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 #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                                 || block == get_irg_end_block(env->irg));
 152                 /* must be the start block (or end-block for endless loops), nothing to
 153                  * do here */
 154                 return;
 155         } else if (arity == 1) {
 156                 edge.block            = block;
 157                 edge.pos              = 0;
 158                 edge.execfreq         = get_block_execfreq(env->execfreqs, block);
 159                 edge.highest_execfreq = 1;
 160                 ARR_APP1(edge_t, env->edges, edge);
 161         } else {
 162                 int    i;
 163                 double highest_execfreq = -1.0;
 164                 int    highest_edge_num = -1;
 165
 166                 edge.block = block;
 167                 for (i = 0; i < arity; ++i) {
 168                         double  execfreq;
 169                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 170
 171                         execfreq = get_block_execfreq(env->execfreqs, pred_block);
 172
 173                         edge.pos              = i;
 174                         edge.execfreq         = execfreq;
 175                         edge.highest_execfreq = 0;
 176                         ARR_APP1(edge_t, env->edges, edge);
 177
 178                         if (execfreq > highest_execfreq) {
 179                                 highest_execfreq = execfreq;
 180                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 181                         }
 182                 }
 183
 184                 if(highest_edge_num >= 0)
 185                         env->edges[highest_edge_num].highest_execfreq = 1;
 186         }
 187 }
 188
 189 static int cmp_edges(const void *d1, const void *d2)
 190 {
 191         const edge_t *e1 = d1;
 192         const edge_t *e2 = d2;
 193
 194         return QSORT_CMP(e2->execfreq, e1->execfreq);
 195 }
 196
 197 static void coalesce_blocks(blocksched_env_t *env)
 198 {
 199         int i;
 200         int edge_count = ARR_LEN(env->edges);
 201
 202         /* run1: only look at jumps */
 203         for (i = 0; i < edge_count; ++i) {
 204                 const edge_t *edge  = &env->edges[i];
 205                 ir_node      *block = edge->block;
 206                 int           pos   = edge->pos;
 207                 ir_node      *pred_block;
 208                 blocksched_entry_t *entry, *pred_entry;
 209
 210                 /* only check edge with highest frequency */
 211                 if (! edge->highest_execfreq)
 212                         continue;
 213
 214                 /* the block might have been removed already... */
 215                 if (is_Bad(get_Block_cfgpred(block, 0)))
 216                         continue;
 217
 218                 pred_block = get_Block_cfgpred_block(block, pos);
 219                 entry      = get_irn_link(block);
 220                 pred_entry = get_irn_link(pred_block);
 221
 222                 if (pred_entry->next != NULL || entry->prev != NULL)
 223                         continue;
 224
 225                 /* only coalesce jumps */
 226                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 227                         continue;
 228
 229                 /* schedule the 2 blocks behind each other */
 230                 DBG((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 231                            pred_entry->block, entry->block, edge->execfreq));
 232                 pred_entry->next = entry;
 233                 entry->prev      = pred_entry;
 234         }
 235
 236         /* run2: remaining edges */
 237         for (i = 0; i < edge_count; ++i) {
 238                 const edge_t *edge  = &env->edges[i];
 239                 ir_node      *block = edge->block;
 240                 int           pos   = edge->pos;
 241                 ir_node      *pred_block;
 242                 blocksched_entry_t *entry, *pred_entry;
 243
 244                 /* the block might have been removed already... */
 245                 if (is_Bad(get_Block_cfgpred(block, 0)))
 246                         continue;
 247
 248                 /* we can't do fallthroughs in backedges */
 249                 if (is_backedge(block, pos))
 250                         continue;
 251
 252                 pred_block = get_Block_cfgpred_block(block, pos);
 253                 entry      = get_irn_link(block);
 254                 pred_entry = get_irn_link(pred_block);
 255
