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 #ifdef HAVE_CONFIG_H
  28 #include "config.h"
  29 #endif
  30
  31 #include <stdlib.h>
  32 #include <stdbool.h>
  33
  34 #include "pset.h"
  35 #include "irnode_t.h"
  36 #include "ircons_t.h"
  37 #include "iredges_t.h"
  38 #include "irbackedge_t.h"
  39 #include "irprintf.h"
  40 #include "ident_t.h"
  41 #include "type_t.h"
  42 #include "entity_t.h"
  43 #include "debug.h"
  44 #include "irgwalk.h"
  45 #include "array.h"
  46 #include "pdeq.h"
  47 #include "execfreq.h"
  48 #include "irnodeset.h"
  49 #include "error.h"
  50
  51 #include "bearch_t.h"
  52 #include "belive_t.h"
  53 #include "besched_t.h"
  54 #include "bespill.h"
  55 #include "belive_t.h"
  56 #include "benode_t.h"
  57 #include "bechordal_t.h"
  58 #include "bespilloptions.h"
  59 #include "bestatevent.h"
  60 #include "bessaconstr.h"
  61 #include "beirg_t.h"
  62 #include "beintlive_t.h"
  63 #include "bemodule.h"
  64 #include "be_t.h"
  65
  66 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  67
  68 #define REMAT_COST_INFINITE  1000
  69
  70 typedef struct reloader_t reloader_t;
  71 struct reloader_t {
  72         reloader_t *next;
  73         ir_node    *can_spill_after;
  74         ir_node    *reloader;
  75         ir_node    *rematted_node;
  76         int         remat_cost_delta; /** costs needed for rematerialization,
  77                                            compared to placing a reload */
  78 };
  79
  80 typedef struct spill_t spill_t;
  81 struct spill_t {
  82         spill_t *next;
  83         ir_node *after;  /**< spill has to be placed after this node (or earlier) */
  84         ir_node *spill;
  85 };
  86
  87 typedef struct spill_info_t spill_info_t;
  88 struct spill_info_t {
  89         ir_node    *to_spill;  /**< the value that should get spilled */
  90         reloader_t *reloaders; /**< list of places where the value should get
  91                                     reloaded */
  92         spill_t    *spills;    /**< list of latest places where spill must be
  93                                     placed */
  94         double      spill_costs; /**< costs needed for spilling the value */
  95         const arch_register_class_t *reload_cls; /** the register class in which the
  96                                                      reload should be placed */
  97 };
  98
  99 struct spill_env_t {
 100         const arch_env_t *arch_env;
 101         ir_graph         *irg;
 102         struct obstack    obst;
 103         be_irg_t         *birg;
 104         int               spill_cost;     /**< the cost of a single spill node */
 105         int               reload_cost;    /**< the cost of a reload node */
 106         set              *spills;         /**< all spill_info_t's, which must be
 107                                                placed */
 108         ir_nodeset_t      mem_phis;       /**< set of all spilled phis. */
 109         ir_exec_freq     *exec_freq;
 110
 111 #ifdef FIRM_STATISTICS
 112         unsigned          spill_count;
 113         unsigned          reload_count;
 114         unsigned          remat_count;
 115         unsigned          spilled_phi_count;
 116 #endif
 117 };
 118
 119 /**
 120  * Compare two spill infos.
