nsz Git - libfirm/blob - ir/be/bespillbelady3.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       MIN-algorithm with some twists (aka belady spiller v3)
  23  * @author      Matthias Braun
  24  * @date        15.01.2008
  25  * @version     $Id$
  26  *
  27  * TODO:
  28  *   - handle phis correctly, decide whether we should spill them ~ok
  29  *   - merge multiple start worksets of blocks ~ok
  30  *   - how to and when to do the tentative phase...
  31  */
  32 #ifdef HAVE_CONFIG_H
  33 #include "config.h"
  34 #endif
  35
  36 #include <stdbool.h>
  37
  38 #include "debug.h"
  39 #include "list.h"
  40 #include "pdeq.h"
  41
  42 #include "irnode_t.h"
  43 #include "irprintf.h"
  44 #include "iredges_t.h"
  45 #include "irloop_t.h"
  46 #include "irgwalk.h"
  47 #include "execfreq.h"
  48
  49 #include "bemodule.h"
  50 #include "bespill.h"
  51 #include "beutil.h"
  52 #include "bespilloptions.h"
  53 #include "besched_t.h"
  54 #include "be_t.h"
  55
  56 #define EXPENSIVE_CHECKS
  57
  58 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  59
  60 typedef struct loop_edge_t loop_edge_t;
  61 struct loop_edge_t {
  62         ir_node     *block;
  63         int          pos;
  64         loop_edge_t *next;
  65 };
  66
  67 typedef struct loop_info_t loop_info_t;
  68 struct loop_info_t {
  69         ir_loop     *loop;
  70         loop_edge_t *entry_edges;
  71         loop_edge_t *exit_edges;
  72         size_t       max_register_pressure;
  73 };
  74
  75 typedef struct worklist_entry_t worklist_entry_t;
  76 struct worklist_entry_t {
  77         struct list_head  head;
  78         ir_node          *value;
  79         ir_node          *reload_point;
  80         ir_loop          *unused_livethrough_loop;
  81 };
  82
  83 typedef struct worklist_t worklist_t;
  84 struct worklist_t {
  85         struct list_head  live_values;
  86         size_t            n_live_values;
  87         ir_visited_t      visited;
  88 };
  89
  90 typedef struct block_info_t block_info_t;
  91 struct block_info_t {
  92         worklist_t *start_worklist;
  93         worklist_t *end_worklist;
  94 };
  95
  96 static const arch_env_t            *arch_env;
  97 static const arch_register_class_t *cls;
  98 static struct obstack               obst;
  99 static spill_env_t                 *senv;
 100 static size_t                       n_regs;
 101 static size_t                       max_register_pressure;
 102 static bool                         tentative_mode;
 103 static bool                         should_have_reached_fixpoint;
 104 static bool                         do_push_unused_livethroughs;
 105 static ir_exec_freq                *exec_freq;
 106 static ir_visited_t                 worklist_visited;
 107
 108 static worklist_t *new_worklist(void)
 109 {
 110         worklist_t *worklist = obstack_alloc(&obst, sizeof(worklist[0]));
 111         memset(worklist, 0, sizeof(worklist[0]));
 112
 113         INIT_LIST_HEAD(&worklist->live_values);
 114         worklist->n_live_values    = 0;
 115
 116         return worklist;
 117 }
 118
 119 static void mark_irn_not_visited(ir_node *node)
 120 {
 121         set_irn_visited(node, get_irg_visited(current_ir_graph) - 1);
 122 }
 123
 124 static bool worklist_contains(const ir_node *node)
 125 {
 126         return irn_visited(node);
 127 }
 128
 129 static block_info_t *get_block_info(ir_node *block)
 130 {
 131         block_info_t *info = get_irn_link(block);
 132         if (info != NULL)
 133                 return info;
 134
 135         info = obstack_alloc(&obst, sizeof(info[0]));
 136         memset(info, 0, sizeof(info[0]));
 137         set_irn_link(block, info);
 138         return info;
 139 }
 140
 141 static void deactivate_worklist(const worklist_t *worklist)
 142 {
 143         struct list_head *entry;
 144
 145         list_for_each(entry, &worklist->live_values) {
 146                 worklist_entry_t *wl_entry
 147                         = list_entry(entry, worklist_entry_t, head);
 148                 assert(worklist_contains(wl_entry->value));
 149                 mark_irn_not_visited(wl_entry->value);
 150                 set_irn_link(wl_entry->value, NULL);
 151         }
 152 }
 153
 154 static void activate_worklist(const worklist_t *worklist)
 155 {
 156         struct list_head *entry;
 157
 158         list_for_each(entry, &worklist->live_values) {
 159                 worklist_entry_t *wl_entry
 160                         = list_entry(entry, worklist_entry_t, head);
