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