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 "beutil.h"
  41 #include "becopyopt_t.h"
  42 #include "becopystat.h"
  43 #include "beirg_t.h"
  44 #include "bemodule.h"
  45 #include "beintlive_t.h"
  46
  47 #define DEBUG_LVL SET_LEVEL_1
  48 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  49
  50 #define MAX_ARITY 20
  51 #define MAX_CLS_SIZE 20
  52 #define MAX_CLS_PHIS 20
  53
  54 /**
  55  * For an explanation of these values see the code of copystat_dump_pretty
  56  */
  57 enum vals_t {
  58         /* FROM HERE: PROBLEM CHARACTERIZATION */
  59
  60         I_ALL_NODES = 0,
  61         I_BLOCKS,
  62
  63         /* phi nodes */
  64         I_PHI_CNT,                      /* number of phi nodes */
  65         I_PHI_ARG_CNT,          /* number of arguments of phis */
  66         I_PHI_ARG_SELF,         /* number of arguments of phis being the phi itself */
  67         I_PHI_ARG_CONST,        /* number of arguments of phis being consts */
  68         I_PHI_ARG_PRED,         /* ... being defined in a cf-pred */
  69         I_PHI_ARG_GLOB,         /* ... being defined elsewhere */
  70         I_PHI_ARITY_S,
  71         I_PHI_ARITY_E    = I_PHI_ARITY_S+MAX_ARITY,
  72
  73         /* copy nodes */
  74         I_CPY_CNT,                      /* number of copynodes */
  75
  76         /* phi classes */
  77         I_CLS_CNT,                      /* number of phi classes */
  78         I_CLS_IF_FREE,          /* number of pc having no interference */
  79         I_CLS_IF_MAX,           /* number of possible interferences in all classes */
  80         I_CLS_IF_CNT,           /* number of actual interferences in all classes */
  81         I_CLS_SIZE_S,
  82         I_CLS_SIZE_E = I_CLS_SIZE_S+MAX_CLS_SIZE,
  83         I_CLS_PHIS_S,
  84         I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
  85
  86         /* FROM HERE: RESULT VLAUES */
  87         /* all of them are external set */
  88
  89         /* ilp values */
  90         I_HEUR_TIME,            /* solving time in milli seconds */
  91         I_ILP_TIME,                     /* solving time in milli seconds */
  92         I_ILP_VARS,
  93         I_ILP_CSTR,
  94         I_ILP_ITER,                     /* number of simplex iterations */
  95
  96         /* copy instructions */
  97         I_COPIES_MAX,           /* max possible costs of copies*/
  98         I_COPIES_INIT,          /* number of copies in initial allocation */
  99         I_COPIES_HEUR,          /* number of copies after heuristic */
 100         I_COPIES_5SEC,          /* number of copies after ilp with max n sec */
 101         I_COPIES_30SEC,         /* number of copies after ilp with max n sec */
 102         I_COPIES_OPT,           /* number of copies after ilp */
 103         I_COPIES_IF,            /* number of copies inevitable due to root-arg-interf */
 104
 105         ASIZE
 106 };
 107
 108 /**
 109  * Holds current values. Values are added till next copystat_reset
 110  */
 111 int curr_vals[ASIZE];
 112
 113 static ir_nodeset_t *all_phi_nodes;
 114 static ir_nodeset_t *all_copy_nodes;
 115 static ir_graph *last_irg;
 116
 117 void be_init_copystat(void) {
 118         FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
 119
 120         all_phi_nodes  = ir_nodeset_new(64);
 121         all_copy_nodes = ir_nodeset_new(64);
 122         memset(curr_vals, 0, sizeof(curr_vals));
 123 }
 124 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copystat);
 125
 126 void be_quit_copystat(void) {
 127         ir_nodeset_del(all_phi_nodes);
 128         ir_nodeset_del(all_copy_nodes);
 129 }
 130 BE_REGISTER_MODULE_DESTRUCTOR(be_quit_copystat);
 131
 132 void copystat_reset(void) {
 133         be_quit_copystat();
 134         be_init_copystat();
 135 }
 136
 137 /**
 138  * Collect general data
