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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       Block-scheduling strategies.
  23  * @author      Matthias Braun, Christoph Mallon
  24  * @date        27.09.2006
  25  * @version     $Id$
  26  *
  27  * The goals of the greedy (and ILP) algorithm here works by assuming that
  28  * we want to change as many jumps to fallthroughs as possible (executed jumps
  29  * actually, we have to look at the execution frequencies). The algorithms
  30  * do this by collecting execution frequencies of all branches (which is easily
  31  * possible when all critical edges are split) then removes critical edges where
  32  * possible as we don't need and want them anymore now. The algorithms then try
  33  * to change as many edges to fallthroughs as possible, this is done by setting
  34  * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
  35  * execution frequencies and tries to transform them to fallthroughs in this order
  36  */
  37 #include "config.h"
  38
  39 #include "beblocksched.h"
  40
  41 #include <stdlib.h>
  42
  43 #include "array.h"
  44 #include "pdeq.h"
  45
  46 #include "iredges.h"
  47 #include "irgwalk.h"
  48 #include "irnode_t.h"
  49 #include "irgraph_t.h"
  50 #include "irloop.h"
  51 #include "irprintf.h"
  52 #include "execfreq.h"
  53 #include "irdump_t.h"
  54 #include "irtools.h"
  55 #include "debug.h"
  56 #include "beirgmod.h"
  57 #include "bemodule.h"
  58 #include "be.h"
  59 #include "error.h"
  60
  61 #include "lc_opts.h"
  62 #include "lc_opts_enum.h"
  63
  64 #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         ir_exec_freq   *execfreqs;
 125         edge_t         *edges;
 126         pdeq           *worklist;
 127         int            blockcount;
 128 };
 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 = (blocksched_env_t*)data;
 137         int                arity;
 138         edge_t             edge;
 139         blocksched_entry_t *entry;
 140         ir_loop            *loop;
 141
 142         memset(&edge, 0, sizeof(edge));
 143
 144         entry = OALLOCZ(env->obst, blocksched_entry_t);
 145         entry->block = block;
 146         set_irn_link(block, entry);
 147
 148         loop = get_irn_loop(block);
 149
 150         arity = get_Block_n_cfgpreds(block);
 151
 152         if (arity == 0) {
 153                 /* must be the start block (or end-block for endless loops),
 154                  * everything else is dead code and should be removed by now */
 155                 assert(block == get_irg_start_block(env->irg)
 156                                 || block == get_irg_end_block(env->irg));
 157                 /* nothing to do here */
 158                 return;
 159         } else if (arity == 1) {
 160                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 161                 ir_loop *pred_loop  = get_irn_loop(pred_block);
 162                 float    freq       = (float)get_block_execfreq(env->execfreqs, block);
 163
 164                 /* is it an edge leaving a loop */
 165                 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
 166                         float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
 167                         edge.outedge_penalty_freq = -(pred_freq - freq);
 168                 }
 169
 170                 edge.block            = block;
 171                 edge.pos              = 0;
 172                 edge.execfreq         = freq;
 173                 edge.highest_execfreq = 1;
 174                 ARR_APP1(edge_t, env->edges, edge);
 175         } else {
 176                 int    i;
 177                 double highest_execfreq = -1.0;
 178                 int    highest_edge_num = -1;
 179
 180                 edge.block = block;
 181                 for (i = 0; i < arity; ++i) {
 182                         double  execfreq;
 183                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 184
 185                         execfreq = get_block_execfreq(env->execfreqs, pred_block);
 186
 187                         edge.pos              = i;
 188                         edge.execfreq         = execfreq;
 189                         edge.highest_execfreq = 0;
 190                         ARR_APP1(edge_t, env->edges, edge);
 191
 192                         if (execfreq > highest_execfreq) {
 193                                 highest_execfreq = execfreq;
 194                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 195                         }
 196                 }
 197
 198                 if (highest_edge_num >= 0)
 199                         env->edges[highest_edge_num].highest_execfreq = 1;
 200         }
 201 }
 202
 203 static int cmp_edges(const void *d1, const void *d2)
 204 {
 205         const edge_t *e1 = (const edge_t*)d1;
 206         const edge_t *e2 = (const edge_t*)d2;
 207
 208         return QSORT_CMP(e2->execfreq, e1->execfreq);
 209 }
 210
 211 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 212 {
