nsz Git - libfirm/blob - ir/be/beblocksched.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       Block-scheduling strategies.
  23  * @author      Matthias Braun, Christoph Mallon
  24  * @date        27.09.2006
  25  * @version     $Id$
  26  *
  27  * The goals of the greedy (and ILP) algorithm here works by assuming that
  28  * we want to change as many jumps to fallthroughs as possible (executed jumps
  29  * actually, we have to look at the execution frequencies). The algorithms
  30  * do this by collecting execution frequencies of all branches (which is easily
  31  * possible when all critical edges are split) then removes critical edges where
  32  * possible as we don't need and want them anymore now. The algorithms then try
  33  * to change as many edges to fallthroughs as possible, this is done by setting
  34  * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
  35  * execution frequencies and tries to transform them to fallthroughs in this order
  36  */
  37 #include "config.h"
  38
  39 #include "beblocksched.h"
  40
  41 #include <stdlib.h>
  42
  43 #include "array.h"
  44 #include "pdeq.h"
  45
  46 #include "iredges.h"
  47 #include "irgwalk.h"
  48 #include "irnode_t.h"
  49 #include "irgraph_t.h"
  50 #include "irloop.h"
  51 #include "irprintf.h"
  52 #include "execfreq.h"
  53 #include "irdump_t.h"
  54 #include "irtools.h"
  55 #include "debug.h"
  56 #include "beirgmod.h"
  57 #include "bemodule.h"
  58 #include "be.h"
  59 #include "error.h"
  60
  61 #include "lc_opts.h"
  62 #include "lc_opts_enum.h"
  63
  64 #ifdef WITH_ILP
  65 #include <lpp/lpp.h>
  66 #include <lpp/lpp_net.h>
  67 #endif /* WITH_ILP */
  68
  69 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  70
  71 typedef enum _blocksched_algos_t {
  72         BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
  73 } blocksched_algos_t;
  74
  75 static int algo = BLOCKSCHED_GREEDY;
  76
  77 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
  78         { "naiv",       BLOCKSCHED_NAIV },
  79         { "extbb",      BLOCKSCHED_EXTBB },
  80         { "greedy", BLOCKSCHED_GREEDY },
  81 #ifdef WITH_ILP
  82         { "ilp",    BLOCKSCHED_ILP },
  83 #endif /* WITH_ILP */
  84         { NULL,     0 }
  85 };
  86
  87 static lc_opt_enum_int_var_t algo_var = {
  88         &algo, blockschedalgo_items
  89 };
  90
  91 static const lc_opt_table_entry_t be_blocksched_options[] = {
  92         LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
  93         LC_OPT_LAST
  94 };
  95
  96 /*
  97  *   ____                   _
  98  *  / ___|_ __ ___  ___  __| |_   _
  99  * | |  _| '__/ _ \/ _ \/ _` | | | |
 100  * | |_| | | |  __/  __/ (_| | |_| |
 101  *  \____|_|  \___|\___|\__,_|\__, |
 102  *                            |___/
 103  */
 104
 105 typedef struct _blocksched_entry_t {
 106         ir_node *block;
 107         struct _blocksched_entry_t *next;
 108         struct _blocksched_entry_t *prev;
 109 } blocksched_entry_t;
 110
 111 typedef struct _edge_t {
 112         ir_node *block;             /**< source block */
 113         int     pos;                /**< number of cfg predecessor (target) */
 114         double  execfreq;           /**< the frequency */
 115         int     highest_execfreq;   /**< flag that indicates whether this edge is the edge with the highest
 116                                          execfreq pointing away from this block */
 117 } edge_t;
 118
 119 typedef struct _blocksched_env_t {
 120         ir_graph       *irg;
 121         struct obstack *obst;
 122         ir_exec_freq   *execfreqs;
 123         edge_t         *edges;
 124         pdeq           *worklist;
 125         int            blockcount;
 126 } blocksched_env_t;
 127
 128 /**
 129  * Collect cfg frequencies of all edges between blocks.
