- stack alignment is now power of two
[libfirm] / ir / be / becopystat.c
1 /*
2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
3  *
4  * This file is part of libFirm.
5  *
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.
10  *
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.
14  *
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
17  * PURPOSE.
18  */
19
20 /**
21  * @file
22  * @brief       Copy node statistics.
23  * @author      Daniel Grund
24  * @date        19.04.2005
25  * @version     $Id$
26  */
27 #ifdef HAVE_CONFIG_H
28 #include "config.h"
29 #endif
30
31 #include <string.h>
32
33 #include "timing.h"
34 #include "xmalloc.h"
35 #include "irgraph.h"
36 #include "irgwalk.h"
37 #include "irprog.h"
38 #include "iredges_t.h"
39 #include "phiclass.h"
40 #include "irnodeset.h"
41
42 #include "bechordal_t.h"
43 #include "beutil.h"
44 #include "becopyopt_t.h"
45 #include "becopystat.h"
46 #include "beirg_t.h"
47 #include "bemodule.h"
48 #include "beintlive_t.h"
49
50 #define DEBUG_LVL SET_LEVEL_1
51 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
52
53 #define MAX_ARITY 20
54 #define MAX_CLS_SIZE 20
55 #define MAX_CLS_PHIS 20
56
57 /**
58  * For an explanation of these values see the code of copystat_dump_pretty
59  */
60 enum vals_t {
61         /* FROM HERE: PROBLEM CHARACTERIZATION */
62
63         I_ALL_NODES = 0,
64         I_BLOCKS,
65
66         /* phi nodes */
67         I_PHI_CNT,                      /* number of phi nodes */
68         I_PHI_ARG_CNT,          /* number of arguments of phis */
69         I_PHI_ARG_SELF,         /* number of arguments of phis being the phi itself */
70         I_PHI_ARG_CONST,        /* number of arguments of phis being consts */
71         I_PHI_ARG_PRED,         /* ... being defined in a cf-pred */
72         I_PHI_ARG_GLOB,         /* ... being defined elsewhere */
73         I_PHI_ARITY_S,
74         I_PHI_ARITY_E    = I_PHI_ARITY_S+MAX_ARITY,
75
76         /* copy nodes */
77         I_CPY_CNT,                      /* number of copynodes */
78
79         /* phi classes */
80         I_CLS_CNT,                      /* number of phi classes */
81         I_CLS_IF_FREE,          /* number of pc having no interference */
82         I_CLS_IF_MAX,           /* number of possible interferences in all classes */
83         I_CLS_IF_CNT,           /* number of actual interferences in all classes */
84         I_CLS_SIZE_S,
85         I_CLS_SIZE_E = I_CLS_SIZE_S+MAX_CLS_SIZE,
86         I_CLS_PHIS_S,
87         I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
88
89         /* FROM HERE: RESULT VLAUES */
90         /* all of them are external set */
91
92         /* ilp values */
93         I_HEUR_TIME,            /* solving time in milli seconds */
94         I_ILP_TIME,                     /* solving time in milli seconds */
95         I_ILP_VARS,
96         I_ILP_CSTR,
97         I_ILP_ITER,                     /* number of simplex iterations */
98
99         /* copy instructions */
100         I_COPIES_MAX,           /* max possible costs of copies*/
101         I_COPIES_INIT,          /* number of copies in initial allocation */
102         I_COPIES_HEUR,          /* number of copies after heuristic */
103         I_COPIES_5SEC,          /* number of copies after ilp with max n sec */
104         I_COPIES_30SEC,         /* number of copies after ilp with max n sec */
105         I_COPIES_OPT,           /* number of copies after ilp */
106         I_COPIES_IF,            /* number of copies inevitable due to root-arg-interf */
107
108         ASIZE
109 };
110
111 /**
112  * Holds current values. Values are added till next copystat_reset
113  */
114 int curr_vals[ASIZE];
115
116 static ir_nodeset_t *all_phi_nodes;
117 static ir_nodeset_t *all_copy_nodes;
118 static ir_graph *last_irg;
119
120 void be_init_copystat(void) {
121         FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
