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 "util.h"
  56 #include "debug.h"
  57 #include "beirgmod.h"
  58 #include "bemodule.h"
  59 #include "be.h"
  60 #include "error.h"
  61
  62 #include "lc_opts.h"
  63 #include "lc_opts_enum.h"
  64
  65 #include "lpp.h"
  66 #include "lpp_net.h"
  67
  68 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  69
  70 typedef enum blocksched_algos_t {
  71         BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
  72 } blocksched_algos_t;
  73
  74 static int algo = BLOCKSCHED_GREEDY;
  75
  76 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
  77         { "naiv",   BLOCKSCHED_NAIV },
  78         { "greedy", BLOCKSCHED_GREEDY },
  79         { "ilp",    BLOCKSCHED_ILP },
  80         { NULL,     0 }
  81 };
  82
  83 static lc_opt_enum_int_var_t algo_var = {
  84         &algo, blockschedalgo_items
  85 };
  86
  87 static const lc_opt_table_entry_t be_blocksched_options[] = {
  88         LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
  89         LC_OPT_LAST
  90 };
  91
  92 /*
  93  *   ____                   _
  94  *  / ___|_ __ ___  ___  __| |_   _
  95  * | |  _| '__/ _ \/ _ \/ _` | | | |
  96  * | |_| | | |  __/  __/ (_| | |_| |
  97  *  \____|_|  \___|\___|\__,_|\__, |
  98  *                            |___/
  99  */
 100
 101 typedef struct blocksched_entry_t blocksched_entry_t;
 102 struct blocksched_entry_t {
 103         ir_node            *block;
 104         blocksched_entry_t *next;
 105         blocksched_entry_t *prev;
 106 };
 107
 108 typedef struct edge_t edge_t;
 109 struct edge_t {
 110         ir_node *block;             /**< source block */
 111         int     pos;                /**< number of cfg predecessor (target) */
 112         double  execfreq;           /**< the frequency */
 113         double  outedge_penalty_freq; /**< for edges leaving the loop this is the
 114                                            penality when we make them a
 115                                            fallthrough. */
 116         int     highest_execfreq;   /**< flag that indicates whether this edge is
 117                                          the edge with the highest execfreq pointing
 118                                          away from this block */
 119 };
 120
 121 typedef struct blocksched_env_t blocksched_env_t;
 122 struct blocksched_env_t {
 123         ir_graph       *irg;
 124         struct obstack *obst;
 125         ir_exec_freq   *execfreqs;
 126         edge_t         *edges;
 127         pdeq           *worklist;
 128         int            blockcount;
 129 };
 130
 131 /**
 132  * Collect cfg frequencies of all edges between blocks.
 133  * Also determines edge with highest frequency.
 134  */
 135 static void collect_egde_frequency(ir_node *block, void *data)
 136 {
 137         blocksched_env_t   *env = (blocksched_env_t*)data;
 138         int                arity;
 139         edge_t             edge;
 140         blocksched_entry_t *entry;
 141         ir_loop            *loop;
 142
 143         memset(&edge, 0, sizeof(edge));
 144
 145         entry = OALLOCZ(env->obst, blocksched_entry_t);
 146         entry->block = block;
 147         set_irn_link(block, entry);
 148
 149         loop = get_irn_loop(block);
 150
 151         arity = get_Block_n_cfgpreds(block);
 152
 153         if (arity == 0) {
 154                 /* must be the start block (or end-block for endless loops),
 155                  * everything else is dead code and should be removed by now */
 156                 assert(block == get_irg_start_block(env->irg)
 157                                 || block == get_irg_end_block(env->irg));
 158                 /* nothing to do here */
 159                 return;
 160         } else if (arity == 1) {
 161                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 162                 ir_loop *pred_loop  = get_irn_loop(pred_block);
 163                 float    freq       = (float)get_block_execfreq(env->execfreqs, block);
 164
 165                 /* is it an edge leaving a loop */
 166                 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
 167                         float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
 168                         edge.outedge_penalty_freq = -(pred_freq - freq);
 169                 }
 170
 171                 edge.block            = block;
 172                 edge.pos              = 0;
 173                 edge.execfreq         = freq;
 174                 edge.highest_execfreq = 1;
 175                 ARR_APP1(edge_t, env->edges, edge);
 176         } else {
 177                 int    i;
 178                 double highest_execfreq = -1.0;
 179                 int    highest_edge_num = -1;
 180
 181                 edge.block = block;
 182                 for (i = 0; i < arity; ++i) {
 183                         double  execfreq;
 184                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 185
 186                         execfreq = get_block_execfreq(env->execfreqs, pred_block);
 187
 188                         edge.pos              = i;
 189                         edge.execfreq         = execfreq;
 190                         edge.highest_execfreq = 0;
