULL);
540 ir_timer_reset_and_start(timer);
542 co_solve_park_moon(co);
544 ir_timer_stop(timer);
546 costs_solved = co_get_copy_costs(co);
547 DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
548 copystat_add_heur_time(ir_timer_elapsed_msec(timer));
549 copystat_add_heur_costs(costs_solved);
550 assert(lower_bound <= costs_solved);
555 /* ILP 1 is not yet implemented, so it makes no sense to compare */
559 co_solve_ilp1(co, 60.0);
561 costs_solved = co_get_copy_costs(co);
562 DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
563 copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
564 assert(lower_bound <= costs_solved);
572 costs_solved = co_get_copy_costs(co);
573 DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
574 copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
575 assert(lower_bound <= costs_solved);
577 #endif /* WITH_ILP */
579 /* free memory for statistic structures */
580 pmap_destroy(saver.saved_colors);
581 co_free_graph_structure(co);
582 co_free_ou_structure(co);