nsz Git - libfirm/blob - 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 #include "config.h"
  28
  29 #include <string.h>
  30
  31 #include "timing.h"
  32 #include "irgraph.h"
  33 #include "irgwalk.h"
  34 #include "irprog.h"
  35 #include "iredges_t.h"
  36 #include "phiclass.h"
  37 #include "irnodeset.h"
  38
  39 #include "bechordal_t.h"
  40 #include "benode.h"
  41 #include "beutil.h"
  42 #include "becopyopt_t.h"
  43 #include "becopystat.h"
  44 #include "beirg.h"
  45 #include "bemodule.h"
  46 #include "beintlive_t.h"
  47
  48 #define DEBUG_LVL SET_LEVEL_1
  49 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  50
  51 #define MAX_ARITY 20
  52 #define MAX_CLS_SIZE 20
  53 #define MAX_CLS_PHIS 20
  54
  55 /**
  56  * For an explanation of these values see the code of copystat_dump_pretty
  57  */
  58 enum vals_t {
  59         /* FROM HERE: PROBLEM CHARACTERIZATION */
  60
  61         I_ALL_NODES = 0,
  62         I_BLOCKS,
  63
  64         /* phi nodes */
  65         I_PHI_CNT,                      /* number of phi nodes */
  66         I_PHI_ARG_CNT,          /* number of arguments of phis */
  67         I_PHI_ARG_SELF,         /* number of arguments of phis being the phi itself */
  68         I_PHI_ARG_CONST,        /* number of arguments of phis being consts */
  69         I_PHI_ARG_PRED,         /* ... being defined in a cf-pred */
  70         I_PHI_ARG_GLOB,         /* ... being defined elsewhere */
  71         I_PHI_ARITY_S,
  72         I_PHI_ARITY_E    = I_PHI_ARITY_S+MAX_ARITY,
  73
  74         /* copy nodes */
  75         I_CPY_CNT,                      /* number of copynodes */
  76
  77         /* phi classes */
  78         I_CLS_CNT,                      /* number of phi classes */
  79         I_CLS_IF_FREE,          /* number of pc having no interference */
  80         I_CLS_IF_MAX,           /* number of possible interferences in all classes */
  81         I_CLS_IF_CNT,           /* number of actual interferences in all classes */
  82         I_CLS_SIZE_S,
  83         I_CLS_SIZE_E = I_CLS_SIZE_S+MAX_CLS_SIZE,
  84         I_CLS_PHIS_S,
  85         I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
  86
  87         /* FROM HERE: RESULT VLAUES */
  88         /* all of them are external set */
  89
  90         /* ilp values */
  91         I_HEUR_TIME,            /* solving time in milli seconds */
  92         I_ILP_TIME,                     /* solving time in milli seconds */
  93         I_ILP_VARS,
  94         I_ILP_CSTR,
  95         I_ILP_ITER,                     /* number of simplex iterations */
  96
  97         /* copy instructions */
  98         I_COPIES_MAX,           /* max possible costs of copies*/
  99         I_COPIES_INIT,          /* number of copies in initial allocation */
 100         I_COPIES_HEUR,          /* number of copies after heuristic */
 101         I_COPIES_5SEC,          /* number of copies after ilp with max n sec */
 102         I_COPIES_30SEC,         /* number of copies after ilp with max n sec */
 103         I_COPIES_OPT,           /* number of copies after ilp */
 104         I_COPIES_IF,            /* number of copies inevitable due to root-arg-interf */
 105
 106         ASIZE
 107 };
 108
 109 /**
 110  * Holds current values. Values are added till next copystat_reset
 111  */
 112 int curr_vals[ASIZE];
 113
 114 static ir_nodeset_t *all_phi_nodes;
 115 static ir_nodeset_t *all_copy_nodes;
 116 static ir_graph *last_irg;
 117
 118 void be_init_copystat(void) {
 119         FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
 120
 121         all_phi_nodes  = ir_nodeset_new(64);
 122         all_copy_nodes = ir_nodeset_new(64);
 123         memset(curr_vals, 0, sizeof(curr_vals));
 124 }
 125 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copystat);
 126
 127 void be_quit_copystat(void) {
 128         ir_nodeset_del(all_phi_nodes);
 129         ir_nodeset_del(all_copy_nodes);
 130 }
 131 BE_REGISTER_MODULE_DESTRUCTOR(be_quit_copystat);
 132
 133 void copystat_reset(void) {
 134         be_quit_copystat();
 135         be_init_copystat();
 136 }
 137
 138 /**
 139  * Collect general data
 140  */
 141 static void irg_stat_walker(ir_node *node, void *env)
 142 {
 143         (void)env;
 144
 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(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)
 160 {
 161         irg_walk_graph(irg, irg_stat_walker, NULL, NULL);
 162         last_irg = irg;
 163 }
 164
 165 /**
 166  * @return 1 if the block at pos @p pos removed a critical edge
 167  *                 0 else
 168  */
 169 static inline int was_edge_critical(const ir_node *bl, int pos) {
 170         const ir_edge_t *edge;
