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