122
123         all_phi_nodes  = ir_nodeset_new(64);
124         all_copy_nodes = ir_nodeset_new(64);
125         memset(curr_vals, 0, sizeof(curr_vals));
126 }
127 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copystat);
128
129 void be_quit_copystat(void) {
130         ir_nodeset_del(all_phi_nodes);
131         ir_nodeset_del(all_copy_nodes);
132 }
133 BE_REGISTER_MODULE_DESTRUCTOR(be_quit_copystat);
134
135 void copystat_reset(void) {
136         be_quit_copystat();
137         be_init_copystat();
138 }
139
140 /**
141  * Collect general data
142  */
143 static void irg_stat_walker(ir_node *node, void *env) {
144         arch_env_t *arch_env = env;
145         curr_vals[I_ALL_NODES]++; /* count all nodes */
146
147         if (is_Block(node)) /* count all blocks */
148                 curr_vals[I_BLOCKS]++;
149
150         if (is_Reg_Phi(node)) /* collect phis */
151                 ir_nodeset_insert(all_phi_nodes, node);
152
153         if (is_Perm_Proj(arch_env, node))
154                 ir_nodeset_insert(all_copy_nodes, node);
155
156         /* TODO: Add 2-Addr-Code nodes */
157 }
158
159 static void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
160         irg_walk_graph(irg, irg_stat_walker, NULL, arch_env);
161         last_irg = irg;
162 }
163
164 /**
165  * @return 1 if the block at pos @p pos removed a critical edge
166  *                 0 else
167  */
168 static INLINE int was_edge_critical(const ir_node *bl, int pos) {
169         const ir_edge_t *edge;
170         const ir_node *bl_at_pos, *bl_before;
171         assert(is_Block(bl));
172
173         /* Does bl have several predecessors ?*/
174         if (get_irn_arity(bl) <= 1)
175                 return 0;
176
177         /* Does the pred have exactly one predecessor */
178         bl_at_pos = get_irn_n(bl, pos);
179         if (get_irn_arity(bl_at_pos) != 1)
180                 return 0;
181
182         /* Does the pred of the pred have several successors */
183         bl_before = get_irn_n(bl_at_pos, 0);
184         edge = get_block_succ_first(bl_before);
185         return get_block_succ_next(bl_before, edge) ? 1 : 0;
186 }
187
188 /**
189  * Collect phi node data
190  */
191 static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi)
192 {
193         int arity, i;
194         ir_node *phi_bl;
195         assert(is_Phi(phi));
196         (void) chordal_env;
197
198         /* count all phi phis */
199         curr_vals[I_PHI_CNT]++;
200
201         /* argument count */
202         arity = get_irn_arity(phi);
203         curr_vals[I_PHI_ARG_CNT] += arity;
204         if (arity > MAX_ARITY)
205                 curr_vals[I_PHI_ARITY_E]++;
206         else
207                 curr_vals[I_PHI_ARITY_S + arity]++;
208
209         phi_bl = get_nodes_block(phi);
210         /* type of argument {self, const, pred, glob} */
211         for (i = 0; i < arity; i++) {
212         ir_node *block_of_arg, *block_ith_pred;
213                 ir_node *arg = get_irn_n(phi, i);
214
215                 if (arg == phi) {
216                         curr_vals[I_PHI_ARG_SELF]++;
217                         continue;
218                 }
219
220                 if (iro_Const == get_irn_opcode(arg)) {
221                         curr_vals[I_PHI_ARG_CONST]++;
222                         continue;
223                 }
224
225                 /* get the pred block skipping blocks on critical edges */
226                 block_ith_pred = get_Block_cfgpred_block(phi_bl, i);
227                 if (was_edge_critical(phi_bl, i))
228                         block_ith_pred = get_Block_cfgpred_block(block_ith_pred, 0);
229
230                 block_of_arg = get_nodes_block(arg);
231                 if (block_of_arg == block_ith_pred) {
232                         curr_vals[I_PHI_ARG_PRED]++;
233                         continue;
234                 }
235
236                 curr_vals[I_PHI_ARG_GLOB]++;