 191                         ARR_APP1(edge_t, env->edges, edge);
 192
 193                         if (execfreq > highest_execfreq) {
 194                                 highest_execfreq = execfreq;
 195                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 196                         }
 197                 }
 198
 199                 if (highest_edge_num >= 0)
 200                         env->edges[highest_edge_num].highest_execfreq = 1;
 201         }
 202 }
 203
 204 static int cmp_edges(const void *d1, const void *d2)
 205 {
 206         const edge_t *e1 = (const edge_t*)d1;
 207         const edge_t *e2 = (const edge_t*)d2;
 208
 209         return QSORT_CMP(e2->execfreq, e1->execfreq);
 210 }
 211
 212 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 213 {
 214         const edge_t *e1 = (const edge_t*)d1;
 215         const edge_t *e2 = (const edge_t*)d2;
 216         /* reverse sorting as penalties are negative */
 217         return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq);
 218 }
 219
 220 static void clear_loop_links(ir_loop *loop)
 221 {
 222         int i, n;
 223
 224         set_loop_link(loop, NULL);
 225         n = get_loop_n_elements(loop);
 226         for (i = 0; i < n; ++i) {
 227                 loop_element elem = get_loop_element(loop, i);
 228                 if (*elem.kind == k_ir_loop) {
 229                         clear_loop_links(elem.son);
 230                 }
 231         }
 232 }
 233
 234 static void coalesce_blocks(blocksched_env_t *env)
 235 {
 236         int i;
 237         int edge_count = ARR_LEN(env->edges);
 238         edge_t *edges = env->edges;
 239
 240         /* sort interblock edges by execution frequency */
 241         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 242
 243         /* run1: only look at jumps */
 244         for (i = 0; i < edge_count; ++i) {
 245                 const edge_t *edge  = &edges[i];
 246                 ir_node      *block = edge->block;
 247                 int           pos   = edge->pos;
 248                 ir_node      *pred_block;
 249                 blocksched_entry_t *entry, *pred_entry;
 250
 251                 /* only check edge with highest frequency */
 252                 if (! edge->highest_execfreq)
 253                         continue;
 254
 255                 /* the block might have been removed already... */
 256                 if (is_Bad(get_Block_cfgpred(block, 0)))
 257                         continue;
 258
 259                 pred_block = get_Block_cfgpred_block(block, pos);
 260                 entry      = (blocksched_entry_t*)get_irn_link(block);
 261                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 262
 263                 if (pred_entry->next != NULL || entry->prev != NULL)
 264                         continue;
 265
 266                 /* only coalesce jumps */
 267                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 268                         continue;
 269
 270                 /* schedule the 2 blocks behind each other */
 271                 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 272                            pred_entry->block, entry->block, edge->execfreq));
 273                 pred_entry->next = entry;
 274                 entry->prev      = pred_entry;
 275         }
 276
 277         /* run2: pick loop fallthroughs */
 278         clear_loop_links(get_irg_loop(env->irg));
 279
 280         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
 281         for (i = 0; i < edge_count; ++i) {
 282                 const edge_t *edge  = &edges[i];
 283                 ir_node      *block = edge->block;
 284                 int           pos   = edge->pos;
 285                 ir_node      *pred_block;
 286                 blocksched_entry_t *entry, *pred_entry;
 287                 ir_loop      *loop;
 288                 ir_loop      *outer_loop;
 289
 290                 /* already seen all loop outedges? */
 291                 if (edge->outedge_penalty_freq == 0)
 292                         break;
 293
 294                 /* the block might have been removed already... */
 295                 if (is_Bad(get_Block_cfgpred(block, pos)))
 296                         continue;
 297
 298                 pred_block = get_Block_cfgpred_block(block, pos);
 299                 entry      = (blocksched_entry_t*)get_irn_link(block);
 300                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 301
 302                 if (pred_entry->next != NULL || entry->prev != NULL)
 303                         continue;
 304
 305                 /* we want at most 1 outedge fallthrough per loop */
 306                 loop = get_irn_loop(pred_block);
 307                 if (get_loop_link(loop) != NULL)
 308                         continue;
 309
 310                 /* schedule the 2 blocks behind each other */
 311                 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
 312                            pred_entry->block, entry->block, edge->execfreq));
 313                 pred_entry->next = entry;
 314                 entry->prev      = pred_entry;
 315
 316                 /* all loops left have an outedge now */
 317                 outer_loop = get_irn_loop(block);
