nsz Git - libfirm/blob - ir/be/bespillutil.c

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       implementation of the spill/reload placement abstraction layer
  23  * @author      Daniel Grund, Sebastian Hack, Matthias Braun
  24  * @date        29.09.2005
  25  * @version     $Id$
  26  */
  27 #include "config.h"
  28
  29 #include <stdlib.h>
  30 #include <stdbool.h>
  31
  32 #include "pset.h"
  33 #include "irnode_t.h"
  34 #include "ircons_t.h"
  35 #include "iredges_t.h"
  36 #include "irbackedge_t.h"
  37 #include "irprintf.h"
  38 #include "ident_t.h"
  39 #include "type_t.h"
  40 #include "entity_t.h"
  41 #include "debug.h"
  42 #include "irgwalk.h"
  43 #include "array.h"
  44 #include "pdeq.h"
  45 #include "execfreq.h"
  46 #include "irnodeset.h"
  47 #include "error.h"
  48
  49 #include "bearch.h"
  50 #include "belive_t.h"
  51 #include "besched.h"
  52 #include "bespill.h"
  53 #include "bespillutil.h"
  54 #include "belive_t.h"
  55 #include "benode.h"
  56 #include "bechordal_t.h"
  57 #include "bestatevent.h"
  58 #include "bessaconstr.h"
  59 #include "beirg.h"
  60 #include "beirgmod.h"
  61 #include "beintlive_t.h"
  62 #include "bemodule.h"
  63 #include "be_t.h"
  64
  65 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  66
  67 #define REMAT_COST_INFINITE  1000
  68
  69 typedef struct reloader_t reloader_t;
  70 struct reloader_t {
  71         reloader_t *next;
  72         ir_node    *can_spill_after;
  73         ir_node    *reloader;
  74         ir_node    *rematted_node;
  75         int         remat_cost_delta; /** costs needed for rematerialization,
  76                                            compared to placing a reload */
  77 };
  78
  79 typedef struct spill_t spill_t;
  80 struct spill_t {
  81         spill_t *next;
  82         ir_node *after;  /**< spill has to be placed after this node (or earlier) */
  83         ir_node *spill;
  84 };
  85
  86 typedef struct spill_info_t spill_info_t;
  87 struct spill_info_t {
  88         ir_node    *to_spill;  /**< the value that should get spilled */
  89         reloader_t *reloaders; /**< list of places where the value should get
  90                                     reloaded */
  91         spill_t    *spills;    /**< list of latest places where spill must be
  92                                     placed */
  93         double      spill_costs; /**< costs needed for spilling the value */
  94         const arch_register_class_t *reload_cls; /** the register class in which the
  95                                                      reload should be placed */
  96 };
  97
  98 struct spill_env_t {
  99         const arch_env_t *arch_env;
 100         ir_graph         *irg;
 101         struct obstack    obst;
 102         int               spill_cost;     /**< the cost of a single spill node */
 103         int               reload_cost;    /**< the cost of a reload node */
 104         set              *spills;         /**< all spill_info_t's, which must be
 105                                                placed */
 106         ir_nodeset_t      mem_phis;       /**< set of all spilled phis. */
 107         ir_exec_freq     *exec_freq;
 108
 109 #ifdef FIRM_STATISTICS
 110         unsigned          spill_count;
 111         unsigned          reload_count;
 112         unsigned          remat_count;
 113         unsigned          spilled_phi_count;
 114 #endif
 115 };
 116
 117 /**
 118  * Compare two spill infos.
