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 static blocksched_entry_t* get_blocksched_entry(const ir_node *block)
 131 {
 132         return (blocksched_entry_t*)get_irn_link(block);
 133 }
 134
 135 /**
 136  * Collect cfg frequencies of all edges between blocks.
 137  * Also determines edge with highest frequency.
 138  */
 139 static void collect_egde_frequency(ir_node *block, void *data)
 140 {
 141         blocksched_env_t   *env = (blocksched_env_t*)data;
 142         int                arity;
 143         edge_t             edge;
 144         blocksched_entry_t *entry;
 145         ir_loop            *loop;
 146
 147         memset(&edge, 0, sizeof(edge));
 148
 149         entry = OALLOCZ(env->obst, blocksched_entry_t);
 150         entry->block = block;
 151         set_irn_link(block, entry);
 152
 153         loop = get_irn_loop(block);
 154
 155         arity = get_Block_n_cfgpreds(block);
 156
 157         if (arity == 0) {
 158                 /* must be the start block (or end-block for endless loops),
 159                  * everything else is dead code and should be removed by now */
 160                 assert(block == get_irg_start_block(env->irg)
 161                                 || block == get_irg_end_block(env->irg));
 162                 /* nothing to do here */
 163                 return;
 164         } else if (arity == 1) {
 165                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 166                 ir_loop *pred_loop  = get_irn_loop(pred_block);
 167                 float    freq       = (float)get_block_execfreq(env->execfreqs, block);
 168
 169                 /* is it an edge leaving a loop */
 170                 if (get_loop_depth(pred_loop) > get_loop_depth(loop)) {
 171                         float pred_freq = (float)get_block_execfreq(env->execfreqs, pred_block);
 172                         edge.outedge_penalty_freq = -(pred_freq - freq);
 173                 }
 174
 175                 edge.block            = block;
 176                 edge.pos              = 0;
 177                 edge.execfreq         = freq;
 178                 edge.highest_execfreq = 1;
 179                 ARR_APP1(edge_t, env->edges, edge);
 180         } else {
 181                 int    i;
 182                 double highest_execfreq = -1.0;
 183                 int    highest_edge_num = -1;
 184
 185                 edge.block = block;
 186                 for (i = 0; i < arity; ++i) {
 187                         double  execfreq;
 188                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 189
 190                         execfreq = get_block_execfreq(env->execfreqs, pred_block);
 191
 192                         edge.pos              = i;
 193                         edge.execfreq         = execfreq;
 194                         edge.highest_execfreq = 0;
 195                         ARR_APP1(edge_t, env->edges, edge);
 196
 197                         if (execfreq > highest_execfreq) {
 198                                 highest_execfreq = execfreq;
 199                                 highest_edge_num = ARR_LEN(env->edges) - 1;
 200                         }
 201                 }
 202
 203                 if (highest_edge_num >= 0)
 204                         env->edges[highest_edge_num].highest_execfreq = 1;
 205         }
 206 }
 207
 208 static int cmp_edges_base(const edge_t *e1, const edge_t *e2)
 209 {
 210         long nr1 = get_irn_node_nr(e1->block);
 211         long nr2 = get_irn_node_nr(e2->block);
 212         if (nr1 < nr2) {
 213                 return 1;
 214         } else if (nr1 > nr2) {
 215                 return -1;
 216         } else {
 217                 if (e1->pos < e2->pos) {
 218                         return 1;
 219                 } else if (e1->pos > e2->pos) {
 220                         return -1;
 221                 } else {
 222                         return 0;
 223                 }
 224         }
 225 }
 226
 227 static int cmp_edges(const void *d1, const void *d2)
 228 {
 229         const edge_t *e1 = (const edge_t*)d1;
 230         const edge_t *e2 = (const edge_t*)d2;
 231         double        freq1 = e1->execfreq;
 232         double        freq2 = e2->execfreq;
 233         if (freq1 < freq2) {
 234                 return 1;
 235         } else if (freq1 > freq2) {
 236                 return -1;
 237         } else {
 238                 return cmp_edges_base(e1, e2);
 239         }
 240 }
 241
 242 static int cmp_edges_outedge_penalty(const void *d1, const void *d2)
 243 {
 244         const edge_t *e1   = (const edge_t*)d1;
 245         const edge_t *e2   = (const edge_t*)d2;
 246         double        pen1 = e1->outedge_penalty_freq;
 247         double        pen2 = e2->outedge_penalty_freq;
 248         if (pen1 > pen2) {
 249                 return 1;
 250         } else if (pen1 < pen2) {
 251                 return -1;
 252         } else {
 253                 return cmp_edges_base(e1, e2);
 254         }
 255 }
 256
