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         const ir_node *insn       = skip_Proj_const(to_spill);
 195         spill_t       *spill;
 196         spill_t       *s;
 197         spill_t       *last;
 198
 199         assert(!arch_irn_is(insn, 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_remat(spill_env_t *env, ir_node *to_spill, ir_node *before,
 236                   ir_node *rematted_node)
 237 {
 238         spill_info_t *spill_info;
 239         reloader_t *reloader;
 240
 241         spill_info = get_spillinfo(env, to_spill);
 242
 243         /* add the remat information */
 244         reloader                   = OALLOC(&env->obst, reloader_t);
 245         reloader->next             = spill_info->reloaders;
 246         reloader->reloader         = before;
 247         reloader->rematted_node    = rematted_node;
 248         reloader->remat_cost_delta = 0; /* We will never have a cost win over a
 249                                            reload since we're not even allowed to
 250                                            create a reload */
 251
 252         spill_info->reloaders  = reloader;
 253
 254         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be rematerialized before %+F\n",
 255                 to_spill, before));
 256 }
 257
 258 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 259                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 260                 int allow_remat)
 261 {
 262         spill_info_t  *info;
 263         reloader_t    *rel;
 264         const ir_node *insn = skip_Proj_const(to_spill);
 265
 266         assert(!arch_irn_is(insn, dont_spill));
 267
 268         info = get_spillinfo(env, to_spill);
 269
 270         if (is_Phi(to_spill)) {
 271                 int i, arity;
 272
 273                 /* create spillinfos for the phi arguments */
 274                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 275                         ir_node *arg = get_irn_n(to_spill, i);
 276                         get_spillinfo(env, arg);
 277                 }
 278         }
 279
 280         assert(!is_Proj(before) && !be_is_Keep(before));
 281
 282         /* put reload into list */
 283         rel                   = OALLOC(&env->obst, reloader_t);
 284         rel->next             = info->reloaders;
 285         rel->reloader         = before;
 286         rel->rematted_node    = NULL;
 287         rel->can_spill_after  = can_spill_after;
 288         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 289
 290         info->reloaders  = rel;
 291         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 292         info->reload_cls = reload_cls;
 293
 294         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 295                 to_spill, before, allow_remat ? "" : " not"));
 296 }
 297
 298 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 299                    const arch_register_class_t *reload_cls, int allow_remat)
 300 {
 301         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 302
 303 }
 304
 305 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 306 {
 307         ir_node *last = sched_last(block);
 308
 309         /* we might have keeps behind the jump... */
 310         while (be_is_Keep(last)) {
 311                 last = sched_prev(last);
 312                 assert(!sched_is_end(last));
 313         }
 314
 315         assert(is_cfop(last));
 316
 317         /* add the reload before the (cond-)jump */
 318         return last;
 319 }
 320
 321 static ir_node *skip_keeps_phis(ir_node *node)
 322 {
 323         while (true) {
 324                 ir_node *next = sched_next(node);
 325                 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
 326                         break;
 327                 node = next;
 328         }
 329         return node;
 330 }
 331
 332 /**
 333  * Returns the point at which you can insert a node that should be executed
 334  * before block @p block when coming from pred @p pos.
 335  */
 336 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 337 {
 338         ir_node *predblock;
 339
 340         /* simply add the reload to the beginning of the block if we only have 1
 341          * predecessor. We don't need to check for phis as there can't be any in a
 342          * block with only 1 pred. */
 343         if (get_Block_n_cfgpreds(block) == 1) {
 344                 assert(!is_Phi(sched_first(block)));
 345                 return sched_first(block);
 346         }
 347
 348         /* We have to reload the value in pred-block */
 349         predblock = get_Block_cfgpred_block(block, pos);
 350         return be_get_end_of_block_insertion_point(predblock);
 351 }
 352
 353 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
 354                           const ir_node *block,
 355                           const arch_register_class_t *reload_cls,
 356                           int allow_remat)
 357 {
 358         ir_node *before = be_get_end_of_block_insertion_point(block);
 359         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 360 }
 361
 362 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 363                            int pos,     const arch_register_class_t *reload_cls,
 364                            int allow_remat)
 365 {
 366         ir_node *before = get_block_insertion_point(block, pos);
 367         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 368 }
 369
 370 void be_spill_phi(spill_env_t *env, ir_node *node)
 371 {
 372         ir_node *block;
 373         int i, arity;
 374
 375         assert(is_Phi(node));
 376
 377         ir_nodeset_insert(&env->mem_phis, node);
 378
 379         /* create spills for the phi arguments */
 380         block = get_nodes_block(node);
 381         for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 382                 ir_node *arg = get_irn_n(node, i);
 383                 ir_node *insert;
 384
 385                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 386                  * and simply always available. */
 387                 if (!sched_is_scheduled(arg)) {
 388                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 389                         insert = be_get_end_of_block_insertion_point(pred_block);
 390                         insert = sched_prev(insert);
 391                 } else {
 392                         insert = skip_keeps_phis(arg);
 393                 }
 394
 395                 be_add_spill(env, arg, insert);
 396         }
 397 }
 398
 399 /*
 400  *   ____                _         ____        _ _ _
 401  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 402  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 403  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 404  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 405  *                                      |_|
 406  */
 407
 408 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 409
 410 /**
 411  * Creates a spill.
