fix belistsched for new scheduling API
[libfirm] / ir / be / becopystat.c
1 /**
2  * Author:      Daniel Grund
3  * Date:                19.04.2005
4  * Copyright:   (c) Universitaet Karlsruhe
5  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
6  */
7 #ifdef HAVE_CONFIG_H
8 #include "config.h"
9 #endif
10
11 #include <string.h>
12 #include <libcore/lc_timing.h>
13
14 #include "xmalloc.h"
15 #include "irgraph.h"
16 #include "irgwalk.h"
17 #include "irprog.h"
18 #include "iredges_t.h"
19 #include "phiclass_t.h"
20 #include "bechordal_t.h"
21 #include "beutil.h"
22 #include "becopyopt_t.h"
23 #include "becopystat.h"
24 #include "beirg_t.h"
25
26 #define DEBUG_LVL SET_LEVEL_1
27 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
28
29 #define MAX_ARITY 20
30 #define MAX_CLS_SIZE 20
31 #define MAX_CLS_PHIS 20
32
33 /**
34  * For an explanation of these values see the code of copystat_dump_pretty
35  */
36 enum vals_t {
37         /* FROM HERE: PROBLEM CHARACTERIZATION */
38
39         I_ALL_NODES = 0,
40         I_BLOCKS,
41
42         /* phi nodes */
43         I_PHI_CNT,                      /* number of phi nodes */
44         I_PHI_ARG_CNT,          /* number of arguments of phis */
45         I_PHI_ARG_SELF,         /* number of arguments of phis being the phi itself */
46         I_PHI_ARG_CONST,        /* number of arguments of phis being consts */
47         I_PHI_ARG_PRED,         /* ... being defined in a cf-pred */
48         I_PHI_ARG_GLOB,         /* ... being defined elsewhere */
49         I_PHI_ARITY_S,
50         I_PHI_ARITY_E    = I_PHI_ARITY_S+MAX_ARITY,
51
52         /* copy nodes */
53         I_CPY_CNT,                      /* number of copynodes */
54
55         /* phi classes */
56         I_CLS_CNT,                      /* number of phi classes */
57         I_CLS_IF_FREE,          /* number of pc having no interference */
58         I_CLS_IF_MAX,           /* number of possible interferences in all classes */
59         I_CLS_IF_CNT,           /* number of actual interferences in all classes */
60         I_CLS_SIZE_S,
61         I_CLS_SIZE_E = I_CLS_SIZE_S+MAX_CLS_SIZE,
62         I_CLS_PHIS_S,
63         I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
64
65         /* FROM HERE: RESULT VLAUES */
66         /* all of them are external set */
67
68         /* ilp values */
69         I_HEUR_TIME,            /* solving time in milli seconds */
70         I_ILP_TIME,                     /* solving time in milli seconds */
71         I_ILP_VARS,
72         I_ILP_CSTR,
73         I_ILP_ITER,                     /* number of simplex iterations */
74
75         /* copy instructions */
76         I_COPIES_MAX,           /* max possible costs of copies*/
77         I_COPIES_INIT,          /* number of copies in initial allocation */
78         I_COPIES_HEUR,          /* number of copies after heuristic */
79         I_COPIES_5SEC,          /* number of copies after ilp with max n sec */
80         I_COPIES_30SEC,         /* number of copies after ilp with max n sec */
81         I_COPIES_OPT,           /* number of copies after ilp */
82         I_COPIES_IF,            /* number of copies inevitable due to root-arg-interf */
83
84         ASIZE
85 };
86
87 /**
88  * Holds current values. Values are added till next copystat_reset
89  */
90 int curr_vals[ASIZE];
91
92 static pset *all_phi_nodes;
93 static pset *all_phi_classes;
94 static pset *all_copy_nodes;
95 static ir_graph *last_irg;
96
97 void copystat_init(void) {
98         FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
99
100         all_phi_nodes   = pset_new_ptr_default();
101         all_phi_classes = pset_new_ptr_default();
102         all_copy_nodes  = pset_new_ptr_default();
103 }
104
105 void copystat_reset(void) {
106         int i;
107         for (i = 0; i < ASIZE; ++i)
108                 curr_vals[i] = 0;