 257 static void clear_loop_links(ir_loop *loop)
 258 {
 259         int i, n;
 260
 261         set_loop_link(loop, NULL);
 262         n = get_loop_n_elements(loop);
 263         for (i = 0; i < n; ++i) {
 264                 loop_element elem = get_loop_element(loop, i);
 265                 if (*elem.kind == k_ir_loop) {
 266                         clear_loop_links(elem.son);
 267                 }
 268         }
 269 }
 270
 271 static void coalesce_blocks(blocksched_env_t *env)
 272 {
 273         int i;
 274         int edge_count = ARR_LEN(env->edges);
 275         edge_t *edges = env->edges;
 276
 277         /* sort interblock edges by execution frequency */
 278         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 279
 280         /* run1: only look at jumps */
 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
 288                 /* only check edge with highest frequency */
 289                 if (! edge->highest_execfreq)
 290                         continue;
 291
 292                 /* the block might have been removed already... */
 293                 if (is_Bad(get_Block_cfgpred(block, 0)))
 294                         continue;
 295
 296                 pred_block = get_Block_cfgpred_block(block, pos);
 297                 entry      = get_blocksched_entry(block);
 298                 pred_entry = get_blocksched_entry(pred_block);
 299
 300                 if (pred_entry->next != NULL || entry->prev != NULL)
 301                         continue;
 302
 303                 /* only coalesce jumps */
 304                 if (get_block_succ_next(pred_block, get_block_succ_first(pred_block)) != NULL)
 305                         continue;
 306
 307                 /* schedule the 2 blocks behind each other */
 308                 DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n",
 309                            pred_entry->block, entry->block, edge->execfreq));
 310                 pred_entry->next = entry;
 311                 entry->prev      = pred_entry;
 312         }
 313
 314         /* run2: pick loop fallthroughs */
 315         clear_loop_links(get_irg_loop(env->irg));
 316
 317         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty);
 318         for (i = 0; i < edge_count; ++i) {
 319                 const edge_t *edge  = &edges[i];
 320                 ir_node      *block = edge->block;
 321                 int           pos   = edge->pos;
 322                 ir_node      *pred_block;
 323                 blocksched_entry_t *entry, *pred_entry;
 324                 ir_loop      *loop;
 325                 ir_loop      *outer_loop;
 326
 327                 /* already seen all loop outedges? */
 328                 if (edge->outedge_penalty_freq == 0)
 329                         break;
 330
 331                 /* the block might have been removed already... */
 332                 if (is_Bad(get_Block_cfgpred(block, pos)))
 333                         continue;
 334
 335                 pred_block = get_Block_cfgpred_block(block, pos);
 336                 entry      = get_blocksched_entry(block);
 337                 pred_entry = get_blocksched_entry(pred_block);
 338
 339                 if (pred_entry->next != NULL || entry->prev != NULL)
 340                         continue;
 341
 342                 /* we want at most 1 outedge fallthrough per loop */
 343                 loop = get_irn_loop(pred_block);
 344                 if (get_loop_link(loop) != NULL)
 345                         continue;
 346
 347                 /* schedule the 2 blocks behind each other */
 348                 DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n",
 349                            pred_entry->block, entry->block, edge->execfreq));
 350                 pred_entry->next = entry;
 351                 entry->prev      = pred_entry;
 352
 353                 /* all loops left have an outedge now */
 354                 outer_loop = get_irn_loop(block);
 355                 do {
 356                         /* we set loop link to loop to mark it */
 357                         set_loop_link(loop, loop);
 358                         loop = get_loop_outer_loop(loop);
 359                 } while (loop != outer_loop);
 360         }
 361
 362         /* sort interblock edges by execution frequency */
 363         qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges);
 364
 365         /* run3: remaining edges */
 366         for (i = 0; i < edge_count; ++i) {
 367                 const edge_t *edge  = &edges[i];
 368                 ir_node      *block = edge->block;
 369                 int           pos   = edge->pos;
 370                 ir_node      *pred_block;
 371                 blocksched_entry_t *entry, *pred_entry;
 372
 373                 /* the block might have been removed already... */
 374                 if (is_Bad(get_Block_cfgpred(block, pos)))
 375                         continue;
 376
 377                 pred_block = get_Block_cfgpred_block(block, pos);