237         }
238 }
239
240 /**
241  * Collect register-constrained node data
242  */
243 static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
244         curr_vals[I_CPY_CNT]++;
245         curr_vals[I_COPIES_MAX]++;
246         if (values_interfere(chordal_env->birg, root, get_Perm_src(root))) {
247                 curr_vals[I_COPIES_IF]++;
248                 assert(0 && "A Perm pair (in/out) should never interfere!");
249         }
250 }
251
252 /**
253  * Collect phi class data
254  */
255 static void stat_phi_class(be_chordal_env_t *chordal_env, ir_node **pc) {
256         int i, o, size, if_free, phis;
257
258         /* phi class count */
259         curr_vals[I_CLS_CNT]++;
260
261         /* phi class size */
262         size = ARR_LEN(pc);
263         if (size > MAX_CLS_SIZE)
264                 curr_vals[I_CLS_SIZE_E]++;
265         else
266                 curr_vals[I_CLS_SIZE_S + size]++;
267
268         /* determine number of phis on this class */
269         for (phis = i = 0; i < size; ++i)
270                 if (is_Phi(pc[i]))
271                         phis++;
272
273         if (phis > MAX_CLS_PHIS)
274                 curr_vals[I_CLS_PHIS_E]++;
275         else
276                 curr_vals[I_CLS_PHIS_S + phis]++;
277
278         /* determine interference of phi class members */
279         curr_vals[I_CLS_IF_MAX] += size * (size - 1) / 2;
280         for (if_free = 1, i = 0; i < size - 1; ++i)
281                 for (o = i + 1; o < size; ++o)
282                         if (values_interfere(chordal_env->birg, pc[i], pc[o])) {
283                                 if_free = 0;
284                                 curr_vals[I_CLS_IF_CNT]++;
285                         }
286
287         /* Does this phi class have an inner interference? */
288         curr_vals[I_CLS_IF_FREE] += if_free;
289 }
290
291 static void copystat_collect_cls(be_chordal_env_t *cenv) {
292         ir_graph              *irg  = cenv->irg;
293         arch_env_t            *aenv = cenv->birg->main_env->arch_env;
294         ir_node               *n, **pc;
295         phi_classes_t         *pc_obj;
296         pset                  *all_phi_classes;
297         ir_nodeset_iterator_t iter;
298
299         copystat_reset();
300         copystat_collect_irg(irg, aenv);
301
302         /* compute the Phi classes of the collected Phis */
303         pc_obj          = phi_class_new_from_set(cenv->irg, all_phi_nodes, 0);
304         all_phi_classes = get_all_phi_classes(pc_obj);
305
306         foreach_ir_nodeset(all_phi_nodes, n, iter) {
307                 if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
308                         stat_phi_node(cenv, n);
309         }
310
311         foreach_ir_nodeset(all_copy_nodes, n, iter) {
312                 if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
313                         stat_copy_node(cenv, n);
314         }
315
316         foreach_pset(all_phi_classes, pc) {
317                 ir_node *member = pc[0];
318                 if (arch_get_irn_reg_class(aenv, member, -1) == cenv->cls)
319                         stat_phi_class(cenv, pc);
320         }
321
322         /* free the phi class object */
323         phi_class_free(pc_obj);
324 }
325
326 void copystat_add_max_costs(int costs) {
327         curr_vals[I_COPIES_MAX] += costs;
328 }
329 void copystat_add_inevit_costs(int costs) {
330         curr_vals[I_COPIES_IF] += costs;
331 }
332 void copystat_add_init_costs(int costs) {
333         curr_vals[I_COPIES_INIT] += costs;
334 }
335 void copystat_add_heur_costs(int costs) {
336         curr_vals[I_COPIES_HEUR] += costs;
337 }
338 void copystat_add_ilp_5_sec_costs(int costs) {
339         curr_vals[I_COPIES_5SEC] += costs;
340 }
341 void copystat_add_ilp_30_sec_costs(int costs) {
342         curr_vals[I_COPIES_30SEC] += costs;
343 }
344 void copystat_add_opt_costs(int costs) {