 130  * Also determines edge with highest frequency.
 131  */
 132 static void collect_egde_frequency(ir_node *block, void *data)
 133 {
 134         blocksched_env_t   *env = data;
 135         int                arity;
 136         edge_t             edge;
 137         blocksched_entry_t *entry;
 138
 139         entry        = obstack_alloc(env->obst, sizeof(entry[0]));
 140         entry->block = block;
 141         entry->next  = NULL;
 142         entry->prev  = NULL;
 143         set_irn_link(block, entry);
 144
 145         arity = get_Block_n_cfgpreds(block);
 146
 147         if (arity == 0) {
 148                 assert(block == get_irg_start_block(env->irg)
 149                                 || block == get_irg_end_block(env->irg));
 150                 /* must be the start block (or end-block for endless loops), nothing to
 151                  * do here */
 152                 return;
 153         } else if (arity == 1) {
 154                 edge.block            = block;
 155                 edge.pos              = 0;
 156                 edge.execfreq         = get_block_execfreq(env->execfreqs, block);
 157                 edge.highest_execfreq = 1;
 158                 ARR_APP1(edge_t, env->edges, edge);
 159         } else {
 160                 int    i;
 161                 double highest_execfreq = -1.0;
 162                 int    highest_edge_num = -1;
 163
 164                 edge.block = block;
 165                 for (i = 0; i < arity; ++i) {
 166                         double  execfreq;
 167                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 168
 169                         execfreq = get_block_execfreq(env->execfreqs, pred_block);
 170
 171                         edge.pos              = i;
 172                         edge.execfreq         = execfreq;
 173                         edge.highest_execfreq = 0;
 174                         ARR_APP1(edge_t, env->edges, edge);
 175
 176                         if (execfreq > highest_execfreq) {
 177                                 highest_execfreq = execfreq;
 178                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 179                         }
 180                 }
 181
 182                 if(highest_edge_num >= 0)
 183                         env->edges[highest_edge_num].highest_execfreq = 1;
 184         }
 185 }
 186
 187 static int cmp_edges(const void *d1, const void *d2)
 188 {
 189         const edge_t *e1 = d1;
 190         const edge_t *e2 = d2;
 191
 192         return QSORT_CMP(e2->execfreq, e1->execfreq);
 193 }
 194
 195 static void coalesce_blocks(blocksched_env_t *env)
 196 {
 197         int i;
 198         int edge_count = ARR_LEN(env->edges);
 199
 200         /* run1: only look at jumps */
 201         for (i = 0; i < edge_count; ++i) {
 202                 const edge_t *edge  = &env->edges[i];
 203                 ir_node      *block = edge->block;
 204                 int           pos   = edge->pos;
 205                 ir_node      *pred_block;
 206                 blocksched_entry_t *entry, *pred_entry;
 207
 208                 /* only check edge with highest frequency */
 209                 if (! edge->highest_execfreq)
 210                         continue;
 211
 212                 /* the block might have been removed already... */
 213                 if (is_Bad(get_Block_cfgpred(block, 0)))
 214                         continue;
 215
 216                 pred_block = get_Block_cfgpred_block(block, pos);
 217                 entry      = get_irn_link(block);
 218                 pred_entry = get_irn_link(pred_block);
 219
 220                 if (pred_entry->next != NULL || entry->prev != NULL)
 221                         continue;
 222
 223                 /* only coalesce jumps */
 224                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 225                         continue;
 226
 227                 /* schedule the 2 blocks behind each other */
 228                 DBG((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 229                            pred_entry->block, entry->block, edge->execfreq));
 230                 pred_entry->next = entry;
 231                 entry->prev      = pred_entry;
 232         }
 233
 234         /* run2: remaining edges */
 235         for (i = 0; i < edge_count; ++i) {
 236                 const edge_t *edge  = &env->edges[i];
 237                 ir_node      *block = edge->block;
 238                 int           pos   = edge->pos;
 239                 ir_node      *pred_block;
 240                 blocksched_entry_t *entry, *pred_entry;
 241
 242                 /* the block might have been removed already... */