 378                 entry      = get_blocksched_entry(block);
 379                 pred_entry = get_blocksched_entry(pred_block);
 380
 381                 /* is 1 of the blocks already attached to another block? */
 382                 if (pred_entry->next != NULL || entry->prev != NULL)
 383                         continue;
 384
 385                 /* schedule the 2 blocks behind each other */
 386                 DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n",
 387                            pred_entry->block, entry->block, edge->execfreq));
 388                 pred_entry->next = entry;
 389                 entry->prev      = pred_entry;
 390         }
 391 }
 392
 393 static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *env)
 394 {
 395         ir_node            *block = entry->block;
 396         ir_node            *succ  = NULL;
 397         blocksched_entry_t *succ_entry;
 398         double              best_succ_execfreq;
 399
 400         if (irn_visited_else_mark(block))
 401                 return;
 402
 403         env->blockcount++;
 404
 405         DB((dbg, LEVEL_1, "Pick succ of %+F\n", block));
 406
 407         /* put all successors into the worklist */
 408         foreach_block_succ(block, edge) {
 409                 ir_node *succ_block = get_edge_src_irn(edge);
 410
 411                 if (irn_visited(succ_block))
 412                         continue;
 413
 414                 /* we only need to put the first of a series of already connected
 415                  * blocks into the worklist */
 416                 succ_entry = get_blocksched_entry(succ_block);
 417                 while (succ_entry->prev != NULL) {
 418                         /* break cycles... */
 419                         if (succ_entry->prev->block == succ_block) {
 420                                 succ_entry->prev->next = NULL;
 421                                 succ_entry->prev       = NULL;
 422                                 break;
 423                         }
 424                         succ_entry = succ_entry->prev;
 425                 }
 426
 427                 if (irn_visited(succ_entry->block))
 428                         continue;
 429
 430                 DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block));
 431                 pdeq_putr(env->worklist, succ_entry->block);
 432         }
 433
 434         if (entry->next != NULL) {
 435                 pick_block_successor(entry->next, env);
 436                 return;
 437         }
 438
 439         DB((dbg, LEVEL_1, "deciding...\n"));
 440         best_succ_execfreq = -1;
 441
 442         /* no successor yet: pick the successor block with the highest execution
 443          * frequency which has no predecessor yet */
 444
 445         foreach_block_succ(block, edge) {
 446                 ir_node *succ_block = get_edge_src_irn(edge);
 447                 double  execfreq;
 448
 449                 if (irn_visited(succ_block))
 450                         continue;
 451
 452                 succ_entry = get_blocksched_entry(succ_block);
 453                 if (succ_entry->prev != NULL)
 454                         continue;
 455
 456                 execfreq = get_block_execfreq(env->execfreqs, succ_block);
 457                 if (execfreq > best_succ_execfreq) {
 458                         best_succ_execfreq = execfreq;
 459                         succ = succ_block;
 460                 }
 461         }
 462
 463         if (succ == NULL) {
 464                 DB((dbg, LEVEL_1, "pick from worklist\n"));
 465
 466                 do {
 467                         if (pdeq_empty(env->worklist)) {
 468                                 DB((dbg, LEVEL_1, "worklist empty\n"));
 469                                 return;
 470                         }
 471                         succ = (ir_node*)pdeq_getl(env->worklist);
 472                 } while (irn_visited(succ));
 473         }
 474
 475         succ_entry       = get_blocksched_entry(succ);
 476         entry->next      = succ_entry;
 477         succ_entry->prev = entry;
 478
 479         pick_block_successor(succ_entry, env);
 480 }
 481
 482 static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
 483 {
 484         ir_graph           *irg        = env->irg;
 485         ir_node            *startblock = get_irg_start_block(irg);
 486         blocksched_entry_t *entry      = get_blocksched_entry(startblock);
 487
 488         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 489         inc_irg_visited(irg);
 490
 491         env->worklist = new_pdeq();
 492         pick_block_successor(entry, env);
 493         assert(pdeq_empty(env->worklist));
 494         del_pdeq(env->worklist);
 495
 496         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 497
 498         return entry;
 499 }
 500
 501 static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry_t *first,
 502                                                                                 int count, struct obstack* obst)
 503 {
