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
  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         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_base(const edge_t *e1, const edge_t *e2)
 204 {
 205         long nr1 = get_irn_node_nr(e1->block);
 206         long nr2 = get_irn_node_nr(e2->block);
 207         if (nr1 < nr2) {
 208                 return 1;
 209         } else if (nr1 > nr2) {
 210                 return -1;
 211         } else {
 212                 if (e1->pos < e2->pos) {
 213                         return 1;
 214                 } else if (e1->pos > e2->pos) {
 215                         return -1;
 216                 } else {
 217                         return 0;
 218                 }
 219         }
 220 }
 221
 222 static int cmp_edges(const void *d1, const void *d2)
 223 {
 224         const edge_t *e1 = (const edge_t*)d1;
 225         const edge_t *e2 = (const edge_t*)d2;
 226         double        freq1 = e1->execfreq;
 227         double        freq2 = e2->execfreq;
 228         if (freq1 < freq2) {
 229                 return 1;
 230         } else if (freq1 > freq2) {
 231                 return -1;
 232         } else {
 233                 return cmp_edges_base(e1, e2);
 234         }
 235 }
 236
 237 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 238 {
 239         const edge_t *e1   = (const edge_t*)d1;
 240         const edge_t *e2   = (const edge_t*)d2;
 241         double        pen1 = e1->outedge_penalty_freq;
 242         double        pen2 = e2->outedge_penalty_freq;
 243         if (pen1 > pen2) {
 244                 return 1;
 245         } else if (pen1 < pen2) {
 246                 return -1;
 247         } else {
 248                 return cmp_edges_base(e1, e2);
 249         }
 250 }
 251
 252 static void clear_loop_links(ir_loop *loop)
 253 {
 254         int i, n;
 255
 256         set_loop_link(loop, NULL);
 257         n = get_loop_n_elements(loop);
 258         for (i = 0; i < n; ++i) {
 259                 loop_element elem = get_loop_element(loop, i);
 260                 if (*elem.kind == k_ir_loop) {
 261                         clear_loop_links(elem.son);
 262                 }
 263         }
 264 }
 265
 266 static void coalesce_blocks(blocksched_env_t *env)
 267 {
 268         int i;
 269         int edge_count = ARR_LEN(env->edges);
 270         edge_t *edges = env->edges;
 271
 272         /* sort interblock edges by execution frequency */
 273         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 274
 275         /* run1: only look at jumps */
 276         for (i = 0; i < edge_count; ++i) {
 277                 const edge_t *edge  = &edges[i];
 278                 ir_node      *block = edge->block;
 279                 int           pos   = edge->pos;
 280                 ir_node      *pred_block;
 281                 blocksched_entry_t *entry, *pred_entry;
 282
 283                 /* only check edge with highest frequency */
 284                 if (! edge->highest_execfreq)
 285                         continue;
 286
 287                 /* the block might have been removed already... */
 288                 if (is_Bad(get_Block_cfgpred(block, 0)))
 289                         continue;
 290
 291                 pred_block = get_Block_cfgpred_block(block, pos);
 292                 entry      = (blocksched_entry_t*)get_irn_link(block);
 293                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 294
 295                 if (pred_entry->next != NULL || entry->prev != NULL)
 296                         continue;
 297
 298                 /* only coalesce jumps */
 299                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 300                         continue;
 301
 302                 /* schedule the 2 blocks behind each other */
 303                 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 304                            pred_entry->block, entry->block, edge->execfreq));
 305                 pred_entry->next = entry;
 306                 entry->prev      = pred_entry;
 307         }
 308
 309         /* run2: pick loop fallthroughs */
 310         clear_loop_links(get_irg_loop(env->irg));
 311
 312         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
 313         for (i = 0; i < edge_count; ++i) {
 314                 const edge_t *edge  = &edges[i];
 315                 ir_node      *block = edge->block;
 316                 int           pos   = edge->pos;
 317                 ir_node      *pred_block;
 318                 blocksched_entry_t *entry, *pred_entry;
 319                 ir_loop      *loop;
 320                 ir_loop      *outer_loop;
 321
 322                 /* already seen all loop outedges? */
 323                 if (edge->outedge_penalty_freq == 0)
 324                         break;
 325
 326                 /* the block might have been removed already... */
 327                 if (is_Bad(get_Block_cfgpred(block, pos)))
 328                         continue;
 329
 330                 pred_block = get_Block_cfgpred_block(block, pos);
 331                 entry      = (blocksched_entry_t*)get_irn_link(block);
 332                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 333
 334                 if (pred_entry->next != NULL || entry->prev != NULL)
 335                         continue;
 336
 337                 /* we want at most 1 outedge fallthrough per loop */
