nsz Git - libfirm/blob - ir/be/beblocksched.c

   1 /*
   2  * Copyright (C) 1995-2007 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 #ifdef HAVE_CONFIG_H
  38 #include "config.h"
  39 #endif
  40
  41 #include "beblocksched.h"
  42
  43 #include <stdlib.h>
  44
  45 #include "array.h"
  46 #include "pdeq.h"
  47
  48 #include "iredges.h"
  49 #include "irgwalk.h"
  50 #include "irnode_t.h"
  51 #include "irgraph_t.h"
  52 #include "irloop.h"
  53 #include "irprintf.h"
  54 #include "execfreq.h"
  55 #include "irdump_t.h"
  56 #include "irtools.h"
  57 #include "debug.h"
  58 #include "beirgmod.h"
  59 #include "bemodule.h"
  60 #include "be.h"
  61
  62 #include <libcore/lc_opts.h>
  63 #include <libcore/lc_opts_enum.h>
  64 #include <libcore/lc_timing.h>
  65
  66 #ifdef WITH_ILP
  67 #include <lpp/lpp.h>
  68 #include <lpp/lpp_net.h>
  69 #endif /* WITH_ILP */
  70
  71 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  72
  73 typedef enum _blocksched_algos_t {
  74         BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
  75 } blocksched_algos_t;
  76
  77 static int algo = BLOCKSCHED_GREEDY;
  78
  79 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
  80         { "naiv",       BLOCKSCHED_NAIV },
  81         { "extbb",      BLOCKSCHED_EXTBB },
  82         { "greedy", BLOCKSCHED_GREEDY },
  83 #ifdef WITH_ILP
  84         { "ilp",    BLOCKSCHED_ILP },
  85 #endif /* WITH_ILP */
  86         { NULL,     0 }
  87 };
  88
  89 static lc_opt_enum_int_var_t algo_var = {
  90         &algo, blockschedalgo_items
  91 };
  92
  93 static const lc_opt_table_entry_t be_blocksched_options[] = {
  94         LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var),
  95         LC_OPT_LAST
  96 };
  97
  98 /*
  99  *   ____                   _
 100  *  / ___|_ __ ___  ___  __| |_   _
 101  * | |  _| '__/ _ \/ _ \/ _` | | | |
 102  * | |_| | | |  __/  __/ (_| | |_| |
 103  *  \____|_|  \___|\___|\__,_|\__, |
 104  *                            |___/
 105  */
 106
 107 typedef struct _blocksched_entry_t {
 108         ir_node *block;
 109         struct _blocksched_entry_t *next;
 110         struct _blocksched_entry_t *prev;
 111 } blocksched_entry_t;
 112
 113 typedef struct _edge_t {
 114         ir_node *block;             /**< source block */
 115         int     pos;                /**< number of cfg predecessor (target) */
 116         double  execfreq;           /**< the frequency */
 117         int     highest_execfreq;   /**< flag that indicates whether this edge is the edge with the highest
 118                                                                      execfreq pointing away from this block */
 119 } edge_t;
 120
 121 typedef struct _blocksched_env_t {
 122         ir_graph       *irg;
 123         struct obstack *obst;
 124         ir_exec_freq   *execfreqs;
 125         edge_t         *edges;
 126         pdeq           *worklist;
 127         int            blockcount;
 128 } blocksched_env_t;
 129
 130 /**
 131  * Collect cfg frequencies of all edges between blocks.
 132  * Also determines edge with highest frequency.