 121  */
 122 static int cmp_spillinfo(const void *x, const void *y, size_t size)
 123 {
 124         const spill_info_t *xx = x;
 125         const spill_info_t *yy = y;
 126         (void) size;
 127
 128         return xx->to_spill != yy->to_spill;
 129 }
 130
 131 /**
 132  * Returns spill info for a specific value (the value that is to be spilled)
 133  */
 134 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
 135 {
 136         spill_info_t info, *res;
 137         int hash = hash_irn(value);
 138
 139         info.to_spill = value;
 140         res = set_find(env->spills, &info, sizeof(info), hash);
 141
 142         if (res == NULL) {
 143                 info.reloaders   = NULL;
 144                 info.spills      = NULL;
 145                 info.spill_costs = -1;
 146                 info.reload_cls  = NULL;
 147                 res = set_insert(env->spills, &info, sizeof(info), hash);
 148         }
 149
 150         return res;
 151 }
 152
 153 spill_env_t *be_new_spill_env(be_irg_t *birg)
 154 {
 155         const arch_env_t *arch_env = birg->main_env->arch_env;
 156
 157         spill_env_t *env = XMALLOC(spill_env_t);
 158         env->spills                     = new_set(cmp_spillinfo, 1024);
 159         env->irg            = be_get_birg_irg(birg);
 160         env->birg           = birg;
 161         env->arch_env       = arch_env;
 162         ir_nodeset_init(&env->mem_phis);
 163         env->spill_cost     = arch_env->spill_cost;
 164         env->reload_cost    = arch_env->reload_cost;
 165         env->exec_freq      = be_get_birg_exec_freq(birg);
 166         obstack_init(&env->obst);
 167
 168 #ifdef FIRM_STATISTICS
 169         env->spill_count       = 0;
 170         env->reload_count      = 0;
 171         env->remat_count       = 0;
 172         env->spilled_phi_count = 0;
 173 #endif
 174
 175         return env;
 176 }
 177
 178 void be_delete_spill_env(spill_env_t *env)
 179 {
 180         del_set(env->spills);
 181         ir_nodeset_destroy(&env->mem_phis);
 182         obstack_free(&env->obst, NULL);
 183         free(env);
 184 }
 185
 186 /*
 187  *  ____  _                  ____      _                 _
 188  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 189  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 190  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 191  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 192  *
 193  */
 194
 195 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 196 {
 197         spill_info_t *spill_info = get_spillinfo(env, to_spill);
 198         spill_t      *spill;
 199         spill_t      *s;
 200         spill_t      *last;
 201
 202         assert(!arch_irn_is(to_spill, dont_spill));
 203         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 204
 205         /* Just for safety make sure that we do not insert the spill in front of a phi */
 206         assert(!is_Phi(sched_next(after)));
 207
 208         /* spills that are dominated by others are not needed */
 209         last = NULL;
 210         s    = spill_info->spills;
 211         for( ; s != NULL; s = s->next) {
 212                 /* no need to add this spill if it is dominated by another */
 213                 if(value_dominates(s->after, after)) {
 214                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 215                         return;
 216                 }
 217                 /* remove spills that we dominate */
 218                 if(value_dominates(after, s->after)) {
 219                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 220                         if(last != NULL) {
 221                                 last->next         = s->next;
 222                         } else {
 223                                 spill_info->spills = s->next;
 224                         }
 225                 } else {
 226                         last = s;
 227                 }
 228         }
 229
 230         spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
 231         spill->after  = after;
 232         spill->next   = spill_info->spills;
 233         spill->spill  = NULL;
 234
 235         spill_info->spills = spill;
 236 }
 237
 238 void be_add_remat(spill_env_t *env, ir_node *to_spill, ir_node *before,
 239                   ir_node *rematted_node)
 240 {
 241         spill_info_t *spill_info;
 242         reloader_t *reloader;
 243
 244         spill_info = get_spillinfo(env, to_spill);
 245
 246         /* add the remat information */
 247         reloader                   = obstack_alloc(&env->obst, sizeof(reloader[0]));
 248         reloader->next             = spill_info->reloaders;
 249         reloader->reloader         = before;
 250         reloader->rematted_node    = rematted_node;
 251         reloader->remat_cost_delta = 0; /* We will never have a cost win over a
 252                                            reload since we're not even allowed to
 253                                            create a reload */
 254
 255         spill_info->reloaders  = reloader;
 256
 257         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be rematerialized before %+F\n",
 258                 to_spill, before));
 259 }
 260
 261 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 262                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 263                 int allow_remat)
 264 {
 265         spill_info_t *info;
 266         reloader_t *rel;
 267
 268         assert(!arch_irn_is(to_spill, dont_spill));