 119  */
 120 static int cmp_spillinfo(const void *x, const void *y, size_t size)
 121 {
 122         const spill_info_t *xx = (const spill_info_t*)x;
 123         const spill_info_t *yy = (const spill_info_t*)y;
 124         (void) size;
 125
 126         return xx->to_spill != yy->to_spill;
 127 }
 128
 129 /**
 130  * Returns spill info for a specific value (the value that is to be spilled)
 131  */
 132 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
 133 {
 134         spill_info_t info, *res;
 135         int hash = hash_irn(value);
 136
 137         info.to_spill = value;
 138         res = (spill_info_t*)set_find(env->spills, &info, sizeof(info), hash);
 139
 140         if (res == NULL) {
 141                 info.reloaders   = NULL;
 142                 info.spills      = NULL;
 143                 info.spill_costs = -1;
 144                 info.reload_cls  = NULL;
 145                 res = (spill_info_t*)set_insert(env->spills, &info, sizeof(info), hash);
 146         }
 147
 148         return res;
 149 }
 150
 151 spill_env_t *be_new_spill_env(ir_graph *irg)
 152 {
 153         const arch_env_t *arch_env = be_get_irg_arch_env(irg);
 154
 155         spill_env_t *env = XMALLOC(spill_env_t);
 156         env->spills         = new_set(cmp_spillinfo, 1024);
 157         env->irg            = irg;
 158         env->arch_env       = arch_env;
 159         ir_nodeset_init(&env->mem_phis);
 160         env->spill_cost     = arch_env->spill_cost;
 161         env->reload_cost    = arch_env->reload_cost;
 162         env->exec_freq      = be_get_irg_exec_freq(irg);
 163         obstack_init(&env->obst);
 164
 165 #ifdef FIRM_STATISTICS
 166         env->spill_count       = 0;
 167         env->reload_count      = 0;
 168         env->remat_count       = 0;
 169         env->spilled_phi_count = 0;
 170 #endif
 171
 172         return env;
 173 }
 174
 175 void be_delete_spill_env(spill_env_t *env)
 176 {
 177         del_set(env->spills);
 178         ir_nodeset_destroy(&env->mem_phis);
 179         obstack_free(&env->obst, NULL);
 180         free(env);
 181 }
 182
 183 /*
 184  *  ____  _                  ____      _                 _
 185  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 186  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 187  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 188  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 189  *
 190  */
 191
 192 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 193 {
 194         spill_info_t  *spill_info = get_spillinfo(env, to_spill);
 195         spill_t       *spill;
 196         spill_t       *s;
 197         spill_t       *last;
 198
 199         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 200         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 201
 202         /* Just for safety make sure that we do not insert the spill in front of a phi */
 203         assert(!is_Phi(sched_next(after)));
 204
 205         /* spills that are dominated by others are not needed */
 206         last = NULL;
 207         s    = spill_info->spills;
 208         for ( ; s != NULL; s = s->next) {
 209                 /* no need to add this spill if it is dominated by another */
 210                 if (value_dominates(s->after, after)) {
 211                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 212                         return;
 213                 }
 214                 /* remove spills that we dominate */
 215                 if (value_dominates(after, s->after)) {
 216                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 217                         if (last != NULL) {
 218                                 last->next         = s->next;
 219                         } else {
 220                                 spill_info->spills = s->next;
 221                         }
 222                 } else {
 223                         last = s;
 224                 }
 225         }
 226
 227         spill         = OALLOC(&env->obst, spill_t);
 228         spill->after  = after;
 229         spill->next   = spill_info->spills;
 230         spill->spill  = NULL;
 231
 232         spill_info->spills = spill;
 233 }
 234
 235 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 236                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 237                 int allow_remat)
 238 {
 239         spill_info_t  *info;
 240         reloader_t    *rel;
 241
 242         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 243
 244         info = get_spillinfo(env, to_spill);
 245
 246         if (is_Phi(to_spill)) {
 247                 int i, arity;
 248
