nsz Git - libfirm/blob - ir/be/bespill.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       implementation of the spill/reload placement abstraction layer
  23  * @author      Daniel Grund, Sebastian Hack, Matthias Braun
  24  * @date                29.09.2005
  25  * @version     $Id$
  26  */
  27 #include "config.h"
  28
  29 #include <stdlib.h>
  30 #include <stdbool.h>
  31
  32 #include "pset.h"
  33 #include "irnode_t.h"
  34 #include "ircons_t.h"
  35 #include "iredges_t.h"
  36 #include "irbackedge_t.h"
  37 #include "irprintf.h"
  38 #include "ident_t.h"
  39 #include "type_t.h"
  40 #include "entity_t.h"
  41 #include "debug.h"
  42 #include "irgwalk.h"
  43 #include "array.h"
  44 #include "pdeq.h"
  45 #include "execfreq.h"
  46 #include "irnodeset.h"
  47 #include "error.h"
  48
  49 #include "bearch_t.h"
  50 #include "belive_t.h"
  51 #include "besched_t.h"
  52 #include "bespill.h"
  53 #include "belive_t.h"
  54 #include "benode_t.h"
  55 #include "bechordal_t.h"
  56 #include "bespilloptions.h"
  57 #include "bestatevent.h"
  58 #include "bessaconstr.h"
  59 #include "beirg_t.h"
  60 #include "beintlive_t.h"
  61 #include "bemodule.h"
  62 #include "be_t.h"
  63
  64 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  65
  66 #define REMAT_COST_INFINITE  1000
  67
  68 typedef struct reloader_t reloader_t;
  69 struct reloader_t {
  70         reloader_t *next;
  71         ir_node    *can_spill_after;
  72         ir_node    *reloader;
  73         ir_node    *rematted_node;
  74         int         remat_cost_delta; /** costs needed for rematerialization,
  75                                            compared to placing a reload */
  76 };
  77
  78 typedef struct spill_t spill_t;
  79 struct spill_t {
  80         spill_t *next;
  81         ir_node *after;  /**< spill has to be placed after this node (or earlier) */
  82         ir_node *spill;
  83 };
  84
  85 typedef struct spill_info_t spill_info_t;
  86 struct spill_info_t {
  87         ir_node    *to_spill;  /**< the value that should get spilled */
  88         reloader_t *reloaders; /**< list of places where the value should get
  89                                     reloaded */
  90         spill_t    *spills;    /**< list of latest places where spill must be
  91                                     placed */
  92         double      spill_costs; /**< costs needed for spilling the value */
  93         const arch_register_class_t *reload_cls; /** the register class in which the
  94                                                      reload should be placed */
  95 };
  96
  97 struct spill_env_t {
  98         const arch_env_t *arch_env;
  99         ir_graph         *irg;
 100         struct obstack    obst;
 101         be_irg_t         *birg;
 102         int               spill_cost;     /**< the cost of a single spill node */
 103         int               reload_cost;    /**< the cost of a reload node */
 104         set              *spills;         /**< all spill_info_t's, which must be
 105                                                placed */
 106         ir_nodeset_t      mem_phis;       /**< set of all spilled phis. */
 107         ir_exec_freq     *exec_freq;
 108
 109 #ifdef FIRM_STATISTICS
 110         unsigned          spill_count;
 111         unsigned          reload_count;
 112         unsigned          remat_count;
 113         unsigned          spilled_phi_count;
 114 #endif
 115 };
 116
 117 /**
 118  * Compare two spill infos.