 269
 270         info = get_spillinfo(env, to_spill);
 271
 272         if (is_Phi(to_spill)) {
 273                 int i, arity;
 274
 275                 /* create spillinfos for the phi arguments */
 276                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 277                         ir_node *arg = get_irn_n(to_spill, i);
 278                         get_spillinfo(env, arg);
 279                 }
 280         }
 281
 282         assert(!is_Proj(before) && !be_is_Keep(before));
 283
 284         /* put reload into list */
 285         rel                   = obstack_alloc(&env->obst, sizeof(rel[0]));
 286         rel->next             = info->reloaders;
 287         rel->reloader         = before;
 288         rel->rematted_node    = NULL;
 289         rel->can_spill_after  = can_spill_after;
 290         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 291
 292         info->reloaders  = rel;
 293         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 294         info->reload_cls = reload_cls;
 295
 296         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 297                 to_spill, before, allow_remat ? "" : " not"));
 298 }
 299
 300 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 301                    const arch_register_class_t *reload_cls, int allow_remat)
 302 {
 303         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 304
 305 }
 306
 307 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 308 {
 309         ir_node *last = sched_last(block);
 310
 311         /* we might have keeps behind the jump... */
 312         while (be_is_Keep(last)) {
 313                 last = sched_prev(last);
 314                 assert(!sched_is_end(last));
 315         }
 316
 317         assert(is_cfop(last));
 318
 319         /* add the reload before the (cond-)jump */
 320         return last;
 321 }
 322
 323 static ir_node *skip_keeps_phis(ir_node *node)
 324 {
 325         while(true) {
 326                 ir_node *next = sched_next(node);
 327                 if(!is_Phi(next) && !be_is_Keep(next))
 328                         break;
 329                 node = next;
 330         }
 331         return node;
 332 }
 333
 334 /**
 335  * Returns the point at which you can insert a node that should be executed
 336  * before block @p block when coming from pred @p pos.
 337  */
 338 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 339 {
 340         ir_node *predblock;
 341
 342         /* simply add the reload to the beginning of the block if we only have 1
 343          * predecessor. We don't need to check for phis as there can't be any in a
 344          * block with only 1 pred. */
 345         if(get_Block_n_cfgpreds(block) == 1) {
 346                 assert(!is_Phi(sched_first(block)));
 347                 return sched_first(block);
 348         }
 349
 350         /* We have to reload the value in pred-block */
 351         predblock = get_Block_cfgpred_block(block, pos);
 352         return be_get_end_of_block_insertion_point(predblock);
 353 }
 354
 355 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
 356                           const ir_node *block,
 357                           const arch_register_class_t *reload_cls,
 358                           int allow_remat)
 359 {
 360         ir_node *before = be_get_end_of_block_insertion_point(block);
 361         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 362 }
 363
 364 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 365                            int pos,     const arch_register_class_t *reload_cls,
 366                            int allow_remat)
 367 {
 368         ir_node *before = get_block_insertion_point(block, pos);
 369         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 370 }
 371
 372 void be_spill_phi(spill_env_t *env, ir_node *node)
 373 {
 374         ir_node *block;
 375         spill_info_t* spill;
 376         int i, arity;
 377
 378         assert(is_Phi(node));
 379
 380         ir_nodeset_insert(&env->mem_phis, node);
 381
 382         /* create spills for the phi arguments */
 383         block = get_nodes_block(node);
 384         spill = get_spillinfo(env, node);
 385         for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 386                 ir_node *arg = get_irn_n(node, i);
 387                 ir_node *insert;
 388                 //get_spillinfo(env, arg);
 389
 390                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 391                  * and simply always available. */
 392                 if(!sched_is_scheduled(arg)) {
 393                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 394                         insert = be_get_end_of_block_insertion_point(pred_block);
 395                         insert = sched_prev(insert);
 396                 } else {
 397                         insert = skip_keeps_phis(arg);
 398                 }
 399
 400                 be_add_spill(env, arg, insert);
 401         }
 402 }
 403
 404 /*
 405  *   ____                _         ____        _ _ _
 406  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 407  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 408  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 409  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 410  *                                      |_|
 411  */
 412
 413 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 414
 415 /**
 416  * Creates a spill.
