nsz Git - libfirm/blob - ir/be/bechordal_main.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       Driver for the chordal register allocator.
  23  * @author      Sebastian Hack
  24  * @date        29.11.2005
  25  * @version     $Id$
  26  */
  27 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <stdlib.h>
  32 #include <time.h>
  33
  34 #include "obst.h"
  35 #include "pset.h"
  36 #include "list.h"
  37 #include "bitset.h"
  38 #include "iterator.h"
  39 #include "firm_config.h"
  40
  41 #include "lc_opts.h"
  42 #include "lc_opts_enum.h"
  43
  44 #include "ircons_t.h"
  45 #include "irmode_t.h"
  46 #include "irgraph_t.h"
  47 #include "irprintf_t.h"
  48 #include "irgwalk.h"
  49 #include "ircons.h"
  50 #include "irdump.h"
  51 #include "irdom.h"
  52 #include "ircons.h"
  53 #include "irbitset.h"
  54 #include "irnode.h"
  55 #include "ircons.h"
  56 #include "debug.h"
  57 #include "xmalloc.h"
  58 #include "execfreq.h"
  59 #include "iredges_t.h"
  60
  61 #include "bechordal_t.h"
  62 #include "beabi.h"
  63 #include "bejavacoal.h"
  64 #include "beutil.h"
  65 #include "besched.h"
  66 #include "besched_t.h"
  67 #include "belive_t.h"
  68 #include "bearch_t.h"
  69 #include "beifg_t.h"
  70 #include "beifg_impl.h"
  71 #include "benode_t.h"
  72 #include "bestatevent.h"
  73 #include "bestat.h"
  74 #include "bemodule.h"
  75 #include "be_t.h"
  76 #include "bera.h"
  77 #include "beirg_t.h"
  78
  79 #include "bespillslots.h"
  80 #include "bespilloptions.h"
  81 #include "belower.h"
  82
  83 #ifdef WITH_ILP
  84 #include "bespillremat.h"
  85 #endif /* WITH_ILP */
  86
  87 #include "bejavacoal.h"
  88 #include "becopystat.h"
  89 #include "becopyopt.h"
  90 #include "bessadestr.h"
  91 #include "beverify.h"
  92 #include "benode_t.h"
  93
  94 static be_ra_chordal_opts_t options = {
  95         BE_CH_DUMP_NONE,
  96         BE_CH_LOWER_PERM_SWAP,
  97         BE_CH_VRFY_WARN,
  98         "",
  99         ""
 100 };
 101
 102 typedef struct _post_spill_env_t {
 103         be_chordal_env_t            cenv;
 104         be_irg_t                    *birg;
 105         const arch_register_class_t *cls;
 106         double                      pre_spill_cost;
 107 } post_spill_env_t;
 108
 109 static be_options_t  *main_opts;
 110
 111 static const lc_opt_enum_int_items_t lower_perm_items[] = {
 112         { "copy", BE_CH_LOWER_PERM_COPY },
 113         { "swap", BE_CH_LOWER_PERM_SWAP },
 114         { NULL, 0 }
 115 };
 116
 117 static const lc_opt_enum_int_items_t lower_perm_stat_items[] = {
 118         { NULL, 0 }
 119 };
 120
 121 static const lc_opt_enum_int_items_t dump_items[] = {
 122         { "none",       BE_CH_DUMP_NONE       },
 123         { "spill",      BE_CH_DUMP_SPILL      },
 124         { "live",       BE_CH_DUMP_LIVE       },
 125         { "color",      BE_CH_DUMP_COLOR      },
 126         { "copymin",    BE_CH_DUMP_COPYMIN    },
 127         { "ssadestr",   BE_CH_DUMP_SSADESTR   },
 128         { "tree",       BE_CH_DUMP_TREE_INTV  },
 129         { "constr",     BE_CH_DUMP_CONSTR     },
 130         { "lower",      BE_CH_DUMP_LOWER      },
 131         { "spillslots", BE_CH_DUMP_SPILLSLOTS },
 132         { "appel",      BE_CH_DUMP_APPEL      },