 243                 if (is_Bad(get_Block_cfgpred(block, 0)))
 244                         continue;
 245
 246                 /* we can't do fallthroughs in backedges */
 247                 if (is_backedge(block, pos))
 248                         continue;
 249
 250                 pred_block = get_Block_cfgpred_block(block, pos);
 251                 entry      = get_irn_link(block);
 252                 pred_entry = get_irn_link(pred_block);
 253
 254                 /* is 1 of the blocks already attached to another block? */
 255                 if (pred_entry->next != NULL || entry->prev != NULL)
 256                         continue;
 257
 258                 /* schedule the 2 blocks behind each other */
 259                 DBG((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 260                            pred_entry->block, entry->block, edge->execfreq));
 261                 pred_entry->next = entry;
 262                 entry->prev      = pred_entry;
 263         }
 264 }
 265
 266 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 267 {
 268         ir_node            *block = entry->block;
 269         ir_node            *succ  = NULL;
 270         blocksched_entry_t *succ_entry;
 271         const ir_edge_t    *edge;
 272         double             best_succ_execfreq;
 273
 274         if (irn_visited_else_mark(block))
 275                 return;
 276
 277         env->blockcount++;
 278
 279         DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 280
 281         /* put all successors into the worklist */
 282         foreach_block_succ(block, edge) {
 283                 ir_node *succ_block = get_edge_src_irn(edge);
 284
 285                 if (irn_visited(succ_block))
 286                         continue;
 287
 288                 /* we only need to put the first of a series of already connected
 289                  * blocks into the worklist */
 290                 succ_entry = get_irn_link(succ_block);
 291                 while (succ_entry->prev != NULL) {
 292                         /* break cycles... */
 293                         if (succ_entry->prev->block == succ_block) {
 294                                 succ_entry->prev->next = NULL;
 295                                 succ_entry->prev       = NULL;
 296                                 break;
 297                         }
 298                         succ_entry = succ_entry->prev;
 299                 };
 300
 301                 if (irn_visited(succ_entry->block))
 302                         continue;
 303
 304                 DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 305                 pdeq_putr(env->worklist, succ_entry->block);
 306         }
 307
 308         if (entry->next != NULL) {
 309                 pick_block_successor(entry->next, env);
 310                 return;
 311         }
 312
 313         DBG((dbg, LEVEL_1, "deciding...\n"));
 314         best_succ_execfreq = -1;
 315
 316         /* no successor yet: pick the successor block with the highest execution
 317          * frequency which has no predecessor yet */
 318
 319         foreach_block_succ(block, edge) {
 320                 ir_node *succ_block = get_edge_src_irn(edge);
 321                 double  execfreq;
 322
 323                 if (irn_visited(succ_block))
 324                         continue;
 325
 326                 succ_entry = get_irn_link(succ_block);
 327                 if (succ_entry->prev != NULL)
 328                         continue;
 329
 330                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 331                 if (execfreq > best_succ_execfreq) {
 332                         best_succ_execfreq = execfreq;
 333                         succ = succ_block;
 334                 }
 335         }
 336
 337         if (succ == NULL) {
 338                 DBG((dbg, LEVEL_1, "pick from worklist\n"));
 339
 340                 do {
 341                         if (pdeq_empty(env->worklist)) {
 342                                 DBG((dbg, LEVEL_1, "worklist empty\n"));
 343                                 return;
 344                         }
 345                         succ = pdeq_getl(env->worklist);
 346                 } while (irn_visited(succ));
 347         }
 348
 349         succ_entry       = get_irn_link(succ);
 350         entry->next      = succ_entry;
 351         succ_entry->prev = entry;
 352
 353         pick_block_successor(succ_entry, env);
 354 }
 355
 356 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 357 {
 358         ir_graph           *irg        = env->irg;
 359         ir_node            *startblock = get_irg_start_block(irg);
 360         blocksched_entry_t *entry      = get_irn_link(startblock);