 412  *
 413  * @param senv      the spill environment
 414  * @param irn       the node that should be spilled
 415  * @param ctx_irn   an user of the spilled node
 416  *
 417  * @return a be_Spill node
 418  */
 419 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 420 {
 421         ir_node       *to_spill = spillinfo->to_spill;
 422         const ir_node *insn     = skip_Proj_const(to_spill);
 423         spill_t *spill;
 424
 425         /* determine_spill_costs must have been run before */
 426         assert(spillinfo->spill_costs >= 0);
 427
 428         /* some backends have virtual noreg/unknown nodes that are not scheduled
 429          * and simply always available. */
 430         if (!sched_is_scheduled(insn)) {
 431                 /* override spillinfos or create a new one */
 432                 spillinfo->spills->spill = new_NoMem();
 433                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 434                 return;
 435         }
 436
 437         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 438         spill = spillinfo->spills;
 439         for ( ; spill != NULL; spill = spill->next) {
 440                 ir_node *after = spill->after;
 441                 ir_node *block = get_block(after);
 442
 443                 after = skip_keeps_phis(after);
 444
 445                 spill->spill = be_spill(block, to_spill);
 446                 sched_add_after(skip_Proj(after), spill->spill);
 447                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 448 #ifdef FIRM_STATISTICS
 449                 env->spill_count++;
 450 #endif
 451         }
 452         DBG((dbg, LEVEL_1, "\n"));
 453 }
 454
 455 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 456
 457 /**
 458  * If the first usage of a Phi result would be out of memory
 459  * there is no sense in allocating a register for it.
 460  * Thus we spill it and all its operands to the same spill slot.
 461  * Therefore the phi/dataB becomes a phi/Memory
 462  *
 463  * @param senv      the spill environment
 464  * @param phi       the Phi node that should be spilled
 465  * @param ctx_irn   an user of the spilled node
 466  */
 467 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 468 {
 469         ir_graph *irg   = env->irg;
 470         ir_node  *phi   = spillinfo->to_spill;
 471         ir_node  *block = get_nodes_block(phi);
 472         ir_node  *unknown;
 473         ir_node **ins;
 474         spill_t  *spill;
 475         int       i;
 476         int       arity;
 477
 478         assert(is_Phi(phi));
 479         assert(!get_opt_cse());
 480         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 481
 482         /* build a new PhiM */
 483         arity   = get_irn_arity(phi);
 484         ins     = ALLOCAN(ir_node*, arity);
 485         unknown = new_r_Unknown(irg, mode_M);
 486         for (i = 0; i < arity; ++i) {
 487                 ins[i] = unknown;
 488         }
 489
 490         /* override or replace spills list... */
 491         spill         = OALLOC(&env->obst, spill_t);
 492         spill->after  = skip_keeps_phis(phi);
 493         spill->spill  = new_r_Phi(block, arity, ins, mode_M);
 494         spill->next   = NULL;
 495
 496         spillinfo->spills = spill;
 497 #ifdef FIRM_STATISTICS
 498         env->spilled_phi_count++;
 499 #endif
 500
 501         for (i = 0; i < arity; ++i) {
 502                 ir_node      *arg      = get_irn_n(phi, i);
 503                 spill_info_t *arg_info = get_spillinfo(env, arg);
 504
 505                 determine_spill_costs(env, arg_info);
 506                 spill_node(env, arg_info);
 507
 508                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 509         }
 510         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 511              spill->spill));
 512 }
 513
 514 /**
 515  * Spill a node.