 133         { "all",        BE_CH_DUMP_ALL        },
 134         { NULL, 0 }
 135 };
 136
 137 static const lc_opt_enum_int_items_t be_ch_vrfy_items[] = {
 138         { "off",    BE_CH_VRFY_OFF    },
 139         { "warn",   BE_CH_VRFY_WARN   },
 140         { "assert", BE_CH_VRFY_ASSERT },
 141         { NULL, 0 }
 142 };
 143
 144 static lc_opt_enum_int_var_t lower_perm_var = {
 145         &options.lower_perm_opt, lower_perm_items
 146 };
 147
 148 static lc_opt_enum_int_var_t dump_var = {
 149         &options.dump_flags, dump_items
 150 };
 151
 152 static lc_opt_enum_int_var_t be_ch_vrfy_var = {
 153         &options.vrfy_option, be_ch_vrfy_items
 154 };
 155
 156 static const lc_opt_table_entry_t be_chordal_options[] = {
 157         LC_OPT_ENT_ENUM_PTR ("perm",          "perm lowering options", &lower_perm_var),
 158         LC_OPT_ENT_ENUM_MASK("dump",          "select dump phases", &dump_var),
 159         LC_OPT_ENT_ENUM_PTR ("vrfy",          "verify options", &be_ch_vrfy_var),
 160         LC_OPT_LAST
 161 };
 162
 163 static void dump(unsigned mask, ir_graph *irg,
 164                                  const arch_register_class_t *cls,
 165                                  const char *suffix,
 166                                  void (*dump_func)(ir_graph *, const char *))
 167 {
 168         if((options.dump_flags & mask) == mask) {
 169                 if (cls) {
 170                         char buf[256];
 171                         snprintf(buf, sizeof(buf), "-%s%s", cls->name, suffix);
 172                         be_dump(irg, buf, dump_func);
 173                 }
 174                 else
 175                         be_dump(irg, suffix, dump_func);
 176         }
 177 }
 178
 179 /**
 180  * Checks for every reload if it's user can perform the load on itself.
 181  */
 182 static void memory_operand_walker(ir_node *irn, void *env) {
 183         be_chordal_env_t *cenv = env;
 184         const arch_env_t *aenv = cenv->birg->main_env->arch_env;
 185         const ir_edge_t  *edge, *ne;
 186         ir_node          *block;
 187         ir_node          *spill;
 188
 189         if (! be_is_Reload(irn))
 190                 return;
 191
 192         /* only use memory operands, if the reload is only used by 1 node */
 193         if(get_irn_n_edges(irn) > 1)
 194                 return;
 195
 196         spill = be_get_Reload_mem(irn);
 197         block = get_nodes_block(irn);
 198
 199         foreach_out_edge_safe(irn, edge, ne) {
 200                 ir_node *src = get_edge_src_irn(edge);
 201                 int     pos  = get_edge_src_pos(edge);
 202
 203                 assert(src && "outedges broken!");
 204
 205                 if (get_nodes_block(src) == block && arch_possible_memory_operand(aenv, src, pos)) {
 206                         arch_perform_memory_operand(aenv, src, spill, pos);
 207                 }
 208         }
 209
 210         /* kill the Reload */
 211         if (get_irn_n_edges(irn) == 0) {
 212                 sched_remove(irn);
 213                 set_irn_n(irn, be_pos_Reload_mem, new_Bad());
 214                 set_irn_n(irn, be_pos_Reload_frame, new_Bad());
 215         }
 216 }
 217
 218 /**
 219  * Starts a walk for memory operands if supported by the backend.