 139  */
 140 static void irg_stat_walker(ir_node *node, void *env)
 141 {
 142         (void)env;
 143
 144         curr_vals[I_ALL_NODES]++; /* count all nodes */
 145
 146         if (is_Block(node)) /* count all blocks */
 147                 curr_vals[I_BLOCKS]++;
 148
 149         if (is_Reg_Phi(node)) /* collect phis */
 150                 ir_nodeset_insert(all_phi_nodes, node);
 151
 152         if (is_Perm_Proj(node))
 153                 ir_nodeset_insert(all_copy_nodes, node);
 154
 155         /* TODO: Add 2-Addr-Code nodes */
 156 }
 157
 158 static void copystat_collect_irg(ir_graph *irg)
 159 {
 160         irg_walk_graph(irg, irg_stat_walker, NULL, NULL);
 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         ir_node               *n, **pc;
 294         phi_classes_t         *pc_obj;
 295         pset                  *all_phi_classes;
 296         ir_nodeset_iterator_t iter;
 297
 298         copystat_reset();
 299         copystat_collect_irg(irg);
 300
 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);
 304
 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);
 308         }
 309
 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);
 313         }
 314
 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);
 319         }
 320
 321         /* free the phi class object */
 322         phi_class_free(pc_obj);
 323 }
 324
 325 void copystat_add_max_costs(int costs) {
 326         curr_vals[I_COPIES_MAX] += costs;
 327 }
 328 void copystat_add_inevit_costs(int costs) {
 329         curr_vals[I_COPIES_IF] += costs;
 330 }
 331 void copystat_add_init_costs(int costs) {
 332         curr_vals[I_COPIES_INIT] += costs;
 333 }
 334 void copystat_add_heur_costs(int costs) {
 335         curr_vals[I_COPIES_HEUR] += costs;
 336 }
 337 void copystat_add_ilp_5_sec_costs(int costs) {
 338         curr_vals[I_COPIES_5SEC] += costs;
 339 }
 340 void copystat_add_ilp_30_sec_costs(int costs) {
 341         curr_vals[I_COPIES_30SEC] += costs;
 342 }
 343 void copystat_add_opt_costs(int costs) {
 344         curr_vals[I_COPIES_OPT] += costs;
 345 }
 346 void copystat_add_heur_time(int time) {
 347         curr_vals[I_HEUR_TIME] += time;
 348 }
 349
 350 #ifdef WITH_ILP
 351
 352 void copystat_add_ilp_time(int time) {
 353         curr_vals[I_ILP_TIME] += time;
 354 }
 355 void copystat_add_ilp_vars(int vars) {
 356         curr_vals[I_ILP_VARS] += vars;
 357 }
 358 void copystat_add_ilp_csts(int csts) {
 359         curr_vals[I_ILP_CSTR] += csts;
 360 }
 361 void copystat_add_ilp_iter(int iters) {
 362         curr_vals[I_ILP_ITER] += iters;
 363 }
 364
 365 #endif /* WITH_ILP */
 366
 367 void copystat_dump(ir_graph *irg) {
 368         int i;
 369         char buf[1024];
 370         FILE *out;
 371
 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");
 375
 376         fprintf(out, "%d\n", ASIZE);
 377         for (i = 0; i < ASIZE; i++) {
 378 #if 0
 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]);
 383                 else
 384 #endif
 385                         fprintf(out, "%i\n", curr_vals[i]);
 386         }
 387
 388         fclose(out);
 389 }
 390
 391 void copystat_dump_pretty(ir_graph *irg) {
 392         int i;
 393         char buf[1024];
 394         FILE *out;
 395
 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");
 399
 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]);
 403
 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]);
 414
 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]);
 424
 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]);
 428
 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]);
 435
 436         fclose(out);
 437 }
 438
 439 /**
 440  * Helpers for saving and restoring colors of nodes.
 441  * Used to get dependable and comparable benchmark results.