 249                 /* create spillinfos for the phi arguments */
 250                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 251                         ir_node *arg = get_irn_n(to_spill, i);
 252                         get_spillinfo(env, arg);
 253                 }
 254         }
 255
 256         assert(!is_Proj(before) && !be_is_Keep(before));
 257
 258         /* put reload into list */
 259         rel                   = OALLOC(&env->obst, reloader_t);
 260         rel->next             = info->reloaders;
 261         rel->reloader         = before;
 262         rel->rematted_node    = NULL;
 263         rel->can_spill_after  = can_spill_after;
 264         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 265
 266         info->reloaders  = rel;
 267         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 268         info->reload_cls = reload_cls;
 269
 270         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 271                 to_spill, before, allow_remat ? "" : " not"));
 272 }
 273
 274 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 275                    const arch_register_class_t *reload_cls, int allow_remat)
 276 {
 277         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 278
 279 }
 280
 281 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 282 {
 283         ir_node *last = sched_last(block);
 284
 285         /* we might have keeps behind the jump... */
 286         while (be_is_Keep(last)) {
 287                 last = sched_prev(last);
 288                 assert(!sched_is_end(last));
 289         }
 290
 291         assert(is_cfop(last));
 292
 293         /* add the reload before the (cond-)jump */
 294         return last;
 295 }
 296
 297 static ir_node *skip_keeps_phis(ir_node *node)
 298 {
 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 = skip_keeps_phis(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 = new_r_NoMem(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                 ir_node *block = get_block(after);
 419
 420                 after = skip_keeps_phis(after);
 421
 422                 spill->spill = be_spill(block, to_spill);
 423                 sched_add_after(skip_Proj(after), spill->spill);
 424                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 425 #ifdef FIRM_STATISTICS
 426                 env->spill_count++;
 427 #endif
 428         }
 429         DBG((dbg, LEVEL_1, "\n"));
 430 }
 431
 432 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 433
 434 /**
 435  * If the first usage of a Phi result would be out of memory
 436  * there is no sense in allocating a register for it.
 437  * Thus we spill it and all its operands to the same spill slot.
 438  * Therefore the phi/dataB becomes a phi/Memory
 439  *
 440  * @param senv      the spill environment
 441  * @param phi       the Phi node that should be spilled
 442  * @param ctx_irn   an user of the spilled node
 443  */
 444 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 445 {
 446         ir_graph *irg   = env->irg;
 447         ir_node  *phi   = spillinfo->to_spill;
 448         ir_node  *block = get_nodes_block(phi);
 449         ir_node  *unknown;
 450         ir_node **ins;
 451         spill_t  *spill;
 452         int       i;
 453         int       arity;
 454
 455         assert(is_Phi(phi));
 456         assert(!get_opt_cse());
 457         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 458
 459         /* build a new PhiM */
 460         arity   = get_irn_arity(phi);
 461         ins     = ALLOCAN(ir_node*, arity);
 462         unknown = new_r_Unknown(irg, mode_M);
 463         for (i = 0; i < arity; ++i) {
 464                 ins[i] = unknown;
 465         }
 466
 467         /* override or replace spills list... */
 468         spill         = OALLOC(&env->obst, spill_t);
 469         spill->after  = skip_keeps_phis(phi);
 470         spill->spill  = be_new_Phi(block, arity, ins, mode_M, NULL);
 471         spill->next   = NULL;
 472         sched_add_after(block, spill->spill);
 473
 474         spillinfo->spills = spill;
 475 #ifdef FIRM_STATISTICS
 476         env->spilled_phi_count++;
 477 #endif
 478
 479         for (i = 0; i < arity; ++i) {
 480                 ir_node      *arg      = get_irn_n(phi, i);
 481                 spill_info_t *arg_info = get_spillinfo(env, arg);
 482
 483                 determine_spill_costs(env, arg_info);
 484                 spill_node(env, arg_info);
 485
 486                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 487         }
 488         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 489              spill->spill));
 490 }
 491
 492 /**
 493  * Spill a node.