 417  *
 418  * @param senv      the spill environment
 419  * @param irn       the node that should be spilled
 420  * @param ctx_irn   an user of the spilled node
 421  *
 422  * @return a be_Spill node
 423  */
 424 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 425 {
 426         ir_node *to_spill = spillinfo->to_spill;
 427         spill_t *spill;
 428
 429         /* determine_spill_costs must have been run before */
 430         assert(spillinfo->spill_costs >= 0);
 431
 432         /* some backends have virtual noreg/unknown nodes that are not scheduled
 433          * and simply always available. */
 434         if(!sched_is_scheduled(to_spill)) {
 435                 /* override spillinfos or create a new one */
 436                 spillinfo->spills->spill = new_NoMem();
 437                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 438                 return;
 439         }
 440
 441         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 442         spill = spillinfo->spills;
 443         for( ; spill != NULL; spill = spill->next) {
 444                 ir_node *after = spill->after;
 445                 ir_node *block = get_block(after);
 446
 447                 after = skip_keeps_phis(after);
 448
 449                 spill->spill   = be_spill(env->arch_env, block, to_spill);
 450                 sched_add_after(after, spill->spill);
 451                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 452 #ifdef FIRM_STATISTICS
 453                 env->spill_count++;
 454 #endif
 455         }
 456         DBG((dbg, LEVEL_1, "\n"));
 457 }
 458
 459 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 460
 461 /**
 462  * If the first usage of a Phi result would be out of memory
 463  * there is no sense in allocating a register for it.
 464  * Thus we spill it and all its operands to the same spill slot.
 465  * Therefore the phi/dataB becomes a phi/Memory
 466  *
 467  * @param senv      the spill environment
 468  * @param phi       the Phi node that should be spilled
 469  * @param ctx_irn   an user of the spilled node
 470  */
 471 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 472 {
 473         ir_graph *irg   = env->irg;
 474         ir_node  *phi   = spillinfo->to_spill;
 475         ir_node  *block = get_nodes_block(phi);
 476         ir_node  *unknown;
 477         ir_node **ins;
 478         spill_t  *spill;
 479         int       i;
 480         int       arity;
 481
 482         assert(is_Phi(phi));
 483         assert(!get_opt_cse());
 484         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 485
 486         /* build a new PhiM */
 487         arity   = get_irn_arity(phi);
 488         ins     = alloca(sizeof(ir_node*) * arity);
 489         unknown = new_r_Unknown(irg, mode_M);
 490         for(i = 0; i < arity; ++i) {
 491                 ins[i] = unknown;
 492         }
 493
 494         /* override or replace spills list... */
 495         spill         = obstack_alloc(&env->obst, sizeof(spill[0]));
 496         spill->after  = skip_keeps_phis(phi);
 497         spill->spill  = new_r_Phi(irg, block, arity, ins, mode_M);
 498         spill->next   = NULL;
 499
 500         spillinfo->spills = spill;
 501 #ifdef FIRM_STATISTICS
 502         env->spilled_phi_count++;
 503 #endif
 504
 505         for(i = 0; i < arity; ++i) {
 506                 ir_node      *arg      = get_irn_n(phi, i);
 507                 spill_info_t *arg_info = get_spillinfo(env, arg);
 508
 509                 determine_spill_costs(env, arg_info);
 510                 spill_node(env, arg_info);
 511
 512                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 513         }
 514         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 515              spill->spill));
 516 }
 517
 518 /**
 519  * Spill a node.
 520  *
 521  * @param senv      the spill environment
 522  * @param to_spill  the node that should be spilled
 523  */
 524 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 525 {
 526         ir_node *to_spill;
 527
 528         /* node is already spilled */
 529         if(spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 530                 return;
 531
 532         to_spill = spillinfo->to_spill;
 533
 534         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 535                 spill_phi(env, spillinfo);
 536         } else {
 537                 spill_irn(env, spillinfo);
 538         }
 539 }
 540
 541 /*
 542  *
 543  *  ____                      _            _       _ _
 544  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 545  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 546  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 547  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 548  *
 549  */
 550
 551 /**
 552  * Tests whether value @p arg is available before node @p reloader
 553  * @returns 1 if value is available, 0 otherwise
 554  */
 555 static int is_value_available(spill_env_t *env, const ir_node *arg,
 556                               const ir_node *reloader)
 557 {
 558         if(is_Unknown(arg) || arg == new_NoMem())
 559                 return 1;
 560
 561         if(be_is_Spill(arg))
 562                 return 1;
 563
 564         if(arg == get_irg_frame(env->irg))
 565                 return 1;
 566
 567         /* hack for now (happens when command should be inserted at end of block) */
 568         if(is_Block(reloader)) {
 569                 return 0;
 570         }
 571
 572         /*
 573          * Ignore registers are always available
 574          */
 575         if (arch_irn_is(arg, ignore)) {
 576                 return 1;
 577         }
 578
 579         /* the following test does not work while spilling,
 580          * because the liveness info is not adapted yet to the effects of the
 581          * additional spills/reloads.