 161                 assert(!worklist_contains(wl_entry->value));
 162                 mark_irn_visited(wl_entry->value);
 163                 set_irn_link(wl_entry->value, wl_entry);
 164         }
 165 }
 166
 167 /**
 168  * duplicate a worklist and directly activates it
 169  */
 170 static void fill_and_activate_worklist(worklist_t *new_worklist,
 171                 const worklist_t *worklist, ir_node *block, ir_node *succ_block,
 172                 int succ_pos)
 173 {
 174         ir_node          *reload_point  = NULL;
 175         size_t            n_live_values = 0;
 176         struct list_head *entry;
 177
 178         if (succ_block != NULL &&
 179                         (get_Block_n_cfgpreds(succ_block) > 1
 180                          || get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) > 1)) {
 181                 reload_point = be_get_end_of_block_insertion_point(block);
 182         }
 183
 184         list_for_each(entry, &worklist->live_values) {
 185                 worklist_entry_t *wl_entry  = list_entry(entry, worklist_entry_t, head);
 186                 ir_node          *value     = wl_entry->value;
 187                 worklist_entry_t *new_entry;
 188
 189                 if (new_worklist->n_live_values >= n_regs)
 190                         break;
 191
 192                 if (is_Phi(value) && get_nodes_block(value) == succ_block) {
 193                         value = get_Phi_pred(value, succ_pos);
 194
 195                         /* can happen for unknown phi preds */
 196                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, value))
 197                                 continue;
 198                 }
 199
 200                 if (irn_visited(value))
 201                         continue;
 202
 203                 new_entry = obstack_alloc(&obst, sizeof(new_entry[0]));
 204                 memset(new_entry, 0, sizeof(new_entry[0]));
 205
 206                 new_entry->value = value;
 207                 if (reload_point != NULL) {
 208                         new_entry->reload_point = reload_point;
 209                 } else {
 210                         new_entry->reload_point = wl_entry->reload_point;
 211                 }
 212
 213                 list_add_tail(&new_entry->head, &new_worklist->live_values);
 214                 ++n_live_values;
 215
 216                 mark_irn_visited(value);
 217                 set_irn_link(value, new_entry);
 218                 new_worklist->n_live_values++;
 219         }
 220 }
 221
 222 static worklist_t *duplicate_worklist(const worklist_t *worklist)
 223 {
 224         worklist_t       *new_worklist;
 225         struct list_head *entry;
 226
 227         new_worklist = obstack_alloc(&obst, sizeof(new_worklist[0]));
 228         memset(new_worklist, 0, sizeof(new_worklist[0]));
 229         INIT_LIST_HEAD(&new_worklist->live_values);
 230         new_worklist->n_live_values = worklist->n_live_values;
 231
 232         list_for_each(entry, &worklist->live_values) {
 233                 worklist_entry_t *wl_entry  = list_entry(entry, worklist_entry_t, head);
 234                 worklist_entry_t *new_entry
 235                         = obstack_alloc(&obst, sizeof(new_entry[0]));
 236
 237                 memcpy(new_entry, wl_entry, sizeof(new_entry[0]));
 238                 list_add_tail(&new_entry->head, &new_worklist->live_values);
 239         }
 240
 241         return new_worklist;
 242 }
 243
 244 #ifdef DEBUG_libfirm
 245 static void print_worklist(const worklist_t *worklist, int level)
 246 {
 247         struct list_head *entry;
 248
 249         DB((dbg, level, "%d values: ", worklist->n_live_values));
 250         list_for_each(entry, &worklist->live_values) {
 251                 worklist_entry_t *wl_entry
 252                         = list_entry(entry, worklist_entry_t, head);
 253
 254                 DB((dbg, level, "%+F ", wl_entry->value));
 255         }
 256 }
 257 #endif
 258
 259
 260
 261 static void clear_loop_info(ir_loop *loop)
 262 {
 263         int n_elements = get_loop_n_elements(loop);
 264         int i;
 265
 266         loop->link = NULL;
 267         for (i = 0; i < n_elements; ++i) {
 268                 loop_element element = get_loop_element(loop, i);
 269                 if (*element.kind != k_ir_loop)
 270                         continue;
 271
 272                 clear_loop_info(element.son);
 273         }
 274 }
 275
 276 static loop_info_t *get_loop_info(ir_loop *loop)
 277 {
 278         loop_info_t *info = loop->link;