 213         const edge_t *e1 = (const edge_t*)d1;
 214         const edge_t *e2 = (const edge_t*)d2;
 215         /* reverse sorting as penalties are negative */
 216         return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq);
 217 }
 218
 219 static void clear_loop_links(ir_loop *loop)
 220 {
 221         int i, n;
 222
 223         set_loop_link(loop, NULL);
 224         n = get_loop_n_elements(loop);
 225         for (i = 0; i < n; ++i) {
 226                 loop_element elem = get_loop_element(loop, i);
 227                 if (*elem.kind == k_ir_loop) {
 228                         clear_loop_links(elem.son);
 229                 }
 230         }
 231 }
 232
 233 static void coalesce_blocks(blocksched_env_t *env)
 234 {
 235         int i;
 236         int edge_count = ARR_LEN(env->edges);
 237         edge_t *edges = env->edges;
 238
 239         /* sort interblock edges by execution frequency */
 240         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 241
 242         /* run1: only look at jumps */
 243         for (i = 0; i < edge_count; ++i) {
 244                 const edge_t *edge  = &edges[i];
 245                 ir_node      *block = edge->block;
 246                 int           pos   = edge->pos;
 247                 ir_node      *pred_block;
 248                 blocksched_entry_t *entry, *pred_entry;
 249
 250                 /* only check edge with highest frequency */
 251                 if (! edge->highest_execfreq)
 252                         continue;
 253
 254                 /* the block might have been removed already... */
 255                 if (is_Bad(get_Block_cfgpred(block, 0)))
 256                         continue;
 257
 258                 pred_block = get_Block_cfgpred_block(block, pos);
 259                 entry      = (blocksched_entry_t*)get_irn_link(block);
 260                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 261
 262                 if (pred_entry->next != NULL || entry->prev != NULL)
 263                         continue;
 264
 265                 /* only coalesce jumps */
 266                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 267                         continue;
 268
 269                 /* schedule the 2 blocks behind each other */
 270                 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 271                            pred_entry->block, entry->block, edge->execfreq));
 272                 pred_entry->next = entry;
 273                 entry->prev      = pred_entry;
 274         }
 275
 276         /* run2: pick loop fallthroughs */
 277         clear_loop_links(get_irg_loop(env->irg));
 278
 279         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
 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                 ir_loop      *loop;
 287                 ir_loop      *outer_loop;
 288
 289                 /* already seen all loop outedges? */
 290                 if (edge->outedge_penalty_freq == 0)
 291                         break;
 292
 293                 /* the block might have been removed already... */
 294                 if (is_Bad(get_Block_cfgpred(block, pos)))
 295                         continue;
 296
 297                 pred_block = get_Block_cfgpred_block(block, pos);
 298                 entry      = (blocksched_entry_t*)get_irn_link(block);
 299                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 300
 301                 if (pred_entry->next != NULL || entry->prev != NULL)
 302                         continue;
 303
 304                 /* we want at most 1 outedge fallthrough per loop */
 305                 loop = get_irn_loop(pred_block);
 306                 if (get_loop_link(loop) != NULL)
 307                         continue;
 308
 309                 /* schedule the 2 blocks behind each other */
 310                 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
 311                            pred_entry->block, entry->block, edge->execfreq));
 312                 pred_entry->next = entry;
 313                 entry->prev      = pred_entry;
 314
 315                 /* all loops left have an outedge now */
 316                 outer_loop = get_irn_loop(block);
 317                 do {
 318                         /* we set loop link to loop to mark it */
 319                         set_loop_link(loop, loop);
 320                         loop = get_loop_outer_loop(loop);
 321                 } while (loop != outer_loop);
 322         }
 323
 324         /* sort interblock edges by execution frequency */
 325         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 326
 327         /* run3: remaining edges */
 328         for (i = 0; i < edge_count; ++i) {
 329                 const edge_t *edge  = &edges[i];
 330                 ir_node      *block = edge->block;
 331                 int           pos   = edge->pos;
 332                 ir_node      *pred_block;
 333                 blocksched_entry_t *entry, *pred_entry;
 334
 335                 /* the block might have been removed already... */