 256                 /* is 1 of the blocks already attached to another block? */
 257                 if (pred_entry->next != NULL || entry->prev != NULL)
 258                         continue;
 259
 260                 /* schedule the 2 blocks behind each other */
 261                 DBG((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 262                            pred_entry->block, entry->block, edge->execfreq));
 263                 pred_entry->next = entry;
 264                 entry->prev      = pred_entry;
 265         }
 266 }
 267
 268 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 269 {
 270         ir_node            *block = entry->block;
 271         ir_node            *succ  = NULL;
 272         blocksched_entry_t *succ_entry;
 273         const ir_edge_t    *edge;
 274         double             best_succ_execfreq;
 275
 276         if (irn_visited(block))
 277                 return;
 278
 279         env->blockcount++;
 280         mark_irn_visited(block);
 281
 282         DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 283
 284         /* put all successors into the worklist */
 285         foreach_block_succ(block, edge) {
 286                 ir_node *succ_block = get_edge_src_irn(edge);
 287
 288                 if (irn_visited(succ_block))
 289                         continue;
 290
 291                 /* we only need to put the first of a series of already connected
 292                  * blocks into the worklist */
 293                 succ_entry = get_irn_link(succ_block);
 294                 while (succ_entry->prev != NULL) {
 295                         /* break cycles... */
 296                         if (succ_entry->prev->block == succ_block) {
 297                                 succ_entry->prev->next = NULL;
 298                                 succ_entry->prev       = NULL;
 299                                 break;
 300                         }
 301                         succ_entry = succ_entry->prev;
 302                 };
 303
 304                 if (irn_visited(succ_entry->block))
 305                         continue;
 306
 307                 DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 308                 pdeq_putr(env->worklist, succ_entry->block);
 309         }
 310
 311         if (entry->next != NULL) {
 312                 pick_block_successor(entry->next, env);
 313                 return;
 314         }
 315
 316         DBG((dbg, LEVEL_1, "deciding...\n"));
 317         best_succ_execfreq = -1;
 318
 319         /* no successor yet: pick the successor block with the highest execution
 320          * frequency which has no predecessor yet */
 321
 322         foreach_block_succ(block, edge) {
 323                 ir_node *succ_block = get_edge_src_irn(edge);
 324                 double  execfreq;
 325
 326                 if (irn_visited(succ_block))
 327                         continue;
 328
 329                 succ_entry = get_irn_link(succ_block);
 330                 if (succ_entry->prev != NULL)
 331                         continue;
 332
 333                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 334                 if (execfreq > best_succ_execfreq) {
 335                         best_succ_execfreq = execfreq;
 336                         succ = succ_block;
 337                 }
 338         }
 339
 340         if (succ == NULL) {
 341                 DBG((dbg, LEVEL_1, "pick from worklist\n"));
 342
 343                 do {
 344                         if (pdeq_empty(env->worklist)) {
 345                                 DBG((dbg, LEVEL_1, "worklist empty\n"));
 346                                 return;
 347                         }
 348                         succ = pdeq_getl(env->worklist);
 349                 } while (irn_visited(succ));
 350         }
 351
 352         succ_entry       = get_irn_link(succ);
 353         entry->next      = succ_entry;
 354         succ_entry->prev = entry;
 355
 356         pick_block_successor(succ_entry, env);
 357 }
 358
 359 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 360 {
 361         ir_graph           *irg        = env->irg;
 362         ir_node            *startblock = get_irg_start_block(irg);
 363         blocksched_entry_t *entry      = get_irn_link(startblock);
 364
 365         set_using_irn_visited(irg);
 366         inc_irg_visited(irg);
 367
 368         env->worklist = new_pdeq();
 369         pick_block_successor(entry, env);
 370         assert(pdeq_empty(env->worklist));
 371         del_pdeq(env->worklist);
 372
 373         clear_using_irn_visited(irg);
 374