 582          */
 583 #if 0
 584         /* we want to remat before the insn reloader
 585          * thus an arguments is alive if
 586          *   - it interferes with the reloaders result
 587          *   - or it is (last-) used by reloader itself
 588          */
 589         if (values_interfere(env->birg->lv, reloader, arg)) {
 590                 return 1;
 591         }
 592
 593         arity = get_irn_arity(reloader);
 594         for (i = 0; i < arity; ++i) {
 595                 ir_node *rel_arg = get_irn_n(reloader, i);
 596                 if (rel_arg == arg)
 597                         return 1;
 598         }
 599 #endif
 600
 601         return 0;
 602 }
 603
 604 /**
 605  * Checks whether the node can principally be rematerialized
 606  */
 607 static int is_remat_node(const ir_node *node)
 608 {
 609         assert(!be_is_Spill(node));
 610
 611         if (arch_irn_is(node, rematerializable))
 612                 return 1;
 613
 614         return 0;
 615 }
 616
 617 /**
 618  * Check if a node is rematerializable. This tests for the following conditions:
 619  *
 620  * - The node itself is rematerializable
 621  * - All arguments of the node are available or also rematerialisable
 622  * - The costs for the rematerialisation operation is less or equal a limit
 623  *
 624  * Returns the costs needed for rematerialisation or something
 625  * >= REMAT_COST_INFINITE if remat is not possible.
 626  */
 627 static int check_remat_conditions_costs(spill_env_t *env,
 628                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 629 {
 630         int i, arity;
 631         int argremats;
 632         int costs = 0;
 633
 634         if (!is_remat_node(spilled))
 635                 return REMAT_COST_INFINITE;
 636
 637         if(be_is_Reload(spilled)) {
 638                 costs += 2;
 639         } else {
 640                 costs += arch_get_op_estimated_cost(spilled);
 641         }
 642         if(parentcosts + costs >= env->reload_cost + env->spill_cost) {
 643                 return REMAT_COST_INFINITE;
 644         }
 645         if (arch_irn_is(spilled, modify_flags)) {
 646                 return REMAT_COST_INFINITE;
 647         }
 648
 649         argremats = 0;
 650         for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 651                 ir_node *arg = get_irn_n(spilled, i);
 652
 653                 if(is_value_available(env, arg, reloader))
 654                         continue;
 655
 656                 /* we have to rematerialize the argument as well */
 657                 if(argremats >= 1) {
 658                         /* we only support rematerializing 1 argument at the moment,
 659                          * so that we don't have to care about register pressure
 660                          */
 661                         return REMAT_COST_INFINITE;
 662                 }
 663                 argremats++;
 664
 665                 costs += check_remat_conditions_costs(env, arg, reloader,
 666                                                       parentcosts + costs);
 667                 if(parentcosts + costs >= env->reload_cost + env->spill_cost)
 668                         return REMAT_COST_INFINITE;
 669         }
 670
 671         return costs;
 672 }
 673
 674 /**
 675  * Re-materialize a node.