 336                 if (is_Bad(get_Block_cfgpred(block, pos)))
 337                         continue;
 338
 339                 pred_block = get_Block_cfgpred_block(block, pos);
 340                 entry      = (blocksched_entry_t*)get_irn_link(block);
 341                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 342
 343                 /* is 1 of the blocks already attached to another block? */
 344                 if (pred_entry->next != NULL || entry->prev != NULL)
 345                         continue;
 346
 347                 /* schedule the 2 blocks behind each other */
 348                 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 349                            pred_entry->block, entry->block, edge->execfreq));
 350                 pred_entry->next = entry;
 351                 entry->prev      = pred_entry;
 352         }
 353 }
 354
 355 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 356 {
 357         ir_node            *block = entry->block;
 358         ir_node            *succ  = NULL;
 359         blocksched_entry_t *succ_entry;
 360         const ir_edge_t    *edge;
 361         double             best_succ_execfreq;
 362
 363         if (irn_visited_else_mark(block))
 364                 return;
 365
 366         env->blockcount++;
 367
 368         DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 369
 370         /* put all successors into the worklist */
 371         foreach_block_succ(block, edge) {
 372                 ir_node *succ_block = get_edge_src_irn(edge);
 373
 374                 if (irn_visited(succ_block))
 375                         continue;
 376
 377                 /* we only need to put the first of a series of already connected
 378                  * blocks into the worklist */
 379                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 380                 while (succ_entry->prev != NULL) {
 381                         /* break cycles... */
 382                         if (succ_entry->prev->block == succ_block) {
 383                                 succ_entry->prev->next = NULL;
 384                                 succ_entry->prev       = NULL;
 385                                 break;
 386                         }
 387                         succ_entry = succ_entry->prev;
 388                 }
 389
 390                 if (irn_visited(succ_entry->block))
 391                         continue;
 392
 393                 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 394                 pdeq_putr(env->worklist, succ_entry->block);
 395         }
 396
 397         if (entry->next != NULL) {
 398                 pick_block_successor(entry->next, env);
 399                 return;
 400         }
 401
 402         DB((dbg, LEVEL_1, "deciding...\n"));
 403         best_succ_execfreq = -1;
 404
 405         /* no successor yet: pick the successor block with the highest execution
 406          * frequency which has no predecessor yet */
 407
 408         foreach_block_succ(block, edge) {
 409                 ir_node *succ_block = get_edge_src_irn(edge);
 410                 double  execfreq;
 411
 412                 if (irn_visited(succ_block))
 413                         continue;
 414
 415                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 416                 if (succ_entry->prev != NULL)
 417                         continue;
 418
 419                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 420                 if (execfreq > best_succ_execfreq) {
 421                         best_succ_execfreq = execfreq;
 422                         succ = succ_block;
 423                 }
 424         }
 425
 426         if (succ == NULL) {
 427                 DB((dbg, LEVEL_1, "pick from worklist\n"));
 428
 429                 do {
 430                         if (pdeq_empty(env->worklist)) {
 431                                 DB((dbg, LEVEL_1, "worklist empty\n"));
 432                                 return;
 433                         }
 434                         succ = (ir_node*)pdeq_getl(env->worklist);
 435                 } while (irn_visited(succ));
 436         }
 437
 438         succ_entry       = (blocksched_entry_t*)get_irn_link(succ);
 439         entry->next      = succ_entry;
 440         succ_entry->prev = entry;
 441
 442         pick_block_successor(succ_entry, env);
 443 }
 444
 445 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 446 {
 447         ir_graph           *irg        = env->irg;
 448         ir_node            *startblock = get_irg_start_block(irg);
 449         blocksched_entry_t *entry      = (blocksched_entry_t*)get_irn_link(startblock);
 450
 451         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 452         inc_irg_visited(irg);
 453
 454         env->worklist = new_pdeq();
 455         pick_block_successor(entry, env);
 456         assert(pdeq_empty(env->worklist));
 457         del_pdeq(env->worklist);
 458
 459         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 460