 279         if (info != NULL)
 280                 return info;
 281
 282         info = obstack_alloc(&obst, sizeof(info[0]));
 283         memset(info, 0, sizeof(info[0]));
 284         info->loop = loop;
 285         loop->link = info;
 286         return info;
 287 }
 288
 289 /**
 290  * constructs loop in+out edges when called in a block walker
 291  */
 292 static void construct_loop_edges(ir_node *block, void *data)
 293 {
 294         int      n_cfgpreds = get_Block_n_cfgpreds(block);
 295         ir_loop *loop       = get_irn_loop(block);
 296         int      i;
 297
 298         (void) data;
 299
 300         for (i = 0; i < n_cfgpreds; ++i) {
 301                 ir_node     *cfgpred_block = get_Block_cfgpred_block(block, i);
 302                 ir_loop     *cfgpred_loop  = get_irn_loop(cfgpred_block);
 303                 loop_edge_t *edge;
 304                 bool        is_exit_edge;
 305                 ir_loop     *l, *goal;
 306
 307                 if (cfgpred_loop == loop)
 308                         continue;
 309
 310                 /* critical edges are splitted, so we can't jump from 1 loop
 311                  * directly into another without going out of it */
 312                 assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
 313
 314                 /* edge out of loop */
 315                 if (get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
 316                         is_exit_edge = true;
 317                         l            = cfgpred_loop;
 318                         goal         = loop;
 319                 } else {
 320                         is_exit_edge = false;
 321                         l            = loop;
 322                         goal         = cfgpred_loop;
 323                 }
 324
 325                 /* this might be a jump out/in multiple loops */
 326                 do {
 327                         loop_info_t *l_info = get_loop_info(l);
 328
 329                         edge = obstack_alloc(&obst, sizeof(edge[0]));
 330                         memset(edge, 0, sizeof(edge[0]));
 331                         edge->block = block;
 332                         edge->pos   = i;
 333
 334                         if (is_exit_edge) {
 335                                 edge->next         = l_info->exit_edges;
 336                                 l_info->exit_edges = edge;
 337                                 assert(get_loop_depth(loop) < get_loop_depth(l));
 338                         } else {
 339                                 edge->next          = l_info->entry_edges;
 340                                 l_info->entry_edges = edge;
 341                                 assert(get_loop_depth(cfgpred_loop) < get_loop_depth(l));
 342                         }
 343                         l = get_loop_outer_loop(l);
 344                 } while(l != goal);
 345         }
 346 }
 347
 348
 349
 350
 351 static void place_reload(worklist_entry_t *entry)
 352 {
 353         if (tentative_mode)
 354                 return;
 355
 356         DB((dbg, LEVEL_1, "reload of %+F before %+F\n", entry->value,
 357                 entry->reload_point));
 358         assert(entry->reload_point != NULL);
 359         be_add_reload(senv, entry->value, entry->reload_point, cls, 1);
 360         entry->reload_point = NULL;
 361 }
 362
 363 /**
 364  * makes sure the worklist contains not more than n_regs - room_needed entries
 365  */
 366 static void make_room(worklist_t *worklist, size_t room_needed)
 367 {
 368         int               i;
 369         int               spills_needed;
 370         struct list_head *entry;
 371
 372         spills_needed = worklist->n_live_values + room_needed - n_regs;
 373         if (spills_needed <= 0)
 374                 return;
 375
 376         entry = worklist->live_values.next;
 377         for(i = spills_needed; i > 0; --i) {
 378                 struct list_head *next = entry->next;
 379                 worklist_entry_t *wl_entry
 380                         = list_entry(entry, worklist_entry_t, head);
 381
 382                 assert(worklist_contains(wl_entry->value));
 383                 mark_irn_not_visited(wl_entry->value);
 384                 place_reload(wl_entry);
 385                 list_del(entry);
 386
 387                 entry = next;
 388         }
 389         worklist->n_live_values -= spills_needed;
 390 }
 391
 392 /**
 393  * a value was used, so bring it to the back of the worklist (which might
 394  * result in a spill of another value).