 171         const ir_node *bl_at_pos, *bl_before;
 172         assert(is_Block(bl));
 173
 174         /* Does bl have several predecessors ?*/
 175         if (get_irn_arity(bl) <= 1)
 176                 return 0;
 177
 178         /* Does the pred have exactly one predecessor */
 179         bl_at_pos = get_irn_n(bl, pos);
 180         if (get_irn_arity(bl_at_pos) != 1)
 181                 return 0;
 182
 183         /* Does the pred of the pred have several successors */
 184         bl_before = get_irn_n(bl_at_pos, 0);
 185         edge = get_block_succ_first(bl_before);
 186         return get_block_succ_next(bl_before, edge) ? 1 : 0;
 187 }
 188
 189 /**
 190  * Collect phi node data
 191  */
 192 static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi)
 193 {
 194         int arity, i;
 195         ir_node *phi_bl;
 196         assert(is_Phi(phi));
 197         (void) chordal_env;
 198
 199         /* count all phi phis */
 200         curr_vals[I_PHI_CNT]++;
 201
 202         /* argument count */
 203         arity = get_irn_arity(phi);
 204         curr_vals[I_PHI_ARG_CNT] += arity;
 205         if (arity > MAX_ARITY)
 206                 curr_vals[I_PHI_ARITY_E]++;
 207         else
 208                 curr_vals[I_PHI_ARITY_S + arity]++;
 209
 210         phi_bl = get_nodes_block(phi);
 211         /* type of argument {self, const, pred, glob} */
 212         for (i = 0; i < arity; i++) {
 213         ir_node *block_of_arg, *block_ith_pred;
 214                 ir_node *arg = get_irn_n(phi, i);
 215
 216                 if (arg == phi) {
 217                         curr_vals[I_PHI_ARG_SELF]++;
 218                         continue;
 219                 }
 220
 221                 if (iro_Const == get_irn_opcode(arg)) {
 222                         curr_vals[I_PHI_ARG_CONST]++;
 223                         continue;
 224                 }
 225
 226                 /* get the pred block skipping blocks on critical edges */
 227                 block_ith_pred = get_Block_cfgpred_block(phi_bl, i);
 228                 if (was_edge_critical(phi_bl, i))
 229                         block_ith_pred = get_Block_cfgpred_block(block_ith_pred, 0);
 230
 231                 block_of_arg = get_nodes_block(arg);
 232                 if (block_of_arg == block_ith_pred) {
 233                         curr_vals[I_PHI_ARG_PRED]++;
 234                         continue;
 235                 }
 236
 237                 curr_vals[I_PHI_ARG_GLOB]++;
 238         }
 239 }
 240
 241 /**
 242  * Collect register-constrained node data
 243  */
 244 static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
 245         curr_vals[I_CPY_CNT]++;
 246         curr_vals[I_COPIES_MAX]++;
 247         if (be_values_interfere(chordal_env->birg->lv, root, get_Perm_src(root))) {
 248                 curr_vals[I_COPIES_IF]++;
 249                 assert(0 && "A Perm pair (in/out) should never interfere!");
 250         }
 251 }
 252
 253 /**
 254  * Collect phi class data
 255  */
 256 static void stat_phi_class(be_chordal_env_t *chordal_env, ir_node **pc) {
 257         int i, o, size, if_free, phis;
 258
 259         /* phi class count */
 260         curr_vals[I_CLS_CNT]++;
 261
 262         /* phi class size */
 263         size = ARR_LEN(pc);
 264         if (size > MAX_CLS_SIZE)
 265                 curr_vals[I_CLS_SIZE_E]++;
 266         else
 267                 curr_vals[I_CLS_SIZE_S + size]++;
 268
 269         /* determine number of phis on this class */
 270         for (phis = i = 0; i < size; ++i)
 271                 if (is_Phi(pc[i]))
 272                         phis++;
 273
 274         if (phis > MAX_CLS_PHIS)
 275                 curr_vals[I_CLS_PHIS_E]++;
 276         else
 277                 curr_vals[I_CLS_PHIS_S + phis]++;
 278
 279         /* determine interference of phi class members */
 280         curr_vals[I_CLS_IF_MAX] += size * (size - 1) / 2;
 281         for (if_free = 1, i = 0; i < size - 1; ++i)
 282                 for (o = i + 1; o < size; ++o)
 283                         if (be_values_interfere(chordal_env->birg->lv, pc[i], pc[o])) {
 284                                 if_free = 0;
 285                                 curr_vals[I_CLS_IF_CNT]++;
 286                         }
 287
 288         /* Does this phi class have an inner interference? */
 289         curr_vals[I_CLS_IF_FREE] += if_free;
 290 }
 291
 292 static void copystat_collect_cls(be_chordal_env_t *cenv) {
 293         ir_graph              *irg  = cenv->irg;
 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);