 516  *
 517  * @param senv      the spill environment
 518  * @param to_spill  the node that should be spilled
 519  */
 520 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 521 {
 522         ir_node *to_spill;
 523
 524         /* node is already spilled */
 525         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 526                 return;
 527
 528         to_spill = spillinfo->to_spill;
 529
 530         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 531                 spill_phi(env, spillinfo);
 532         } else {
 533                 spill_irn(env, spillinfo);
 534         }
 535 }
 536
 537 /*
 538  *
 539  *  ____                      _            _       _ _
 540  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 541  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 542  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 543  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 544  *
 545  */
 546
 547 /**
 548  * Tests whether value @p arg is available before node @p reloader
 549  * @returns 1 if value is available, 0 otherwise
 550  */
 551 static int is_value_available(spill_env_t *env, const ir_node *arg,
 552                               const ir_node *reloader)
 553 {
 554         if (is_Unknown(arg) || arg == new_NoMem())
 555                 return 1;
 556
 557         if (be_is_Spill(skip_Proj_const(arg)))
 558                 return 1;
 559
 560         if (arg == get_irg_frame(env->irg))
 561                 return 1;
 562
 563         (void)reloader;
 564
 565         if (get_irn_mode(arg) == mode_T)
 566                 return 0;
 567
 568         /*
 569          * Ignore registers are always available
 570          */
 571         if (arch_irn_is_ignore(arg))
 572                 return 1;
 573
 574         return 0;
 575 }
 576
 577 /**
 578  * Check if a node is rematerializable. This tests for the following conditions:
 579  *
 580  * - The node itself is rematerializable
 581  * - All arguments of the node are available or also rematerialisable
 582  * - The costs for the rematerialisation operation is less or equal a limit
 583  *
 584  * Returns the costs needed for rematerialisation or something
 585  * >= REMAT_COST_INFINITE if remat is not possible.
 586  */
 587 static int check_remat_conditions_costs(spill_env_t *env,
 588                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 589 {
 590         int i, arity;
 591         int argremats;
 592         int costs = 0;
 593         const ir_node *insn = skip_Proj_const(spilled);
 594
 595         assert(!be_is_Spill(insn));
 596         if (!arch_irn_is(insn, rematerializable))
 597                 return REMAT_COST_INFINITE;
 598
 599         if (be_is_Reload(insn)) {
 600                 costs += 2;
 601         } else {
 602                 costs += arch_get_op_estimated_cost(insn);
 603         }
 604         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 605                 return REMAT_COST_INFINITE;
 606         }
 607         /* never rematerialize a node which modifies the flags.
 608          * (would be better to test wether the flags are actually live at point
 609          * reloader...)
 610          */
 611         if (arch_irn_is(insn, modify_flags)) {
 612                 return REMAT_COST_INFINITE;
 613         }
 614
 615         argremats = 0;
 616         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 617                 ir_node *arg = get_irn_n(insn, i);
 618
 619                 if (is_value_available(env, arg, reloader))
 620                         continue;
 621
 622                 /* we have to rematerialize the argument as well */
 623                 ++argremats;
 624                 if (argremats > 1) {
 625                         /* we only support rematerializing 1 argument at the moment,
 626                          * as multiple arguments could increase register pressure */
 627                         return REMAT_COST_INFINITE;
 628                 }
 629
 630                 costs += check_remat_conditions_costs(env, arg, reloader,
 631                                                       parentcosts + costs);
 632                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 633                         return REMAT_COST_INFINITE;
 634         }
 635
 636         return costs;
 637 }
 638
 639 /**
 640  * Re-materialize a node.