 220  */
 221 static INLINE void check_for_memory_operands(be_chordal_env_t *chordal_env) {
 222         irg_walk_graph(chordal_env->irg, NULL, memory_operand_walker, chordal_env);
 223 }
 224
 225 /**
 226  * Sorry for doing stats again...
 227  */
 228 typedef struct _node_stat_t {
 229         unsigned int n_phis;      /**< Phis of the current register class. */
 230         unsigned int n_mem_phis;  /**< Memory Phis (Phis with spill operands). */
 231         unsigned int n_copies;    /**< Copies */
 232         unsigned int n_perms;     /**< Perms */
 233         unsigned int n_spills;    /**< Spill nodes */
 234         unsigned int n_reloads;   /**< Reloads */
 235 } node_stat_t;
 236
 237 struct node_stat_walker {
 238         node_stat_t      *stat;
 239         const arch_env_t *arch_env;
 240 };
 241
 242 static void node_stat_walker(ir_node *irn, void *data)
 243 {
 244         struct node_stat_walker *env  = data;
 245         const arch_env_t        *aenv = env->arch_env;
 246
 247         /* if the node is a normal phi */
 248         if(is_Phi(irn)) {
 249                 if (get_irn_mode(irn) == mode_M) {
 250                         env->stat->n_mem_phis++;
 251                 } else {
 252                         env->stat->n_phis++;
 253                 }
 254         } else {
 255                 arch_irn_class_t classify = arch_irn_classify(aenv, irn);
 256
 257                 if(classify & arch_irn_class_spill)
 258                         ++env->stat->n_spills;
 259                 if(classify & arch_irn_class_reload)
 260                         ++env->stat->n_reloads;
 261                 if(classify & arch_irn_class_copy)
 262                         ++env->stat->n_copies;
 263                 if(classify & arch_irn_class_perm)
 264                         ++env->stat->n_perms;
 265         }
 266 }
 267
 268 static void node_stats(be_irg_t *birg, node_stat_t *stat)
 269 {
 270         struct node_stat_walker env;
 271
 272         memset(stat, 0, sizeof(*stat));
 273         env.arch_env = birg->main_env->arch_env;
 274         env.stat     = stat;
 275         irg_walk_graph(birg->irg, NULL, node_stat_walker, &env);
 276 }
 277
 278 static void insn_count_walker(ir_node *irn, void *data)
 279 {
 280         unsigned long *cnt = data;
 281
 282         switch(get_irn_opcode(irn)) {
 283         case iro_Proj:
 284         case iro_Phi:
 285         case iro_Start:
 286         case iro_End:
 287                 break;
 288         default:
 289                 (*cnt)++;
 290         }
 291 }
 292
 293 static unsigned long count_insns(ir_graph *irg)
 294 {
 295         unsigned long cnt = 0;
 296         irg_walk_graph(irg, insn_count_walker, NULL, &cnt);
 297         return cnt;
 298 }
 299
 300 static void block_count_walker(ir_node *node, void *data)
 301 {
 302         unsigned long *cnt = data;
 303         if (node == get_irg_end_block(current_ir_graph))
 304                 return;
 305         (*cnt)++;
 306 }
 307
 308 static unsigned long count_blocks(ir_graph *irg)
 309 {
 310         unsigned long cnt = 0;
 311         irg_block_walk_graph(irg, block_count_walker, NULL, &cnt);
 312         return cnt;
 313 }
 314
 315 static node_stat_t last_node_stat;
 316
 317 /**
 318  * Perform things which need to be done per register class before spilling.
