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