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         unsigned int n_remats;    /**< Remats */
 236 } node_stat_t;
 237
 238 struct node_stat_walker {
 239         node_stat_t      *stat;
 240         const arch_env_t *arch_env;
 241 };
 242
 243 static void node_stat_walker(ir_node *irn, void *data)
 244 {
 245         struct node_stat_walker *env  = data;
 246         const arch_env_t        *aenv = env->arch_env;
 247
 248         /* if the node is a normal phi */
 249         if(is_Phi(irn)) {
 250                 if (get_irn_mode(irn) == mode_M) {
 251                         env->stat->n_mem_phis++;
 252                 } else {
 253                         env->stat->n_phis++;
 254                 }
 255         } else {
 256                 arch_irn_class_t classify = arch_irn_classify(aenv, irn);
 257
 258                 if(classify & arch_irn_class_spill)
 259                         ++env->stat->n_spills;
 260                 if(classify & arch_irn_class_reload)
 261                         ++env->stat->n_reloads;
 262                 if(classify & arch_irn_class_remat)
 263                         ++env->stat->n_remats;
 264                 if(classify & arch_irn_class_copy)
 265                         ++env->stat->n_copies;
 266                 if(classify & arch_irn_class_perm)
 267                         ++env->stat->n_perms;
 268         }
 269 }
 270
 271 static void node_stats(be_irg_t *birg, node_stat_t *stat)
 272 {
 273         struct node_stat_walker env;
 274
 275         memset(stat, 0, sizeof(*stat));
 276         env.arch_env = birg->main_env->arch_env;
 277         env.stat     = stat;
 278         irg_walk_graph(birg->irg, NULL, node_stat_walker, &env);
 279 }
 280
 281 static void insn_count_walker(ir_node *irn, void *data)
 282 {
 283         unsigned long *cnt = data;
 284
 285         switch(get_irn_opcode(irn)) {
 286         case iro_Proj:
 287         case iro_Phi:
 288         case iro_Start:
 289         case iro_End:
 290                 break;
 291         default:
 292                 (*cnt)++;
 293         }
 294 }
 295
 296 static unsigned long count_insns(ir_graph *irg)
 297 {
 298         unsigned long cnt = 0;
 299         irg_walk_graph(irg, insn_count_walker, NULL, &cnt);
 300         return cnt;
 301 }
 302
 303 static void block_count_walker(ir_node *node, void *data)
 304 {
 305         unsigned long *cnt = data;
 306         if (node == get_irg_end_block(current_ir_graph))
 307                 return;
 308         (*cnt)++;
 309 }
 310
 311 static unsigned long count_blocks(ir_graph *irg)
 312 {
 313         unsigned long cnt = 0;
 314         irg_block_walk_graph(irg, block_count_walker, NULL, &cnt);
 315         return cnt;
 316 }
 317
 318 static node_stat_t last_node_stat;
 319
 320 /**
 321  * Perform things which need to be done per register class before spilling.
 322  */
 323 static void pre_spill(post_spill_env_t *pse, const arch_register_class_t *cls)
 324 {
 325         be_chordal_env_t    *chordal_env = &pse->cenv;
 326         be_irg_t            *birg        = pse->birg;
 327         ir_graph            *irg         = be_get_birg_irg(birg);
 328         const be_main_env_t *main_env    = birg->main_env;
 329
 330         pse->cls                   = cls;
 331         chordal_env->cls           = cls;
 332         chordal_env->border_heads  = pmap_create();
 333         chordal_env->ignore_colors = bitset_malloc(chordal_env->cls->n_regs);
 334
 335         be_assure_liveness(birg);
 336         be_liveness_assure_chk(be_get_birg_liveness(birg));
 337
 338         stat_ev_do(pse->pre_spill_cost = be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq));
 339
 340         /* put all ignore registers into the ignore register set. */
 341         be_put_ignore_regs(birg, pse->cls, chordal_env->ignore_colors);
 342
 343         BE_TIMER_PUSH(t_ra_constr);
 344         be_pre_spill_prepare_constr(chordal_env);
 345         BE_TIMER_POP(t_ra_constr);
 346
 347         dump(BE_CH_DUMP_CONSTR, birg->irg, pse->cls, "-constr-pre", dump_ir_block_graph_sched);
 348 }
 349
 350 /**
 351  * Perform things which need to be done per register class after spilling.
 352  */
 353 static void post_spill(post_spill_env_t *pse, int iteration) {
 354         be_chordal_env_t    *chordal_env = &pse->cenv;
 355         be_irg_t            *birg        = pse->birg;
 356         ir_graph            *irg         = birg->irg;
 357         const be_main_env_t *main_env    = birg->main_env;
 358         int                  colors_n     = arch_register_class_n_regs(chordal_env->cls);
 359         int             allocatable_regs = colors_n - be_put_ignore_regs(birg, chordal_env->cls, NULL);
 360
 361         /* some special classes contain only ignore regs, no work to be done */
 362         if (allocatable_regs > 0) {
 363                 stat_ev_dbl("bechordal_spillcosts", be_estimate_irg_costs(irg, main_env->arch_env, birg->exec_freq) - pse->pre_spill_cost);
 364
 365                 /*
 366                         If we have a backend provided spiller, post spill is
 367                         called in a loop after spilling for each register class.