 319  */
 320 static void pre_spill(post_spill_env_t *pse, const arch_register_class_t *cls)
 321 {
 322         be_chordal_env_t    *chordal_env = &pse->cenv;
 323         be_irg_t            *birg        = pse->birg;
 324         ir_graph            *irg         = be_get_birg_irg(birg);
 325         const be_main_env_t *main_env    = birg->main_env;
 326
 327         pse->cls                   = cls;
 328         chordal_env->cls           = cls;
 329         chordal_env->border_heads  = pmap_create();
 330         chordal_env->ignore_colors = bitset_malloc(chordal_env->cls->n_regs);
 331
 332         be_assure_liveness(birg);
 333         be_liveness_assure_chk(be_get_birg_liveness(birg));
 334
 335         stat_ev_do(pse->pre_spill_cost = be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq));
 336
 337         /* put all ignore registers into the ignore register set. */
 338         be_put_ignore_regs(birg, pse->cls, chordal_env->ignore_colors);
 339
 340         BE_TIMER_PUSH(t_ra_constr);
 341         be_pre_spill_prepare_constr(chordal_env);
 342         BE_TIMER_POP(t_ra_constr);
 343
 344         dump(BE_CH_DUMP_CONSTR, birg->irg, pse->cls, "-constr-pre", dump_ir_block_graph_sched);
 345 }
 346
 347 /**
 348  * Perform things which need to be done per register class after spilling.
 349  */
 350 static void post_spill(post_spill_env_t *pse, int iteration) {
 351         be_chordal_env_t    *chordal_env = &pse->cenv;
 352         be_irg_t            *birg        = pse->birg;
 353         ir_graph            *irg         = birg->irg;
 354         const be_main_env_t *main_env    = birg->main_env;
 355         int                  colors_n     = arch_register_class_n_regs(chordal_env->cls);
 356         int             allocatable_regs = colors_n - be_put_ignore_regs(birg, chordal_env->cls, NULL);
 357
 358         /* some special classes contain only ignore regs, no work to be done */
 359         if (allocatable_regs > 0) {
 360                 stat_ev_dbl("bechordal_spillcosts", be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq) - pse->pre_spill_cost);
 361
 362                 /*
 363                         If we have a backend provided spiller, post spill is
 364                         called in a loop after spilling for each register class.
 365                         But we only need to fix stack nodes once in this case.
 366                 */
 367                 BE_TIMER_PUSH(t_ra_spill);
 368                 check_for_memory_operands(chordal_env);
 369                 if (iteration == 0) {
 370                         be_abi_fix_stack_nodes(birg->abi);
 371                 }
 372                 BE_TIMER_POP(t_ra_spill);
 373
 374                 BE_TIMER_PUSH(t_verify);
 375
 376                 /* verify schedule and register pressure */
 377                 if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 378                         be_verify_schedule(birg);
 379                         be_verify_register_pressure(birg, pse->cls, irg);
 380                 } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 381                         assert(be_verify_schedule(birg) && "Schedule verification failed");
 382                         assert(be_verify_register_pressure(birg, pse->cls, irg)
 383                                 && "Register pressure verification failed");
 384                 }
 385                 BE_TIMER_POP(t_verify);
 386
 387                 /* Color the graph. */
 388                 BE_TIMER_PUSH(t_ra_color);
 389                 be_ra_chordal_color(chordal_env);
 390                 BE_TIMER_POP(t_ra_color);
 391
 392                 dump(BE_CH_DUMP_CONSTR, irg, pse->cls, "-color", dump_ir_block_graph_sched);
 393
 394                 /* Create the ifg with the selected flavor */
 395                 BE_TIMER_PUSH(t_ra_ifg);
 396                 chordal_env->ifg = be_create_ifg(chordal_env);
 397                 BE_TIMER_POP(t_ra_ifg);
 398
 399                 stat_ev_if {
 400                         be_ifg_stat_t stat;
 401                         node_stat_t   node_stat;
 402
