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  *
  26  * The goals of the greedy (and ILP) algorithm here works by assuming that
  27  * we want to change as many jumps to fallthroughs as possible (executed jumps
  28  * actually, we have to look at the execution frequencies). The algorithms
  29  * do this by collecting execution frequencies of all branches (which is easily
  30  * possible when all critical edges are split) then removes critical edges where
  31  * possible as we don't need and want them anymore now. The algorithms then try
  32  * to change as many edges to fallthroughs as possible, this is done by setting
  33  * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
  34  * execution frequencies and tries to transform them to fallthroughs in this order
  35  */
  36 #include "config.h"
  37
  38 #include "beblocksched.h"
  39
  40 #include <stdlib.h>
  41
  42 #include "array.h"
  43 #include "pdeq.h"
  44 #include "beirg.h"
  45 #include "iredges.h"
  46 #include "irgwalk.h"
  47 #include "irnode_t.h"
  48 #include "irgraph_t.h"
  49 #include "irloop.h"
  50 #include "irprintf.h"
  51 #include "execfreq.h"
  52 #include "irdump_t.h"
  53 #include "irtools.h"
  54 #include "util.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 #include "lpp.h"
  65 #include "lpp_net.h"
  66
  67 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  68
  69 typedef enum blocksched_algos_t {
  70         BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP
  71 } blocksched_algos_t;
  72
  73 static int algo = BLOCKSCHED_GREEDY;
  74
  75 static const lc_opt_enum_int_items_t blockschedalgo_items[] = {
  76         { "naiv",   BLOCKSCHED_NAIV },
  77         { "greedy", BLOCKSCHED_GREEDY },
  78         { "ilp",    BLOCKSCHED_ILP },
  79         { NULL,     0 }
  80 };
  81
  82 static lc_opt_enum_int_var_t algo_var = {
  83         &algo, blockschedalgo_items
  84 };
  85
  86 static const lc_opt_table_entry_t be_blocksched_options[] = {
  87         LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var),
  88         LC_OPT_LAST
  89 };
  90
  91 /*
  92  *   ____                   _
  93  *  / ___|_ __ ___  ___  __| |_   _
  94  * | |  _| '__/ _ \/ _ \/ _` | | | |
  95  * | |_| | | |  __/  __/ (_| | |_| |
  96  *  \____|_|  \___|\___|\__,_|\__, |
  97  *                            |___/
  98  */
  99
 100 typedef struct blocksched_entry_t blocksched_entry_t;
 101 struct blocksched_entry_t {
 102         ir_node            *block;
 103         blocksched_entry_t *next;
 104         blocksched_entry_t *prev;
 105 };
 106
 107 typedef struct edge_t edge_t;
 108 struct edge_t {
 109         ir_node *block;             /**< source block */
 110         int     pos;                /**< number of cfg predecessor (target) */
 111         double  execfreq;           /**< the frequency */
 112         double  outedge_penalty_freq; /**< for edges leaving the loop this is the
 113                                            penality when we make them a
 114                                            fallthrough. */
 115         int     highest_execfreq;   /**< flag that indicates whether this edge is
 116                                          the edge with the highest execfreq pointing
 117                                          away from this block */
 118 };
 119
 120 typedef struct blocksched_env_t blocksched_env_t;
 121 struct blocksched_env_t {
 122         ir_graph       *irg;
 123         struct obstack  obst;
 124         edge_t         *edges;
 125         pdeq           *worklist;
 126         int            blockcount;
 127 };
 128
 129 static blocksched_entry_t* get_blocksched_entry(const ir_node *block)
 130 {
 131         return (blocksched_entry_t*)get_irn_link(block);
 132 }
 133
 134 /**
 135  * Collect cfg frequencies of all edges between blocks.
 136  * Also determines edge with highest frequency.