 119  */
 120 static int cmp_spillinfo(const void *x, const void *y, size_t size)
 121 {
 122         const spill_info_t *xx = x;
 123         const spill_info_t *yy = y;
 124         (void) size;
 125
 126         return xx->to_spill != yy->to_spill;
 127 }
 128
 129 /**
 130  * Returns spill info for a specific value (the value that is to be spilled)
 131  */
 132 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
 133 {
 134         spill_info_t info, *res;
 135         int hash = hash_irn(value);
 136
 137         info.to_spill = value;
 138         res = set_find(env->spills, &info, sizeof(info), hash);
 139
 140         if (res == NULL) {
 141                 info.reloaders   = NULL;
 142                 info.spills      = NULL;
 143                 info.spill_costs = -1;
 144                 info.reload_cls  = NULL;
 145                 res = set_insert(env->spills, &info, sizeof(info), hash);
 146         }
 147
 148         return res;
 149 }
 150
 151 spill_env_t *be_new_spill_env(be_irg_t *birg)
 152 {
 153         const arch_env_t *arch_env = birg->main_env->arch_env;
 154
 155         spill_env_t *env = XMALLOC(spill_env_t);
 156         env->spills                     = new_set(cmp_spillinfo, 1024);
 157         env->irg            = be_get_birg_irg(birg);
 158         env->birg           = birg;
 159         env->arch_env       = arch_env;
 160         ir_nodeset_init(&env->mem_phis);
 161         env->spill_cost     = arch_env->spill_cost;
 162         env->reload_cost    = arch_env->reload_cost;
 163         env->exec_freq      = be_get_birg_exec_freq(birg);
 164         obstack_init(&env->obst);
 165
 166 #ifdef FIRM_STATISTICS
 167         env->spill_count       = 0;
 168         env->reload_count      = 0;
 169         env->remat_count       = 0;
 170         env->spilled_phi_count = 0;
 171 #endif
 172
 173         return env;
 174 }
 175
 176 void be_delete_spill_env(spill_env_t *env)
 177 {
 178         del_set(env->spills);
 179         ir_nodeset_destroy(&env->mem_phis);
 180         obstack_free(&env->obst, NULL);
 181         free(env);
 182 }
 183
 184 /*
 185  *  ____  _                  ____      _                 _
 186  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 187  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 188  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 189  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 190  *
 191  */
 192
 193 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 194 {
 195         spill_info_t *spill_info = get_spillinfo(env, to_spill);
 196         spill_t      *spill;
 197         spill_t      *s;
 198         spill_t      *last;
 199
 200         assert(!arch_irn_is(to_spill, dont_spill));
 201         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 202
 203         /* Just for safety make sure that we do not insert the spill in front of a phi */
 204         assert(!is_Phi(sched_next(after)));
 205
 206         /* spills that are dominated by others are not needed */
 207         last = NULL;
 208         s    = spill_info->spills;
 209         for( ; s != NULL; s = s->next) {
 210                 /* no need to add this spill if it is dominated by another */
 211                 if(value_dominates(s->after, after)) {
 212                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 213                         return;
 214                 }
 215                 /* remove spills that we dominate */
 216                 if(value_dominates(after, s->after)) {
 217                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 218                         if(last != NULL) {
 219                                 last->next         = s->next;
 220                         } else {
 221                                 spill_info->spills = s->next;
 222                         }
 223                 } else {
 224                         last = s;
 225                 }
 226         }
 227
 228         spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
 229         spill->after  = after;
 230         spill->next   = spill_info->spills;
 231         spill->spill  = NULL;
 232
 233         spill_info->spills = spill;
 234 }
 235
 236 void be_add_remat(spill_env_t *env, ir_node *to_spill, ir_node *before,
 237                   ir_node *rematted_node)
 238 {
 239         spill_info_t *spill_info;
 240         reloader_t *reloader;
 241
 242         spill_info = get_spillinfo(env, to_spill);
 243
 244         /* add the remat information */
 245         reloader                   = obstack_alloc(&env->obst, sizeof(reloader[0]));
 246         reloader->next             = spill_info->reloaders;
 247         reloader->reloader         = before;
 248         reloader->rematted_node    = rematted_node;
 249         reloader->remat_cost_delta = 0; /* We will never have a cost win over a
 250                                            reload since we're not even allowed to
 251                                            create a reload */
 252
 253         spill_info->reloaders  = reloader;
 254
 255         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be rematerialized before %+F\n",
 256                 to_spill, before));
 257 }
 258
 259 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 260                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 261                 int allow_remat)
 262 {
 263         spill_info_t *info;
 264         reloader_t *rel;
 265
 266         assert(!arch_irn_is(to_spill, dont_spill));