 338                 loop = get_irn_loop(pred_block);
 339                 if (get_loop_link(loop) != NULL)
 340                         continue;
 341
 342                 /* schedule the 2 blocks behind each other */
 343                 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
 344                            pred_entry->block, entry->block, edge->execfreq));
 345                 pred_entry->next = entry;
 346                 entry->prev      = pred_entry;
 347
 348                 /* all loops left have an outedge now */
 349                 outer_loop = get_irn_loop(block);
 350                 do {
 351                         /* we set loop link to loop to mark it */
 352                         set_loop_link(loop, loop);
 353                         loop = get_loop_outer_loop(loop);
 354                 } while (loop != outer_loop);
 355         }
 356
 357         /* sort interblock edges by execution frequency */
 358         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 359
 360         /* run3: remaining edges */
 361         for (i = 0; i < edge_count; ++i) {
 362                 const edge_t *edge  = &edges[i];
 363                 ir_node      *block = edge->block;
 364                 int           pos   = edge->pos;
 365                 ir_node      *pred_block;
 366                 blocksched_entry_t *entry, *pred_entry;
 367
 368                 /* the block might have been removed already... */
 369                 if (is_Bad(get_Block_cfgpred(block, pos)))
 370                         continue;
 371
 372                 pred_block = get_Block_cfgpred_block(block, pos);
 373                 entry      = (blocksched_entry_t*)get_irn_link(block);
 374                 pred_entry = (blocksched_entry_t*)get_irn_link(pred_block);
 375
 376                 /* is 1 of the blocks already attached to another block? */
 377                 if (pred_entry->next != NULL || entry->prev != NULL)
 378                         continue;
 379
 380                 /* schedule the 2 blocks behind each other */
 381                 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 382                            pred_entry->block, entry->block, edge->execfreq));
 383                 pred_entry->next = entry;
 384                 entry->prev      = pred_entry;
 385         }
 386 }
 387
 388 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 389 {
 390         ir_node            *block = entry->block;
 391         ir_node            *succ  = NULL;
 392         blocksched_entry_t *succ_entry;
 393         const ir_edge_t    *edge;
 394         double             best_succ_execfreq;
 395
 396         if (irn_visited_else_mark(block))
 397                 return;
 398
 399         env->blockcount++;
 400
 401         DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 402
 403         /* put all successors into the worklist */
 404         foreach_block_succ(block, edge) {
 405                 ir_node *succ_block = get_edge_src_irn(edge);
 406
 407                 if (irn_visited(succ_block))
 408                         continue;
 409
 410                 /* we only need to put the first of a series of already connected
 411                  * blocks into the worklist */
 412                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 413                 while (succ_entry->prev != NULL) {
 414                         /* break cycles... */
 415                         if (succ_entry->prev->block == succ_block) {
 416                                 succ_entry->prev->next = NULL;
 417                                 succ_entry->prev       = NULL;
 418                                 break;
 419                         }
 420                         succ_entry = succ_entry->prev;
 421                 }
 422
 423                 if (irn_visited(succ_entry->block))
 424                         continue;
 425
 426                 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 427                 pdeq_putr(env->worklist, succ_entry->block);
 428         }
 429
 430         if (entry->next != NULL) {
 431                 pick_block_successor(entry->next, env);
 432                 return;
 433         }
 434
 435         DB((dbg, LEVEL_1, "deciding...\n"));
 436         best_succ_execfreq = -1;
 437
 438         /* no successor yet: pick the successor block with the highest execution
 439          * frequency which has no predecessor yet */
 440
 441         foreach_block_succ(block, edge) {
 442                 ir_node *succ_block = get_edge_src_irn(edge);
 443                 double  execfreq;
 444
 445                 if (irn_visited(succ_block))
 446                         continue;
 447
 448                 succ_entry = (blocksched_entry_t*)get_irn_link(succ_block);
 449                 if (succ_entry->prev != NULL)
 450                         continue;
 451
 452                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 453                 if (execfreq > best_succ_execfreq) {
 454                         best_succ_execfreq = execfreq;
 455                         succ = succ_block;
 456                 }
 457         }
 458
 459         if (succ == NULL) {
 460                 DB((dbg, LEVEL_1, "pick from worklist\n"));
 461
 462                 do {
 463                         if (pdeq_empty(env->worklist)) {
 464                                 DB((dbg, LEVEL_1, "worklist empty\n"));
 465                                 return;
 466                         }
 467                         succ = (ir_node*)pdeq_getl(env->worklist);
 468                 } while (irn_visited(succ));
 469         }
 470
 471         succ_entry       = (blocksched_entry_t*)get_irn_link(succ);
 472         entry->next      = succ_entry;
 473         succ_entry->prev = entry;
 474
 475         pick_block_successor(succ_entry, env);
 476 }
 477
 478 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 479 {
 480         ir_graph           *irg        = env->irg;
 481         ir_node            *startblock = get_irg_start_block(irg);
 482         blocksched_entry_t *entry      = (blocksched_entry_t*)get_irn_link(startblock);
 483
 484         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 485         inc_irg_visited(irg);
 486
 487         env->worklist = new_pdeq();
 488         pick_block_successor(entry, env);
 489         assert(pdeq_empty(env->worklist));
 490         del_pdeq(env->worklist);
 491
 492         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 493
 494         return entry;
 495 }
 496
 497 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 498                                                                                 int count, struct obstack* obst)
 499 {
 500         int                i = 0;
 501         ir_node            **block_list;
 502         blocksched_entry_t *entry;
 503         (void) env;
 504
 505         block_list = NEW_ARR_D(ir_node *, obst, count);
 506         DB((dbg, LEVEL_1, "Blockschedule:\n"));
 507
 508         for (entry = first; entry != NULL; entry = entry->next) {
 509                 assert(i < count);
 510                 block_list[i++] = entry->block;
 511                 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
 512         }
 513         assert(i == count);
 514
 515         return block_list;
 516 }
 517
 518 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 519 {
 520         blocksched_env_t   env;
 521         struct obstack     obst;
 522         blocksched_entry_t *start_entry;
 523         ir_node            **block_list;
 524
 525         obstack_init(&obst);
 526
 527         env.irg        = irg;
 528         env.obst       = &obst;
 529         env.execfreqs  = execfreqs;
 530         env.edges      = NEW_ARR_F(edge_t, 0);
 531         env.worklist   = NULL;
 532         env.blockcount = 0;
 533
 534         assure_loopinfo(irg);
 535
 536         // collect edge execution frequencies
 537         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 538
 539         (void)be_remove_empty_blocks(irg);
 540
 541         if (algo != BLOCKSCHED_NAIV)
 542                 coalesce_blocks(&env);
 543
 544         start_entry = finish_block_schedule(&env);
 545         block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
 546                                               be_get_be_obst(irg));
 547
 548         DEL_ARR_F(env.edges);
 549         obstack_free(&obst, NULL);
 550
 551         return block_list;
 552 }
 553
 554 /*
 555  *  ___ _     ____
 556  * |_ _| |   |  _ \
 557  *  | || |   | |_) |
 558  *  | || |___|  __/
 559  * |___|_____|_|
 560  *
 561  */
 562
 563 typedef struct ilp_edge_t {
 564         ir_node *block;   /**< source block */
 565         int     pos;      /**< number of cfg predecessor (target) */
 566         int     ilpvar;
 567 } ilp_edge_t;
 568
 569 typedef struct blocksched_ilp_env_t {
 570         blocksched_env_t env;
 571         ilp_edge_t       *ilpedges;
 572         lpp_t            *lpp;
 573 } blocksched_ilp_env_t;
 574
 575 typedef struct blocksched_ilp_entry_t {
 576         ir_node *block;
 577         struct blocksched_entry_t *next;
 578         struct blocksched_entry_t *prev;
 579
 580         int out_cst;
 581 } blocksched_ilp_entry_t;
 582
 583 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 584 {
 585         char       name[64];
 586         ilp_edge_t edge;
 587         int        edgeidx = ARR_LEN(env->ilpedges);
 588
 589         snprintf(name, sizeof(name), "edge%d", edgeidx);
 590
 591         edge.block  = block;
 592         edge.pos    = pos;
 593         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 594
 595         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 596         return edgeidx;
 597 }
 598
 599 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 600 {
 601         blocksched_ilp_env_t *env        = data;
 602         ir_graph             *irg        = env->env.irg;
 603         ir_node              *startblock = get_irg_start_block(irg);
 604         int                  arity;
 605         lpp_cst_t            cst;
 606         char                 name[64];
 607         int                  out_count;
 608         blocksched_ilp_entry_t *entry;
 609
 610         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 611         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 612
 613         entry          = OALLOC(env->env.obst, blocksched_ilp_entry_t);
 614         entry->block   = block;
 615         entry->next    = NULL;
 616         entry->prev    = NULL;
 617         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
 618         set_irn_link(block, entry);
 619
 620         if (block == startblock)
 621                 return;
 622
 623         arity = get_irn_arity(block);
 624         if (arity == 1) {
 625                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 626                 add_ilp_edge(block, 0, execfreq, env);
 627         }
 628         else {
 629                 int i;
 630
 631                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 632                 cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
 633
 634                 for (i = 0; i < arity; ++i) {
 635                         double     execfreq;
 636                         int        edgenum;
 637                         ilp_edge_t *edge;
 638                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 639
 640                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 641                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 642                         edge     = &env->ilpedges[edgenum];
 643                         lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0);
 644                 }
 645         }
 646 }
 647
 648
 649 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 650 {
 651         int           edge_count = ARR_LEN(env->ilpedges);
 652         int           i;
 653
 654         /* complete out constraints */
 655         for (i = 0; i < edge_count; ++i) {
 656                 const ilp_edge_t *edge  = &env->ilpedges[i];
 657                 ir_node          *block = edge->block;
 658                 ir_node          *pred;
 659                 blocksched_ilp_entry_t *entry;
 660
 661                 /* the block might have been removed already... */
 662                 if (is_Bad(get_Block_cfgpred(block, 0)))
 663                         continue;
 664
 665                 pred  = get_Block_cfgpred_block(block, edge->pos);
 666                 entry = get_irn_link(pred);
 667
 668                 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 669                                   pred, block, edge->pos));
 670                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 671         }
 672
 673         lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
 674         assert(lpp_is_sol_valid(env->lpp));
 675
 676         /* Apply results to edges */
 677         for (i = 0; i < edge_count; ++i) {
 678                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 679                 ir_node            *block = edge->block;
 680                 ir_node            *pred;
 681                 int                is_jump;
 682                 blocksched_entry_t *entry;
 683                 blocksched_entry_t *pred_entry;
 684
 685                 /* the block might have been removed already... */
 686                 if (is_Bad(get_Block_cfgpred(block, 0)))
 687                         continue;
 688
 689                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 690                 if (is_jump)
 691                         continue;
 692
 693                 pred       = get_Block_cfgpred_block(block, edge->pos);
 694                 entry      = get_irn_link(block);
 695                 pred_entry = get_irn_link(pred);
 696
 697                 assert(entry->prev == NULL && pred_entry->next == NULL);
 698                 entry->prev      = pred_entry;
 699                 pred_entry->next = entry;
 700         }
 701 }
 702
 703 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 704 {
 705         blocksched_ilp_env_t env;
 706         struct obstack       obst;
 707         blocksched_entry_t   *start_entry;
 708         ir_node              **block_list;
 709
 710         obstack_init(&obst);
 711
 712         env.env.irg        = irg;
 713         env.env.obst       = &obst;
 714         env.env.execfreqs  = execfreqs;
 715         env.env.worklist   = NULL;
 716         env.env.blockcount = 0;
 717         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 718
 719         env.lpp = lpp_new("blockschedule", lpp_minimize);
 720         lpp_set_time_limit(env.lpp, 20);
 721         lpp_set_log(env.lpp, stdout);
 722
 723         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 724
 725         (void)be_remove_empty_blocks(irg);
 726         coalesce_blocks_ilp(&env);
 727
 728         start_entry = finish_block_schedule(&env.env);
 729         block_list  = create_blocksched_array(&env.env, start_entry,
 730                                               env.env.blockcount,
 731                                               be_get_be_obst(irg));
 732
 733         DEL_ARR_F(env.ilpedges);
 734         lpp_free(env.lpp);
 735         obstack_free(&obst, NULL);
 736
 737         return block_list;
 738 }
 739
 740 /*
 741  *  __  __       _
 742  * |  \/  | __ _(_)_ __
 743  * | |\/| |/ _` | | '_ \
 744  * | |  | | (_| | | | | |
 745  * |_|  |_|\__,_|_|_| |_|
 746  *
 747  */
 748 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 749 void be_init_blocksched(void)
 750 {
 751         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 752
 753         lc_opt_add_table(be_grp, be_blocksched_options);
 754
 755         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 756 }
 757
 758 ir_node **be_create_block_schedule(ir_graph *irg)
 759 {
 760         ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
 761
 762         switch (algo) {
 763         case BLOCKSCHED_GREEDY:
 764         case BLOCKSCHED_NAIV:
 765                 return create_block_schedule_greedy(irg, execfreqs);
 766         case BLOCKSCHED_ILP:
 767                 return create_block_schedule_ilp(irg, execfreqs);
 768         }
 769
 770         panic("unknown blocksched algo");
 771 }