 361
 362         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 363         inc_irg_visited(irg);
 364
 365         env->worklist = new_pdeq();
 366         pick_block_successor(entry, env);
 367         assert(pdeq_empty(env->worklist));
 368         del_pdeq(env->worklist);
 369
 370         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 371
 372         return entry;
 373 }
 374
 375 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 376                                                                                 int count, struct obstack* obst)
 377 {
 378         int                i = 0;
 379         ir_node            **block_list;
 380         blocksched_entry_t *entry;
 381         (void) env;
 382
 383         block_list = NEW_ARR_D(ir_node *, obst, count);
 384         DBG((dbg, LEVEL_1, "Blockschedule:\n"));
 385
 386         for (entry = first; entry != NULL; entry = entry->next) {
 387                 assert(i < count);
 388                 block_list[i++] = entry->block;
 389                 DBG((dbg, LEVEL_1, "\t%+F\n", entry->block));
 390         }
 391         assert(i == count);
 392
 393         return block_list;
 394 }
 395
 396 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 397 {
 398         blocksched_env_t   env;
 399         struct obstack     obst;
 400         blocksched_entry_t *start_entry;
 401         ir_node            **block_list;
 402
 403         obstack_init(&obst);
 404
 405         env.irg        = irg;
 406         env.obst       = &obst;
 407         env.execfreqs  = execfreqs;
 408         env.edges      = NEW_ARR_F(edge_t, 0);
 409         env.worklist   = NULL;
 410         env.blockcount = 0;
 411
 412         // collect edge execution frequencies
 413         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 414
 415         // sort interblock edges by execution frequency
 416         qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
 417
 418         (void)be_remove_empty_blocks(irg);
 419
 420         if (algo != BLOCKSCHED_NAIV)
 421                 coalesce_blocks(&env);
 422
 423         start_entry = finish_block_schedule(&env);
 424         block_list  = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
 425
 426         DEL_ARR_F(env.edges);
 427         obstack_free(&obst, NULL);
 428
 429         return block_list;
 430 }
 431
 432 /*
 433  *  ___ _     ____
 434  * |_ _| |   |  _ \
 435  *  | || |   | |_) |
 436  *  | || |___|  __/
 437  * |___|_____|_|
 438  *
 439  */
 440
 441 #ifdef WITH_ILP
 442 typedef struct _ilp_edge_t {
 443         ir_node *block;   /**< source block */
 444         int     pos;      /**< number of cfg predecessor (target) */
 445         int     ilpvar;
 446 } ilp_edge_t;
 447
 448 typedef struct _blocksched_ilp_env_t {
 449         blocksched_env_t env;
 450         ilp_edge_t       *ilpedges;
 451         lpp_t            *lpp;
 452 } blocksched_ilp_env_t;
 453
 454 typedef struct _blocksched_ilp_entry_t {
 455         ir_node *block;
 456         struct _blocksched_entry_t *next;
 457         struct _blocksched_entry_t *prev;
 458
 459         int out_cst;
 460 } blocksched_ilp_entry_t;
 461
 462 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 463 {
 464         char       name[64];
 465         ilp_edge_t edge;
 466         int        edgeidx = ARR_LEN(env->ilpedges);
 467
 468         snprintf(name, sizeof(name), "edge%d", edgeidx);
 469
 470         edge.block  = block;
 471         edge.pos    = pos;
 472         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 473
 474         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 475         return edgeidx;
 476 }
 477
 478 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 479 {
 480         blocksched_ilp_env_t *env        = data;
 481         ir_graph             *irg        = env->env.irg;
 482         ir_node              *startblock = get_irg_start_block(irg);
 483         int                  arity;
 484         lpp_cst_t            cst;
 485         char                 name[64];
 486         int                  out_count;
 487         blocksched_ilp_entry_t *entry;
 488
 489         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 490         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 491
 492         entry          = obstack_alloc(env->env.obst, sizeof(entry[0]));
 493         entry->block   = block;
 494         entry->next    = NULL;
 495         entry->prev    = NULL;
 496         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
 497         set_irn_link(block, entry);
 498
 499         if (block == startblock)
 500                 return;
 501
 502         arity = get_irn_arity(block);