 494  *
 495  * @param senv      the spill environment
 496  * @param to_spill  the node that should be spilled
 497  */
 498 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 499 {
 500         ir_node *to_spill;
 501
 502         /* node is already spilled */
 503         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 504                 return;
 505
 506         to_spill = spillinfo->to_spill;
 507
 508         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 509                 spill_phi(env, spillinfo);
 510         } else {
 511                 spill_irn(env, spillinfo);
 512         }
 513 }
 514
 515 /*
 516  *
 517  *  ____                      _            _       _ _
 518  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 519  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 520  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 521  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 522  *
 523  */
 524
 525 /**
 526  * Tests whether value @p arg is available before node @p reloader
 527  * @returns 1 if value is available, 0 otherwise
 528  */
 529 static int is_value_available(spill_env_t *env, const ir_node *arg,
 530                               const ir_node *reloader)
 531 {
 532         if (is_Unknown(arg) || is_NoMem(arg))
 533                 return 1;
 534
 535         if (be_is_Spill(skip_Proj_const(arg)))
 536                 return 1;
 537
 538         if (arg == get_irg_frame(env->irg))
 539                 return 1;
 540
 541         (void)reloader;
 542
 543         if (get_irn_mode(arg) == mode_T)
 544                 return 0;
 545
 546         /*
 547          * Ignore registers are always available
 548          */
 549         if (arch_irn_is_ignore(arg))
 550                 return 1;
 551
 552         return 0;
 553 }
 554
 555 /**
 556  * Check if a node is rematerializable. This tests for the following conditions:
 557  *
 558  * - The node itself is rematerializable
 559  * - All arguments of the node are available or also rematerialisable
 560  * - The costs for the rematerialisation operation is less or equal a limit
 561  *
 562  * Returns the costs needed for rematerialisation or something
 563  * >= REMAT_COST_INFINITE if remat is not possible.
 564  */
 565 static int check_remat_conditions_costs(spill_env_t *env,
 566                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 567 {
 568         int i, arity;
 569         int argremats;
 570         int costs = 0;
 571         const ir_node *insn = skip_Proj_const(spilled);
 572
 573         assert(!be_is_Spill(insn));
 574         if (!arch_irn_is(insn, rematerializable))
 575                 return REMAT_COST_INFINITE;
 576
 577         if (be_is_Reload(insn)) {
 578                 costs += 2;
 579         } else {
 580                 costs += arch_get_op_estimated_cost(insn);
 581         }
 582         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 583                 return REMAT_COST_INFINITE;
 584         }
 585         /* never rematerialize a node which modifies the flags.
 586          * (would be better to test wether the flags are actually live at point
 587          * reloader...)
 588          */
 589         if (arch_irn_is(insn, modify_flags)) {
 590                 return REMAT_COST_INFINITE;
 591         }
 592
 593         argremats = 0;
 594         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 595                 ir_node *arg = get_irn_n(insn, i);
 596
 597                 if (is_value_available(env, arg, reloader))
 598                         continue;
 599
 600                 /* we have to rematerialize the argument as well */
 601                 ++argremats;
 602                 if (argremats > 1) {
 603                         /* we only support rematerializing 1 argument at the moment,
 604                          * as multiple arguments could increase register pressure */
 605                         return REMAT_COST_INFINITE;
 606                 }
 607
 608                 costs += check_remat_conditions_costs(env, arg, reloader,
 609                                                       parentcosts + costs);
 610                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 611                         return REMAT_COST_INFINITE;
 612         }
 613
 614         return costs;
 615 }
 616
 617 /**
 618  * Re-materialize a node.