 395  */
 396 static void val_used(worklist_t *worklist, ir_node *value, ir_node *sched_point)
 397 {
 398         /* already in the worklist? move around, otherwise add at back */
 399         worklist_entry_t *entry = get_irn_link(value);
 400
 401         assert(arch_irn_consider_in_reg_alloc(arch_env, cls, value));
 402
 403         if (worklist_contains(value)) {
 404                 assert(entry != NULL);
 405
 406                 list_del(&entry->head);
 407         } else {
 408                 if (entry == NULL) {
 409                         entry        = obstack_alloc(&obst, sizeof(entry[0]));
 410                         memset(entry, 0, sizeof(entry[0]));
 411
 412                         entry->value = value;
 413                         set_irn_link(value, entry);
 414                 }
 415
 416                 ++worklist->n_live_values;
 417                 mark_irn_visited(value);
 418         }
 419
 420         entry->reload_point = sched_point;
 421         list_add_tail(&entry->head, &worklist->live_values);
 422 }
 423
 424 static void worklist_remove(worklist_t *worklist, ir_node *value)
 425 {
 426         worklist_entry_t *entry = get_irn_link(value);
 427         assert(entry != NULL);
 428         list_del(&entry->head);
 429         --worklist->n_live_values;
 430
 431         assert(worklist_contains(value));
 432         mark_irn_not_visited(value);
 433 }
 434
 435 static void update_max_pressure(worklist_t *worklist)
 436 {
 437         if (worklist->n_live_values > max_register_pressure)
 438                 max_register_pressure = worklist->n_live_values;
 439 }
 440
 441 static void do_spilling(ir_node *block, worklist_t *worklist)
 442 {
 443         ir_node *node;
 444
 445         assert(worklist != NULL);
 446
 447         sched_foreach_reverse(block, node) {
 448                 int    i, arity;
 449                 size_t n_defs = 0;
 450
 451                 DB((dbg, LEVEL_2, "\t%+F... ", node));
 452                 update_max_pressure(worklist);
 453
 454                 if (is_Phi(node)) {
 455                         ir_node *node2;
 456                         /* TODO: if we have some free registers, then we could decide to
 457                          * not spill some phis (but not for phis where at least 1 input is
 458                          * themselfes) */
 459
 460                         /* we have to spill all phis that are not live */
 461                         sched_foreach_reverse_from(node, node2) {
 462                                 assert(is_Phi(node2));
 463
 464                                 if (worklist_contains(node2))
 465                                         continue;
 466                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node2))
 467                                         continue;
 468
 469                                 if (!tentative_mode)
 470                                         be_spill_phi(senv, node2);
 471                         }
 472                         DB((dbg, LEVEL_2, "\n"));
 473                         break;
 474                 }
 475
 476                 /* remove values defined by this instruction from the workset. Values
 477                  * defined but not in the workset need free registers */
 478                 if (get_irn_mode(node) == mode_T) {
 479                         const ir_edge_t *edge;
 480
 481                         foreach_out_edge(node, edge) {
 482                                 ir_node *proj = get_edge_src_irn(edge);
 483                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj))
 484                                         continue;
 485                                 if (worklist_contains(proj)) {
 486                                         worklist_remove(worklist, proj);
 487                                 } else {
 488                                         ++n_defs;
 489                                 }
 490                         }
 491                 } else if (arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
 492                         if (worklist_contains(node)) {
 493                                 worklist_remove(worklist, node);
 494                         } else {
 495                                 n_defs = 1;
 496                         }
 497                 }
 498
 499                 /* make sure we have enough free registers for the spills */
 500                 make_room(worklist, n_defs);
 501
 502                 /* put all values used by the instruction into the workset */
 503                 arity = get_irn_arity(node);
 504                 for(i = 0; i < arity; ++i) {
 505                         ir_node *use = get_irn_n(node, i);
 506
 507                         if (!arch_irn_consider_in_reg_alloc(arch_env, cls, use))
 508                                 continue;
 509
 510                         val_used(worklist, use, node);
 511                 }
 512
 513                 /* we might have too many values in the worklist now and need to spill
 514                  * some */
 515                 make_room(worklist, 0);
 516
 517 #ifdef DEBUG_libfirm
 518                 print_worklist(worklist, LEVEL_2);
 519                 DB((dbg, LEVEL_2, "\n"));
 520 #endif
 521         }
 522
 523         update_max_pressure(worklist);
 524
 525 #ifdef DEBUG_libfirm
 526         DB((dbg, LEVEL_1, "worklist at begin of %+F:", block));
 527         print_worklist(worklist, LEVEL_1);
 528         DB((dbg, LEVEL_1, "\n"));
 529 #endif
 530 }
 531
 532 static bool worklists_equal(const worklist_t *wl1, const worklist_t *wl2)
 533 {
 534         const struct list_head *i1 = &wl1->live_values;
 535         const struct list_head *i2 = &wl2->live_values;
 536
 537         for ( ; i1 != &wl1->live_values && i2 != &wl2->live_values;
 538                         i1 = i1->next, i2 = i2->next) {
 539                 worklist_entry_t *entry1 = list_entry(i1, worklist_entry_t, head);
 540                 worklist_entry_t *entry2 = list_entry(i2, worklist_entry_t, head);
 541
 542                 if (entry1->value != entry2->value)
 543                         return false;
 544         }
 545         /* both lists at the end */
 546         if (i1 != &wl1->live_values || i2 != &wl2->live_values)
 547                 return false;
 548
 549         return true;
 550 }
 551
 552 static bool fill_start_worklist(worklist_t *new_worklist, ir_node *block)
 553 {
 554         double           best_execfreq   = -1;