 137  */
 138 static void collect_egde_frequency(ir_node *block, void *data)
 139 {
 140         blocksched_env_t   *env = (blocksched_env_t*)data;
 141         int                arity;
 142         edge_t             edge;
 143         blocksched_entry_t *entry;
 144         ir_loop            *loop;
 145
 146         memset(&edge, 0, sizeof(edge));
 147
 148         entry = OALLOCZ(&env->obst, blocksched_entry_t);
 149         entry->block = block;
 150         set_irn_link(block, entry);
 151
 152         loop = get_irn_loop(block);
 153
 154         arity = get_Block_n_cfgpreds(block);
 155
 156         if (arity == 0) {
 157                 /* must be the start block (or end-block for endless loops),
 158                  * everything else is dead code and should be removed by now */
 159                 assert(block == get_irg_start_block(env->irg)
 160                                 || block == get_irg_end_block(env->irg));
 161                 /* nothing to do here */
 162                 return;
 163         } else if (arity == 1) {
 164                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 165                 ir_loop *pred_loop  = get_irn_loop(pred_block);
 166                 float    freq       = (float)get_block_execfreq(block);
 167
 168                 /* is it an edge leaving a loop */
 169                 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
 170                         float pred_freq = (float)get_block_execfreq(pred_block);
 171                         edge.outedge_penalty_freq = -(pred_freq - freq);
 172                 }
 173
 174                 edge.block            = block;
 175                 edge.pos              = 0;
 176                 edge.execfreq         = freq;
 177                 edge.highest_execfreq = 1;
 178                 ARR_APP1(edge_t, env->edges, edge);
 179         } else {
 180                 int    i;
 181                 double highest_execfreq = -1.0;
 182                 int    highest_edge_num = -1;
 183
 184                 edge.block = block;
 185                 for (i = 0; i < arity; ++i) {
 186                         double  execfreq;
 187                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 188
 189                         execfreq = get_block_execfreq(pred_block);
 190
 191                         edge.pos              = i;
 192                         edge.execfreq         = execfreq;
 193                         edge.highest_execfreq = 0;
 194                         ARR_APP1(edge_t, env->edges, edge);
 195
 196                         if (execfreq > highest_execfreq) {
 197                                 highest_execfreq = execfreq;
 198                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 199                         }
 200                 }
 201
 202                 if (highest_edge_num >= 0)
 203                         env->edges[highest_edge_num].highest_execfreq = 1;
 204         }
 205 }
 206
 207 static int cmp_edges_base(const edge_t *e1, const edge_t *e2)
 208 {
 209         long nr1 = get_irn_node_nr(e1->block);
 210         long nr2 = get_irn_node_nr(e2->block);
 211         if (nr1 < nr2) {
 212                 return 1;
 213         } else if (nr1 > nr2) {
 214                 return -1;
 215         } else {
 216                 if (e1->pos < e2->pos) {
 217                         return 1;
 218                 } else if (e1->pos > e2->pos) {
 219                         return -1;
 220                 } else {
 221                         return 0;
 222                 }
 223         }
 224 }
 225
 226 static int cmp_edges(const void *d1, const void *d2)
 227 {
 228         const edge_t *e1 = (const edge_t*)d1;
 229         const edge_t *e2 = (const edge_t*)d2;
 230         double        freq1 = e1->execfreq;
 231         double        freq2 = e2->execfreq;
 232         if (freq1 < freq2) {
 233                 return 1;
 234         } else if (freq1 > freq2) {
 235                 return -1;
 236         } else {
 237                 return cmp_edges_base(e1, e2);
 238         }
 239 }
 240
 241 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 242 {
 243         const edge_t *e1   = (const edge_t*)d1;
 244         const edge_t *e2   = (const edge_t*)d2;
 245         double        pen1 = e1->outedge_penalty_freq;
 246         double        pen2 = e2->outedge_penalty_freq;
 247         if (pen1 > pen2) {
 248                 return 1;
 249         } else if (pen1 < pen2) {
 250                 return -1;
 251         } else {
 252                 return cmp_edges_base(e1, e2);
 253         }
 254 }
 255
 256 static void clear_loop_links(ir_loop *loop)
 257 {
 258         int i, n;
 259
 260         set_loop_link(loop, NULL);
 261         n = get_loop_n_elements(loop);
 262         for (i = 0; i < n; ++i) {
 263                 loop_element elem = get_loop_element(loop, i);