109         del_pset(all_phi_nodes);
110         del_pset(all_phi_classes);
111         del_pset(all_copy_nodes);
112         all_phi_nodes = pset_new_ptr_default();
113         all_phi_classes = pset_new_ptr_default();
114         all_copy_nodes = pset_new_ptr_default();
115 }
116
117 /**
118  * Collect general data
119  */
120 static void irg_stat_walker(ir_node *node, void *env) {
121         arch_env_t *arch_env = env;
122         curr_vals[I_ALL_NODES]++; /* count all nodes */
123
124         if (is_Block(node)) /* count all blocks */
125                 curr_vals[I_BLOCKS]++;
126
127         if (is_Reg_Phi(node)) /* collect phis */
128                 pset_insert_ptr(all_phi_nodes, node);
129
130         if (is_Perm_Proj(arch_env, node))
131                 pset_insert_ptr(all_copy_nodes, node);
132
133         /* TODO: Add 2-Addr-Code nodes */
134 }
135
136 static void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
137         irg_walk_graph(irg, irg_stat_walker, NULL, arch_env);
138         all_phi_classes = phi_class_compute_by_phis(all_phi_nodes);
139         last_irg = irg;
140 }
141
142 /**
143  * @return 1 if the block at pos @p pos removed a critical edge
144  *                 0 else
145  */
146 static INLINE int was_edge_critical(const ir_node *bl, int pos) {
147         const ir_edge_t *edge;
148         const ir_node *bl_at_pos, *bl_before;
149         assert(is_Block(bl));
150
151         /* Does bl have several predecessors ?*/
152         if (get_irn_arity(bl) <= 1)
153                 return 0;
154
155         /* Does the pred have exactly one predecessor */
156         bl_at_pos = get_irn_n(bl, pos);
157         if (get_irn_arity(bl_at_pos) != 1)
158                 return 0;
159
160         /* Does the pred of the pred have several successors */
161         bl_before = get_irn_n(bl_at_pos, 0);
162         edge = get_block_succ_first(bl_before);
163         return get_block_succ_next(bl_before, edge) ? 1 : 0;
164 }
165
166 /**
167  * Collect phi node data
168  */
169 static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi) {
170         int arity, i;
171         ir_node *phi_bl;
172         assert(is_Phi(phi));
173
174         /* count all phi phis */
175         curr_vals[I_PHI_CNT]++;
176
177         /* argument count */
178         arity = get_irn_arity(phi);
179         curr_vals[I_PHI_ARG_CNT] += arity;
180         if (arity > MAX_ARITY)
181                 curr_vals[I_PHI_ARITY_E]++;
182         else
183                 curr_vals[I_PHI_ARITY_S + arity]++;
184
185         phi_bl = get_nodes_block(phi);
186         /* type of argument {self, const, pred, glob} */
187         for (i = 0; i < arity; i++) {
188         ir_node *block_of_arg, *block_ith_pred;
189                 ir_node *arg = get_irn_n(phi, i);
190
191                 if (arg == phi) {
192                         curr_vals[I_PHI_ARG_SELF]++;
193                         continue;
194                 }
195
196                 if (iro_Const == get_irn_opcode(arg)) {
197                         curr_vals[I_PHI_ARG_CONST]++;
198                         continue;
199                 }
200
201                 /* get the pred block skipping blocks on critical edges */
202                 block_ith_pred = get_Block_cfgpred_block(phi_bl, i);
203                 if (was_edge_critical(phi_bl, i))
204                         block_ith_pred = get_Block_cfgpred_block(block_ith_pred, 0);
205
206                 block_of_arg = get_nodes_block(arg);
207                 if (block_of_arg == block_ith_pred) {
208                         curr_vals[I_PHI_ARG_PRED]++;
209                         continue;
210                 }
211
212                 curr_vals[I_PHI_ARG_GLOB]++;
213         }
214 }
215
216 /**
217  * Collect register-constrained node data
218  */
219 static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