 375         return entry;
 376 }
 377
 378 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 379                                                                                 int count, struct obstack* obst)
 380 {
 381         int                i = 0;
 382         ir_node            **block_list;
 383         blocksched_entry_t *entry;
 384         (void) env;
 385
 386         block_list = NEW_ARR_D(ir_node *, obst, count);
 387         DBG((dbg, LEVEL_1, "Blockschedule:\n"));
 388
 389         for (entry = first; entry != NULL; entry = entry->next) {
 390                 assert(i < count);
 391                 block_list[i++] = entry->block;
 392                 DBG((dbg, LEVEL_1, "\t%+F\n", entry->block));
 393         }
 394         assert(i == count);
 395
 396         return block_list;
 397 }
 398
 399 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 400 {
 401         blocksched_env_t   env;
 402         struct obstack     obst;
 403         blocksched_entry_t *start_entry;
 404         ir_node            **block_list;
 405
 406         obstack_init(&obst);
 407
 408         env.irg        = irg;
 409         env.obst       = &obst;
 410         env.execfreqs  = execfreqs;
 411         env.edges      = NEW_ARR_F(edge_t, 0);
 412         env.worklist   = NULL;
 413         env.blockcount = 0;
 414
 415         // collect edge execution frequencies
 416         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 417
 418         // sort interblock edges by execution frequency
 419         qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges);
 420
 421         (void)be_remove_empty_blocks(irg);
 422
 423         if (algo != BLOCKSCHED_NAIV)
 424                 coalesce_blocks(&env);
 425
 426         start_entry = finish_block_schedule(&env);
 427         block_list  = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg));
 428
 429         DEL_ARR_F(env.edges);
 430         obstack_free(&obst, NULL);
 431
 432         return block_list;
 433 }
 434
 435 /*
 436  *  ___ _     ____
 437  * |_ _| |   |  _ \
 438  *  | || |   | |_) |
 439  *  | || |___|  __/
 440  * |___|_____|_|
 441  *
 442  */
 443
 444 #ifdef WITH_ILP
 445 typedef struct _ilp_edge_t {
 446         ir_node *block;   /**< source block */
 447         int     pos;      /**< number of cfg predecessor (target) */
 448         int     ilpvar;
 449 } ilp_edge_t;
 450
 451 typedef struct _blocksched_ilp_env_t {
 452         blocksched_env_t env;
 453         ilp_edge_t       *ilpedges;
 454         lpp_t            *lpp;
 455 } blocksched_ilp_env_t;
 456
 457 typedef struct _blocksched_ilp_entry_t {
 458         ir_node *block;
 459         struct _blocksched_entry_t *next;
 460         struct _blocksched_entry_t *prev;
 461
 462         int out_cst;
 463 } blocksched_ilp_entry_t;
 464
 465 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 466 {
 467         char       name[64];
 468         ilp_edge_t edge;
 469         int        edgeidx = ARR_LEN(env->ilpedges);
 470
 471         snprintf(name, sizeof(name), "edge%d", edgeidx);
 472
 473         edge.block  = block;
 474         edge.pos    = pos;
 475         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 476
 477         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 478         return edgeidx;
 479 }
 480
 481 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 482 {
 483         blocksched_ilp_env_t *env        = data;
 484         ir_graph             *irg        = env->env.irg;
 485         ir_node              *startblock = get_irg_start_block(irg);
 486         int                  arity;
 487         lpp_cst_t            cst;
 488         char                 name[64];
 489         int                  out_count;
 490         blocksched_ilp_entry_t *entry;
 491
 492         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 493         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 494
 495         entry          = obstack_alloc(env->env.obst, sizeof(entry[0]));
 496         entry->block   = block;
 497         entry->next    = NULL;
 498         entry->prev    = NULL;
 499         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
 500         set_irn_link(block, entry);
 501
 502         if (block == startblock)
 503                 return;
 504
 505         arity = get_irn_arity(block);
 506         if (arity == 1) {
 507                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 508                 add_ilp_edge(block, 0, execfreq, env);
 509         }
 510         else {
 511                 int i;
 512                 int *edgenums = alloca(sizeof(edgenums[0]) * arity);