 619  *
 620  * @param senv      the spill environment
 621  * @param spilled   the node that was spilled
 622  * @param reloader  a irn that requires a reload
 623  */
 624 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 625 {
 626         int i, arity;
 627         ir_node *res;
 628         ir_node *bl;
 629         ir_node **ins;
 630
 631         if (is_Block(reloader)) {
 632                 bl = reloader;
 633         } else {
 634                 bl = get_nodes_block(reloader);
 635         }
 636
 637         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 638         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 639                 ir_node *arg = get_irn_n(spilled, i);
 640
 641                 if (is_value_available(env, arg, reloader)) {
 642                         ins[i] = arg;
 643                 } else {
 644                         ins[i] = do_remat(env, arg, reloader);
 645 #ifdef FIRM_STATISTICS
 646                         /* don't count the recursive call as remat */
 647                         env->remat_count--;
 648 #endif
 649                 }
 650         }
 651
 652         /* create a copy of the node */
 653         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 654                           get_irn_op(spilled), get_irn_mode(spilled),
 655                           get_irn_arity(spilled), ins);
 656         copy_node_attr(env->irg, spilled, res);
 657         arch_env_mark_remat(env->arch_env, res);
 658
 659         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 660
 661         if (! is_Proj(res)) {
 662                 /* insert in schedule */
 663                 sched_reset(res);
 664                 sched_add_before(reloader, res);
 665 #ifdef FIRM_STATISTICS
 666                 env->remat_count++;
 667 #endif
 668         }
 669
 670         return res;
 671 }
 672
 673 double be_get_spill_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(env->exec_freq, block);
 677         (void) to_spill;
 678
 679         return env->spill_cost * freq;
 680 }
 681
 682 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 683                                        const ir_node *before)
 684 {
 685         if (be_do_remats) {
 686                 /* is the node rematerializable? */
 687                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 688                 if (costs < (unsigned) env->reload_cost)
 689                         return costs;
 690         }
 691
 692         return env->reload_cost;
 693 }
 694
 695 double be_get_reload_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
 700         if (be_do_remats) {
 701                 /* is the node rematerializable? */
 702                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 703                 if (costs < env->reload_cost)
 704                         return costs * freq;
 705         }
 706
 707         return env->reload_cost * freq;
 708 }
 709
 710 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 711                            const ir_node *before)
 712 {
 713         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 714 }
 715
 716 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 717                                    ir_node *block, int pos)
 718 {
 719         ir_node *before = get_block_insertion_point(block, pos);
 720         return be_get_reload_costs(env, to_spill, before);
 721 }
 722
 723 /*
 724  *  ___                     _     ____      _                 _
 725  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 726  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 727  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 728  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 729  *
 730  */
 731
 732 /**
 733  * analyzes how to best spill a node and determine costs for that
 734  */
 735 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 736 {
 737         ir_node       *to_spill = spillinfo->to_spill;
 738         const ir_node *insn     = skip_Proj_const(to_spill);
 739         ir_node       *spill_block;
 740         spill_t       *spill;
 741         double         spill_execfreq;
 742
 743         /* already calculated? */
 744         if (spillinfo->spill_costs >= 0)
 745                 return;
 746
 747         assert(!arch_irn_is(insn, dont_spill));
 748         assert(!be_is_Reload(insn));
 749
 750         /* some backends have virtual noreg/unknown nodes that are not scheduled
 751          * and simply always available.
 752          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 753          * predecessor (of a PhiM) but this test might match other things too...
 754          */
 755         if (!sched_is_scheduled(insn)) {
 756                 ir_graph *irg = get_irn_irg(to_spill);
 757                 /* override spillinfos or create a new one */
 758                 spill_t *spill = OALLOC(&env->obst, spill_t);
 759                 spill->after = NULL;
 760                 spill->next  = NULL;
 761                 spill->spill = new_r_NoMem(irg);
 762
 763                 spillinfo->spills      = spill;
 764                 spillinfo->spill_costs = 0;
 765
 766                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 767                 return;
 768         }
 769
 770         spill_block    = get_nodes_block(insn);
 771         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 772
 773         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 774                 /* TODO calculate correct costs...