 503         if (arity == 1) {
 504                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 505                 add_ilp_edge(block, 0, execfreq, env);
 506         }
 507         else {
 508                 int i;
 509                 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
 510
 511                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 512                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
 513
 514                 for (i = 0; i < arity; ++i) {
 515                         double     execfreq;
 516                         int        edgenum;
 517                         ilp_edge_t *edge;
 518                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 519
 520                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 521                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 522                         edge     = &env->ilpedges[edgenum];
 523                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 524                 }
 525         }
 526 }
 527
 528
 529 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 530 {
 531         int  i;
 532         int  edge_count = ARR_LEN(env->ilpedges);
 533
 534         /* complete out constraints */
 535         for(i = 0; i < edge_count; ++i) {
 536                 const ilp_edge_t *edge  = &env->ilpedges[i];
 537                 ir_node          *block = edge->block;
 538                 ir_node          *pred;
 539                 blocksched_ilp_entry_t *entry;
 540
 541                 /* the block might have been removed already... */
 542                 if (is_Bad(get_Block_cfgpred(block, 0)))
 543                         continue;
 544
 545                 pred  = get_Block_cfgpred_block(block, edge->pos);
 546                 entry = get_irn_link(pred);
 547
 548                 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 549                                   pred, block, edge->pos));
 550                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 551         }
 552
 553 #if 0
 554         {
 555                 FILE *f;
 556                 char fname[256];
 557                 lpp_dump(env->lpp, "lpp.out");
 558                 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
 559                 f = fopen(fname, "w");
 560                 lpp_dump_plain(env->lpp, f);
 561                 fclose(f);
 562         }
 563 #endif
 564
 565         //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
 566         lpp_solve_net(env->lpp, "i44pc52", "cplex");
 567         assert(lpp_is_sol_valid(env->lpp));
 568
 569         /* Apply results to edges */
 570         for (i = 0; i < edge_count; ++i) {
 571                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 572                 ir_node            *block = edge->block;
 573                 ir_node            *pred;
 574                 int                is_jump;
 575                 blocksched_entry_t *entry;
 576                 blocksched_entry_t *pred_entry;
 577
 578                 /* the block might have been removed already... */
 579                 if (is_Bad(get_Block_cfgpred(block, 0)))
 580                         continue;
 581
 582                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 583                 if (is_jump)
 584                         continue;
 585
 586                 pred       = get_Block_cfgpred_block(block, edge->pos);
 587                 entry      = get_irn_link(block);
 588                 pred_entry = get_irn_link(pred);
 589
 590                 assert(entry->prev == NULL && pred_entry->next == NULL);
 591                 entry->prev      = pred_entry;
 592                 pred_entry->next = entry;
 593         }
 594 }
 595
 596 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 597 {
 598         blocksched_ilp_env_t env;
 599         struct obstack       obst;
 600         blocksched_entry_t   *start_entry;
 601         ir_node              **block_list;
 602
 603         obstack_init(&obst);
 604
 605         env.env.irg        = irg;
 606         env.env.obst       = &obst;
 607         env.env.execfreqs  = execfreqs;
 608         env.env.worklist   = NULL;
 609         env.env.blockcount = 0;
 610         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 611
 612         env.lpp = new_lpp("blockschedule", lpp_minimize);
 613         lpp_set_time_limit(env.lpp, 20);
 614         lpp_set_log(env.lpp, stdout);
 615
 616         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 617
 618         (void)be_remove_empty_blocks(irg);
 619         coalesce_blocks_ilp(&env);
 620
 621         start_entry = finish_block_schedule(&env.env);