 267
 268         info = get_spillinfo(env, to_spill);
 269
 270         if (is_Phi(to_spill)) {
 271                 int i, arity;
 272
 273                 /* create spillinfos for the phi arguments */
 274                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 275                         ir_node *arg = get_irn_n(to_spill, i);
 276                         get_spillinfo(env, arg);
 277                 }
 278         }
 279
 280         assert(!is_Proj(before) && !be_is_Keep(before));
 281
 282         /* put reload into list */
 283         rel                   = obstack_alloc(&env->obst, sizeof(rel[0]));
 284         rel->next             = info->reloaders;
 285         rel->reloader         = before;
 286         rel->rematted_node    = NULL;
 287         rel->can_spill_after  = can_spill_after;
 288         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 289
 290         info->reloaders  = rel;
 291         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 292         info->reload_cls = reload_cls;
 293
 294         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 295                 to_spill, before, allow_remat ? "" : " not"));
 296 }
 297
 298 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 299                    const arch_register_class_t *reload_cls, int allow_remat)
 300 {
 301         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 302
 303 }
 304
 305 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 306 {
 307         ir_node *last = sched_last(block);
 308
 309         /* we might have keeps behind the jump... */
 310         while (be_is_Keep(last)) {
 311                 last = sched_prev(last);
 312                 assert(!sched_is_end(last));
 313         }
 314
 315         assert(is_cfop(last));
 316
 317         /* add the reload before the (cond-)jump */
 318         return last;
 319 }
 320
 321 static ir_node *skip_keeps_phis(ir_node *node)
 322 {
 323         while(true) {
 324                 ir_node *next = sched_next(node);
 325                 if(!is_Phi(next) && !be_is_Keep(next))
 326                         break;
 327                 node = next;
 328         }
 329         return node;
 330 }
 331
 332 /**
 333  * Returns the point at which you can insert a node that should be executed
 334  * before block @p block when coming from pred @p pos.
 335  */
 336 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 337 {
 338         ir_node *predblock;
 339
 340         /* simply add the reload to the beginning of the block if we only have 1
 341          * predecessor. We don't need to check for phis as there can't be any in a
 342          * block with only 1 pred. */
 343         if(get_Block_n_cfgpreds(block) == 1) {
 344                 assert(!is_Phi(sched_first(block)));
 345                 return sched_first(block);
 346         }
 347
 348         /* We have to reload the value in pred-block */
 349         predblock = get_Block_cfgpred_block(block, pos);
 350         return be_get_end_of_block_insertion_point(predblock);
 351 }
 352
 353 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
 354                           const ir_node *block,
 355                           const arch_register_class_t *reload_cls,
 356                           int allow_remat)
 357 {
 358         ir_node *before = be_get_end_of_block_insertion_point(block);
 359         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 360 }
 361
 362 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 363                            int pos,     const arch_register_class_t *reload_cls,
 364                            int allow_remat)
 365 {
 366         ir_node *before = get_block_insertion_point(block, pos);
 367         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 368 }
 369
 370 void be_spill_phi(spill_env_t *env, ir_node *node)
 371 {
 372         ir_node *block;
 373         spill_info_t* spill;
 374         int i, arity;
 375
 376         assert(is_Phi(node));
 377
 378         ir_nodeset_insert(&env->mem_phis, node);
 379
 380         /* create spills for the phi arguments */
 381         block = get_nodes_block(node);
 382         spill = get_spillinfo(env, node);
 383         for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 384                 ir_node *arg = get_irn_n(node, i);
 385                 ir_node *insert;
 386                 //get_spillinfo(env, arg);
 387
 388                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 389                  * and simply always available. */
 390                 if(!sched_is_scheduled(arg)) {
 391                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 392                         insert = be_get_end_of_block_insertion_point(pred_block);
 393                         insert = sched_prev(insert);
 394                 } else {
 395                         insert = skip_keeps_phis(arg);
 396                 }
 397
 398                 be_add_spill(env, arg, insert);
 399         }
 400 }
 401
 402 /*
 403  *   ____                _         ____        _ _ _
 404  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 405  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 406  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 407  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 408  *                                      |_|
 409  */
 410
 411 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 412
 413 /**
 414  * Creates a spill.