 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_out(n) == 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_out(n) == 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_out(member) == 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_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_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         be_chordal_env_t *chordal_env;
 446         pmap *saved_colors;
 447         int flag; /* 0 save, 1 load */
 448 } color_save_t;
 449
 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_out(irn)) {
 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);
 456                 } else { /*load */
 457                         arch_register_t *reg = pmap_get(saver->saved_colors, irn);
 458                         arch_set_irn_register(irn, reg);
 459                 }
 460         }
 461 }
 462
 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);
 466 }
 467
 468 #ifdef WITH_ILP
 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);
 472 }
 473 #endif
 474
 475 /**
 476  * Main compare routine
 477  */
 478 void co_compare_solvers(be_chordal_env_t *chordal_env) {
 479         copy_opt_t    *co;
 480         ir_timer_t    *timer;
 481         color_save_t  saver;
 482         int costs_inevit, costs_init, costs_solved, lower_bound;
 483
 484         copystat_collect_cls(chordal_env);
 485
 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));
 490
 491         /* save colors */
 492         saver.chordal_env  = chordal_env;
 493         saver.saved_colors = pmap_create();
 494         save_colors(&saver);
 495
 496         /* initial values */
 497         costs_inevit = co_get_inevit_copy_costs(co);
 498         lower_bound  = co_get_lower_bound(co);
 499         costs_init   = co_get_copy_costs(co);
 500
 501         DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
 502         DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
 503         DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
 504
 505         copystat_add_inevit_costs(costs_inevit);
 506         copystat_add_init_costs(costs_init);
 507         copystat_add_max_costs(co_get_max_copy_costs(co));
 508
 509         /* heuristic 1 (Daniel Grund) */
 510         timer = ir_timer_register("heur1", NULL);
 511         ir_timer_reset_and_start(timer);
 512
 513         co_solve_heuristic(co);
 514
 515         ir_timer_stop(timer);
 516
 517         costs_solved = co_get_copy_costs(co);
 518         DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
 519         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 520         copystat_add_heur_costs(costs_solved);
 521         assert(lower_bound <= costs_solved);
 522
 523         /* heuristic 2 (Sebastian Hack) */
 524         timer = ir_timer_register("heur2", NULL);
 525         ir_timer_reset_and_start(timer);
 526
 527         co_solve_heuristic_new(co);
 528
 529         ir_timer_stop(timer);
 530
 531         costs_solved = co_get_copy_costs(co);
 532         DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
 533         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 534         copystat_add_heur_costs(costs_solved);
 535         assert(lower_bound <= costs_solved);
 536
 537         /* Park & Moon register coalescing (Kimon Hoffmann) */
 538         timer = ir_timer_register("park", NULL);
 539         ir_timer_reset_and_start(timer);
 540
 541         co_solve_park_moon(co);
 542
 543         ir_timer_stop(timer);
 544
 545         costs_solved = co_get_copy_costs(co);
 546         DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
 547         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 548         copystat_add_heur_costs(costs_solved);
 549         assert(lower_bound <= costs_solved);
 550
 551
 552 #ifdef WITH_ILP
 553
 554         /* ILP 1 is not yet implemented, so it makes no sense to compare */
 555 #if 0
 556         load_colors(&saver);
 557
 558         co_solve_ilp1(co, 60.0);
 559
 560         costs_solved = co_get_copy_costs(co);
 561         DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
 562         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
 563         assert(lower_bound <= costs_solved);
 564 #endif /* 0 */
 565
 566         /* ILP 2 */
 567         load_colors(&saver);
 568
 569         co_solve_ilp2(co);
 570
 571         costs_solved = co_get_copy_costs(co);
 572         DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
 573         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
 574         assert(lower_bound <= costs_solved);
 575
 576 #endif /* WITH_ILP */
 577
 578         /* free memory for statistic structures */
 579         pmap_destroy(saver.saved_colors);
 580         co_free_graph_structure(co);
 581         co_free_ou_structure(co);
 582         free_copy_opt(co);
 583 }