220         curr_vals[I_CPY_CNT]++;
221         curr_vals[I_COPIES_MAX]++;
222         if (nodes_interfere(chordal_env, root, get_Perm_src(root))) {
223                 curr_vals[I_COPIES_IF]++;
224                 assert(0 && "A Perm pair (in/out) should never interfere!");
225         }
226 }
227
228 /**
229  * Collect phi class data
230  */
231 static void stat_phi_class(be_chordal_env_t *chordal_env, pset *pc) {
232         int i, o, size, if_free, phis;
233         ir_node **members, *p;
234
235         /* phi class count */
236         curr_vals[I_CLS_CNT]++;
237
238         /* phi class size */
239         size = pset_count(pc);
240         if (size > MAX_CLS_SIZE)
241                 curr_vals[I_CLS_SIZE_E]++;
242         else
243                 curr_vals[I_CLS_SIZE_S + size]++;
244
245         /* get an array of all members for double iterating */
246         members = xmalloc(size * sizeof(*members));
247         DBG((dbg, LEVEL_2, "Phi-class:\n"));
248         for (i = 0, p = pset_first(pc); p; p = pset_next(pc)) {
249                 DBG((dbg, LEVEL_2, "  %+F\n", p));
250                 members[i++] = p;
251         }
252         assert(i == size);
253
254         /* determine number of phis on this class */
255         phis = 0;
256         for (i = 0; i < size; ++i)
257                 if (is_Phi(members[i]))
258                         phis++;
259         if (phis > MAX_CLS_PHIS)
260                 curr_vals[I_CLS_PHIS_E]++;
261         else
262                 curr_vals[I_CLS_PHIS_S + phis]++;
263
264         /* determine interference of phi class members */
265         curr_vals[I_CLS_IF_MAX] += size*(size-1)/2;
266         if_free = 1;
267         for (i = 0; i < size-1; ++i)
268                 for (o = i+1; o < size; ++o)
269                         if (nodes_interfere(chordal_env, members[i], members[o])) {
270                                 if_free = 0;
271                                 curr_vals[I_CLS_IF_CNT]++;
272                         }
273
274         /* Does this phi class have an inner interference? */
275         curr_vals[I_CLS_IF_FREE] += if_free;
276
277         xfree(members);
278 }
279
280 void copystat_collect_cls(be_chordal_env_t *cenv) {
281         ir_node *n;
282         pset *pc;
283         ir_graph *irg = cenv->irg;
284         arch_env_t *aenv = cenv->birg->main_env->arch_env;
285
286         copystat_reset();
287         copystat_collect_irg(irg, aenv);
288
289         for (n = pset_first(all_phi_nodes); n; n = pset_next(all_phi_nodes))
290                 if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
291                         stat_phi_node(cenv, n);
292
293         for (n = pset_first(all_copy_nodes); n; n = pset_next(all_copy_nodes))
294                 if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
295                         stat_copy_node(cenv, n);
296
297         for (pc = pset_first(all_phi_classes); pc; pc = pset_next(all_phi_classes)) {
298                 ir_node *member = pset_first(pc);
299                 pset_break(pc);
300                 if (arch_get_irn_reg_class(aenv, member, -1) == cenv->cls)
301                         stat_phi_class(cenv, pc);
302         }
303 }
304
305 void copystat_add_max_costs(int costs) {
306         curr_vals[I_COPIES_MAX] += costs;
307 }
308 void copystat_add_inevit_costs(int costs) {
309         curr_vals[I_COPIES_IF] += costs;
310 }
311 void copystat_add_init_costs(int costs) {
312         curr_vals[I_COPIES_INIT] += costs;
313 }
314 void copystat_add_heur_costs(int costs) {
315         curr_vals[I_COPIES_HEUR] += costs;
316 }
317 void copystat_add_ilp_5_sec_costs(int costs) {
318         curr_vals[I_COPIES_5SEC] += costs;
319 }
320 void copystat_add_ilp_30_sec_costs(int costs) {
321         curr_vals[I_COPIES_30SEC] += costs;
322 }
323 void copystat_add_opt_costs(int costs) {
324         curr_vals[I_COPIES_OPT] += costs;