 641  *
 642  * @param senv      the spill environment
 643  * @param spilled   the node that was spilled
 644  * @param reloader  a irn that requires a reload
 645  */
 646 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 647 {
 648         int i, arity;
 649         ir_node *res;
 650         ir_node *bl;
 651         ir_node **ins;
 652
 653         if (is_Block(reloader)) {
 654                 bl = reloader;
 655         } else {
 656                 bl = get_nodes_block(reloader);
 657         }
 658
 659         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 660         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 661                 ir_node *arg = get_irn_n(spilled, i);
 662
 663                 if (is_value_available(env, arg, reloader)) {
 664                         ins[i] = arg;
 665                 } else {
 666                         ins[i] = do_remat(env, arg, reloader);
 667 #ifdef FIRM_STATISTICS
 668                         /* don't count the recursive call as remat */
 669                         env->remat_count--;
 670 #endif
 671                 }
 672         }
 673
 674         /* create a copy of the node */
 675         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 676                           get_irn_op(spilled), get_irn_mode(spilled),
 677                           get_irn_arity(spilled), ins);
 678         copy_node_attr(env->irg, spilled, res);
 679         arch_env_mark_remat(env->arch_env, res);
 680
 681         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 682
 683         if (! is_Proj(res)) {
 684                 /* insert in schedule */
 685                 sched_reset(res);
 686                 sched_add_before(reloader, res);
 687 #ifdef FIRM_STATISTICS
 688                 env->remat_count++;
 689 #endif
 690         }
 691
 692         return res;
 693 }
 694
 695 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 696 {
 697         ir_node *block = get_nodes_block(before);
 698         double   freq  = get_block_execfreq(env->exec_freq, block);
 699         (void) to_spill;
 700
 701         return env->spill_cost * freq;
 702 }
 703
 704 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 705                                        const ir_node *before)
 706 {
 707         if (be_do_remats) {
 708                 /* is the node rematerializable? */
 709                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 710                 if (costs < (unsigned) env->reload_cost)
 711                         return costs;
 712         }
 713
 714         return env->reload_cost;
 715 }
 716
 717 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 718 {
 719         ir_node      *block = get_nodes_block(before);
 720         double        freq  = get_block_execfreq(env->exec_freq, block);
 721
 722         if (be_do_remats) {
 723                 /* is the node rematerializable? */
 724                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 725                 if (costs < env->reload_cost)
 726                         return costs * freq;
 727         }
 728
 729         return env->reload_cost * freq;
 730 }
 731
 732 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 733                            const ir_node *before)
 734 {
 735         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 736 }
 737
 738 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 739                                    ir_node *block, int pos)
 740 {
 741         ir_node *before = get_block_insertion_point(block, pos);
 742         return be_get_reload_costs(env, to_spill, before);
 743 }
 744
 745 /*
 746  *  ___                     _     ____      _                 _
 747  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 748  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 749  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 750  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 751  *
 752  */
 753
 754 /**
 755  * analyzes how to best spill a node and determine costs for that
 756  */
 757 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 758 {
 759         ir_node       *to_spill = spillinfo->to_spill;
 760         const ir_node *insn     = skip_Proj_const(to_spill);
 761         ir_node       *spill_block;
 762         spill_t       *spill;
 763         double         spill_execfreq;
 764
 765         /* already calculated? */
 766         if (spillinfo->spill_costs >= 0)
 767                 return;
 768
 769         assert(!arch_irn_is(insn, dont_spill));
 770         assert(!be_is_Reload(insn));
 771
 772         /* some backends have virtual noreg/unknown nodes that are not scheduled
 773          * and simply always available.
 774          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 775          * predecessor (of a PhiM) but this test might match other things too...
 776          */
 777         if (!sched_is_scheduled(insn)) {
 778                 /* override spillinfos or create a new one */
 779                 spill_t *spill = OALLOC(&env->obst, spill_t);
 780                 spill->after = NULL;
 781                 spill->next  = NULL;
 782                 spill->spill = new_NoMem();
 783
 784                 spillinfo->spills      = spill;
 785                 spillinfo->spill_costs = 0;
 786
 787                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 788                 return;
 789         }
 790
 791         spill_block    = get_nodes_block(insn);
 792         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 793
 794         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 795                 /* TODO calculate correct costs...