 415  *
 416  * @param senv      the spill environment
 417  * @param irn       the node that should be spilled
 418  * @param ctx_irn   an user of the spilled node
 419  *
 420  * @return a be_Spill node
 421  */
 422 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 423 {
 424         ir_node *to_spill = spillinfo->to_spill;
 425         spill_t *spill;
 426
 427         /* determine_spill_costs must have been run before */
 428         assert(spillinfo->spill_costs >= 0);
 429
 430         /* some backends have virtual noreg/unknown nodes that are not scheduled
 431          * and simply always available. */
 432         if(!sched_is_scheduled(to_spill)) {
 433                 /* override spillinfos or create a new one */
 434                 spillinfo->spills->spill = new_NoMem();
 435                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 436                 return;
 437         }
 438
 439         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 440         spill = spillinfo->spills;
 441         for( ; spill != NULL; spill = spill->next) {
 442                 ir_node *after = spill->after;
 443                 ir_node *block = get_block(after);
 444
 445                 after = skip_keeps_phis(after);
 446
 447                 spill->spill = be_spill(block, to_spill);
 448                 sched_add_after(after, spill->spill);
 449                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 450 #ifdef FIRM_STATISTICS
 451                 env->spill_count++;
 452 #endif
 453         }
 454         DBG((dbg, LEVEL_1, "\n"));
 455 }
 456
 457 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 458
 459 /**
 460  * If the first usage of a Phi result would be out of memory
 461  * there is no sense in allocating a register for it.
 462  * Thus we spill it and all its operands to the same spill slot.
 463  * Therefore the phi/dataB becomes a phi/Memory
 464  *
 465  * @param senv      the spill environment
 466  * @param phi       the Phi node that should be spilled
 467  * @param ctx_irn   an user of the spilled node
 468  */
 469 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 470 {
 471         ir_graph *irg   = env->irg;
 472         ir_node  *phi   = spillinfo->to_spill;
 473         ir_node  *block = get_nodes_block(phi);
 474         ir_node  *unknown;
 475         ir_node **ins;
 476         spill_t  *spill;
 477         int       i;
 478         int       arity;
 479
 480         assert(is_Phi(phi));
 481         assert(!get_opt_cse());
 482         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 483
 484         /* build a new PhiM */
 485         arity   = get_irn_arity(phi);
 486         ins     = ALLOCAN(ir_node*, arity);
 487         unknown = new_r_Unknown(irg, mode_M);
 488         for(i = 0; i < arity; ++i) {
 489                 ins[i] = unknown;
 490         }
 491
 492         /* override or replace spills list... */
 493         spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
 494         spill->after  = skip_keeps_phis(phi);
 495         spill->spill  = new_r_Phi(irg, block, arity, ins, mode_M);
 496         spill->next   = NULL;
 497
 498         spillinfo->spills = spill;
 499 #ifdef FIRM_STATISTICS
 500         env->spilled_phi_count++;
 501 #endif
 502
 503         for(i = 0; i < arity; ++i) {
 504                 ir_node      *arg      = get_irn_n(phi, i);
 505                 spill_info_t *arg_info = get_spillinfo(env, arg);
 506
 507                 determine_spill_costs(env, arg_info);
 508                 spill_node(env, arg_info);
 509
 510                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 511         }
 512         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 513              spill->spill));
 514 }
 515
 516 /**
 517  * Spill a node.
 518  *
 519  * @param senv      the spill environment
 520  * @param to_spill  the node that should be spilled
 521  */
 522 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 523 {
 524         ir_node *to_spill;
 525
 526         /* node is already spilled */
 527         if(spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 528                 return;
 529
 530         to_spill = spillinfo->to_spill;
 531
 532         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 533                 spill_phi(env, spillinfo);
 534         } else {
 535                 spill_irn(env, spillinfo);
 536         }
 537 }
 538
 539 /*
 540  *
 541  *  ____                      _            _       _ _
 542  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 543  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 544  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 545  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 546  *
 547  */
 548
 549 /**
 550  * Tests whether value @p arg is available before node @p reloader
 551  * @returns 1 if value is available, 0 otherwise
 552  */
 553 static int is_value_available(spill_env_t *env, const ir_node *arg,
 554                               const ir_node *reloader)
 555 {
 556         if(is_Unknown(arg) || arg == new_NoMem())
 557                 return 1;
 558
 559         if(be_is_Spill(arg))
 560                 return 1;
 561
 562         if(arg == get_irg_frame(env->irg))
 563                 return 1;
 564
 565         /* hack for now (happens when command should be inserted at end of block) */
 566         if(is_Block(reloader)) {
 567                 return 0;
 568         }
 569
 570         /*
 571          * Ignore registers are always available
 572          */
 573         if (arch_irn_is(arg, ignore)) {
 574                 return 1;
 575         }
 576
 577         /* the following test does not work while spilling,
 578          * because the liveness info is not adapted yet to the effects of the
 579          * additional spills/reloads.