 368                         But we only need to fix stack nodes once in this case.
 369                 */
 370                 BE_TIMER_PUSH(t_ra_spill_apply);
 371                 check_for_memory_operands(chordal_env);
 372                 if (iteration == 0) {
 373                         be_abi_fix_stack_nodes(birg->abi);
 374                 }
 375                 BE_TIMER_POP(t_ra_spill_apply);
 376
 377                 BE_TIMER_PUSH(t_verify);
 378
 379                 /* verify schedule and register pressure */
 380                 if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
 381                         be_verify_schedule(birg);
 382                         be_verify_register_pressure(birg, pse->cls, irg);
 383                 } else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
 384                         assert(be_verify_schedule(birg) && "Schedule verification failed");
 385                         assert(be_verify_register_pressure(birg, pse->cls, irg)
 386                                 && "Register pressure verification failed");
 387                 }
 388                 BE_TIMER_POP(t_verify);
 389
 390                 /* Color the graph. */
 391                 BE_TIMER_PUSH(t_ra_color);
 392                 be_ra_chordal_color(chordal_env);
 393                 BE_TIMER_POP(t_ra_color);
 394
 395                 dump(BE_CH_DUMP_CONSTR, irg, pse->cls, "-color", dump_ir_block_graph_sched);
 396
 397                 /* Create the ifg with the selected flavor */
 398                 BE_TIMER_PUSH(t_ra_ifg);
 399                 chordal_env->ifg = be_create_ifg(chordal_env);
 400                 BE_TIMER_POP(t_ra_ifg);
 401
 402                 stat_ev_if {
 403                         be_ifg_stat_t stat;
 404                         node_stat_t   node_stat;
 405
 406                         be_ifg_stat(birg, chordal_env->ifg, &stat);
 407                         stat_ev_dbl("bechordal_ifg_nodes", stat.n_nodes);
 408                         stat_ev_dbl("bechordal_ifg_edges", stat.n_edges);
 409                         stat_ev_dbl("bechordal_ifg_comps", stat.n_comps);
 410
 411                         node_stats(birg, &node_stat);
 412                         stat_ev_dbl("bechordal_perms_before_coal",
 413                                         node_stat.n_perms - last_node_stat.n_perms);
 414                         stat_ev_dbl("bechordal_copies_before_coal",
 415                                         node_stat.n_copies - last_node_stat.n_copies);
 416                 }
 417
 418                 /* copy minimization */
 419                 BE_TIMER_PUSH(t_ra_copymin);
 420                 co_driver(chordal_env);
 421                 BE_TIMER_POP(t_ra_copymin);
 422
 423                 dump(BE_CH_DUMP_COPYMIN, irg, pse->cls, "-copymin", dump_ir_block_graph_sched);
 424
 425
 426                 /* ssa destruction */
 427                 BE_TIMER_PUSH(t_ra_ssa);
 428                 be_ssa_destruction(chordal_env);
 429                 BE_TIMER_POP(t_ra_ssa);
 430
 431                 dump(BE_CH_DUMP_SSADESTR, irg, pse->cls, "-ssadestr", dump_ir_block_graph_sched);
 432
 433                 if (chordal_env->opts->vrfy_option != BE_CH_VRFY_OFF) {
 434                         BE_TIMER_PUSH(t_verify);
 435                         be_ssa_destruction_check(chordal_env);
 436                         BE_TIMER_POP(t_verify);
 437                 }
 438
 439                 /* the ifg exists only if there are allocatable regs */
 440                 be_ifg_free(chordal_env->ifg);
 441         }
 442
 443         /* free some always allocated data structures */
 444         pmap_destroy(chordal_env->border_heads);
 445         bitset_free(chordal_env->ignore_colors);
 446 }
 447
 448 /**
 449  * Performs chordal register allocation for each register class on given irg.
 450  *
 451  * @param birg  Backend irg object
 452  * @return Structure containing timer for the single phases or NULL if no timing requested.
