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