 580          */
 581 #if 0
 582         /* we want to remat before the insn reloader
 583          * thus an arguments is alive if
 584          *   - it interferes with the reloaders result
 585          *   - or it is (last-) used by reloader itself
 586          */
 587         if (values_interfere(env->birg->lv, reloader, arg)) {
 588                 return 1;
 589         }
 590
 591         arity = get_irn_arity(reloader);
 592         for (i = 0; i < arity; ++i) {
 593                 ir_node *rel_arg = get_irn_n(reloader, i);
 594                 if (rel_arg == arg)
 595                         return 1;
 596         }
 597 #endif
 598
 599         return 0;
 600 }
 601
 602 /**
 603  * Checks whether the node can principally be rematerialized
 604  */
 605 static int is_remat_node(const ir_node *node)
 606 {
 607         assert(!be_is_Spill(node));
 608
 609         if (arch_irn_is(node, rematerializable))
 610                 return 1;
 611
 612         return 0;
 613 }
 614
 615 /**
 616  * Check if a node is rematerializable. This tests for the following conditions:
 617  *
 618  * - The node itself is rematerializable
 619  * - All arguments of the node are available or also rematerialisable
 620  * - The costs for the rematerialisation operation is less or equal a limit
 621  *
 622  * Returns the costs needed for rematerialisation or something
 623  * >= REMAT_COST_INFINITE if remat is not possible.
 624  */
 625 static int check_remat_conditions_costs(spill_env_t *env,
 626                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 627 {
 628         int i, arity;
 629         int argremats;
 630         int costs = 0;
 631
 632         if (!is_remat_node(spilled))
 633                 return REMAT_COST_INFINITE;
 634
 635         if(be_is_Reload(spilled)) {
 636                 costs += 2;
 637         } else {
 638                 costs += arch_get_op_estimated_cost(spilled);
 639         }
 640         if(parentcosts + costs >= env->reload_cost + env->spill_cost) {
 641                 return REMAT_COST_INFINITE;
 642         }
 643         if (arch_irn_is(spilled, modify_flags)) {
 644                 return REMAT_COST_INFINITE;
 645         }
 646
 647         argremats = 0;
 648         for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 649                 ir_node *arg = get_irn_n(spilled, i);
 650
 651                 if(is_value_available(env, arg, reloader))
 652                         continue;
 653
 654                 /* we have to rematerialize the argument as well */
 655                 if(argremats >= 1) {
 656                         /* we only support rematerializing 1 argument at the moment,
 657                          * so that we don't have to care about register pressure
 658                          */
 659                         return REMAT_COST_INFINITE;
 660                 }
 661                 argremats++;
 662
 663                 costs += check_remat_conditions_costs(env, arg, reloader,
 664                                                       parentcosts + costs);
 665                 if(parentcosts + costs >= env->reload_cost + env->spill_cost)
 666                         return REMAT_COST_INFINITE;
 667         }
 668
 669         return costs;
 670 }
 671
 672 /**
 673  * Re-materialize a node.