 133  */
 134 static void collect_egde_frequency(ir_node *block, void *data)
 135 {
 136         blocksched_env_t   *env = data;
 137         int                arity;
 138         edge_t             edge;
 139         blocksched_entry_t *entry;
 140
 141         entry        = obstack_alloc(env->obst, sizeof(entry[0]));
 142         entry->block = block;
 143         entry->next  = NULL;
 144         entry->prev  = NULL;
 145         set_irn_link(block, entry);
 146
 147         arity = get_Block_n_cfgpreds(block);
 148
 149         if (arity == 0) {
 150                 assert(block == get_irg_start_block(env->irg));
 151                 /* must be the start block, nothing to 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(block))
 275                 return;
 276
 277         env->blockcount++;
 278         mark_irn_visited(block);
 279
 280         DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 281
 282         /* put all successors into the worklist */
 283         foreach_block_succ(block, edge) {
 284                 ir_node *succ_block = get_edge_src_irn(edge);
 285
 286                 if (irn_visited(succ_block))
 287                         continue;
 288
 289                 /* we only need to put the first of a series of already connected
 290                  * blocks into the worklist */
 291                 succ_entry = get_irn_link(succ_block);
 292                 while (succ_entry->prev != NULL) {
 293                         /* break cycles... */
 294                         if (succ_entry->prev->block == succ_block) {
 295                                 succ_entry->prev->next = NULL;
 296                                 succ_entry->prev       = NULL;
 297                                 break;
 298                         }
 299                         succ_entry = succ_entry->prev;
 300                 };
 301
 302                 if (irn_visited(succ_entry->block))
 303                         continue;
 304
 305                 DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 306                 pdeq_putr(env->worklist, succ_entry->block);
 307         }
 308
 309         if (entry->next != NULL) {
 310                 pick_block_successor(entry->next, env);
 311                 return;
 312         }
 313
 314         DBG((dbg, LEVEL_1, "deciding...\n"));
 315         best_succ_execfreq = -1;
 316
 317         /* no successor yet: pick the successor block with the highest execution
 318          * frequency which has no predecessor yet */
 319
 320         foreach_block_succ(block, edge) {
 321                 ir_node *succ_block = get_edge_src_irn(edge);
 322                 double  execfreq;
 323
 324                 if (irn_visited(succ_block))
 325                         continue;
 326
 327                 succ_entry = get_irn_link(succ_block);
 328                 if (succ_entry->prev != NULL)
 329                         continue;
 330
 331                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 332                 if (execfreq > best_succ_execfreq) {
 333                         best_succ_execfreq = execfreq;
 334                         succ = succ_block;
 335                 }
 336         }
 337
 338         if (succ == NULL) {
 339                 DBG((dbg, LEVEL_1, "pick from worklist\n"));
 340
 341                 do {
 342                         if (pdeq_empty(env->worklist)) {
 343                                 DBG((dbg, LEVEL_1, "worklist empty\n"));
 344                                 return;
 345                         }
 346                         succ = pdeq_getl(env->worklist);
 347                 } while (irn_visited(succ));
 348         }
 349
 350         succ_entry       = get_irn_link(succ);
 351         entry->next      = succ_entry;
 352         succ_entry->prev = entry;
 353
 354         pick_block_successor(succ_entry, env);
 355 }
 356
 357 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 358 {
 359         ir_graph           *irg        = env->irg;
 360         ir_node            *startblock = get_irg_start_block(irg);
 361         blocksched_entry_t *entry      = get_irn_link(startblock);
 362
 363         set_using_visited(irg);
 364         inc_irg_visited(irg);
 365
 366         env->worklist = new_pdeq();
 367         pick_block_successor(entry, env);
 368         assert(pdeq_empty(env->worklist));
 369         del_pdeq(env->worklist);
 370
 371         clear_using_visited(irg);
 372
 373         return entry;
 374 }
 375
 376 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 377                                                                                 int count, struct obstack* obst)
 378 {
 379         int                i = 0;
 380         ir_node            **block_list;
 381         blocksched_entry_t *entry;
 382         (void) env;
 383
 384         block_list = NEW_ARR_D(ir_node *, obst, count);
 385         DBG((dbg, LEVEL_1, "Blockschedule:\n"));
 386
 387         for (entry = first; entry != NULL; entry = entry->next) {
 388                 assert(i < count);
 389                 block_list[i++] = entry->block;
 390                 DBG((dbg, LEVEL_1, "\t%+F\n", entry->block));
 391         }
 392         assert(i == count);
 393
 394         return block_list;
 395 }
 396
 397 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 398 {
 399         blocksched_env_t   env;
 400         struct obstack     obst;
 401         blocksched_entry_t *start_entry;
 402         ir_node            **block_list;
 403
 404         obstack_init(&obst);
 405
 406         env.irg        = irg;
 407         env.obst       = &obst;
 408         env.execfreqs  = execfreqs;
 409         env.edges      = NEW_ARR_F(edge_t, 0);
 410         env.worklist   = NULL;
 411         env.blockcount = 0;
 412
 413         // collect edge execution frequencies
 414         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 415
 416         // sort interblock edges by execution frequency
 417         qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
 418
 419         (void)be_remove_empty_blocks(irg);
 420
 421         if (algo != BLOCKSCHED_NAIV)
 422                 coalesce_blocks(&env);
 423
 424         start_entry = finish_block_schedule(&env);
 425         block_list  = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
 426
 427         DEL_ARR_F(env.edges);
 428         obstack_free(&obst, NULL);
 429
 430         return block_list;
 431 }
 432
 433 /*
 434  *  ___ _     ____
 435  * |_ _| |   |  _ \
 436  *  | || |   | |_) |
 437  *  | || |___|  __/
 438  * |___|_____|_|
 439  *
 440  */
 441
 442 #ifdef WITH_ILP
 443 typedef struct _ilp_edge_t {
 444         ir_node *block;   /**< source block */
 445         int     pos;      /**< number of cfg predecessor (target) */
 446         int     ilpvar;
 447 } ilp_edge_t;
 448
 449 typedef struct _blocksched_ilp_env_t {
 450         blocksched_env_t env;
 451         ilp_edge_t       *ilpedges;
 452         lpp_t            *lpp;
 453 } blocksched_ilp_env_t;
 454
 455 typedef struct _blocksched_ilp_entry_t {
 456         ir_node *block;
 457         struct _blocksched_entry_t *next;
 458         struct _blocksched_entry_t *prev;
 459
 460         int out_cst;
 461 } blocksched_ilp_entry_t;
 462
 463 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 464 {
 465         char       name[64];
 466         ilp_edge_t edge;
 467         int        edgeidx = ARR_LEN(env->ilpedges);
 468
 469         snprintf(name, sizeof(name), "edge%d", edgeidx);
 470
 471         edge.block  = block;
 472         edge.pos    = pos;
 473         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 474
 475         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 476         return edgeidx;
 477 }
 478
 479 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 480 {
 481         blocksched_ilp_env_t *env        = data;
 482         ir_graph             *irg        = env->env.irg;
 483         ir_node              *startblock = get_irg_start_block(irg);
 484         int                  arity;
 485         lpp_cst_t            cst;
 486         char                 name[64];
 487         int                  out_count;
 488         blocksched_ilp_entry_t *entry;
 489
 490         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 491         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 492
 493         entry          = obstack_alloc(env->env.obst, sizeof(entry[0]));
 494         entry->block   = block;
 495         entry->next    = NULL;
 496         entry->prev    = NULL;
 497         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
 498         set_irn_link(block, entry);
 499
 500         if (block == startblock)
 501                 return;
 502
 503         arity = get_irn_arity(block);
 504         if (arity == 1) {
 505                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 506                 add_ilp_edge(block, 0, execfreq, env);
 507         }
 508         else {
 509                 int i;
 510                 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
 511
 512                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 513                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
 514
 515                 for (i = 0; i < arity; ++i) {
 516                         double     execfreq;
 517                         int        edgenum;
 518                         ilp_edge_t *edge;
 519                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 520
 521                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 522                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 523                         edge     = &env->ilpedges[edgenum];
 524                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 525                 }
 526         }
 527 }
 528
 529
 530 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 531 {
 532         int  i;
 533         int  edge_count = ARR_LEN(env->ilpedges);
 534
 535         /* complete out constraints */
 536         for(i = 0; i < edge_count; ++i) {
 537                 const ilp_edge_t *edge  = &env->ilpedges[i];
 538                 ir_node          *block = edge->block;
 539                 ir_node          *pred;
 540                 blocksched_ilp_entry_t *entry;
 541
 542                 /* the block might have been removed already... */
 543                 if (is_Bad(get_Block_cfgpred(block, 0)))
 544                         continue;
 545
 546                 pred  = get_Block_cfgpred_block(block, edge->pos);
 547                 entry = get_irn_link(pred);
 548
 549                 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 550                                   pred, block, edge->pos));
 551                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 552         }
 553
 554 #if 0
 555         {
 556                 FILE *f;
 557                 char fname[256];
 558                 lpp_dump(env->lpp, "lpp.out");
 559                 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
 560                 f = fopen(fname, "w");
 561                 lpp_dump_plain(env->lpp, f);
 562                 fclose(f);
 563         }
 564 #endif
 565
 566         //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
 567         lpp_solve_net(env->lpp, "i44pc52", "cplex");
 568         assert(lpp_is_sol_valid(env->lpp));
 569
 570         /* Apply results to edges */
 571         for (i = 0; i < edge_count; ++i) {
 572                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 573                 ir_node            *block = edge->block;
 574                 ir_node            *pred;
 575                 int                is_jump;
 576                 blocksched_entry_t *entry;
 577                 blocksched_entry_t *pred_entry;
 578
 579                 /* the block might have been removed already... */
 580                 if (is_Bad(get_Block_cfgpred(block, 0)))
 581                         continue;
 582
 583                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 584                 if (is_jump)
 585                         continue;
 586
 587                 pred       = get_Block_cfgpred_block(block, edge->pos);
 588                 entry      = get_irn_link(block);
 589                 pred_entry = get_irn_link(pred);
 590
 591                 assert(entry->prev == NULL && pred_entry->next == NULL);
 592                 entry->prev      = pred_entry;
 593                 pred_entry->next = entry;
 594         }
 595 }
 596
 597 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 598 {
 599         blocksched_ilp_env_t env;
 600         struct obstack       obst;
 601         blocksched_entry_t   *start_entry;
 602         ir_node              **block_list;
 603
 604         obstack_init(&obst);
 605
 606         env.env.irg        = irg;
 607         env.env.obst       = &obst;
 608         env.env.execfreqs  = execfreqs;
 609         env.env.worklist   = NULL;
 610         env.env.blockcount = 0;
 611         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 612
 613         env.lpp = new_lpp("blockschedule", lpp_minimize);
 614         lpp_set_time_limit(env.lpp, 20);
 615         lpp_set_log(env.lpp, stdout);
 616
 617         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 618
 619         (void)be_remove_empty_blocks(irg);
 620         coalesce_blocks_ilp(&env);
 621
 622         start_entry = finish_block_schedule(&env.env);
 623         block_list  = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
 624
 625         DEL_ARR_F(env.ilpedges);
 626         free_lpp(env.lpp);
 627         obstack_free(&obst, NULL);
 628
 629         return block_list;
 630 }
 631 #endif /* WITH_ILP */
 632
 633 /*
 634  *  _____      _   ____  ____
 635  * | ____|_  _| |_| __ )| __ )
 636  * |  _| \ \/ / __|  _ \|  _ \
 637  * | |___ >  <| |_| |_) | |_) |
 638  * |_____/_/\_\\__|____/|____/
 639  *
 640  */
 641
 642 /** A simple forward single linked list. */
 643 typedef struct {
 644         ir_node  *start;   /**< start of the list */
 645         ir_node  *end;     /**< last block in the list */
 646         unsigned n_blks;  /**< number of blocks in the list */
 647 } anchor;
 648
 649 static void add_block(anchor *list, ir_node *block) {
 650         if (list->start == NULL) {
 651                 list->start = block;
 652                 list->end   = block;
 653         } else {
 654                 set_irn_link(list->end, block);
 655                 list->end = block;
 656         }
 657
 658         list->n_blks++;
 659 }
 660
 661 static void create_block_list(ir_node *leader_block, anchor *list) {
 662         int             i;
 663         const ir_edge_t *edge;
 664         ir_node         *block = NULL;
 665         ir_extblk       *extbb = get_Block_extbb(leader_block);
 666
 667         if (extbb_visited(extbb))
 668                 return;
 669         mark_extbb_visited(extbb);
 670
 671         for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
 672                 block = get_extbb_block(extbb, i);
 673                 add_block(list, block);
 674         }
 675
 676         assert(block != NULL);
 677
 678         /* pick successor extbbs */
 679         foreach_block_succ(block, edge) {
 680                 ir_node *succ = get_edge_src_irn(edge);
 681                 create_block_list(succ, list);
 682         }
 683
 684         for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
 685                 block = get_extbb_block(extbb, i);
 686
 687                 foreach_block_succ(block, edge) {
 688                         ir_node *succ = get_edge_src_irn(edge);
 689                         create_block_list(succ, list);
 690                 }
 691         }
 692 }
 693
 694 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
 695
 696 /*
 697  * Calculates a block schedule. The schedule is stored as a linked
 698  * list starting at the start_block of the irg.
 699  */
 700 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 701 {
 702         anchor list;
 703         ir_node **blk_list, *b, *n;
 704         unsigned i;
 705
 706         /* schedule extended basic blocks */
 707         compute_extbb_execfreqs(irg, execfreqs);
 708         //compute_extbb(irg);
 709
 710         list.start  = NULL;
 711         list.end    = NULL;
 712         list.n_blks = 0;
 713
 714         set_using_irn_link(irg);
 715         set_using_visited(irg);
 716         inc_irg_block_visited(irg);
 717
 718         create_block_list(get_irg_start_block(irg), &list);
 719
 720         /** create an array, so we can go forward and backward */
 721         blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
 722
 723         for (i = 0, b = list.start; b; b = n, ++i) {
 724                 n = get_irn_link(b);
 725                 blk_list[i] = b;
 726         }
 727
 728         clear_using_irn_link(irg);
 729         clear_using_visited(irg);
 730
 731         return blk_list;
 732 }
 733
 734 /*
 735  *  __  __       _
 736  * |  \/  | __ _(_)_ __
 737  * | |\/| |/ _` | | '_ \
 738  * | |  | | (_| | | | | |
 739  * |_|  |_|\__,_|_|_| |_|
 740  *
 741  */
 742 void be_init_blocksched(void)
 743 {
 744         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 745         lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
 746
 747         lc_opt_add_table(blocksched_grp, be_blocksched_options);
 748
 749         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 750 }
 751
 752 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
 753
 754 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 755 {
 756         switch(algo) {
 757         case BLOCKSCHED_GREEDY:
 758         case BLOCKSCHED_NAIV:
 759                 return create_block_schedule_greedy(irg, execfreqs);
 760         case BLOCKSCHED_EXTBB:
 761                 return create_extbb_block_schedule(irg, execfreqs);
 762 #ifdef WITH_ILP
 763         case BLOCKSCHED_ILP:
 764                 return create_block_schedule_ilp(irg, execfreqs);
 765 #endif /* WITH_ILP */
 766         }
 767
 768         assert(0 && "unknown blocksched algo");
 769         return NULL;
 770 }