 442  */
 443 typedef struct color_saver {
 444         be_chordal_env_t *chordal_env;
 445         pmap *saved_colors;
 446         int flag; /* 0 save, 1 load */
 447 } color_save_t;
 448
 449 static void save_load(ir_node *irn, void *env) {
 450         color_save_t *saver = env;
 451         if (saver->chordal_env->cls == arch_get_irn_reg_class(irn, -1)) {
 452                 if (saver->flag == 0) { /* save */
 453                         const arch_register_t *reg = arch_get_irn_register(irn);
 454                         pmap_insert(saver->saved_colors, irn, (void *) reg);
 455                 } else { /*load */
 456                         arch_register_t *reg = pmap_get(saver->saved_colors, irn);
 457                         arch_set_irn_register(irn, reg);
 458                 }
 459         }
 460 }
 461
 462 static void save_colors(color_save_t *color_saver) {
 463         color_saver->flag = 0;
 464         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
 465 }
 466
 467 #ifdef WITH_ILP
 468 static void load_colors(color_save_t *color_saver) {
 469         color_saver->flag = 1;
 470         irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
 471 }
 472 #endif
 473
 474 /**
 475  * Main compare routine
 476  */
 477 void co_compare_solvers(be_chordal_env_t *chordal_env) {
 478         copy_opt_t    *co;
 479         ir_timer_t    *timer;
 480         color_save_t  saver;
 481         int costs_inevit, costs_init, costs_solved, lower_bound;
 482
 483         copystat_collect_cls(chordal_env);
 484
 485         co = new_copy_opt(chordal_env, co_get_costs_loop_depth);
 486         co_build_ou_structure(co);
 487         co_build_graph_structure(co);
 488         DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
 489
 490         /* save colors */
 491         saver.chordal_env  = chordal_env;
 492         saver.saved_colors = pmap_create();
 493         save_colors(&saver);
 494
 495         /* initial values */
 496         costs_inevit = co_get_inevit_copy_costs(co);
 497         lower_bound  = co_get_lower_bound(co);
 498         costs_init   = co_get_copy_costs(co);
 499
 500         DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
 501         DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
 502         DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
 503
 504         copystat_add_inevit_costs(costs_inevit);
 505         copystat_add_init_costs(costs_init);
 506         copystat_add_max_costs(co_get_max_copy_costs(co));
 507
 508         /* heuristic 1 (Daniel Grund) */
 509         timer = ir_timer_register("heur1", NULL);
 510         ir_timer_reset_and_start(timer);
 511
 512         co_solve_heuristic(co);
 513
 514         ir_timer_stop(timer);
 515
 516         costs_solved = co_get_copy_costs(co);
 517         DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
 518         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 519         copystat_add_heur_costs(costs_solved);
 520         assert(lower_bound <= costs_solved);
 521
 522         /* heuristic 2 (Sebastian Hack) */
 523         timer = ir_timer_register("heur2", NULL);
 524         ir_timer_reset_and_start(timer);
 525
 526         co_solve_heuristic_new(co);
 527
 528         ir_timer_stop(timer);
 529
 530         costs_solved = co_get_copy_costs(co);
 531         DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
 532         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 533         copystat_add_heur_costs(costs_solved);
 534         assert(lower_bound <= costs_solved);
 535
 536         /* Park & Moon register coalescing (Kimon Hoffmann) */
 537         timer = ir_timer_register("park", NULL);
 538         ir_timer_reset_and_start(timer);
 539
 540         co_solve_park_moon(co);
 541
 542         ir_timer_stop(timer);
 543
 544         costs_solved = co_get_copy_costs(co);
 545         DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
 546         copystat_add_heur_time(ir_timer_elapsed_msec(timer));
 547         copystat_add_heur_costs(costs_solved);
 548         assert(lower_bound <= costs_solved);
 549
 550
 551 #ifdef WITH_ILP
 552
 553         /* ILP 1 is not yet implemented, so it makes no sense to compare */
 554 #if 0
 555         load_colors(&saver);
 556
 557         co_solve_ilp1(co, 60.0);
 558
 559         costs_solved = co_get_copy_costs(co);
 560         DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
 561         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
 562         assert(lower_bound <= costs_solved);
 563 #endif /* 0 */
 564
 565         /* ILP 2 */
 566         load_colors(&saver);
 567
 568         co_solve_ilp2(co);
 569
 570         costs_solved = co_get_copy_costs(co);
 571         DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
 572         copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
 573         assert(lower_bound <= costs_solved);
 574
 575 #endif /* WITH_ILP */
 576
 577         /* free memory for statistic structures */
 578         pmap_destroy(saver.saved_colors);
 579         co_free_graph_structure(co);
 580         co_free_ou_structure(co);
 581         free_copy_opt(co);
 582 }