 674  *
 675  * @param senv      the spill environment
 676  * @param spilled   the node that was spilled
 677  * @param reloader  a irn that requires a reload
 678  */
 679 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 680 {
 681         int i, arity;
 682         ir_node *res;
 683         ir_node *bl;
 684         ir_node **ins;
 685
 686         if(is_Block(reloader)) {
 687                 bl = reloader;
 688         } else {
 689                 bl = get_nodes_block(reloader);
 690         }
 691
 692         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 693         for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 694                 ir_node *arg = get_irn_n(spilled, i);
 695
 696                 if(is_value_available(env, arg, reloader)) {
 697                         ins[i] = arg;
 698                 } else {
 699                         ins[i] = do_remat(env, arg, reloader);
 700 #ifdef FIRM_STATISTICS
 701                         /* don't count the recursive call as remat */
 702                         env->remat_count--;
 703 #endif
 704                 }
 705         }
 706
 707         /* create a copy of the node */
 708         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 709                           get_irn_op(spilled), get_irn_mode(spilled),
 710                           get_irn_arity(spilled), ins);
 711         copy_node_attr(spilled, res);
 712         arch_env_mark_remat(env->arch_env, res);
 713         new_backedge_info(res);
 714
 715         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 716
 717         if (! is_Proj(res)) {
 718                 /* insert in schedule */
 719                 sched_reset(res);
 720                 sched_add_before(reloader, res);
 721 #ifdef FIRM_STATISTICS
 722                 env->remat_count++;
 723 #endif
 724         }
 725
 726         return res;
 727 }
 728
 729 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 730 {
 731         ir_node *block = get_nodes_block(before);
 732         double   freq  = get_block_execfreq(env->exec_freq, block);
 733         (void) to_spill;
 734
 735         return env->spill_cost * freq;
 736 }
 737
 738 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 739                                        const ir_node *before)
 740 {
 741         if(be_do_remats) {
 742                 /* is the node rematerializable? */
 743                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 744                 if(costs < (unsigned) env->reload_cost)
 745                         return costs;
 746         }
 747
 748         return env->reload_cost;
 749 }
 750
 751 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 752 {
 753         ir_node      *block = get_nodes_block(before);
 754         double        freq  = get_block_execfreq(env->exec_freq, block);
 755
 756         if(be_do_remats) {
 757                 /* is the node rematerializable? */
 758                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 759                 if(costs < env->reload_cost)
 760                         return costs * freq;
 761         }
 762
 763         return env->reload_cost * freq;
 764 }
 765
 766 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 767                            const ir_node *before)
 768 {
 769         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 770 }
 771
 772 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 773                                    ir_node *block, int pos)
 774 {
 775         ir_node *before = get_block_insertion_point(block, pos);
 776         return be_get_reload_costs(env, to_spill, before);
 777 }
 778
 779 /*
 780  *  ___                     _     ____      _                 _
 781  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 782  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 783  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 784  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 785  *
 786  */
 787
 788 /**
 789  * analyzes how to best spill a node and determine costs for that
 790  */
 791 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 792 {
 793         ir_node *to_spill = spillinfo->to_spill;
 794         ir_node *spill_block;
 795         spill_t *spill;
 796         double   spill_execfreq;
 797
 798         /* already calculated? */
 799         if(spillinfo->spill_costs >= 0)
 800                 return;
 801
 802         assert(!arch_irn_is(to_spill, dont_spill));
 803         assert(!be_is_Reload(to_spill));
 804
 805         /* some backends have virtual noreg/unknown nodes that are not scheduled
 806          * and simply always available.
 807          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 808          * predecessor (of a PhiM) but this test might match other things too...
 809          */
 810         if(!sched_is_scheduled(to_spill)) {
 811                 /* override spillinfos or create a new one */
 812                 spill_t *spill = obstack_alloc(&env->obst, sizeof(spill[0]));
 813                 spill->after = NULL;
 814                 spill->next  = NULL;
 815                 spill->spill = new_NoMem();
 816
 817                 spillinfo->spills      = spill;
 818                 spillinfo->spill_costs = 0;
 819
 820                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 821                 return;
 822         }
 823
 824         spill_block    = get_nodes_block(to_spill);
 825         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 826
 827         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 828                 /* TODO calculate correct costs...