 513
 514                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 515                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
 516
 517                 for (i = 0; i < arity; ++i) {
 518                         double     execfreq;
 519                         int        edgenum;
 520                         ilp_edge_t *edge;
 521                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 522
 523                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 524                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 525                         edge     = &env->ilpedges[edgenum];
 526                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 527                 }
 528         }
 529 }
 530
 531
 532 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 533 {
 534         int  i;
 535         int  edge_count = ARR_LEN(env->ilpedges);
 536
 537         /* complete out constraints */
 538         for(i = 0; i < edge_count; ++i) {
 539                 const ilp_edge_t *edge  = &env->ilpedges[i];
 540                 ir_node          *block = edge->block;
 541                 ir_node          *pred;
 542                 blocksched_ilp_entry_t *entry;
 543
 544                 /* the block might have been removed already... */
 545                 if (is_Bad(get_Block_cfgpred(block, 0)))
 546                         continue;
 547
 548                 pred  = get_Block_cfgpred_block(block, edge->pos);
 549                 entry = get_irn_link(pred);
 550
 551                 DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 552                                   pred, block, edge->pos));
 553                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 554         }
 555
 556 #if 0
 557         {
 558                 FILE *f;
 559                 char fname[256];
 560                 lpp_dump(env->lpp, "lpp.out");
 561                 snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg));
 562                 f = fopen(fname, "w");
 563                 lpp_dump_plain(env->lpp, f);
 564                 fclose(f);
 565         }
 566 #endif
 567
 568         //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver);
 569         lpp_solve_net(env->lpp, "i44pc52", "cplex");
 570         assert(lpp_is_sol_valid(env->lpp));
 571
 572         /* Apply results to edges */
 573         for (i = 0; i < edge_count; ++i) {
 574                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 575                 ir_node            *block = edge->block;
 576                 ir_node            *pred;
 577                 int                is_jump;
 578                 blocksched_entry_t *entry;
 579                 blocksched_entry_t *pred_entry;
 580
 581                 /* the block might have been removed already... */
 582                 if (is_Bad(get_Block_cfgpred(block, 0)))
 583                         continue;
 584
 585                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 586                 if (is_jump)
 587                         continue;
 588
 589                 pred       = get_Block_cfgpred_block(block, edge->pos);
 590                 entry      = get_irn_link(block);
 591                 pred_entry = get_irn_link(pred);
 592
 593                 assert(entry->prev == NULL && pred_entry->next == NULL);
 594                 entry->prev      = pred_entry;
 595                 pred_entry->next = entry;
 596         }
 597 }
 598
 599 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 600 {
 601         blocksched_ilp_env_t env;
 602         struct obstack       obst;
 603         blocksched_entry_t   *start_entry;
 604         ir_node              **block_list;
 605
 606         obstack_init(&obst);
 607
 608         env.env.irg        = irg;
 609         env.env.obst       = &obst;
 610         env.env.execfreqs  = execfreqs;
 611         env.env.worklist   = NULL;
 612         env.env.blockcount = 0;
 613         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 614
 615         env.lpp = new_lpp("blockschedule", lpp_minimize);
 616         lpp_set_time_limit(env.lpp, 20);
 617         lpp_set_log(env.lpp, stdout);
 618
 619         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 620
 621         (void)be_remove_empty_blocks(irg);
 622         coalesce_blocks_ilp(&env);
 623
 624         start_entry = finish_block_schedule(&env.env);
 625         block_list  = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg));
 626
 627         DEL_ARR_F(env.ilpedges);
 628         free_lpp(env.lpp);