 264                 if (*elem.kind == k_ir_loop) {
 265                         clear_loop_links(elem.son);
 266                 }
 267         }
 268 }
 269
 270 static void coalesce_blocks(blocksched_env_t *env)
 271 {
 272         int i;
 273         int edge_count = ARR_LEN(env->edges);
 274         edge_t *edges = env->edges;
 275
 276         /* sort interblock edges by execution frequency */
 277         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 278
 279         /* run1: only look at jumps */
 280         for (i = 0; i < edge_count; ++i) {
 281                 const edge_t *edge  = &edges[i];
 282                 ir_node      *block = edge->block;
 283                 int           pos   = edge->pos;
 284                 ir_node      *pred_block;
 285                 blocksched_entry_t *entry, *pred_entry;
 286
 287                 /* only check edge with highest frequency */
 288                 if (! edge->highest_execfreq)
 289                         continue;
 290
 291                 /* the block might have been removed already... */
 292                 if (is_Bad(get_Block_cfgpred(block, 0)))
 293                         continue;
 294
 295                 pred_block = get_Block_cfgpred_block(block, pos);
 296                 entry      = get_blocksched_entry(block);
 297                 pred_entry = get_blocksched_entry(pred_block);
 298
 299                 if (pred_entry->next != NULL || entry->prev != NULL)
 300                         continue;
 301
 302                 /* only coalesce jumps */
 303                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 304                         continue;
 305
 306                 /* schedule the 2 blocks behind each other */
 307                 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 308                            pred_entry->block, entry->block, edge->execfreq));
 309                 pred_entry->next = entry;
 310                 entry->prev      = pred_entry;
 311         }
 312
 313         /* run2: pick loop fallthroughs */
 314         clear_loop_links(get_irg_loop(env->irg));
 315
 316         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
 317         for (i = 0; i < edge_count; ++i) {
 318                 const edge_t *edge  = &edges[i];
 319                 ir_node      *block = edge->block;
 320                 int           pos   = edge->pos;
 321                 ir_node      *pred_block;
 322                 blocksched_entry_t *entry, *pred_entry;
 323                 ir_loop      *loop;
 324                 ir_loop      *outer_loop;
 325
 326                 /* already seen all loop outedges? */
 327                 if (edge->outedge_penalty_freq == 0)
 328                         break;
 329
 330                 /* the block might have been removed already... */
 331                 if (is_Bad(get_Block_cfgpred(block, pos)))
 332                         continue;
 333
 334                 pred_block = get_Block_cfgpred_block(block, pos);
 335                 entry      = get_blocksched_entry(block);
 336                 pred_entry = get_blocksched_entry(pred_block);
 337
 338                 if (pred_entry->next != NULL || entry->prev != NULL)
 339                         continue;
 340
 341                 /* we want at most 1 outedge fallthrough per loop */
 342                 loop = get_irn_loop(pred_block);
 343                 if (get_loop_link(loop) != NULL)
 344                         continue;
 345
 346                 /* schedule the 2 blocks behind each other */
 347                 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
 348                            pred_entry->block, entry->block, edge->execfreq));
 349                 pred_entry->next = entry;
 350                 entry->prev      = pred_entry;
 351
 352                 /* all loops left have an outedge now */
 353                 outer_loop = get_irn_loop(block);
 354                 do {
 355                         /* we set loop link to loop to mark it */
 356                         set_loop_link(loop, loop);
 357                         loop = get_loop_outer_loop(loop);
 358                 } while (loop != outer_loop);
 359         }
 360
 361         /* sort interblock edges by execution frequency */
 362         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 363
 364         /* run3: remaining edges */
 365         for (i = 0; i < edge_count; ++i) {
 366                 const edge_t *edge  = &edges[i];
 367                 ir_node      *block = edge->block;
 368                 int           pos   = edge->pos;
 369                 ir_node      *pred_block;
 370                 blocksched_entry_t *entry, *pred_entry;
 371
 372                 /* the block might have been removed already... */
 373                 if (is_Bad(get_Block_cfgpred(block, pos)))