 676  *
 677  * @param senv      the spill environment
 678  * @param spilled   the node that was spilled
 679  * @param reloader  a irn that requires a reload
 680  */
 681 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 682 {
 683         int i, arity;
 684         ir_node *res;
 685         ir_node *bl;
 686         ir_node **ins;
 687
 688         if(is_Block(reloader)) {
 689                 bl = reloader;
 690         } else {
 691                 bl = get_nodes_block(reloader);
 692         }
 693
 694         ins = alloca(get_irn_arity(spilled) * sizeof(ins[0]));
 695         for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 696                 ir_node *arg = get_irn_n(spilled, i);
 697
 698                 if(is_value_available(env, arg, reloader)) {
 699                         ins[i] = arg;
 700                 } else {
 701                         ins[i] = do_remat(env, arg, reloader);
 702 #ifdef FIRM_STATISTICS
 703                         /* don't count the recursive call as remat */
 704                         env->remat_count--;
 705 #endif
 706                 }
 707         }
 708
 709         /* create a copy of the node */
 710         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 711                           get_irn_op(spilled), get_irn_mode(spilled),
 712                           get_irn_arity(spilled), ins);
 713         copy_node_attr(spilled, res);
 714         arch_env_mark_remat(env->arch_env, res);
 715         new_backedge_info(res);
 716
 717         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 718
 719         if (! is_Proj(res)) {
 720                 /* insert in schedule */
 721                 sched_reset(res);
 722                 sched_add_before(reloader, res);
 723 #ifdef FIRM_STATISTICS
 724                 env->remat_count++;
 725 #endif
 726         }
 727
 728         return res;
 729 }
 730
 731 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 732 {
 733         ir_node *block = get_nodes_block(before);
 734         double   freq  = get_block_execfreq(env->exec_freq, block);
 735         (void) to_spill;
 736
 737         return env->spill_cost * freq;
 738 }
 739
 740 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 741                                        const ir_node *before)
 742 {
 743         if(be_do_remats) {
 744                 /* is the node rematerializable? */
 745                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 746                 if(costs < (unsigned) env->reload_cost)
 747                         return costs;
 748         }
 749
 750         return env->reload_cost;
 751 }
 752
 753 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 754 {
 755         ir_node      *block = get_nodes_block(before);
 756         double        freq  = get_block_execfreq(env->exec_freq, block);
 757
 758         if(be_do_remats) {
 759                 /* is the node rematerializable? */
 760                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 761                 if(costs < env->reload_cost)
 762                         return costs * freq;
 763         }
 764
 765         return env->reload_cost * freq;
 766 }
 767
 768 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 769                            const ir_node *before)
 770 {
 771         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 772 }
 773
 774 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 775                                    ir_node *block, int pos)
 776 {
 777         ir_node *before = get_block_insertion_point(block, pos);
 778         return be_get_reload_costs(env, to_spill, before);
 779 }
 780
 781 /*
 782  *  ___                     _     ____      _                 _
 783  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 784  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 785  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 786  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 787  *
 788  */
 789
 790 /**
 791  * analyzes how to best spill a node and determine costs for that
 792  */
 793 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 794 {
 795         ir_node *to_spill = spillinfo->to_spill;
 796         ir_node *spill_block;
 797         spill_t *spill;
 798         double   spill_execfreq;
 799
 800         /* already calculated? */
 801         if(spillinfo->spill_costs >= 0)
 802                 return;
 803
 804         assert(!arch_irn_is(to_spill, dont_spill));
 805         assert(!be_is_Reload(to_spill));
 806
 807         /* some backends have virtual noreg/unknown nodes that are not scheduled
 808          * and simply always available.
 809          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 810          * predecessor (of a PhiM) but this test might match other things too...
 811          */
 812         if(!sched_is_scheduled(to_spill)) {
 813                 /* override spillinfos or create a new one */
 814                 spill_t *spill = obstack_alloc(&env->obst, sizeof(spill[0]));
 815                 spill->after = NULL;
 816                 spill->next  = NULL;
 817                 spill->spill = new_NoMem();
 818
 819                 spillinfo->spills      = spill;
 820                 spillinfo->spill_costs = 0;
 821
 822                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 823                 return;
 824         }
 825
 826         spill_block    = get_nodes_block(to_spill);
 827         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 828
 829         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 830                 /* TODO calculate correct costs...