 318                 do {
 319                         /* we set loop link to loop to mark it */
 320                         set_loop_link(loop, loop);
 321                         loop = get_loop_outer_loop(loop);
 322                 } while (loop != outer_loop);
 323         }
 324
 325         /* sort interblock edges by execution frequency */
 326         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 327
 328         /* run3: remaining edges */
 329         for (i = 0; i < edge_count; ++i) {
 330                 const edge_t *edge  = &edges[i];
 331                 ir_node      *block = edge->block;
 332                 int           pos   = edge->pos;
 333                 ir_node      *pred_block;
 334                 blocksched_entry_t *entry, *pred_entry;
 335
 336                 /* the block might have been removed already... */
 337                 if (is_Bad(get_Block_cfgpred(block, pos)))
 338                         continue;
 339
 340                 pred_block = get_Block_cfgpred_block(block, pos);
 341                 entry      = (blocksched_entry_t*)get_irn_link(block);
 342                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 343
 344                 /* is 1 of the blocks already attached to another block? */
 345                 if (pred_entry->next != NULL || entry->prev != NULL)
 346                         continue;
 347
 348                 /* schedule the 2 blocks behind each other */
 349                 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 350                            pred_entry->block, entry->block, edge->execfreq));
 351                 pred_entry->next = entry;
 352                 entry->prev      = pred_entry;
 353         }
 354 }
 355
 356 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 357 {
 358         ir_node            *block = entry->block;
 359         ir_node            *succ  = NULL;
 360         blocksched_entry_t *succ_entry;
 361         const ir_edge_t    *edge;
 362         double             best_succ_execfreq;
 363
 364         if (irn_visited_else_mark(block))
 365                 return;
 366
 367         env->blockcount++;
 368
 369         DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 370
 371         /* put all successors into the worklist */
 372         foreach_block_succ(block, edge) {
 373                 ir_node *succ_block = get_edge_src_irn(edge);
 374
 375                 if (irn_visited(succ_block))
 376                         continue;
 377
 378                 /* we only need to put the first of a series of already connected
 379                  * blocks into the worklist */
 380                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 381                 while (succ_entry->prev != NULL) {
 382                         /* break cycles... */
 383                         if (succ_entry->prev->block == succ_block) {
 384                                 succ_entry->prev->next = NULL;
 385                                 succ_entry->prev       = NULL;
 386                                 break;
 387                         }
 388                         succ_entry = succ_entry->prev;
 389                 }
 390
 391                 if (irn_visited(succ_entry->block))
 392                         continue;
 393
 394                 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 395                 pdeq_putr(env->worklist, succ_entry->block);
 396         }
 397
 398         if (entry->next != NULL) {
 399                 pick_block_successor(entry->next, env);
 400                 return;
 401         }
 402
 403         DB((dbg, LEVEL_1, "deciding...\n"));
 404         best_succ_execfreq = -1;
 405
 406         /* no successor yet: pick the successor block with the highest execution
 407          * frequency which has no predecessor yet */
 408
 409         foreach_block_succ(block, edge) {
 410                 ir_node *succ_block = get_edge_src_irn(edge);
 411                 double  execfreq;
 412
 413                 if (irn_visited(succ_block))
 414                         continue;
 415
 416                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 417                 if (succ_entry->prev != NULL)
 418                         continue;
 419
 420                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 421                 if (execfreq > best_succ_execfreq) {
 422                         best_succ_execfreq = execfreq;
 423                         succ = succ_block;
 424                 }
 425         }
 426
 427         if (succ == NULL) {
 428                 DB((dbg, LEVEL_1, "pick from worklist\n"));
 429
 430                 do {
 431                         if (pdeq_empty(env->worklist)) {
 432                                 DB((dbg, LEVEL_1, "worklist empty\n"));
 433                                 return;
 434                         }
 435                         succ = (ir_node*)pdeq_getl(env->worklist);
 436                 } while (irn_visited(succ));
 437         }
 438
 439         succ_entry       = (blocksched_entry_t*)get_irn_link(succ);
 440         entry->next      = succ_entry;
 441         succ_entry->prev = entry;
 442
 443         pick_block_successor(succ_entry, env);
 444 }
 445
 446 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 447 {