 622         block_list  = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
 623
 624         DEL_ARR_F(env.ilpedges);
 625         free_lpp(env.lpp);
 626         obstack_free(&obst, NULL);
 627
 628         return block_list;
 629 }
 630 #endif /* WITH_ILP */
 631
 632 /*
 633  *  _____      _   ____  ____
 634  * | ____|_  _| |_| __ )| __ )
 635  * |  _| \ \/ / __|  _ \|  _ \
 636  * | |___ >  <| |_| |_) | |_) |
 637  * |_____/_/\_\\__|____/|____/
 638  *
 639  */
 640
 641 /** A simple forward single linked list. */
 642 typedef struct {
 643         ir_node  *start;   /**< start of the list */
 644         ir_node  *end;     /**< last block in the list */
 645         unsigned n_blks;  /**< number of blocks in the list */
 646 } anchor;
 647
 648 static void add_block(anchor *list, ir_node *block) {
 649         if (list->start == NULL) {
 650                 list->start = block;
 651                 list->end   = block;
 652         } else {
 653                 set_irn_link(list->end, block);
 654                 list->end = block;
 655         }
 656
 657         list->n_blks++;
 658 }
 659
 660 static void create_block_list(ir_node *leader_block, anchor *list) {
 661         int             i;
 662         const ir_edge_t *edge;
 663         ir_node         *block = NULL;
 664         ir_extblk       *extbb = get_Block_extbb(leader_block);
 665
 666         if (extbb_visited(extbb))
 667                 return;
 668         mark_extbb_visited(extbb);
 669
 670         for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
 671                 block = get_extbb_block(extbb, i);
 672                 add_block(list, block);
 673         }
 674
 675         assert(block != NULL);
 676
 677         /* pick successor extbbs */
 678         foreach_block_succ(block, edge) {
 679                 ir_node *succ = get_edge_src_irn(edge);
 680                 create_block_list(succ, list);
 681         }
 682
 683         for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
 684                 block = get_extbb_block(extbb, i);
 685
 686                 foreach_block_succ(block, edge) {
 687                         ir_node *succ = get_edge_src_irn(edge);
 688                         create_block_list(succ, list);
 689                 }
 690         }
 691 }
 692
 693 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
 694
 695 /*
 696  * Calculates a block schedule. The schedule is stored as a linked
 697  * list starting at the start_block of the irg.
 698  */
 699 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 700 {
 701         anchor list;
 702         ir_node **blk_list, *b, *n;
 703         unsigned i;
 704
 705         /* schedule extended basic blocks */
 706         compute_extbb_execfreqs(irg, execfreqs);
 707         //compute_extbb(irg);
 708
 709         list.start  = NULL;
 710         list.end    = NULL;
 711         list.n_blks = 0;
 712
 713         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 714         inc_irg_block_visited(irg);
 715
 716         create_block_list(get_irg_start_block(irg), &list);
 717
 718         /** create an array, so we can go forward and backward */
 719         blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
 720
 721         for (i = 0, b = list.start; b; b = n, ++i) {
 722                 n = get_irn_link(b);
 723                 blk_list[i] = b;
 724         }
 725
 726         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 727
 728         return blk_list;
 729 }
 730
 731 /*
 732  *  __  __       _
 733  * |  \/  | __ _(_)_ __
 734  * | |\/| |/ _` | | '_ \
 735  * | |  | | (_| | | | | |
 736  * |_|  |_|\__,_|_|_| |_|
 737  *
 738  */
 739 void be_init_blocksched(void)
 740 {
 741         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 742         lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
 743
 744         lc_opt_add_table(blocksched_grp, be_blocksched_options);
 745
 746         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 747 }
 748
 749 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
 750
 751 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 752 {
 753         switch(algo) {
 754         case BLOCKSCHED_GREEDY:
 755         case BLOCKSCHED_NAIV:
 756                 return create_block_schedule_greedy(irg, execfreqs);
 757         case BLOCKSCHED_EXTBB:
 758                 return create_extbb_block_schedule(irg, execfreqs);
 759 #ifdef WITH_ILP
 760         case BLOCKSCHED_ILP:
 761                 return create_block_schedule_ilp(irg, execfreqs);
 762 #endif /* WITH_ILP */
 763         }
 764
 765         panic("unknown blocksched algo");
 766         return NULL;
 767 }