 796                  * (though we can't remat this node anyway so no big problem) */
 797                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 798                 return;
 799         }
 800
 801         if (spillinfo->spills != NULL) {
 802                 spill_t *s;
 803                 double   spills_execfreq;
 804
 805                 /* calculate sum of execution frequencies of individual spills */
 806                 spills_execfreq = 0;
 807                 s               = spillinfo->spills;
 808                 for ( ; s != NULL; s = s->next) {
 809                         ir_node *spill_block = get_block(s->after);
 810                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 811
 812                         spills_execfreq += freq;
 813                 }
 814
 815                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 816                     spills_execfreq * env->spill_cost,
 817                     spill_execfreq * env->spill_cost));
 818
 819                 /* multi-/latespill is advantageous -> return*/
 820                 if (spills_execfreq < spill_execfreq) {
 821                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 822                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 823                         return;
 824                 }
 825         }
 826
 827         /* override spillinfos or create a new one */
 828         spill        = OALLOC(&env->obst, spill_t);
 829         spill->after = skip_keeps_phis(to_spill);
 830         spill->next  = NULL;
 831         spill->spill = NULL;
 832
 833         spillinfo->spills      = spill;
 834         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 835         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 836 }
 837
 838 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 839 {
 840         const spill_info_t *si = get_spillinfo(env, irn);
 841         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 842         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 843         bitset_t *reloads      = bitset_alloca(n_blocks);
 844         reloader_t *r;
 845         spill_t *s;
 846
 847         if (si == NULL)
 848                 return;
 849
 850         /* Fill the bitset with the dominance pre-order numbers
 851          * of the blocks the reloads are located in. */
 852         for (r = si->reloaders; r != NULL; r = r->next) {
 853                 ir_node *bl = get_nodes_block(r->reloader);
 854                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 855         }
 856
 857         /* Now, cancel out all the blocks that are dominated by each spill.
 858          * If the bitset is not empty after that, we have reloads that are
 859          * not dominated by any spill. */
 860         for (s = si->spills; s != NULL; s = s->next) {
 861                 ir_node *bl = get_nodes_block(s->after);
 862                 int start   = get_Block_dom_tree_pre_num(bl);
 863                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 864
 865                 bitset_clear_range(reloads, start, end);
 866         }
 867
 868         if (!bitset_is_empty(reloads))
 869                 be_add_spill(env, si->to_spill, si->to_spill);
 870 }
 871
 872 void be_insert_spills_reloads(spill_env_t *env)
 873 {
 874         const ir_exec_freq    *exec_freq = env->exec_freq;
 875         spill_info_t          *si;
 876         ir_nodeset_iterator_t  iter;
 877         ir_node               *node;
 878
 879         be_timer_push(T_RA_SPILL_APPLY);
 880
 881         /* create all phi-ms first, this is needed so, that phis, hanging on
 882            spilled phis work correctly */
 883         foreach_ir_nodeset(&env->mem_phis, node, iter) {
 884                 spill_info_t *info = get_spillinfo(env, node);
 885                 spill_node(env, info);
 886         }
 887
 888         /* process each spilled node */
 889         for (si = set_first(env->spills); si; si = set_next(env->spills)) {
 890                 reloader_t *rld;
 891                 ir_node  *to_spill        = si->to_spill;
 892                 ir_mode  *mode            = get_irn_mode(to_spill);
 893                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 894                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 895                 int       force_remat     = 0;
 896
 897                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 898
 899                 determine_spill_costs(env, si);
 900
 901                 /* determine possibility of rematerialisations */
 902                 if (be_do_remats) {
 903                         /* calculate cost savings for each indivial value when it would
 904                            be rematted instead of reloaded */
 905                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 906                                 double   freq;
 907                                 int      remat_cost;
 908                                 int      remat_cost_delta;
 909                                 ir_node *block;
 910                                 ir_node *reloader = rld->reloader;
 911
 912                                 if (rld->rematted_node != NULL) {
 913                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 914                                              rld->rematted_node, reloader));
 915                                         continue;
 916                                 }
 917                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 918                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 919                                              reloader));
 920                                         all_remat_costs = REMAT_COST_INFINITE;
 921                                         continue;
 922                                 }
 923
 924                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 925                                                                            reloader, 0);
 926                                 if (remat_cost >= REMAT_COST_INFINITE) {
 927                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 928                                              reloader));
 929                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 930                                         all_remat_costs       = REMAT_COST_INFINITE;
 931                                         continue;