 831                  * (though we can't remat this node anyway so no big problem) */
 832                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 833                 return;
 834         }
 835
 836         if(spillinfo->spills != NULL) {
 837                 spill_t *s;
 838                 double   spills_execfreq;
 839
 840                 /* calculate sum of executaion frequencies of individual spills */
 841                 spills_execfreq = 0;
 842                 s               = spillinfo->spills;
 843                 for( ; s != NULL; s = s->next) {
 844                         ir_node *spill_block = get_block(s->after);
 845                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 846
 847                         spills_execfreq += freq;
 848                 }
 849
 850                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 851                     spills_execfreq * env->spill_cost,
 852                     spill_execfreq * env->spill_cost));
 853
 854                 /* multi-/latespill is advantageous -> return*/
 855                 if(spills_execfreq < spill_execfreq) {
 856                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 857                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 858                         return;
 859                 }
 860         }
 861
 862         /* override spillinfos or create a new one */
 863         spill        = obstack_alloc(&env->obst, sizeof(spill[0]));
 864         spill->after = skip_keeps_phis(to_spill);
 865         spill->next  = NULL;
 866         spill->spill = NULL;
 867
 868         spillinfo->spills      = spill;
 869         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 870         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 871 }
 872
 873 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 874 {
 875         const spill_info_t *si = get_spillinfo(env, irn);
 876         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 877         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 878         bitset_t *reloads      = bitset_alloca(n_blocks);
 879         reloader_t *r;
 880         spill_t *s;
 881
 882         if (si == NULL)
 883                 return;
 884
 885         /* Fill the bitset with the dominance pre-order numbers
 886          * of the blocks the reloads are located in. */
 887         for (r = si->reloaders; r != NULL; r = r->next) {
 888                 ir_node *bl = get_nodes_block(r->reloader);
 889                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 890         }
 891
 892         /* Now, cancel out all the blocks that are dominated by each spill.
 893          * If the bitset is not empty after that, we have reloads that are
 894          * not dominated by any spill. */
 895         for (s = si->spills; s != NULL; s = s->next) {
 896                 ir_node *bl = get_nodes_block(s->after);
 897                 int start   = get_Block_dom_tree_pre_num(bl);
 898                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 899
 900                 bitset_clear_range(reloads, start, end);
 901         }
 902
 903         if (!bitset_is_empty(reloads))
 904                 be_add_spill(env, si->to_spill, si->to_spill);
 905 }
 906
 907 void be_insert_spills_reloads(spill_env_t *env)
 908 {
 909         const arch_env_t      *arch_env  = env->arch_env;
 910         const ir_exec_freq    *exec_freq = env->exec_freq;
 911         spill_info_t          *si;
 912         ir_nodeset_iterator_t  iter;
 913         ir_node               *node;
 914
 915         BE_TIMER_PUSH(t_ra_spill_apply);
 916
 917         /* create all phi-ms first, this is needed so, that phis, hanging on
 918            spilled phis work correctly */
 919         foreach_ir_nodeset(&env->mem_phis, node, iter) {
 920                 spill_info_t *info = get_spillinfo(env, node);
 921                 spill_node(env, info);
 922         }
 923
 924         /* process each spilled node */
 925         for (si = set_first(env->spills); si; si = set_next(env->spills)) {
 926                 reloader_t *rld;
 927                 ir_node  *to_spill        = si->to_spill;
 928                 ir_mode  *mode            = get_irn_mode(to_spill);
 929                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 930                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 931                 int       force_remat     = 0;
 932
 933                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 934
 935                 determine_spill_costs(env, si);
 936
 937                 /* determine possibility of rematerialisations */
 938                 if(be_do_remats) {
 939                         /* calculate cost savings for each indivial value when it would
 940                            be rematted instead of reloaded */
 941                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 942                                 double   freq;
 943                                 int      remat_cost;
 944                                 int      remat_cost_delta;
 945                                 ir_node *block;
 946                                 ir_node *reloader = rld->reloader;
 947
 948                                 if(rld->rematted_node != NULL) {
 949                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 950                                              rld->rematted_node, reloader));
 951                                         continue;
 952                                 }
 953                                 if(rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 954                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 955                                              reloader));
 956                                         all_remat_costs = REMAT_COST_INFINITE;
 957                                         continue;
 958                                 }
 959
 960                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 961                                                                            reloader, 0);
 962                                 if(remat_cost >= REMAT_COST_INFINITE) {
 963                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 964                                              reloader));
 965                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 966                                         all_remat_costs       = REMAT_COST_INFINITE;