 374                         continue;
 375
 376                 pred_block = get_Block_cfgpred_block(block, pos);
 377                 entry      = get_blocksched_entry(block);
 378                 pred_entry = get_blocksched_entry(pred_block);
 379
 380                 /* is 1 of the blocks already attached to another block? */
 381                 if (pred_entry->next != NULL || entry->prev != NULL)
 382                         continue;
 383
 384                 /* schedule the 2 blocks behind each other */
 385                 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 386                            pred_entry->block, entry->block, edge->execfreq));
 387                 pred_entry->next = entry;
 388                 entry->prev      = pred_entry;
 389         }
 390 }
 391
 392 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 393 {
 394         ir_node            *block = entry->block;
 395         ir_node            *succ  = NULL;
 396         blocksched_entry_t *succ_entry;
 397         double              best_succ_execfreq;
 398
 399         if (irn_visited_else_mark(block))
 400                 return;
 401
 402         env->blockcount++;
 403
 404         DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 405
 406         /* put all successors into the worklist */
 407         foreach_block_succ(block, edge) {
 408                 ir_node *succ_block = get_edge_src_irn(edge);
 409
 410                 if (irn_visited(succ_block))
 411                         continue;
 412
 413                 /* we only need to put the first of a series of already connected
 414                  * blocks into the worklist */
 415                 succ_entry = get_blocksched_entry(succ_block);
 416                 while (succ_entry->prev != NULL) {
 417                         /* break cycles... */
 418                         if (succ_entry->prev->block == succ_block) {
 419                                 succ_entry->prev->next = NULL;
 420                                 succ_entry->prev       = NULL;
 421                                 break;
 422                         }
 423                         succ_entry = succ_entry->prev;
 424                 }
 425
 426                 if (irn_visited(succ_entry->block))
 427                         continue;
 428
 429                 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 430                 pdeq_putr(env->worklist, succ_entry->block);
 431         }
 432
 433         if (entry->next != NULL) {
 434                 pick_block_successor(entry->next, env);
 435                 return;
 436         }
 437
 438         DB((dbg, LEVEL_1, "deciding...\n"));
 439         best_succ_execfreq = -1;
 440
 441         /* no successor yet: pick the successor block with the highest execution
 442          * frequency which has no predecessor yet */
 443
 444         foreach_block_succ(block, edge) {
 445                 ir_node *succ_block = get_edge_src_irn(edge);
 446
 447                 if (irn_visited(succ_block))
 448                         continue;
 449
 450                 succ_entry = get_blocksched_entry(succ_block);
 451                 if (succ_entry->prev != NULL)
 452                         continue;
 453
 454                 double execfreq = get_block_execfreq(succ_block);
 455                 if (execfreq > best_succ_execfreq) {
 456                         best_succ_execfreq = execfreq;
 457                         succ = succ_block;
 458                 }
 459         }
 460
 461         if (succ == NULL) {
 462                 DB((dbg, LEVEL_1, "pick from worklist\n"));
 463
 464                 do {
 465                         if (pdeq_empty(env->worklist)) {
 466                                 DB((dbg, LEVEL_1, "worklist empty\n"));
 467                                 return;
 468                         }
 469                         succ = (ir_node*)pdeq_getl(env->worklist);
 470                 } while (irn_visited(succ));
 471         }
 472
 473         succ_entry       = get_blocksched_entry(succ);
 474         entry->next      = succ_entry;
 475         succ_entry->prev = entry;
 476
 477         pick_block_successor(succ_entry, env);
 478 }
 479
 480 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 481 {
 482         ir_graph           *irg        = env->irg;
 483         ir_node            *startblock = get_irg_start_block(irg);
 484         blocksched_entry_t *entry      = get_blocksched_entry(startblock);
 485
 486         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 487         inc_irg_visited(irg);
 488
 489         env->worklist = new_pdeq();
 490         pick_block_successor(entry, env);
 491         assert(pdeq_empty(env->worklist));
 492         del_pdeq(env->worklist);
 493
 494         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 495
 496         return entry;
 497 }
 498
 499 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 500                                                                                 int count, struct obstack* obst)
 501 {
 502         int                i = 0;
 503         ir_node            **block_list;
 504         blocksched_entry_t *entry;
 505         (void) env;
 506
 507         block_list = NEW_ARR_D(ir_node *, obst, count);
 508         DB((dbg, LEVEL_1, "Blockschedule:\n"));
 509
 510         for (entry = first; entry != NULL; entry = entry->next) {
 511                 assert(i < count);
 512                 block_list[i++] = entry->block;
 513                 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
 514         }
 515         assert(i == count);
 516
 517         return block_list;
 518 }
 519
 520 static ir_node **create_block_schedule_greedy(ir_graph *irg)
 521 {
 522         blocksched_env_t   env;
 523         blocksched_entry_t *start_entry;
 524         ir_node            **block_list;
 525
 526         env.irg        = irg;
 527         env.edges      = NEW_ARR_F(edge_t, 0);
 528         env.worklist   = NULL;
 529         env.blockcount = 0;
 530         obstack_init(&env.obst);
 531
 532         assure_loopinfo(irg);
 533
 534         // collect edge execution frequencies
 535         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 536
 537         (void)be_remove_empty_blocks(irg);
 538
 539         if (algo != BLOCKSCHED_NAIV)
 540                 coalesce_blocks(&env);
 541
 542         start_entry = finish_block_schedule(&env);
 543         block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
 544                                               be_get_be_obst(irg));
 545
 546         DEL_ARR_F(env.edges);
 547         obstack_free(&env.obst, NULL);
 548
 549         return block_list;
 550 }
 551
 552 /*
 553  *  ___ _     ____
 554  * |_ _| |   |  _ \
 555  *  | || |   | |_) |
 556  *  | || |___|  __/
 557  * |___|_____|_|
 558  *
 559  */
 560
 561 typedef struct ilp_edge_t {
 562         ir_node *block;   /**< source block */
 563         int     pos;      /**< number of cfg predecessor (target) */
 564         int     ilpvar;
 565 } ilp_edge_t;
 566
 567 typedef struct blocksched_ilp_env_t {
 568         blocksched_env_t env;
 569         ilp_edge_t       *ilpedges;
 570         lpp_t            *lpp;
 571 } blocksched_ilp_env_t;
 572
 573 typedef struct blocksched_ilp_entry_t {
 574         ir_node *block;
 575         struct blocksched_entry_t *next;
 576         struct blocksched_entry_t *prev;
 577
 578         int out_cst;
 579 } blocksched_ilp_entry_t;
 580
 581 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 582 {
 583         char       name[64];
 584         ilp_edge_t edge;
 585         int        edgeidx = ARR_LEN(env->ilpedges);
 586
 587         snprintf(name, sizeof(name), "edge%d", edgeidx);
 588
 589         edge.block  = block;
 590         edge.pos    = pos;
 591         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 592
 593         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 594         return edgeidx;
 595 }
 596
 597 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 598 {
 599         blocksched_ilp_env_t *env        = (blocksched_ilp_env_t*)data;
 600         ir_graph             *irg        = env->env.irg;
 601         ir_node              *startblock = get_irg_start_block(irg);
 602         int                  arity;
 603         char                 name[64];
 604         int                  out_count;
 605         blocksched_ilp_entry_t *entry;
 606
 607         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 608         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 609
 610         entry          = OALLOC(&env->env.obst, blocksched_ilp_entry_t);
 611         entry->block   = block;
 612         entry->next    = NULL;
 613         entry->prev    = NULL;
 614         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
 615         set_irn_link(block, entry);
 616
 617         if (block == startblock)
 618                 return;
 619
 620         arity = get_irn_arity(block);
 621         if (arity == 1) {
 622                 double execfreq = get_block_execfreq(block);
 623                 add_ilp_edge(block, 0, execfreq, env);
 624         }
 625         else {
 626                 int i;
 627                 int cst_idx;
 628
 629                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 630                 cst_idx = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
 631
 632                 for (i = 0; i < arity; ++i) {
 633                         double     execfreq;
 634                         int        edgenum;
 635                         ilp_edge_t *edge;
 636                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 637
 638                         execfreq = get_block_execfreq(pred_block);