 932                                 }
 933
 934                                 remat_cost_delta      = remat_cost - env->reload_cost;
 935                                 rld->remat_cost_delta = remat_cost_delta;
 936                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 937                                 freq                  = get_block_execfreq(exec_freq, block);
 938                                 all_remat_costs      += remat_cost_delta * freq;
 939                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 940                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 941                                      remat_cost_delta * freq));
 942                         }
 943                         if (all_remat_costs < REMAT_COST_INFINITE) {
 944                                 /* we don't need the costs for the spill if we can remat
 945                                    all reloaders */
 946                                 all_remat_costs -= si->spill_costs;
 947
 948                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 949                                      env->spill_cost, si->spill_costs));
 950                         }
 951
 952                         if (all_remat_costs < 0) {
 953                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 954                                      all_remat_costs));
 955                                 force_remat = 1;
 956                         }
 957                 }
 958
 959                 /* go through all reloads for this spill */
 960                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 961                         ir_node *copy; /* a reload is a "copy" of the original value */
 962
 963                         if (rld->rematted_node != NULL) {
 964                                 copy = rld->rematted_node;
 965                                 sched_add_before(rld->reloader, copy);
 966                         } else if (be_do_remats &&
 967                                         (force_remat || rld->remat_cost_delta < 0)) {
 968                                 copy = do_remat(env, to_spill, rld->reloader);
 969                         } else {
 970                                 /* make sure we have a spill */
 971                                 spill_node(env, si);
 972
 973                                 /* create a reload, use the first spill for now SSA
 974                                  * reconstruction for memory comes below */
 975                                 assert(si->spills != NULL);
 976                                 copy = be_reload(si->reload_cls, rld->reloader, mode,
 977                                                  si->spills->spill);
 978 #ifdef FIRM_STATISTICS
 979                                 env->reload_count++;
 980 #endif
 981                         }
 982
 983                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 984                              copy, to_spill, rld->reloader));
 985                         ARR_APP1(ir_node*, copies, copy);
 986                 }
 987
 988                 /* if we had any reloads or remats, then we need to reconstruct the
 989                  * SSA form for the spilled value */
 990                 if (ARR_LEN(copies) > 0) {
 991                         be_ssa_construction_env_t senv;
 992                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 993
 994                         be_ssa_construction_init(&senv, env->irg);
 995                         be_ssa_construction_add_copy(&senv, to_spill);
 996                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 997                         be_ssa_construction_fix_users(&senv, to_spill);
 998
 999 #if 0
1000                         /* no need to enable this as long as we invalidate liveness
1001                            after this function... */
1002                         be_ssa_construction_update_liveness_phis(&senv);
1003                         be_liveness_update(to_spill);
1004                         len = ARR_LEN(copies);
1005                         for (i = 0; i < len; ++i) {
1006                                 be_liveness_update(lv, copies[i]);
1007                         }
1008 #endif
1009                         be_ssa_construction_destroy(&senv);
1010                 }
1011                 /* need to reconstruct SSA form if we had multiple spills */
1012                 if (si->spills != NULL && si->spills->next != NULL) {
1013                         spill_t *spill;
1014                         int      spill_count = 0;
1015
1016                         be_ssa_construction_env_t senv;
1017
1018                         be_ssa_construction_init(&senv, env->irg);
1019                         spill = si->spills;
1020                         for ( ; spill != NULL; spill = spill->next) {
1021                                 /* maybe we rematerialized the value and need no spill */
1022                                 if (spill->spill == NULL)
1023                                         continue;
1024                                 be_ssa_construction_add_copy(&senv, spill->spill);
1025                                 spill_count++;
1026                         }
1027                         if (spill_count > 1) {
1028                                 /* all reloads are attached to the first spill, fix them now */
1029                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1030                         }
1031
1032                         be_ssa_construction_destroy(&senv);
1033                 }
1034
1035                 DEL_ARR_F(copies);
1036                 si->reloaders = NULL;
1037         }
1038
1039         stat_ev_dbl("spill_spills", env->spill_count);
1040         stat_ev_dbl("spill_reloads", env->reload_count);
1041         stat_ev_dbl("spill_remats", env->remat_count);
1042         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1043
1044         /* Matze: In theory be_ssa_construction should take care of the liveness...
1045          * try to disable this again in the future */
1046         be_liveness_invalidate(be_get_irg_liveness(env->irg));
1047
1048         be_remove_dead_nodes_from_schedule(env->irg);
1049
1050         be_timer_pop(T_RA_SPILL_APPLY);
1051 }
1052
1053 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill);
1054 void be_init_spill(void)
1055 {
1056         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1057 }