 629         obstack_free(&obst, NULL);
 630
 631         return block_list;
 632 }
 633 #endif /* WITH_ILP */
 634
 635 /*
 636  *  _____      _   ____  ____
 637  * | ____|_  _| |_| __ )| __ )
 638  * |  _| \ \/ / __|  _ \|  _ \
 639  * | |___ >  <| |_| |_) | |_) |
 640  * |_____/_/\_\\__|____/|____/
 641  *
 642  */
 643
 644 /** A simple forward single linked list. */
 645 typedef struct {
 646         ir_node  *start;   /**< start of the list */
 647         ir_node  *end;     /**< last block in the list */
 648         unsigned n_blks;  /**< number of blocks in the list */
 649 } anchor;
 650
 651 static void add_block(anchor *list, ir_node *block) {
 652         if (list->start == NULL) {
 653                 list->start = block;
 654                 list->end   = block;
 655         } else {
 656                 set_irn_link(list->end, block);
 657                 list->end = block;
 658         }
 659
 660         list->n_blks++;
 661 }
 662
 663 static void create_block_list(ir_node *leader_block, anchor *list) {
 664         int             i;
 665         const ir_edge_t *edge;
 666         ir_node         *block = NULL;
 667         ir_extblk       *extbb = get_Block_extbb(leader_block);
 668
 669         if (extbb_visited(extbb))
 670                 return;
 671         mark_extbb_visited(extbb);
 672
 673         for (i = 0; i < get_extbb_n_blocks(extbb); ++i) {
 674                 block = get_extbb_block(extbb, i);
 675                 add_block(list, block);
 676         }
 677
 678         assert(block != NULL);
 679
 680         /* pick successor extbbs */
 681         foreach_block_succ(block, edge) {
 682                 ir_node *succ = get_edge_src_irn(edge);
 683                 create_block_list(succ, list);
 684         }
 685
 686         for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) {
 687                 block = get_extbb_block(extbb, i);
 688
 689                 foreach_block_succ(block, edge) {
 690                         ir_node *succ = get_edge_src_irn(edge);
 691                         create_block_list(succ, list);
 692                 }
 693         }
 694 }
 695
 696 void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs);
 697
 698 /*
 699  * Calculates a block schedule. The schedule is stored as a linked
 700  * list starting at the start_block of the irg.
 701  */
 702 static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 703 {
 704         anchor list;
 705         ir_node **blk_list, *b, *n;
 706         unsigned i;
 707
 708         /* schedule extended basic blocks */
 709         compute_extbb_execfreqs(irg, execfreqs);
 710         //compute_extbb(irg);
 711
 712         list.start  = NULL;
 713         list.end    = NULL;
 714         list.n_blks = 0;
 715
 716         set_using_irn_link(irg);
 717         set_using_irn_visited(irg);
 718         inc_irg_block_visited(irg);
 719
 720         create_block_list(get_irg_start_block(irg), &list);
 721
 722         /** create an array, so we can go forward and backward */
 723         blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks);
 724
 725         for (i = 0, b = list.start; b; b = n, ++i) {
 726                 n = get_irn_link(b);
 727                 blk_list[i] = b;
 728         }
 729
 730         clear_using_irn_link(irg);
 731         clear_using_irn_visited(irg);
 732
 733         return blk_list;
 734 }
 735
 736 /*
 737  *  __  __       _
 738  * |  \/  | __ _(_)_ __
 739  * | |\/| |/ _` | | '_ \
 740  * | |  | | (_| | | | | |
 741  * |_|  |_|\__,_|_|_| |_|
 742  *
 743  */
 744 void be_init_blocksched(void)
 745 {
 746         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 747         lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched");
 748
 749         lc_opt_add_table(blocksched_grp, be_blocksched_options);
 750
 751         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 752 }
 753
 754 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched);
 755
 756 ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs)
 757 {
 758         switch(algo) {
 759         case BLOCKSCHED_GREEDY:
 760         case BLOCKSCHED_NAIV:
 761                 return create_block_schedule_greedy(irg, execfreqs);
 762         case BLOCKSCHED_EXTBB:
 763                 return create_extbb_block_schedule(irg, execfreqs);
 764 #ifdef WITH_ILP
 765         case BLOCKSCHED_ILP:
 766                 return create_block_schedule_ilp(irg, execfreqs);
 767 #endif /* WITH_ILP */
 768         }
 769
 770         assert(0 && "unknown blocksched algo");
 771         return NULL;
 772 }