 403                         be_ifg_stat(birg, chordal_env->ifg, &stat);
 404                         stat_ev_dbl("bechordal_ifg_nodes", stat.n_nodes);
 405                         stat_ev_dbl("bechordal_ifg_edges", stat.n_edges);
 406                         stat_ev_dbl("bechordal_ifg_comps", stat.n_comps);
 407
 408                         node_stats(birg, &node_stat);
 409                         stat_ev_dbl("bechordal_perms_before_coal",
 410                                         node_stat.n_perms - last_node_stat.n_perms);
 411                         stat_ev_dbl("bechordal_copies_before_coal",
 412                                         node_stat.n_copies - last_node_stat.n_copies);
 413                 }
 414
 415                 /* copy minimization */
 416                 BE_TIMER_PUSH(t_ra_copymin);
 417                 co_driver(chordal_env);
 418                 BE_TIMER_POP(t_ra_copymin);
 419
 420                 dump(BE_CH_DUMP_COPYMIN, irg, pse->cls, "-copymin", dump_ir_block_graph_sched);
 421
 422
 423                 /* ssa destruction */
 424                 BE_TIMER_PUSH(t_ra_ssa);
 425                 be_ssa_destruction(chordal_env);
 426                 BE_TIMER_POP(t_ra_ssa);
 427
 428                 dump(BE_CH_DUMP_SSADESTR, irg, pse->cls, "-ssadestr", dump_ir_block_graph_sched);
 429
 430                 if (chordal_env->opts->vrfy_option != BE_CH_VRFY_OFF) {
 431                         BE_TIMER_PUSH(t_verify);
 432                         be_ssa_destruction_check(chordal_env);
 433                         BE_TIMER_POP(t_verify);
 434                 }
 435
 436                 /* the ifg exists only if there are allocatable regs */
 437                 be_ifg_free(chordal_env->ifg);
 438         }
 439
 440         /* free some always allocated data structures */
 441         pmap_destroy(chordal_env->border_heads);
 442         bitset_free(chordal_env->ignore_colors);
 443 }
 444
 445 /**
 446  * Performs chordal register allocation for each register class on given irg.
 447  *
 448  * @param birg  Backend irg object
 449  * @return Structure containing timer for the single phases or NULL if no timing requested.
 450  */
 451 static void be_ra_chordal_main(be_irg_t *birg)
 452 {
 453         const be_main_env_t *main_env = birg->main_env;
 454         const arch_env_t    *arch_env = main_env->arch_env;
 455         ir_graph            *irg      = birg->irg;
 456         int                 j, m;
 457         be_chordal_env_t    chordal_env;
 458         struct obstack      obst;
 459
 460         main_opts = main_env->options;
 461
 462         BE_TIMER_PUSH(t_ra_other);
 463
 464         BE_TIMER_PUSH(t_ra_prolog);
 465
 466         be_assure_liveness(birg);
 467
 468         chordal_env.obst          = &obst;
 469         chordal_env.opts          = &options;
 470         chordal_env.irg           = irg;
 471         chordal_env.birg          = birg;
 472         chordal_env.border_heads  = NULL;
 473         chordal_env.ifg           = NULL;
 474         chordal_env.ignore_colors = NULL;
 475
 476         obstack_init(&obst);
 477
 478         BE_TIMER_POP(t_ra_prolog);
 479
 480         stat_ev_if {
 481                 be_stat_ev("bechordal_insns_before", count_insns(irg));
 482                 be_stat_ev("bechordal_blocks_before", count_blocks(irg));
 483                 node_stats(birg, &last_node_stat);
 484         }
 485
 486         if (! arch_code_generator_has_spiller(birg->cg)) {
 487                 /* use one of the generic spiller */
 488
 489                 /* Perform the following for each register class. */
 490                 for (j = 0, m = arch_env_get_n_reg_class(arch_env); j < m; ++j) {
 491                         post_spill_env_t pse;
 492                         const arch_register_class_t *cls
 493                                 = arch_env_get_reg_class(arch_env, j);
 494
 495                         if(arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
 496                                 continue;
 497
 498
 499                         stat_ev_ctx_push_str("bechordal_cls", cls->name);
 500
 501                         memcpy(&pse.cenv, &chordal_env, sizeof(chordal_env));
 502                         pse.birg = birg;
 503                         pre_spill(&pse, cls);
 504
 505                         BE_TIMER_PUSH(t_ra_spill);
 506                         be_do_spill(birg, cls);
 507                         BE_TIMER_POP(t_ra_spill);
 508
 509                         dump(BE_CH_DUMP_SPILL, irg, pse.cls, "-spill",
 510                              dump_ir_block_graph_sched);
 511
 512                         post_spill(&pse, 0);
 513
 514                         stat_ev_if {
 515                                 node_stat_t node_stat;
 516
 517                                 node_stats(birg, &node_stat);
 518                                 stat_ev_dbl("bechordal_phis",
 519                                                 node_stat.n_phis - last_node_stat.n_phis);
 520                                 stat_ev_dbl("bechordal_mem_phis",
 521                                                 node_stat.n_mem_phis - last_node_stat.n_mem_phis);
 522                                 stat_ev_dbl("bechordal_reloads",
 523                                                 node_stat.n_reloads - last_node_stat.n_reloads);
 524                                 stat_ev_dbl("bechordal_spills",
 525                                                 node_stat.n_spills - last_node_stat.n_spills);
 526                                 stat_ev_dbl("bechordal_perms_after_coal",
 527                                                 node_stat.n_perms - last_node_stat.n_perms);
 528                                 stat_ev_dbl("bechordal_copies_after_coal",
 529                                                 node_stat.n_copies - last_node_stat.n_copies);
 530
 531                                 last_node_stat = node_stat;
 532                                 stat_ev_ctx_pop("bechordal_cls");
 533                         }
 534                 }
 535         } else {
 536                 post_spill_env_t *pse;
 537
 538                 /* the backend has its own spiller */
 539                 m = arch_env_get_n_reg_class(arch_env);
 540
 541                 pse = alloca(m * sizeof(pse[0]));
 542
 543                 for (j = 0; j < m; ++j) {
 544                         memcpy(&pse[j].cenv, &chordal_env, sizeof(chordal_env));
 545                         pse[j].birg = birg;
 546                         pre_spill(&pse[j], pse[j].cls);
 547                 }
 548
 549                 BE_TIMER_PUSH(t_ra_spill);
 550                 arch_code_generator_spill(birg->cg, birg);
 551                 BE_TIMER_POP(t_ra_spill);
 552                 dump(BE_CH_DUMP_SPILL, irg, NULL, "-spill", dump_ir_block_graph_sched);
 553
 554                 for (j = 0; j < m; ++j) {
 555                         post_spill(&pse[j], j);
 556                 }
 557         }
 558
 559         BE_TIMER_PUSH(t_verify);
 560         be_verify_register_allocation(birg);
 561         BE_TIMER_POP(t_verify);
 562
 563         BE_TIMER_PUSH(t_ra_epilog);
 564         lower_nodes_after_ra(birg, options.lower_perm_opt & BE_CH_LOWER_PERM_COPY ? 1 : 0);
 565         dump(BE_CH_DUMP_LOWER, irg, NULL, "-belower-after-ra", dump_ir_block_graph_sched);
 566
 567         obstack_free(&obst, NULL);
 568         be_liveness_invalidate(be_get_birg_liveness(birg));
 569         BE_TIMER_POP(t_ra_epilog);
 570
 571         BE_TIMER_POP(t_ra_other);
 572
 573         stat_ev_if {
 574                 be_stat_ev("bechordal_insns_after", count_insns(irg));
 575         }
 576 }
 577
 578 static be_ra_t be_ra_chordal_allocator = {
 579         be_ra_chordal_main,
 580 };
 581
 582 void be_init_chordal_main(void)
 583 {
 584         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 585         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
 586         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
 587
 588         lc_opt_add_table(chordal_grp, be_chordal_options);
 589
 590         be_register_allocator("chordal", &be_ra_chordal_allocator);
 591 }
 592
 593 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal_main);