 461         return entry;
 462 }
 463
 464 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 465                                                                                 int count, struct obstack* obst)
 466 {
 467         int                i = 0;
 468         ir_node            **block_list;
 469         blocksched_entry_t *entry;
 470         (void) env;
 471
 472         block_list = NEW_ARR_D(ir_node *, obst, count);
 473         DB((dbg, LEVEL_1, "Blockschedule:\n"));
 474
 475         for (entry = first; entry != NULL; entry = entry->next) {
 476                 assert(i < count);
 477                 block_list[i++] = entry->block;
 478                 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
 479         }
 480         assert(i == count);
 481
 482         return block_list;
 483 }
 484
 485 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 486 {
 487         blocksched_env_t   env;
 488         struct obstack     obst;
 489         blocksched_entry_t *start_entry;
 490         ir_node            **block_list;
 491
 492         obstack_init(&obst);
 493
 494         env.irg        = irg;
 495         env.obst       = &obst;
 496         env.execfreqs  = execfreqs;
 497         env.edges      = NEW_ARR_F(edge_t, 0);
 498         env.worklist   = NULL;
 499         env.blockcount = 0;
 500
 501         /* make sure loopinfo is up-to-date */
 502         if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
 503                 construct_cf_backedges(irg);
 504         }
 505
 506         // collect edge execution frequencies
 507         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 508
 509         (void)be_remove_empty_blocks(irg);
 510
 511         if (algo != BLOCKSCHED_NAIV)
 512                 coalesce_blocks(&env);
 513
 514         start_entry = finish_block_schedule(&env);
 515         block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
 516                                               be_get_be_obst(irg));
 517
 518         DEL_ARR_F(env.edges);
 519         obstack_free(&obst, NULL);
 520
 521         return block_list;
 522 }
 523
 524 /*
 525  *  ___ _     ____
 526  * |_ _| |   |  _ \
 527  *  | || |   | |_) |
 528  *  | || |___|  __/
 529  * |___|_____|_|
 530  *
 531  */
 532
 533 typedef struct ilp_edge_t {
 534         ir_node *block;   /**< source block */
 535         int     pos;      /**< number of cfg predecessor (target) */
 536         int     ilpvar;
 537 } ilp_edge_t;
 538
 539 typedef struct blocksched_ilp_env_t {
 540         blocksched_env_t env;
 541         ilp_edge_t       *ilpedges;
 542         lpp_t            *lpp;
 543 } blocksched_ilp_env_t;
 544
 545 typedef struct blocksched_ilp_entry_t {
 546         ir_node *block;
 547         struct blocksched_entry_t *next;
 548         struct blocksched_entry_t *prev;
 549
 550         int out_cst;
 551 } blocksched_ilp_entry_t;
 552
 553 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 554 {
 555         char       name[64];
 556         ilp_edge_t edge;
 557         int        edgeidx = ARR_LEN(env->ilpedges);
 558
 559         snprintf(name, sizeof(name), "edge%d", edgeidx);
 560
 561         edge.block  = block;
 562         edge.pos    = pos;
 563         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 564
 565         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 566         return edgeidx;
 567 }
 568
 569 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 570 {
 571         blocksched_ilp_env_t *env        = data;
 572         ir_graph             *irg        = env->env.irg;
 573         ir_node              *startblock = get_irg_start_block(irg);
 574         int                  arity;
 575         lpp_cst_t            cst;
 576         char                 name[64];
 577         int                  out_count;
 578         blocksched_ilp_entry_t *entry;
 579
 580         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 581         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 582
 583         entry          = OALLOC(env->env.obst, blocksched_ilp_entry_t);
 584         entry->block   = block;
 585         entry->next    = NULL;
 586         entry->prev    = NULL;
 587         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1);
 588         set_irn_link(block, entry);
 589
 590         if (block == startblock)
 591                 return;
 592
 593         arity = get_irn_arity(block);
 594         if (arity == 1) {
 595                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 596                 add_ilp_edge(block, 0, execfreq, env);
 597         }
 598         else {
 599                 int i;
 600
 601                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 602                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1);
 603
 604                 for (i = 0; i < arity; ++i) {
 605                         double     execfreq;
 606                         int        edgenum;
 607                         ilp_edge_t *edge;
