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