345         curr_vals[I_COPIES_OPT] += costs;
346 }
347 void copystat_add_heur_time(int time) {
348         curr_vals[I_HEUR_TIME] += time;
349 }
350
351 #ifdef WITH_ILP
352
353 void copystat_add_ilp_time(int time) {
354         curr_vals[I_ILP_TIME] += time;
355 }
356 void copystat_add_ilp_vars(int vars) {
357         curr_vals[I_ILP_VARS] += vars;
358 }
359 void copystat_add_ilp_csts(int csts) {
360         curr_vals[I_ILP_CSTR] += csts;
361 }
362 void copystat_add_ilp_iter(int iters) {
363         curr_vals[I_ILP_ITER] += iters;
364 }
365
366 #endif /* WITH_ILP */
367
368 void copystat_dump(ir_graph *irg) {
369         int i;
370         char buf[1024];
371         FILE *out;
372
373         snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
374         buf[sizeof(buf) - 1] = '\0';
375         out = be_ffopen(buf, "stat", "wt");
376
377         fprintf(out, "%d\n", ASIZE);
378         for (i = 0; i < ASIZE; i++) {
379 #if 0
380                 if (i >= I_PHI_ARITY_S && i <= I_PHI_ARITY_E)
381                         fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_PHI_CNT]);
382                 else if (i >= I_CLS_SIZE_S && i <= I_CLS_SIZE_E)
383                         fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_CLS_CNT]);
384                 else
385 #endif
386                         fprintf(out, "%i\n", curr_vals[i]);
387         }
388
389         fclose(out);
390 }
391
392 void copystat_dump_pretty(ir_graph *irg) {
393         int i;
394         char buf[1024];
395         FILE *out;
396
397         snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
398         buf[sizeof(buf) - 1] = '\0';
399         out = be_ffopen(buf, "pstat", "wt");
400
401         fprintf(out, "Nodes     %4d\n", curr_vals[I_ALL_NODES]);
402         fprintf(out, "Blocks    %4d\n", curr_vals[I_BLOCKS]);
403         fprintf(out, "CopyIrn   %4d\n", curr_vals[I_CPY_CNT]);
404
405         fprintf(out, "\nPhis      %4d\n", curr_vals[I_PHI_CNT]);
406         fprintf(out, "... argument types\n");
407         fprintf(out, " Total      %4d\n", curr_vals[I_PHI_ARG_CNT]);
408         fprintf(out, " Self       %4d\n", curr_vals[I_PHI_ARG_SELF]);
409         fprintf(out, " Constants  %4d\n", curr_vals[I_PHI_ARG_CONST]);
410         fprintf(out, " CF-Pred    %4d\n", curr_vals[I_PHI_ARG_PRED]);
411         fprintf(out, " Others     %4d\n", curr_vals[I_PHI_ARG_GLOB]);
412         fprintf(out, "... arities\n");
413         for (i = I_PHI_ARITY_S; i<=I_PHI_ARITY_E; i++)
414                 fprintf(out, " %2i %4d\n", i-I_PHI_ARITY_S, curr_vals[i]);
415
416         fprintf(out, "\nPhi classes   %4d\n", curr_vals[I_CLS_CNT]);
417         fprintf(out, " compl. free  %4d\n", curr_vals[I_CLS_IF_FREE]);
418         fprintf(out, " inner intf.  %4d / %4d\n", curr_vals[I_CLS_IF_CNT], curr_vals[I_CLS_IF_MAX]);
419         fprintf(out, "... sizes\n");
420         for (i = I_CLS_SIZE_S; i<=I_CLS_SIZE_E; i++)
421                 fprintf(out, " %2i %4d\n", i-I_CLS_SIZE_S, curr_vals[i]);
422         fprintf(out, "... contained phis\n");
423         for (i = I_CLS_PHIS_S; i<=I_CLS_PHIS_E; i++)
424                 fprintf(out, " %2i %4d\n", i-I_CLS_PHIS_S, curr_vals[i]);
425
426         fprintf(out, "\nILP stat\n");
427         fprintf(out, " Time %8d\n", curr_vals[I_ILP_TIME]);
428         fprintf(out, " Iter %8d\n", curr_vals[I_ILP_ITER]);
429
430         fprintf(out, "\nCopy stat\n");
431         fprintf(out, " Max  %4d\n", curr_vals[I_COPIES_MAX]);
432         fprintf(out, " Init %4d\n", curr_vals[I_COPIES_INIT]);
433         fprintf(out, " Heur %4d\n", curr_vals[I_COPIES_HEUR]);
434         fprintf(out, " Opt  %4d\n", curr_vals[I_COPIES_OPT]);
435         fprintf(out, " Intf %4d\n", curr_vals[I_COPIES_IF]);
436
437         fclose(out);
438 }
439
440 /**
441  * Helpers for saving and restoring colors of nodes.
442  * Used to get dependable and comparable benchmark results.