 504         int                i = 0;
 505         ir_node            **block_list;
 506         blocksched_entry_t *entry;
 507         (void) env;
 508
 509         block_list = NEW_ARR_D(ir_node *, obst, count);
 510         DB((dbg, LEVEL_1, "Blockschedule:\n"));
 511
 512         for (entry = first; entry != NULL; entry = entry->next) {
 513                 assert(i < count);
 514                 block_list[i++] = entry->block;
 515                 DB((dbg, LEVEL_1, "\t%+F\n", entry->block));
 516         }
 517         assert(i == count);
 518
 519         return block_list;
 520 }
 521
 522 static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs)
 523 {
 524         blocksched_env_t   env;
 525         struct obstack     obst;
 526         blocksched_entry_t *start_entry;
 527         ir_node            **block_list;
 528
 529         obstack_init(&obst);
 530
 531         env.irg        = irg;
 532         env.obst       = &obst;
 533         env.execfreqs  = execfreqs;
 534         env.edges      = NEW_ARR_F(edge_t, 0);
 535         env.worklist   = NULL;
 536         env.blockcount = 0;
 537
 538         assure_loopinfo(irg);
 539
 540         // collect edge execution frequencies
 541         irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env);
 542
 543         (void)be_remove_empty_blocks(irg);
 544
 545         if (algo != BLOCKSCHED_NAIV)
 546                 coalesce_blocks(&env);
 547
 548         start_entry = finish_block_schedule(&env);
 549         block_list  = create_blocksched_array(&env, start_entry, env.blockcount,
 550                                               be_get_be_obst(irg));
 551
 552         DEL_ARR_F(env.edges);
 553         obstack_free(&obst, NULL);
 554
 555         return block_list;
 556 }
 557
 558 /*
 559  *  ___ _     ____
 560  * |_ _| |   |  _ \
 561  *  | || |   | |_) |
 562  *  | || |___|  __/
 563  * |___|_____|_|
 564  *
 565  */
 566
 567 typedef struct ilp_edge_t {
 568         ir_node *block;   /**< source block */
 569         int     pos;      /**< number of cfg predecessor (target) */
 570         int     ilpvar;
 571 } ilp_edge_t;
 572
 573 typedef struct blocksched_ilp_env_t {
 574         blocksched_env_t env;
 575         ilp_edge_t       *ilpedges;
 576         lpp_t            *lpp;
 577 } blocksched_ilp_env_t;
 578
 579 typedef struct blocksched_ilp_entry_t {
 580         ir_node *block;
 581         struct blocksched_entry_t *next;
 582         struct blocksched_entry_t *prev;
 583
 584         int out_cst;
 585 } blocksched_ilp_entry_t;
 586
 587 static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp_env_t *env)
 588 {
 589         char       name[64];
 590         ilp_edge_t edge;
 591         int        edgeidx = ARR_LEN(env->ilpedges);
 592
 593         snprintf(name, sizeof(name), "edge%d", edgeidx);
 594
 595         edge.block  = block;
 596         edge.pos    = pos;
 597         edge.ilpvar = lpp_add_var_default(env->lpp, name, lpp_binary, execfreq, 1.0);
 598
 599         ARR_APP1(ilp_edge_t, env->ilpedges, edge);
 600         return edgeidx;
 601 }
 602
 603 static void collect_egde_frequency_ilp(ir_node *block, void *data)
 604 {
 605         blocksched_ilp_env_t *env        = (blocksched_ilp_env_t*)data;
 606         ir_graph             *irg        = env->env.irg;
 607         ir_node              *startblock = get_irg_start_block(irg);
 608         int                  arity;
 609         char                 name[64];
 610         int                  out_count;
 611         blocksched_ilp_entry_t *entry;
 612
 613         snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block));
 614         out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK);
 615
 616         entry          = OALLOC(env->env.obst, blocksched_ilp_entry_t);
 617         entry->block   = block;
 618         entry->next    = NULL;
 619         entry->prev    = NULL;
 620         entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1);
 621         set_irn_link(block, entry);
 622
 623         if (block == startblock)
 624                 return;
 625
 626         arity = get_irn_arity(block);
 627         if (arity == 1) {
 628                 double execfreq = get_block_execfreq(env->env.execfreqs, block);
 629                 add_ilp_edge(block, 0, execfreq, env);
 630         }
 631         else {
 632                 int i;
 633                 int cst_idx;
 634
 635                 snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block));
 636                 cst_idx = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1);
 637
 638                 for (i = 0; i < arity; ++i) {
 639                         double     execfreq;
 640                         int        edgenum;
 641                         ilp_edge_t *edge;
 642                         ir_node    *pred_block = get_Block_cfgpred_block(block, i);
 643
 644                         execfreq = get_block_execfreq(env->env.execfreqs, pred_block);
 645                         edgenum  = add_ilp_edge(block, i, execfreq, env);
 646                         edge     = &env->ilpedges[edgenum];
 647                         lpp_set_factor_fast(env->lpp, cst_idx, edge->ilpvar, 1.0);
 648                 }
 649         }
 650 }
 651
 652 static blocksched_ilp_entry_t *get_blocksched_ilp_entry(const ir_node *block)
 653 {
 654         return (blocksched_ilp_entry_t*)get_irn_link(block);
 655 }
 656
 657 static void coalesce_blocks_ilp(blocksched_ilp_env_t *env)
 658 {
 659         int           edge_count = ARR_LEN(env->ilpedges);
 660         int           i;
 661
 662         /* complete out constraints */
 663         for (i = 0; i < edge_count; ++i) {
 664                 const ilp_edge_t *edge  = &env->ilpedges[i];
 665                 ir_node          *block = edge->block;
 666                 ir_node          *pred;
 667                 blocksched_ilp_entry_t *entry;
 668
 669                 /* the block might have been removed already... */
 670                 if (is_Bad(get_Block_cfgpred(block, 0)))
 671                         continue;
 672
 673                 pred  = get_Block_cfgpred_block(block, edge->pos);
 674                 entry = get_blocksched_ilp_entry(pred);
 675
 676                 DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n",
 677                                   pred, block, edge->pos));
 678                 lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0);
 679         }
 680
 681         lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver);
 682         assert(lpp_is_sol_valid(env->lpp));
 683
 684         /* Apply results to edges */
 685         for (i = 0; i < edge_count; ++i) {
 686                 const ilp_edge_t   *edge  = &env->ilpedges[i];
 687                 ir_node            *block = edge->block;
 688                 ir_node            *pred;
 689                 int                is_jump;
 690                 blocksched_entry_t *entry;
 691                 blocksched_entry_t *pred_entry;
 692
 693                 /* the block might have been removed already... */
 694                 if (is_Bad(get_Block_cfgpred(block, 0)))
 695                         continue;
 696
 697                 is_jump = (int)lpp_get_var_sol(env->lpp, edge->ilpvar);
 698                 if (is_jump)
 699                         continue;
 700
 701                 pred       = get_Block_cfgpred_block(block, edge->pos);
 702                 entry      = get_blocksched_entry(block);
 703                 pred_entry = get_blocksched_entry(pred);
 704
 705                 assert(entry->prev == NULL && pred_entry->next == NULL);
 706                 entry->prev      = pred_entry;
 707                 pred_entry->next = entry;
 708         }
 709 }
 710
 711 static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs)
 712 {
 713         blocksched_ilp_env_t env;
 714         struct obstack       obst;
 715         blocksched_entry_t   *start_entry;
 716         ir_node              **block_list;
 717
 718         obstack_init(&obst);
 719
 720         env.env.irg        = irg;
 721         env.env.obst       = &obst;
 722         env.env.execfreqs  = execfreqs;
 723         env.env.worklist   = NULL;
 724         env.env.blockcount = 0;
 725         env.ilpedges       = NEW_ARR_F(ilp_edge_t, 0);
 726
 727         env.lpp = lpp_new("blockschedule", lpp_minimize);
 728         lpp_set_time_limit(env.lpp, 20);
 729         lpp_set_log(env.lpp, stdout);
 730
 731         irg_block_walk_graph(irg, collect_egde_frequency_ilp, NULL, &env);
 732
 733         (void)be_remove_empty_blocks(irg);
 734         coalesce_blocks_ilp(&env);
 735
 736         start_entry = finish_block_schedule(&env.env);
 737         block_list  = create_blocksched_array(&env.env, start_entry,
 738                                               env.env.blockcount,
 739                                               be_get_be_obst(irg));
 740
 741         DEL_ARR_F(env.ilpedges);
 742         lpp_free(env.lpp);
 743         obstack_free(&obst, NULL);
 744
 745         return block_list;
 746 }
 747
 748 /*
 749  *  __  __       _
 750  * |  \/  | __ _(_)_ __
 751  * | |\/| |/ _` | | '_ \
 752  * | |  | | (_| | | | | |
 753  * |_|  |_|\__,_|_|_| |_|
 754  *
 755  */
 756 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched)
 757 void be_init_blocksched(void)
 758 {
 759         lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
 760
 761         lc_opt_add_table(be_grp, be_blocksched_options);
 762
 763         FIRM_DBG_REGISTER(dbg, "firm.be.blocksched");
 764 }
 765
 766 ir_node **be_create_block_schedule(ir_graph *irg)
 767 {
 768         ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg);
 769
 770         switch (algo) {
 771         case BLOCKSCHED_GREEDY:
 772         case BLOCKSCHED_NAIV:
 773                 return create_block_schedule_greedy(irg, execfreqs);
 774         case BLOCKSCHED_ILP:
 775                 return create_block_schedule_ilp(irg, execfreqs);
 776         }
 777
 778         panic("unknown blocksched algo");
 779 }