 829                  * (though we can't remat this node anyway so no big problem) */
 830                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 831                 return;
 832         }
 833
 834         if(spillinfo->spills != NULL) {
 835                 spill_t *s;
 836                 double   spills_execfreq;
 837
 838                 /* calculate sum of execution frequencies of individual spills */
 839                 spills_execfreq = 0;
 840                 s               = spillinfo->spills;
 841                 for( ; s != NULL; s = s->next) {
 842                         ir_node *spill_block = get_block(s->after);
 843                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 844
 845                         spills_execfreq += freq;
 846                 }
 847
 848                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 849                     spills_execfreq * env->spill_cost,
 850                     spill_execfreq * env->spill_cost));
 851
 852                 /* multi-/latespill is advantageous -> return*/
 853                 if(spills_execfreq < spill_execfreq) {
 854                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 855                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 856                         return;
 857                 }
 858         }
 859
 860         /* override spillinfos or create a new one */
 861         spill        = obstack_alloc(&env->obst, sizeof(spill[0]));
 862         spill->after = skip_keeps_phis(to_spill);
 863         spill->next  = NULL;
 864         spill->spill = NULL;
 865
 866         spillinfo->spills      = spill;
 867         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 868         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 869 }
 870
 871 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 872 {
 873         const spill_info_t *si = get_spillinfo(env, irn);
 874         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 875         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 876         bitset_t *reloads      = bitset_alloca(n_blocks);
 877         reloader_t *r;
 878         spill_t *s;
 879
 880         if (si == NULL)
 881                 return;
 882
 883         /* Fill the bitset with the dominance pre-order numbers
 884          * of the blocks the reloads are located in. */
 885         for (r = si->reloaders; r != NULL; r = r->next) {
 886                 ir_node *bl = get_nodes_block(r->reloader);
 887                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 888         }
 889
 890         /* Now, cancel out all the blocks that are dominated by each spill.
 891          * If the bitset is not empty after that, we have reloads that are
 892          * not dominated by any spill. */
 893         for (s = si->spills; s != NULL; s = s->next) {
 894                 ir_node *bl = get_nodes_block(s->after);
 895                 int start   = get_Block_dom_tree_pre_num(bl);
 896                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 897
 898                 bitset_clear_range(reloads, start, end);
 899         }
 900
 901         if (!bitset_is_empty(reloads))
 902                 be_add_spill(env, si->to_spill, si->to_spill);
 903 }
 904
 905 void be_insert_spills_reloads(spill_env_t *env)
 906 {
 907         const ir_exec_freq    *exec_freq = env->exec_freq;
 908         spill_info_t          *si;
 909         ir_nodeset_iterator_t  iter;
 910         ir_node               *node;
 911
 912         BE_TIMER_PUSH(t_ra_spill_apply);
 913
 914         /* create all phi-ms first, this is needed so, that phis, hanging on
 915            spilled phis work correctly */
 916         foreach_ir_nodeset(&env->mem_phis, node, iter) {
 917                 spill_info_t *info = get_spillinfo(env, node);
 918                 spill_node(env, info);
 919         }
 920
 921         /* process each spilled node */
 922         for (si = set_first(env->spills); si; si = set_next(env->spills)) {
 923                 reloader_t *rld;
 924                 ir_node  *to_spill        = si->to_spill;
 925                 ir_mode  *mode            = get_irn_mode(to_spill);
 926                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 927                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 928                 int       force_remat     = 0;
 929
 930                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 931
 932                 determine_spill_costs(env, si);
 933
 934                 /* determine possibility of rematerialisations */
 935                 if(be_do_remats) {
 936                         /* calculate cost savings for each indivial value when it would
 937                            be rematted instead of reloaded */
 938                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 939                                 double   freq;
 940                                 int      remat_cost;
 941                                 int      remat_cost_delta;
 942                                 ir_node *block;
 943                                 ir_node *reloader = rld->reloader;
 944
 945                                 if(rld->rematted_node != NULL) {
 946                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 947                                              rld->rematted_node, reloader));
 948                                         continue;
 949                                 }
 950                                 if(rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 951                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 952                                              reloader));
 953                                         all_remat_costs = REMAT_COST_INFINITE;
 954                                         continue;
 955                                 }
 956
 957                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 958                                                                            reloader, 0);
 959                                 if(remat_cost >= REMAT_COST_INFINITE) {
 960                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 961                                              reloader));
 962                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 963                                         all_remat_costs       = REMAT_COST_INFINITE;
 964                                         continue;