 608                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 609
 610                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 611                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 612                         edge     = &env->ilpedges[edgenum];
 613                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 614                 }
 615         }
 616 }
 617
 618
 619 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 620 {
 621         int           edge_count = ARR_LEN(env->ilpedges);
 622         be_options_t *options    = be_get_irg_options(env->env.irg);
 623         int           i;
 624
 625         /* complete out constraints */
 626         for (i = 0; i < edge_count; ++i) {
 627                 const ilp_edge_t *edge  = &env->ilpedges[i];
 628                 ir_node          *block = edge->block;
 629                 ir_node          *pred;
 630                 blocksched_ilp_entry_t *entry;
 631
 632                 /* the block might have been removed already... */
 633                 if (is_Bad(get_Block_cfgpred(block, 0)))
 634                         continue;
 635
 636                 pred  = get_Block_cfgpred_block(block, edge->pos);
 637                 entry = get_irn_link(pred);
 638
 639                 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 640                                   pred, block, edge->pos));
 641                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 642         }
 643
 644         lpp_solve_net(env->lpp, options->ilp_server, options->ilp_solver);
 645         assert(lpp_is_sol_valid(env->lpp));
 646
 647         /* Apply results to edges */
 648         for (i = 0; i < edge_count; ++i) {
 649                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 650                 ir_node            *block = edge->block;
 651                 ir_node            *pred;
 652                 int                is_jump;
 653                 blocksched_entry_t *entry;
 654                 blocksched_entry_t *pred_entry;
 655
 656                 /* the block might have been removed already... */
 657                 if (is_Bad(get_Block_cfgpred(block, 0)))
 658                         continue;
 659
 660                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 661                 if (is_jump)
 662                         continue;
 663
 664                 pred       = get_Block_cfgpred_block(block, edge->pos);
 665                 entry      = get_irn_link(block);
 666                 pred_entry = get_irn_link(pred);
 667
 668                 assert(entry->prev == NULL && pred_entry->next == NULL);
 669                 entry->prev      = pred_entry;
 670                 pred_entry->next = entry;
 671         }
 672 }
 673
 674 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 675 {
 676         blocksched_ilp_env_t env;
 677         struct obstack       obst;
 678         blocksched_entry_t   *start_entry;
 679         ir_node              **block_list;
 680
 681         obstack_init(&obst);
 682
 683         env.env.irg        = irg;
 684         env.env.obst       = &obst;
 685         env.env.execfreqs  = execfreqs;
 686         env.env.worklist   = NULL;
 687         env.env.blockcount = 0;
 688         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 689
 690         env.lpp = new_lpp("blockschedule", lpp_minimize);
 691         lpp_set_time_limit(env.lpp, 20);
 692         lpp_set_log(env.lpp, stdout);
 693
 694         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 695
 696         (void)be_remove_empty_blocks(irg);
 697         coalesce_blocks_ilp(&env);
 698
 699         start_entry = finish_block_schedule(&env.env);
 700         block_list  = create_blocksched_array(&env.env, start_entry,
 701                                               env.env.blockcount,
 702                                               be_get_be_obst(irg));
 703
 704         DEL_ARR_F(env.ilpedges);
 705         free_lpp(env.lpp);
 706         obstack_free(&obst, NULL);
 707
 708         return block_list;
 709 }
 710
 711 /*
 712  *  __  __       _
 713  * |  \/  | __ _(_)_ __
 714  * | |\/| |/ _` | | '_ \
 715  * | |  | | (_| | | | | |
 716  * |_|  |_|\__,_|_|_| |_|
 717  *
 718  */
 719 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 720 void be_init_blocksched(void)
 721 {
 722         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 723
 724         lc_opt_add_table(be_grp, be_blocksched_options);
 725
 726         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 727 }
 728
 729 ir_node **be_create_block_schedule(ir_graph *irg)
 730 {
 731         ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
 732
 733         switch (algo) {
 734         case BLOCKSCHED_GREEDY:
 735         case BLOCKSCHED_NAIV:
 736                 return create_block_schedule_greedy(irg, execfreqs);
 737         case BLOCKSCHED_ILP:
 738                 return create_block_schedule_ilp(irg, execfreqs);
 739         }
 740
 741         panic("unknown blocksched algo");
 742 }