 453  */
 454 static void be_ra_chordal_main(be_irg_t *birg)
 455 {
 456         const be_main_env_t *main_env = birg->main_env;
 457         const arch_env_t    *arch_env = main_env->arch_env;
 458         ir_graph            *irg      = birg->irg;
 459         int                 j, m;
 460         be_chordal_env_t    chordal_env;
 461         struct obstack      obst;
 462
 463         main_opts = main_env->options;
 464
 465         BE_TIMER_PUSH(t_ra_other);
 466
 467         BE_TIMER_PUSH(t_ra_prolog);
 468
 469         be_assure_liveness(birg);
 470
 471         chordal_env.obst          = &obst;
 472         chordal_env.opts          = &options;
 473         chordal_env.irg           = irg;
 474         chordal_env.birg          = birg;
 475         chordal_env.border_heads  = NULL;
 476         chordal_env.ifg           = NULL;
 477         chordal_env.ignore_colors = NULL;
 478
 479         obstack_init(&obst);
 480
 481         BE_TIMER_POP(t_ra_prolog);
 482
 483         stat_ev_if {
 484                 be_stat_ev("bechordal_insns_before", count_insns(irg));
 485                 be_stat_ev("bechordal_blocks_before", count_blocks(irg));
 486                 node_stats(birg, &last_node_stat);
 487         }
 488
 489         if (! arch_code_generator_has_spiller(birg->cg)) {
 490                 /* use one of the generic spiller */
 491
 492                 /* Perform the following for each register class. */
 493                 for (j = 0, m = arch_env_get_n_reg_class(arch_env); j < m; ++j) {
 494                         post_spill_env_t pse;
 495                         const arch_register_class_t *cls
 496                                 = arch_env_get_reg_class(arch_env, j);
 497
 498                         if(arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
 499                                 continue;
 500
 501
 502                         stat_ev_ctx_push_str("bechordal_cls", cls->name);
 503
 504                         stat_ev_if {
 505                                 be_do_stat_reg_pressure(birg, cls);
 506                         }
 507
 508                         memcpy(&pse.cenv, &chordal_env, sizeof(chordal_env));
 509                         pse.birg = birg;
 510                         pre_spill(&pse, cls);
 511
 512                         BE_TIMER_PUSH(t_ra_spill);
 513                         be_do_spill(birg, cls);
 514                         BE_TIMER_POP(t_ra_spill);
 515
 516                         dump(BE_CH_DUMP_SPILL, irg, pse.cls, "-spill",
 517                              dump_ir_block_graph_sched);
 518
 519                         post_spill(&pse, 0);
 520
 521                         stat_ev_if {
 522                                 node_stat_t node_stat;
 523
 524                                 node_stats(birg, &node_stat);
 525                                 stat_ev_dbl("bechordal_phis",
 526                                                 node_stat.n_phis - last_node_stat.n_phis);
 527                                 stat_ev_dbl("bechordal_mem_phis",
 528                                                 node_stat.n_mem_phis - last_node_stat.n_mem_phis);
 529                                 stat_ev_dbl("bechordal_reloads",
 530                                                 node_stat.n_reloads - last_node_stat.n_reloads);
 531                                 stat_ev_dbl("bechordal_remats",
 532                                                 node_stat.n_remats - last_node_stat.n_remats);
 533                                 stat_ev_dbl("bechordal_spills",
 534                                                 node_stat.n_spills - last_node_stat.n_spills);
 535                                 stat_ev_dbl("bechordal_perms_after_coal",
 536                                                 node_stat.n_perms - last_node_stat.n_perms);
 537                                 stat_ev_dbl("bechordal_copies_after_coal",
 538                                                 node_stat.n_copies - last_node_stat.n_copies);
 539
 540                                 last_node_stat = node_stat;
 541                                 stat_ev_ctx_pop("bechordal_cls");
 542                         }
 543                 }
 544         } else {
 545                 post_spill_env_t *pse;
 546
 547                 /* the backend has its own spiller */
 548                 m = arch_env_get_n_reg_class(arch_env);
 549
 550                 pse = alloca(m * sizeof(pse[0]));
 551
 552                 for (j = 0; j < m; ++j) {
 553                         memcpy(&pse[j].cenv, &chordal_env, sizeof(chordal_env));
 554                         pse[j].birg = birg;
 555                         pre_spill(&pse[j], pse[j].cls);
 556                 }
 557
 558                 BE_TIMER_PUSH(t_ra_spill);
 559                 arch_code_generator_spill(birg->cg, birg);
 560                 BE_TIMER_POP(t_ra_spill);
 561                 dump(BE_CH_DUMP_SPILL, irg, NULL, "-spill", dump_ir_block_graph_sched);
 562
 563                 for (j = 0; j < m; ++j) {
 564                         post_spill(&pse[j], j);
 565                 }
 566         }
 567
 568         BE_TIMER_PUSH(t_verify);
 569         be_verify_register_allocation(birg);
 570         BE_TIMER_POP(t_verify);
 571
 572         BE_TIMER_PUSH(t_ra_epilog);
 573         lower_nodes_after_ra(birg, options.lower_perm_opt & BE_CH_LOWER_PERM_COPY ? 1 : 0);
 574         dump(BE_CH_DUMP_LOWER, irg, NULL, "-belower-after-ra", dump_ir_block_graph_sched);
 575
 576         obstack_free(&obst, NULL);
 577         be_liveness_invalidate(be_get_birg_liveness(birg));
 578         BE_TIMER_POP(t_ra_epilog);
 579
 580         BE_TIMER_POP(t_ra_other);
 581
 582         stat_ev_if {
 583                 be_stat_ev("bechordal_insns_after", count_insns(irg));
 584         }
 585 }
 586
 587 static be_ra_t be_ra_chordal_allocator = {
 588         be_ra_chordal_main,
 589 };
 590
 591 void be_init_chordal_main(void)
 592 {
 593         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 594         lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
 595         lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
 596
 597         lc_opt_add_table(chordal_grp, be_chordal_options);
 598
 599         be_register_allocator("chordal", &be_ra_chordal_allocator);
 600 }
 601
 602 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal_main);