 639                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 640                         edge     = &env->ilpedges[edgenum];
 641                         lpp_set_factor_fast(env->lpp, cst_idx, edge->ilpvar, 1.0);
 642                 }
 643         }
 644 }
 645
 646 static blocksched_ilp_entry_t *get_blocksched_ilp_entry(const ir_node *block)
 647 {
 648         return (blocksched_ilp_entry_t*)get_irn_link(block);
 649 }
 650
 651 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 652 {
 653         int           edge_count = ARR_LEN(env->ilpedges);
 654         int           i;
 655
 656         /* complete out constraints */
 657         for (i = 0; i < edge_count; ++i) {
 658                 const ilp_edge_t *edge  = &env->ilpedges[i];
 659                 ir_node          *block = edge->block;
 660                 ir_node          *pred;
 661                 blocksched_ilp_entry_t *entry;
 662
 663                 /* the block might have been removed already... */
 664                 if (is_Bad(get_Block_cfgpred(block, 0)))
 665                         continue;
 666
 667                 pred  = get_Block_cfgpred_block(block, edge->pos);
 668                 entry = get_blocksched_ilp_entry(pred);
 669
 670                 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 671                                   pred, block, edge->pos));
 672                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 673         }
 674
 675         lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
 676         assert(lpp_is_sol_valid(env->lpp));
 677
 678         /* Apply results to edges */
 679         for (i = 0; i < edge_count; ++i) {
 680                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 681                 ir_node            *block = edge->block;
 682                 ir_node            *pred;
 683                 int                is_jump;
 684                 blocksched_entry_t *entry;
 685                 blocksched_entry_t *pred_entry;
 686
 687                 /* the block might have been removed already... */
 688                 if (is_Bad(get_Block_cfgpred(block, 0)))
 689                         continue;
 690
 691                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 692                 if (is_jump)
 693                         continue;
 694
 695                 pred       = get_Block_cfgpred_block(block, edge->pos);
 696                 entry      = get_blocksched_entry(block);
 697                 pred_entry = get_blocksched_entry(pred);
 698
 699                 assert(entry->prev == NULL && pred_entry->next == NULL);
 700                 entry->prev      = pred_entry;
 701                 pred_entry->next = entry;
 702         }
 703 }
 704
 705 static ir_node **create_block_schedule_ilp(ir_graph *irg)
 706 {
 707         blocksched_ilp_env_t env;
 708         blocksched_entry_t   *start_entry;
 709         ir_node              **block_list;
 710
 711         env.env.irg        = irg;
 712         env.env.worklist   = NULL;
 713         env.env.blockcount = 0;
 714         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 715         obstack_init(&env.env.obst);
 716
 717         env.lpp = lpp_new("blockschedule", lpp_minimize);
 718         lpp_set_time_limit(env.lpp, 20);
 719         lpp_set_log(env.lpp, stdout);
 720
 721         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 722
 723         (void)be_remove_empty_blocks(irg);
 724         coalesce_blocks_ilp(&env);
 725
 726         start_entry = finish_block_schedule(&env.env);
 727         block_list  = create_blocksched_array(&env.env, start_entry,
 728                                               env.env.blockcount,
 729                                               be_get_be_obst(irg));
 730
 731         DEL_ARR_F(env.ilpedges);
 732         lpp_free(env.lpp);
 733         obstack_free(&env.env.obst, NULL);
 734
 735         return block_list;
 736 }
 737
 738 /*
 739  *  __  __       _
 740  * |  \/  | __ _(_)_ __
 741  * | |\/| |/ _` | | '_ \
 742  * | |  | | (_| | | | | |
 743  * |_|  |_|\__,_|_|_| |_|
 744  *
 745  */
 746 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 747 void be_init_blocksched(void)
 748 {
 749         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 750
 751         lc_opt_add_table(be_grp, be_blocksched_options);
 752
 753         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 754 }
 755
 756 ir_node **be_create_block_schedule(ir_graph *irg)
 757 {
 758         switch (algo) {
 759         case BLOCKSCHED_GREEDY:
 760         case BLOCKSCHED_NAIV:
 761                 return create_block_schedule_greedy(irg);
 762         case BLOCKSCHED_ILP:
 763                 return create_block_schedule_ilp(irg);
 764         }
 765
 766         panic("unknown blocksched algo");
 767 }