 555         worklist_t      *best_worklist   = NULL;
 556         ir_node         *best_succ_block = NULL;
 557         int              best_pos;
 558         const ir_edge_t *edge;
 559
 560         /* construct worklist */
 561         foreach_block_succ(block, edge) {
 562                 ir_node      *succ_block = get_edge_src_irn(edge);
 563                 double       execfreq    = get_block_execfreq(exec_freq, succ_block);
 564                 block_info_t *block_info;
 565                 worklist_t   *succ_worklist;
 566
 567                 if (execfreq < best_execfreq)
 568                         continue;
 569
 570                 block_info    = get_block_info(succ_block);
 571                 succ_worklist = block_info->start_worklist;
 572
 573                 if (succ_worklist == NULL || succ_worklist->visited >= worklist_visited)
 574                         continue;
 575
 576                 best_execfreq   = execfreq;
 577                 best_worklist   = succ_worklist;
 578                 best_succ_block = succ_block;
 579                 best_pos        = get_edge_src_pos(edge);
 580         }
 581
 582         if (best_worklist == NULL)
 583                 return false;
 584         best_worklist->visited = worklist_visited;
 585
 586         fill_and_activate_worklist(new_worklist, best_worklist, block,
 587                         best_succ_block, best_pos);
 588         return true;
 589 }
 590
 591 static worklist_t *construct_start_worklist(ir_node *block)
 592 {
 593         worklist_t *worklist = new_worklist();
 594
 595         ++worklist_visited;
 596
 597         while(fill_start_worklist(worklist, block)) {
 598                 if (worklist->n_live_values >= n_regs)
 599                         break;
 600         }
 601
 602         return worklist;
 603 }
 604
 605 static void process_block(ir_node *block, void *env)
 606 {
 607         block_info_t    *block_info;
 608         worklist_t      *worklist;
 609         int              n_preds;
 610
 611         (void) env;
 612         DB((dbg, LEVEL_1, "Processing %+F\n", block));
 613
 614         worklist   = construct_start_worklist(block);
 615         block_info = get_block_info(block);
 616
 617 #ifdef DEBUG_libfirm
 618         DB((dbg, LEVEL_1, "worklist at end of %+F:", block));
 619         print_worklist(worklist, LEVEL_1);
 620         DB((dbg, LEVEL_1, "\n"));
 621
 622         if (should_have_reached_fixpoint) {
 623                 assert(worklists_equal(block_info->end_worklist, worklist));
 624         }
 625 #endif
 626         block_info->end_worklist = duplicate_worklist(worklist);
 627
 628         do_spilling(block, worklist);
 629         deactivate_worklist(worklist);
 630
 631 #ifdef DEBUG_libfirm
 632         if (should_have_reached_fixpoint) {
 633                 assert(worklists_equal(block_info->start_worklist, worklist));
 634         }
 635 #endif
 636         block_info->start_worklist = worklist;
 637
 638         /* we shouldn't have any live values left at the start block */
 639         n_preds = get_Block_n_cfgpreds(block);
 640         //assert(n_preds != 0 || worklist->n_live_values == 0);
 641 }
 642
 643 typedef struct block_or_loop_t block_or_loop_t;
 644 struct block_or_loop_t {
 645         union {
 646                 ir_node *block;
 647                 ir_loop *loop;
 648         } v;
 649         bool is_loop;
 650 };
 651
 652 static block_or_loop_t *loop_blocks;
 653 static ir_loop         *current_loop;
 654
 655 static void find_blocks(ir_node *block);
 656
 657 static void find_in_loop(ir_loop *loop, ir_node *entry)
 658 {
 659         loop_info_t     *loop_info = get_loop_info(loop);
 660         block_or_loop_t block_or_loop;
 661         loop_edge_t     *edge;
 662
 663         /* simply mark 1 block in the loop to indicate that the loop was already
 664          * processed */
 665         ir_node *some_block = loop_info->entry_edges->block;
 666         if (Block_block_visited(some_block))
 667                 return;
 668
 669         block_or_loop.v.loop  = loop;
 670         block_or_loop.is_loop = true;
 671         ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
 672
 673 #ifndef NDEBUG
 674         {
 675                 /* we should be 1 of the entry blocks */
 676                 loop_edge_t *edge  = loop_info->entry_edges;
 677                 bool         found = false;
 678                 for ( ; edge != NULL; edge = edge->next) {
 679                         if (edge->block == entry)
 680                                 found = true;
 681                 }
 682                 assert(found);
 683         }
 684 #endif
 685         /* check all loop successors */
 686         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 687                 ir_node *succ      = edge->block;
 688                 ir_loop *succ_loop = get_irn_loop(succ);
 689
 690                 if (succ_loop == current_loop) {
 691                         find_blocks(succ);
 692                 } else {
 693                         assert(get_loop_depth(succ_loop) < get_loop_depth(current_loop));
 694                 }
 695         }
 696 }
 697
 698 static void find_blocks(ir_node *block)
 699 {
 700         const ir_edge_t *edge;
 701         block_or_loop_t block_or_loop;
 702
 703         if (Block_block_visited(block))
 704                 return;
 705
 706         block_or_loop.v.block = block;
 707         block_or_loop.is_loop = false;
 708
 709         ARR_APP1(block_or_loop_t, loop_blocks, block_or_loop);
 710         mark_Block_block_visited(block);
 711
 712         foreach_block_succ(block, edge) {
 713                 ir_node *succ = get_edge_src_irn(edge);
 714
 715                 /* is it part of the current loop? */
 716                 ir_loop *loop = get_irn_loop(succ);
 717                 if (loop != current_loop) {
 718                         /* a sub-loop? */
 719                         if (get_loop_depth(loop) > get_loop_depth(current_loop)) {
 720                                 find_in_loop(loop, succ);
 721                         } else {
 722                                 /* parent loop: we're not interested in the block */
 723                         }
 724                 } else {
 725                         find_blocks(succ);
 726                 }
 727         }
 728 }
 729
 730 /**
 731  * append an entry to a worklist. WARNING: The entry must not already be in the
 732  * worklist.
 733  */
 734 static void worklist_append(worklist_t *worklist, ir_node *value,
 735                             ir_node *reload_point,
 736                                                         ir_loop *unused_livethrough_loop)
 737 {
 738         worklist_entry_t *entry     = obstack_alloc(&obst, sizeof(entry[0]));
 739         memset(entry, 0, sizeof(entry[0]));
 740
 741 #ifdef EXPENSIVE_CHECKS
 742         {
 743                 struct list_head *entry;
 744                 list_for_each(entry, &worklist->live_values) {
 745                         worklist_entry_t *wl_entry
 746                                 = list_entry(entry, worklist_entry_t, head);
 747                         assert(wl_entry->value != value);
 748                 }
 749         }
 750 #endif
 751
 752         entry->value                   = value;
 753         entry->reload_point            = reload_point;
 754         entry->unused_livethrough_loop = unused_livethrough_loop;
 755         list_add_tail(&entry->head, &worklist->live_values);
 756         ++worklist->n_live_values;
 757         assert(worklist->n_live_values <= n_regs);
 758 }
 759
 760 static void push_unused_livethrough(loop_info_t *loop_info, ir_node *value)
 761 {
 762         loop_edge_t *edge;
 763         ++worklist_visited;
 764
 765         /* add the value to all loop exit and entry blocks */
 766         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 767                 ir_node            *block
 768                         = get_Block_cfgpred_block(edge->block, edge->pos);
 769                 const block_info_t *info     = get_block_info(block);
 770                 worklist_t         *worklist = info->end_worklist;
 771                 ir_node            *reload_point = NULL;
 772
 773                 if (worklist->visited >= worklist_visited)
 774                         continue;
 775                 worklist->visited = worklist_visited;
 776
 777                 /* TODO: we need a smarter mechanism here, that makes the reloader place
 778                  * reload nodes on all loop exits... */
 779
 780                 worklist_append(worklist, value, reload_point, loop_info->loop);
 781         }
 782         edge = loop_info->entry_edges;
 783         for ( ; edge != NULL; edge = edge->next) {
 784                 ir_node            *entry_block = edge->block;
 785                 const block_info_t *info        = get_block_info(entry_block);
 786                 worklist_t         *worklist    = info->start_worklist;
 787                 ir_node            *pred_block;
 788                 ir_node            *reload_point;
 789
 790                 if (worklist->visited >= worklist_visited)
 791                         continue;
 792                 worklist->visited = worklist_visited;
 793
 794                 pred_block   = get_Block_cfgpred_block(entry_block, edge->pos);
 795                 reload_point = be_get_end_of_block_insertion_point(pred_block);
 796
 797                 worklist_append(worklist, value, reload_point, loop_info->loop);
 798         }
 799
 800         set_irn_link(value, NULL);
 801         ++loop_info->max_register_pressure;
 802 }
 803
 804 static void push_unused_livethroughs(loop_info_t *loop_info)
 805 {
 806         loop_edge_t *edge;
 807
 808         /* we can only push unused livethroughs if register pressure inside the loop
 809          * was low enough */
 810         if (loop_info->max_register_pressure >= n_regs)
 811                 return;
 812
 813         /* find unused livethroughs: register pressure in the loop was low enough
 814          * which means that we had no spills which implies that at every point in
 815          * the loop all*/
 816         for (edge = loop_info->exit_edges; edge != NULL; edge = edge->next) {
 817                 ir_node            *block          = edge->block;
 818                 const block_info_t *info           = get_block_info(block);
 819                 worklist_t         *start_worklist = info->start_worklist;
 820                 ir_node            *exit_block;
 821                 const block_info_t *exit_info;
 822                 worklist_t         *end_worklist;
 823                 struct list_head   *entry;
 824
 825                 if (start_worklist == NULL)
 826                         continue;
 827
 828                 exit_block   = get_Block_cfgpred_block(edge->block, edge->pos);
 829                 exit_info    = get_block_info(exit_block);
 830                 end_worklist = exit_info->end_worklist;
 831
 832                 activate_worklist(end_worklist);
 833                 /* all values contained in the start_worklist, which are not available
 834                  * in the end_worklist, must be unused livethroughs */
 835
 836                 list_for_each(entry, &start_worklist->live_values) {
 837                         worklist_entry_t *wl_entry
 838                                 = list_entry(entry, worklist_entry_t, head);
 839                         ir_node *value = wl_entry->value;
 840
 841                         if (loop_info->max_register_pressure >= n_regs)
 842                                 break;
 843
 844
 845                         if (!worklist_contains(value)) {
 846                                 /* add it to all loop exits */
 847                                 DB((dbg, LEVEL_2, "Found unused livethrough: %+F (regpressure in loop %d)\n", value, loop_info->max_register_pressure));
 848                                 push_unused_livethrough(loop_info, value);
 849                                 /* the value should now be part of end_worklist, so mark it */
 850                                 mark_irn_visited(value);
 851                         }
 852                 }
 853
 854                 deactivate_worklist(end_worklist);
 855         }
 856 }
 857
 858 static void process_block_or_loop(const block_or_loop_t *block_or_loop)
 859 {
 860         if (block_or_loop->is_loop) {
 861                 loop_info_t *loop_info = get_loop_info(block_or_loop->v.loop);
 862
 863                 if (do_push_unused_livethroughs)
 864                         push_unused_livethroughs(loop_info);
 865
 866                 if (loop_info->max_register_pressure > max_register_pressure)
 867                         max_register_pressure = loop_info->max_register_pressure;
 868
 869                 return;
 870         }
 871         process_block(block_or_loop->v.block, NULL);
 872 }
 873
 874 static void process_loop(ir_loop *loop)
 875 {
 876         int         n_elements = get_loop_n_elements(loop);
 877         int         i, len;
 878         loop_info_t *loop_info;
 879         loop_edge_t *edge;
 880         ir_node     *some_block;
 881
 882         /* first handle all sub-loops */
 883         for (i = 0; i < n_elements; ++i) {
 884                 loop_element element = get_loop_element(loop, i);
 885                 if (*element.kind != k_ir_loop)
 886                         continue;
 887
 888                 process_loop(element.son);
 889         }
 890
 891         /* create a postorder of the blocks */
 892         loop_info = get_loop_info(loop);
 893         edge      = loop_info->entry_edges;
 894         if (edge != NULL) {
 895                 some_block = edge->block;
 896         } else {
 897                 assert(loop == get_irg_loop(current_ir_graph));
 898                 some_block = get_irg_start_block(current_ir_graph);
 899         }
 900
 901         loop_blocks  = NEW_ARR_F(block_or_loop_t,0);
 902         current_loop = loop;
 903
 904         ir_reserve_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
 905         inc_irg_block_visited(current_ir_graph);
 906         find_blocks(some_block);
 907         /* for endless loops the end-block might be unreachable */
 908         if (loop == get_irg_loop(current_ir_graph)) {
 909                 find_blocks(get_irg_end_block(current_ir_graph));
 910         }
 911         ir_free_resources(current_ir_graph, IR_RESOURCE_BLOCK_VISITED);
 912
 913         DB((dbg, LEVEL_3, "Block List for loop %p\n", loop));
 914         len = ARR_LEN(loop_blocks);
 915         for (i = 0; i < len; ++i) {
 916                 block_or_loop_t *block_or_loop = &loop_blocks[i];
 917                 if (block_or_loop->is_loop) {
 918                         DB((dbg, LEVEL_3, " L-%p", block_or_loop->v.loop));
 919                 } else {
 920                         DB((dbg, LEVEL_3, " B-%+F", block_or_loop->v.block));
 921                 }
 922         }
 923         DB((dbg, LEVEL_3, "\n"));
 924
 925         max_register_pressure = 0;
 926
 927         /* run1: tentative phase */
 928         tentative_mode = true;
 929         for (i = len-1; i >= 0; --i) {
 930                 process_block_or_loop(&loop_blocks[i]);
 931         }
 932
 933         /* run2: tentative phase - should reach fixpoint */
 934         tentative_mode = true;
 935         for (i = len-1; i >= 0; --i) {
 936                 process_block_or_loop(&loop_blocks[i]);
 937         }
 938
 939 #ifndef NDEBUG
 940         /* run3: tentative phase - check fixpoint */
 941         tentative_mode               = true;
 942         should_have_reached_fixpoint = true;
 943         for (i = len-1; i >= 0; --i) {
 944                 process_block_or_loop(&loop_blocks[i]);
 945         }
 946         should_have_reached_fixpoint = false;
 947 #endif
 948
 949         /* run4: add spills/reloads */
 950         tentative_mode              = false;
 951         do_push_unused_livethroughs = true;
 952         for (i = len-1; i >= 0; --i) {
 953                 process_block_or_loop(&loop_blocks[i]);
 954         }
 955         do_push_unused_livethroughs = false;
 956
 957         loop_info->max_register_pressure = max_register_pressure;
 958         DB((dbg, LEVEL_2, "Regpressure in loop %p: %u\n", loop,
 959                                 (unsigned) max_register_pressure));
 960
 961         DEL_ARR_F(loop_blocks);
 962 }
 963
 964 static void fix_block_borders(ir_node *block, void *data)
 965 {
 966         block_info_t *block_info     = get_block_info(block);
 967         worklist_t   *start_worklist = block_info->start_worklist;
 968         int           n_cfgpreds     = get_Block_n_cfgpreds(block);
 969         ir_loop      *loop           = get_irn_loop(block);
 970         int           i;
 971
 972         (void) data;
 973         assert(start_worklist != NULL);
 974
 975         for (i = 0; i < n_cfgpreds; ++i) {
 976                 ir_node      *pred_block      = get_Block_cfgpred_block(block, i);
 977                 block_info_t *pred_block_info = get_block_info(pred_block);
 978                 worklist_t   *end_worklist    = pred_block_info->end_worklist;
 979                 ir_loop      *pred_loop       = get_irn_loop(pred_block);
 980                 bool          is_loop_entry   = false;
 981                 struct list_head *entry;
 982
 983                 assert(end_worklist != NULL);
 984
 985                 if (get_loop_depth(pred_loop) < get_loop_depth(loop)) {
 986                         is_loop_entry = true;
 987                 }
 988
 989                 /* reload missing values */
 990                 activate_worklist(end_worklist);
 991
 992                 list_for_each(entry, &start_worklist->live_values) {
 993                         worklist_entry_t *wl_entry
 994                                 = list_entry(entry, worklist_entry_t, head);
 995                         ir_node          *value = wl_entry->value;
 996
 997                         if (is_Phi(value) && get_nodes_block(value) == block) {
 998                                 value = get_irn_n(value, i);
 999
1000                                 /* we might have unknowns as argument for the phi */
1001                                 if (!arch_irn_consider_in_reg_alloc(arch_env, cls, value))
1002                                         continue;
1003                         }
1004
1005                         if (worklist_contains(value))
1006                                 continue;
1007                         if (wl_entry->unused_livethrough_loop != NULL && !is_loop_entry)
1008                                 continue;
1009
1010                         be_add_reload_on_edge(senv, value, block, i, cls, 1);
1011                 }
1012
1013                 deactivate_worklist(end_worklist);
1014         }
1015 }
1016
1017 static void be_spill_belady3(be_irg_t *birg, const arch_register_class_t *ncls)
1018 {
1019         ir_graph *irg = be_get_birg_irg(birg);
1020
1021         cls    = ncls;
1022         n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
1023
1024         /* shortcut for register classes with ignore regs only */
1025         if (n_regs == 0)
1026                 return;
1027
1028         worklist_visited = 0;
1029         arch_env         = be_get_birg_arch_env(birg);
1030         exec_freq        = be_get_birg_exec_freq(birg);
1031
1032         be_clear_links(irg);
1033         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1034                         | IR_RESOURCE_LOOP_LINK);
1035         inc_irg_visited(irg);
1036
1037         obstack_init(&obst);
1038         senv = be_new_spill_env(birg);
1039
1040         assure_cf_loop(irg);
1041         clear_loop_info(get_irg_loop(irg));
1042         irg_block_walk_graph(irg, construct_loop_edges, NULL, NULL);
1043
1044         process_loop(get_irg_loop(current_ir_graph));
1045
1046         irg_block_walk_graph(irg, fix_block_borders, NULL, NULL);
1047
1048         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK
1049                         | IR_RESOURCE_LOOP_LINK);
1050
1051         be_insert_spills_reloads(senv);
1052
1053         obstack_free(&obst, NULL);
1054
1055         /* clean up */
1056         be_delete_spill_env(senv);
1057 }
1058
1059 void be_init_spillbelady3(void)
1060 {
1061         static be_spiller_t belady3_spiller = {
1062                 be_spill_belady3
1063         };
1064
1065         be_register_spiller("belady3", &belady3_spiller);
1066         FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady3");
1067 }
1068
1069 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady3);