 448         ir_graph           *irg        = env->irg;
 449         ir_node            *startblock = get_irg_start_block(irg);
 450         blocksched_entry_t *entry      = (blocksched_entry_t*)get_irn_link(startblock);
 451
 452         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 453         inc_irg_visited(irg);
 454
 455         env->worklist = new_pdeq();
 456         pick_block_successor(entry, env);
 457         assert(pdeq_empty(env->worklist));
 458         del_pdeq(env->worklist);
 459
 460         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 461
 462         return entry;
 463 }
 464
 465 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 466                                                                                 int count, struct obstack* obst)
 467 {
 468         int                i = 0;
 469         ir_node            **block_list;
 470         blocksched_entry_t *entry;
 471         (void) env;
 472
 473         block_list = NEW_ARR_D(ir_node *, obst, count);
 474         DB((dbg, LEVEL_1, "Blockschedule:\n"));
 475
 476         for (entry = first; entry != NULL; entry = entry->next) {
 477                 assert(i < count);
 478                 block_list[i++] = entry->block;
 479                 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
 480         }
 481         assert(i == count);
 482
 483         return block_list;
 484 }
 485
 486 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 487 {
 488         blocksched_env_t   env;
 489         struct obstack     obst;
 490         blocksched_entry_t *start_entry;
 491         ir_node            **block_list;
 492
 493         obstack_init(&obst);
 494
 495         env.irg        = irg;
 496         env.obst       = &obst;
 497         env.execfreqs  = execfreqs;
 498         env.edges      = NEW_ARR_F(edge_t, 0);
 499         env.worklist   = NULL;
 500         env.blockcount = 0;
 501
 502         assure_loopinfo(irg);
 503
 504         // collect edge execution frequencies
 505         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 506
 507         (void)be_remove_empty_blocks(irg);
 508
 509         if (algo != BLOCKSCHED_NAIV)
 510                 coalesce_blocks(&env);
 511
 512         start_entry = finish_block_schedule(&env);
 513         block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
 514                                               be_get_be_obst(irg));
 515
 516         DEL_ARR_F(env.edges);
 517         obstack_free(&obst, NULL);
 518
 519         return block_list;
 520 }
 521
 522 /*
 523  *  ___ _     ____
 524  * |_ _| |   |  _ \
 525  *  | || |   | |_) |
 526  *  | || |___|  __/
 527  * |___|_____|_|
 528  *
 529  */
 530
 531 typedef struct ilp_edge_t {
 532         ir_node *block;   /**< source block */
 533         int     pos;      /**< number of cfg predecessor (target) */
 534         int     ilpvar;
 535 } ilp_edge_t;
 536
 537 typedef struct blocksched_ilp_env_t {
 538         blocksched_env_t env;
 539         ilp_edge_t       *ilpedges;
 540         lpp_t            *lpp;
 541 } blocksched_ilp_env_t;
 542
 543 typedef struct blocksched_ilp_entry_t {
 544         ir_node *block;
 545         struct blocksched_entry_t *next;
 546         struct blocksched_entry_t *prev;
 547
 548         int out_cst;
 549 } blocksched_ilp_entry_t;
 550
 551 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 552 {
 553         char       name[64];
 554         ilp_edge_t edge;
 555         int        edgeidx = ARR_LEN(env->ilpedges);
 556
 557         snprintf(name, sizeof(name), "edge%d", edgeidx);
 558
 559         edge.block  = block;
 560         edge.pos    = pos;
 561         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 562
 563         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 564         return edgeidx;
 565 }
 566
 567 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 568 {
 569         blocksched_ilp_env_t *env        = data;
 570         ir_graph             *irg        = env->env.irg;
 571         ir_node              *startblock = get_irg_start_block(irg);
 572         int                  arity;
 573         lpp_cst_t            cst;
 574         char                 name[64];
 575         int                  out_count;
 576         blocksched_ilp_entry_t *entry;
 577
 578         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 579         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 580
 581         entry          = OALLOC(env->env.obst, blocksched_ilp_entry_t);
 582         entry->block   = block;
 583         entry->next    = NULL;
 584         entry->prev    = NULL;
 585         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
 586         set_irn_link(block, entry);
 587
 588         if (block == startblock)
 589                 return;
 590
 591         arity = get_irn_arity(block);
 592         if (arity == 1) {
 593                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 594                 add_ilp_edge(block, 0, execfreq, env);
 595         }
 596         else {
 597                 int i;
 598
 599                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 600                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