 965                                 }
 966
 967                                 remat_cost_delta      = remat_cost - env->reload_cost;
 968                                 rld->remat_cost_delta = remat_cost_delta;
 969                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 970                                 freq                  = get_block_execfreq(exec_freq, block);
 971                                 all_remat_costs      += remat_cost_delta * freq;
 972                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 973                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 974                                      remat_cost_delta * freq));
 975                         }
 976                         if(all_remat_costs < REMAT_COST_INFINITE) {
 977                                 /* we don't need the costs for the spill if we can remat
 978                                    all reloaders */
 979                                 all_remat_costs -= si->spill_costs;
 980
 981                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 982                                      env->spill_cost, si->spill_costs));
 983                         }
 984
 985                         if(all_remat_costs < 0) {
 986                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 987                                      all_remat_costs));
 988                                 force_remat = 1;
 989                         }
 990                 }
 991
 992                 /* go through all reloads for this spill */
 993                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 994                         ir_node *copy; /* a reload is a "copy" of the original value */
 995
 996                         if (rld->rematted_node != NULL) {
 997                                 copy = rld->rematted_node;
 998                                 sched_add_before(rld->reloader, copy);
 999                         } else if (be_do_remats &&
1000                                         (force_remat || rld->remat_cost_delta < 0)) {
1001                                 copy = do_remat(env, to_spill, rld->reloader);
1002                         } else {
1003                                 /* make sure we have a spill */
1004                                 spill_node(env, si);
1005
1006                                 /* create a reload, use the first spill for now SSA
1007                                  * reconstruction for memory comes below */
1008                                 assert(si->spills != NULL);
1009                                 copy = be_reload(si->reload_cls, rld->reloader, mode,
1010                                                  si->spills->spill);
1011 #ifdef FIRM_STATISTICS
1012                                 env->reload_count++;
1013 #endif
1014                         }
1015
1016                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
1017                              copy, to_spill, rld->reloader));
1018                         ARR_APP1(ir_node*, copies, copy);
1019                 }
1020
1021                 /* if we had any reloads or remats, then we need to reconstruct the
1022                  * SSA form for the spilled value */
1023                 if (ARR_LEN(copies) > 0) {
1024                         be_ssa_construction_env_t senv;
1025                         /* be_lv_t *lv = be_get_birg_liveness(env->birg); */
1026
1027                         be_ssa_construction_init(&senv, env->birg);
1028                         be_ssa_construction_add_copy(&senv, to_spill);
1029                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
1030                         be_ssa_construction_fix_users(&senv, to_spill);
1031
1032 #if 0
1033                         /* no need to enable this as long as we invalidate liveness
1034                            after this function... */
1035                         be_ssa_construction_update_liveness_phis(&senv);
1036                         be_liveness_update(to_spill);
1037                         len = ARR_LEN(copies);
1038                         for(i = 0; i < len; ++i) {
1039                                 be_liveness_update(lv, copies[i]);
1040                         }
1041 #endif
1042                         be_ssa_construction_destroy(&senv);
1043                 }
1044                 /* need to reconstruct SSA form if we had multiple spills */
1045                 if (si->spills != NULL && si->spills->next != NULL) {
1046                         spill_t *spill;
1047                         int      spill_count = 0;
1048
1049                         be_ssa_construction_env_t senv;
1050
1051                         be_ssa_construction_init(&senv, env->birg);
1052                         spill = si->spills;
1053                         for( ; spill != NULL; spill = spill->next) {
1054                                 /* maybe we rematerialized the value and need no spill */
1055                                 if(spill->spill == NULL)
1056                                         continue;
1057                                 be_ssa_construction_add_copy(&senv, spill->spill);
1058                                 spill_count++;
1059                         }
1060                         if(spill_count > 1) {
1061                                 /* all reloads are attached to the first spill, fix them now */
1062                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1063                         }
1064
1065                         be_ssa_construction_destroy(&senv);
1066                 }
1067
1068                 DEL_ARR_F(copies);
1069                 si->reloaders = NULL;
1070         }
1071
1072         stat_ev_dbl("spill_spills", env->spill_count);
1073         stat_ev_dbl("spill_reloads", env->reload_count);
1074         stat_ev_dbl("spill_remats", env->remat_count);
1075         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1076
1077         /* Matze: In theory be_ssa_construction should take care of the liveness...
1078          * try to disable this again in the future */
1079         be_liveness_invalidate(env->birg->lv);
1080
1081         be_remove_dead_nodes_from_schedule(env->birg);
1082
1083         BE_TIMER_POP(t_ra_spill_apply);
1084 }
1085
1086 void be_init_spill(void)
1087 {
1088         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1089 }
1090
1091 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill);