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