325 }
326 void copystat_add_heur_time(int time) {
327         curr_vals[I_HEUR_TIME] += time;
328 }
329
330 #ifdef WITH_ILP
331
332 void copystat_add_ilp_time(int time) {
333         curr_vals[I_ILP_TIME] += time;
334 }
335 void copystat_add_ilp_vars(int vars) {
336         curr_vals[I_ILP_VARS] += vars;
337 }
338 void copystat_add_ilp_csts(int csts) {
339         curr_vals[I_ILP_CSTR] += csts;
340 }
341 void copystat_add_ilp_iter(int iters) {
342         curr_vals[I_ILP_ITER] += iters;
343 }
344
345 #endif /* WITH_ILP */
346
347 void copystat_dump(ir_graph *irg) {
348         int i;
349         char buf[1024];
350         FILE *out;
351
352         snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
353         buf[sizeof(buf) - 1] = '\0';
354         out = ffopen(buf, "stat", "wt");
355
356         fprintf(out, "%d\n", ASIZE);
357         for (i = 0; i < ASIZE; i++) {
358 #if 0
359                 if (i >= I_PHI_ARITY_S && i <= I_PHI_ARITY_E)
360                         fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_PHI_CNT]);
361                 else if (i >= I_CLS_SIZE_S && i <= I_CLS_SIZE_E)
362                         fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_CLS_CNT]);
363                 else
364 #endif
365                         fprintf(out, "%i\n", curr_vals[i]);
366         }
367
368         fclose(out);
369 }
370
371 void copystat_dump_pretty(ir_graph *irg) {
372         int i;
373         char buf[1024];
374         FILE *out;
375
376         snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
377         buf[sizeof(buf) - 1] = '\0';
378         out = ffopen(buf, "pstat", "wt");
379
380         fprintf(out, "Nodes     %4d\n", curr_vals[I_ALL_NODES]);
381         fprintf(out, "Blocks    %4d\n", curr_vals[I_BLOCKS]);
382         fprintf(out, "CopyIrn   %4d\n", curr_vals[I_CPY_CNT]);
383
384         fprintf(out, "\nPhis      %4d\n", curr_vals[I_PHI_CNT]);
385         fprintf(out, "... argument types\n");
386         fprintf(out, " Total      %4d\n", curr_vals[I_PHI_ARG_CNT]);
387         fprintf(out, " Self       %4d\n", curr_vals[I_PHI_ARG_SELF]);
388         fprintf(out, " Constants  %4d\n", curr_vals[I_PHI_ARG_CONST]);
389         fprintf(out, " CF-Pred    %4d\n", curr_vals[I_PHI_ARG_PRED]);
390         fprintf(out, " Others     %4d\n", curr_vals[I_PHI_ARG_GLOB]);
391         fprintf(out, "... arities\n");
392         for (i = I_PHI_ARITY_S; i<=I_PHI_ARITY_E; i++)
393                 fprintf(out, " %2i %4d\n", i-I_PHI_ARITY_S, curr_vals[i]);
394
395         fprintf(out, "\nPhi classes   %4d\n", curr_vals[I_CLS_CNT]);
396         fprintf(out, " compl. free  %4d\n", curr_vals[I_CLS_IF_FREE]);
397         fprintf(out, " inner intf.  %4d / %4d\n", curr_vals[I_CLS_IF_CNT], curr_vals[I_CLS_IF_MAX]);
398         fprintf(out, "... sizes\n");
399         for (i = I_CLS_SIZE_S; i<=I_CLS_SIZE_E; i++)
400                 fprintf(out, " %2i %4d\n", i-I_CLS_SIZE_S, curr_vals[i]);
401         fprintf(out, "... contained phis\n");
402         for (i = I_CLS_PHIS_S; i<=I_CLS_PHIS_E; i++)
403                 fprintf(out, " %2i %4d\n", i-I_CLS_PHIS_S, curr_vals[i]);
404
405         fprintf(out, "\nILP stat\n");
406         fprintf(out, " Time %8d\n", curr_vals[I_ILP_TIME]);
407         fprintf(out, " Iter %8d\n", curr_vals[I_ILP_ITER]);
408
409         fprintf(out, "\nCopy stat\n");
410         fprintf(out, " Max  %4d\n", curr_vals[I_COPIES_MAX]);
411         fprintf(out, " Init %4d\n", curr_vals[I_COPIES_INIT]);
412         fprintf(out, " Heur %4d\n", curr_vals[I_COPIES_HEUR]);
413         fprintf(out, " Opt  %4d\n", curr_vals[I_COPIES_OPT]);
414         fprintf(out, " Intf %4d\n", curr_vals[I_COPIES_IF]);
415
416         fclose(out);
417 }
418
419 /**
420  * Helpers for saving and restoring colors of nodes.
421  * Used to get dependable and comparable benchmark results.
422  */
423 typedef struct color_saver {
424         arch_env_t *arch_env;
425         be_chordal_env_t *chordal_env;
426         pmap *saved_colors;
427         int flag; /* 0 save, 1 load */
428 } color_save_t;
429
430 static void save_load(ir_node *irn, void *env) {
431         color_save_t *saver = env;
432         if (saver->chordal_env->cls == arch_get_irn_reg_class(saver->arch_env, irn, -1)) {
433                 if (saver->flag == 0) { /* save */
434                         const arch_register_t *reg = arch_get_irn_register(saver->arch_env, irn);
435                         pmap_insert(saver->saved_colors, irn, (void *) reg);
436                 } else { /*load */
437                         arch_register_t *reg = pmap_get(saver->saved_colors, irn);
438                         arch_set_irn_register(saver->arch_env, irn, reg);
439                 }
440         }
441 }
442
443 static void save_colors(color_save_t *color_saver) {
444         color_saver->flag = 0;
445         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
446 }
447
448 #ifdef WITH_ILP
449 static void load_colors(color_save_t *color_saver) {
450         color_saver->flag = 1;
451         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
452 }
453 #endif
454
455 /**
456  * Main compare routine
457  */
458 void co_compare_solvers(be_chordal_env_t *chordal_env) {
459         copy_opt_t *co;
460         lc_timer_t *timer;
461         color_save_t saver;
462         int costs_inevit, costs_init, costs_solved, lower_bound;
463
464         phi_class_compute(chordal_env->irg);
465         copystat_collect_cls(chordal_env);
466
467         co = new_copy_opt(chordal_env, co_get_costs_loop_depth);
468         co_build_ou_structure(co);
469         co_build_graph_structure(co);
470         DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
471
472         /* save colors */
473         saver.arch_env     = chordal_env->birg->main_env->arch_env;
474         saver.chordal_env  = chordal_env;
475         saver.saved_colors = pmap_create();
476         save_colors(&saver);
477         be_ra_chordal_check(co->cenv);
478
479         /* initial values */
480         costs_inevit = co_get_inevit_copy_costs(co);
481         lower_bound  = co_get_lower_bound(co);
482         costs_init   = co_get_copy_costs(co);
483
484         DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
485         DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
486         DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
487
488         copystat_add_inevit_costs(costs_inevit);
489         copystat_add_init_costs(costs_init);
490         copystat_add_max_costs(co_get_max_copy_costs(co));
491
492         /* heuristic 1 (Daniel Grund) */
493         timer = lc_timer_register("heur1", NULL);
494         lc_timer_reset_and_start(timer);
495
496         co_solve_heuristic(co);
497
498         lc_timer_stop(timer);
499
500         be_ra_chordal_check(co->cenv);
501         costs_solved = co_get_copy_costs(co);
502         DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
503         copystat_add_heur_time(lc_timer_elapsed_msec(timer));
504         copystat_add_heur_costs(costs_solved);
505         assert(lower_bound <= costs_solved);
506
507         /* heuristic 2 (Sebastian Hack) */
508         timer = lc_timer_register("heur2", NULL);
509         lc_timer_reset_and_start(timer);
510
511         co_solve_heuristic_new(co);
512
513         lc_timer_stop(timer);
514
515         be_ra_chordal_check(co->cenv);
516         costs_solved = co_get_copy_costs(co);
517         DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
518         copystat_add_heur_time(lc_timer_elapsed_msec(timer));
519         copystat_add_heur_costs(costs_solved);
520         assert(lower_bound <= costs_solved);
521
522         /* Park & Moon register coalescing (Kimon Hoffmann) */
523         timer = lc_timer_register("park", NULL);
524         lc_timer_reset_and_start(timer);
525
526         co_solve_park_moon(co);
527
528         lc_timer_stop(timer);
529
530         be_ra_chordal_check(co->cenv);
531         costs_solved = co_get_copy_costs(co);
532         DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
533         copystat_add_heur_time(lc_timer_elapsed_msec(timer));
534         copystat_add_heur_costs(costs_solved);
535         assert(lower_bound <= costs_solved);
536
537
538 #ifdef WITH_ILP
539
540         /* ILP 1 is not yet implemented, so it makes no sense to compare */
541 #if 0
542         load_colors(&saver);
543
544         co_solve_ilp1(co, 60.0);
545
546         costs_solved = co_get_copy_costs(co);
547         DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
548         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
549         assert(lower_bound <= costs_solved);
550 #endif /* 0 */
551
552         /* ILP 2 */
553         load_colors(&saver);
554
555         co_solve_ilp2(co);
556
557         be_ra_chordal_check(co->cenv);
558         costs_solved = co_get_copy_costs(co);
559         DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
560         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
561         assert(lower_bound <= costs_solved);
562
563 #endif /* WITH_ILP */
564
565         /* free memory for statistic structures */
566         pmap_destroy(saver.saved_colors);
567         co_free_graph_structure(co);
568         co_free_ou_structure(co);
569         free_copy_opt(co);
570 }