 601
 602                 for (i = 0; i < arity; ++i) {
 603                         double     execfreq;
 604                         int        edgenum;
 605                         ilp_edge_t *edge;
 606                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 607
 608                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 609                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 610                         edge     = &env->ilpedges[edgenum];
 611                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 612                 }
 613         }
 614 }
 615
 616
 617 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 618 {
 619         int           edge_count = ARR_LEN(env->ilpedges);
 620         be_options_t *options    = be_get_irg_options(env->env.irg);
 621         int           i;
 622
 623         /* complete out constraints */
 624         for (i = 0; i < edge_count; ++i) {
 625                 const ilp_edge_t *edge  = &env->ilpedges[i];
 626                 ir_node          *block = edge->block;
 627                 ir_node          *pred;
 628                 blocksched_ilp_entry_t *entry;
 629
 630                 /* the block might have been removed already... */
 631                 if (is_Bad(get_Block_cfgpred(block, 0)))
 632                         continue;
 633
 634                 pred  = get_Block_cfgpred_block(block, edge->pos);
 635                 entry = get_irn_link(pred);
 636
 637                 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 638                                   pred, block, edge->pos));
 639                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 640         }
 641
 642         lpp_solve_net(env->lpp, options->ilp_server, options->ilp_solver);
 643         assert(lpp_is_sol_valid(env->lpp));
 644
 645         /* Apply results to edges */
 646         for (i = 0; i < edge_count; ++i) {
 647                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 648                 ir_node            *block = edge->block;
 649                 ir_node            *pred;
 650                 int                is_jump;
 651                 blocksched_entry_t *entry;
 652                 blocksched_entry_t *pred_entry;
 653
 654                 /* the block might have been removed already... */
 655                 if (is_Bad(get_Block_cfgpred(block, 0)))
 656                         continue;
 657
 658                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 659                 if (is_jump)
 660                         continue;
 661
 662                 pred       = get_Block_cfgpred_block(block, edge->pos);
 663                 entry      = get_irn_link(block);
 664                 pred_entry = get_irn_link(pred);
 665
 666                 assert(entry->prev == NULL && pred_entry->next == NULL);
 667                 entry->prev      = pred_entry;
 668                 pred_entry->next = entry;
 669         }
 670 }
 671
 672 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 673 {
 674         blocksched_ilp_env_t env;
 675         struct obstack       obst;
 676         blocksched_entry_t   *start_entry;
 677         ir_node              **block_list;
 678
 679         obstack_init(&obst);
 680
 681         env.env.irg        = irg;
 682         env.env.obst       = &obst;
 683         env.env.execfreqs  = execfreqs;
 684         env.env.worklist   = NULL;
 685         env.env.blockcount = 0;
 686         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 687
 688         env.lpp = lpp_new("blockschedule", lpp_minimize);
 689         lpp_set_time_limit(env.lpp, 20);
 690         lpp_set_log(env.lpp, stdout);
 691
 692         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 693
 694         (void)be_remove_empty_blocks(irg);
 695         coalesce_blocks_ilp(&env);
 696
 697         start_entry = finish_block_schedule(&env.env);
 698         block_list  = create_blocksched_array(&env.env, start_entry,
 699                                               env.env.blockcount,
 700                                               be_get_be_obst(irg));
 701
 702         DEL_ARR_F(env.ilpedges);
 703         lpp_free(env.lpp);
 704         obstack_free(&obst, NULL);
 705
 706         return block_list;
 707 }
 708
 709 /*
 710  *  __  __       _
 711  * |  \/  | __ _(_)_ __
 712  * | |\/| |/ _` | | '_ \
 713  * | |  | | (_| | | | | |
 714  * |_|  |_|\__,_|_|_| |_|
 715  *
 716  */
 717 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 718 void be_init_blocksched(void)
 719 {
 720         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 721
 722         lc_opt_add_table(be_grp, be_blocksched_options);
 723
 724         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 725 }
 726
 727 ir_node **be_create_block_schedule(ir_graph *irg)
 728 {
 729         ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
 730
 731         switch (algo) {
 732         case BLOCKSCHED_GREEDY:
 733         case BLOCKSCHED_NAIV:
 734                 return create_block_schedule_greedy(irg, execfreqs);
 735         case BLOCKSCHED_ILP:
 736                 return create_block_schedule_ilp(irg, execfreqs);
 737         }
 738
 739         panic("unknown blocksched algo");
 740 }