 967                                         continue;
 968                                 }
 969
 970                                 remat_cost_delta      = remat_cost - env->reload_cost;
 971                                 rld->remat_cost_delta = remat_cost_delta;
 972                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 973                                 freq                  = get_block_execfreq(exec_freq, block);
 974                                 all_remat_costs      += remat_cost_delta * freq;
 975                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 976                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 977                                      remat_cost_delta * freq));
 978                         }
 979                         if(all_remat_costs < REMAT_COST_INFINITE) {
 980                                 /* we don't need the costs for the spill if we can remat
 981                                    all reloaders */
 982                                 all_remat_costs -= si->spill_costs;
 983
 984                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 985                                      env->spill_cost, si->spill_costs));
 986                         }
 987
 988                         if(all_remat_costs < 0) {
 989                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 990                                      all_remat_costs));
 991                                 force_remat = 1;
 992                         }
 993                 }
 994
 995                 /* go through all reloads for this spill */
 996                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 997                         ir_node *copy; /* a reload is a "copy" of the original value */
 998
 999                         if (rld->rematted_node != NULL) {
1000                                 copy = rld->rematted_node;
1001                                 sched_add_before(rld->reloader, copy);
1002                         } else if (be_do_remats &&
1003                                         (force_remat || rld->remat_cost_delta < 0)) {
1004                                 copy = do_remat(env, to_spill, rld->reloader);
1005                         } else {
1006                                 /* make sure we have a spill */
1007                                 spill_node(env, si);
1008
1009                                 /* create a reload, use the first spill for now SSA
1010                                  * reconstruction for memory comes below */
1011                                 assert(si->spills != NULL);
1012                                 copy = be_reload(arch_env, si->reload_cls, rld->reloader, mode,
1013                                                  si->spills->spill);
1014 #ifdef FIRM_STATISTICS
1015                                 env->reload_count++;
1016 #endif
1017                         }
1018
1019                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
1020                              copy, to_spill, rld->reloader));
1021                         ARR_APP1(ir_node*, copies, copy);
1022                 }
1023
1024                 /* if we had any reloads or remats, then we need to reconstruct the
1025                  * SSA form for the spilled value */
1026                 if (ARR_LEN(copies) > 0) {
1027                         be_ssa_construction_env_t senv;
1028                         /* be_lv_t *lv = be_get_birg_liveness(env->birg); */
1029
1030                         be_ssa_construction_init(&senv, env->birg);
1031                         be_ssa_construction_add_copy(&senv, to_spill);
1032                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
1033                         be_ssa_construction_fix_users(&senv, to_spill);
1034
1035 #if 0
1036                         /* no need to enable this as long as we invalidate liveness
1037                            after this function... */
1038                         be_ssa_construction_update_liveness_phis(&senv);
1039                         be_liveness_update(to_spill);
1040                         len = ARR_LEN(copies);
1041                         for(i = 0; i < len; ++i) {
1042                                 be_liveness_update(lv, copies[i]);
1043                         }
1044 #endif
1045                         be_ssa_construction_destroy(&senv);
1046                 }
1047                 /* need to reconstruct SSA form if we had multiple spills */
1048                 if (si->spills != NULL && si->spills->next != NULL) {
1049                         spill_t *spill;
1050                         int      spill_count = 0;
1051
1052                         be_ssa_construction_env_t senv;
1053
1054                         be_ssa_construction_init(&senv, env->birg);
1055                         spill = si->spills;
1056                         for( ; spill != NULL; spill = spill->next) {
1057                                 /* maybe we rematerialized the value and need no spill */
1058                                 if(spill->spill == NULL)
1059                                         continue;
1060                                 be_ssa_construction_add_copy(&senv, spill->spill);
1061                                 spill_count++;
1062                         }
1063                         if(spill_count > 1) {
1064                                 /* all reloads are attached to the first spill, fix them now */
1065                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1066                         }
1067
1068                         be_ssa_construction_destroy(&senv);
1069                 }
1070
1071                 DEL_ARR_F(copies);
1072                 si->reloaders = NULL;
1073         }
1074
1075         stat_ev_dbl("spill_spills", env->spill_count);
1076         stat_ev_dbl("spill_reloads", env->reload_count);
1077         stat_ev_dbl("spill_remats", env->remat_count);
1078         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1079
1080         /* Matze: In theory be_ssa_construction should take care of the liveness...
1081          * try to disable this again in the future */
1082         be_liveness_invalidate(env->birg->lv);
1083
1084         be_remove_dead_nodes_from_schedule(env->birg);
1085
1086         BE_TIMER_POP(t_ra_spill_apply);
1087 }
1088
1089 void be_init_spill(void)
1090 {
1091         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1092 }
1093
1094 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill);