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