443  */
444 typedef struct color_saver {
445         arch_env_t *arch_env;
446         be_chordal_env_t *chordal_env;
447         pmap *saved_colors;
448         int flag; /* 0 save, 1 load */
449 } color_save_t;
450
451 static void save_load(ir_node *irn, void *env) {
452         color_save_t *saver = env;
453         if (saver->chordal_env->cls == arch_get_irn_reg_class(saver->arch_env, irn, -1)) {
454                 if (saver->flag == 0) { /* save */
455                         const arch_register_t *reg = arch_get_irn_register(saver->arch_env, irn);
456                         pmap_insert(saver->saved_colors, irn, (void *) reg);
457                 } else { /*load */
458                         arch_register_t *reg = pmap_get(saver->saved_colors, irn);
459                         arch_set_irn_register(saver->arch_env, irn, reg);
460                 }
461         }
462 }
463
464 static void save_colors(color_save_t *color_saver) {
465         color_saver->flag = 0;
466         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
467 }
468
469 #ifdef WITH_ILP
470 static void load_colors(color_save_t *color_saver) {
471         color_saver->flag = 1;
472         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
473 }
474 #endif
475
476 /**
477  * Main compare routine
478  */
479 void co_compare_solvers(be_chordal_env_t *chordal_env) {
480         copy_opt_t    *co;
481         ir_timer_t    *timer;
482         color_save_t  saver;
483         int costs_inevit, costs_init, costs_solved, lower_bound;
484
485         copystat_collect_cls(chordal_env);
486
487         co = new_copy_opt(chordal_env, co_get_costs_loop_depth);
488         co_build_ou_structure(co);
489         co_build_graph_structure(co);
490         DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
491
492         /* save colors */
493         saver.arch_env     = chordal_env->birg->main_env->arch_env;
494         saver.chordal_env  = chordal_env;
495         saver.saved_colors = pmap_create();
496         save_colors(&saver);
497
498         /* initial values */
499         costs_inevit = co_get_inevit_copy_costs(co);
500         lower_bound  = co_get_lower_bound(co);
501         costs_init   = co_get_copy_costs(co);
502
503         DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
504         DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
505         DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
506
507         copystat_add_inevit_costs(costs_inevit);
508         copystat_add_init_costs(costs_init);
509         copystat_add_max_costs(co_get_max_copy_costs(co));
510
511         /* heuristic 1 (Daniel Grund) */
512         timer = ir_timer_register("heur1", NULL);
513         ir_timer_reset_and_start(timer);
514
515         co_solve_heuristic(co);
516
517         ir_timer_stop(timer);
518
519         costs_solved = co_get_copy_costs(co);
520         DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
521         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
522         copystat_add_heur_costs(costs_solved);
523         assert(lower_bound <= costs_solved);
524
525         /* heuristic 2 (Sebastian Hack) */
526         timer = ir_timer_register("heur2", NULL);
527         ir_timer_reset_and_start(timer);
528
529         co_solve_heuristic_new(co);
530
531         ir_timer_stop(timer);
532
533         costs_solved = co_get_copy_costs(co);
534         DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
535         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
536         copystat_add_heur_costs(costs_solved);
537         assert(lower_bound <= costs_solved);
538
539         /* Park & Moon register coalescing (Kimon Hoffmann) */
540         timer = ir_timer_register("park", NULL);
541         ir_timer_reset_and_start(timer);
542
543         co_solve_park_moon(co);
544
545         ir_timer_stop(timer);
546
547         costs_solved = co_get_copy_costs(co);
548         DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
549         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
550         copystat_add_heur_costs(costs_solved);
551         assert(lower_bound <= costs_solved);
552
553
554 #ifdef WITH_ILP
555
556         /* ILP 1 is not yet implemented, so it makes no sense to compare */
557 #if 0
558         load_colors(&saver);
559
560         co_solve_ilp1(co, 60.0);
561
562         costs_solved = co_get_copy_costs(co);
563         DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
564         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
565         assert(lower_bound <= costs_solved);
566 #endif /* 0 */
567
568         /* ILP 2 */
569         load_colors(&saver);
570
571         co_solve_ilp2(co);
572
573         costs_solved = co_get_copy_costs(co);
574         DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
575         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
576         assert(lower_bound <= costs_solved);
577
578 #endif /* WITH_ILP */
579
580         /* free memory for statistic structures */
581         pmap_destroy(saver.saved_colors);
582         co_free_graph_structure(co);
583         co_free_ou_structure(co);
584         free_copy_opt(co);
585 }