 775                  * (though we can't remat this node anyway so no big problem) */
 776                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 777                 return;
 778         }
 779
 780         if (spillinfo->spills != NULL) {
 781                 spill_t *s;
 782                 double   spills_execfreq;
 783
 784                 /* calculate sum of execution frequencies of individual spills */
 785                 spills_execfreq = 0;
 786                 s               = spillinfo->spills;
 787                 for ( ; s != NULL; s = s->next) {
 788                         ir_node *spill_block = get_block(s->after);
 789                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 790
 791                         spills_execfreq += freq;
 792                 }
 793
 794                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 795                     spills_execfreq * env->spill_cost,
 796                     spill_execfreq * env->spill_cost));
 797
 798                 /* multi-/latespill is advantageous -> return*/
 799                 if (spills_execfreq < spill_execfreq) {
 800                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 801                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 802                         return;
 803                 }
 804         }
 805
 806         /* override spillinfos or create a new one */
 807         spill        = OALLOC(&env->obst, spill_t);
 808         spill->after = skip_keeps_phis(to_spill);
 809         spill->next  = NULL;
 810         spill->spill = NULL;
 811
 812         spillinfo->spills      = spill;
 813         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 814         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 815 }
 816
 817 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 818 {
 819         const spill_info_t *si = get_spillinfo(env, irn);
 820         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 821         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 822         bitset_t *reloads      = bitset_alloca(n_blocks);
 823         reloader_t *r;
 824         spill_t *s;
 825
 826         if (si == NULL)
 827                 return;
 828
 829         /* Fill the bitset with the dominance pre-order numbers
 830          * of the blocks the reloads are located in. */
 831         for (r = si->reloaders; r != NULL; r = r->next) {
 832                 ir_node *bl = get_nodes_block(r->reloader);
 833                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 834         }
 835
 836         /* Now, cancel out all the blocks that are dominated by each spill.
 837          * If the bitset is not empty after that, we have reloads that are
 838          * not dominated by any spill. */
 839         for (s = si->spills; s != NULL; s = s->next) {
 840                 ir_node *bl = get_nodes_block(s->after);
 841                 int start   = get_Block_dom_tree_pre_num(bl);
 842                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 843
 844                 bitset_clear_range(reloads, start, end);
 845         }
 846
 847         if (!bitset_is_empty(reloads))
 848                 be_add_spill(env, si->to_spill, si->to_spill);
 849 }
 850
 851 void be_insert_spills_reloads(spill_env_t *env)
 852 {
 853         const ir_exec_freq    *exec_freq = env->exec_freq;
 854         spill_info_t          *si;
 855         ir_nodeset_iterator_t  iter;
 856         ir_node               *node;
 857
 858         be_timer_push(T_RA_SPILL_APPLY);
 859
 860         /* create all phi-ms first, this is needed so, that phis, hanging on
 861            spilled phis work correctly */
 862         foreach_ir_nodeset(&env->mem_phis, node, iter) {
 863                 spill_info_t *info = get_spillinfo(env, node);
 864                 spill_node(env, info);
 865         }
 866
 867         /* process each spilled node */
 868         foreach_set(env->spills, spill_info_t*, si) {
 869                 reloader_t *rld;
 870                 ir_node  *to_spill        = si->to_spill;
 871                 ir_mode  *mode            = get_irn_mode(to_spill);
 872                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 873                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 874                 int       force_remat     = 0;
 875
 876                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 877
 878                 determine_spill_costs(env, si);
 879
 880                 /* determine possibility of rematerialisations */
 881                 if (be_do_remats) {
 882                         /* calculate cost savings for each indivial value when it would
 883                            be rematted instead of reloaded */
 884                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 885                                 double   freq;
 886                                 int      remat_cost;
 887                                 int      remat_cost_delta;
 888                                 ir_node *block;
 889                                 ir_node *reloader = rld->reloader;
 890
 891                                 if (rld->rematted_node != NULL) {
 892                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 893                                              rld->rematted_node, reloader));
 894                                         continue;
 895                                 }
 896                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 897                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 898                                              reloader));
 899                                         all_remat_costs = REMAT_COST_INFINITE;
 900                                         continue;
 901                                 }
 902
 903                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 904                                                                            reloader, 0);
 905                                 if (remat_cost >= REMAT_COST_INFINITE) {
 906                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 907                                              reloader));
 908                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 909                                         all_remat_costs       = REMAT_COST_INFINITE;
 910                                         continue;
 911                                 }
 912
 913                                 remat_cost_delta      = remat_cost - env->reload_cost;
 914                                 rld->remat_cost_delta = remat_cost_delta;
 915                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 916                                 freq                  = get_block_execfreq(exec_freq, block);
 917                                 all_remat_costs      += remat_cost_delta * freq;
 918                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 919                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 920                                      remat_cost_delta * freq));
 921                         }
 922                         if (all_remat_costs < REMAT_COST_INFINITE) {
 923                                 /* we don't need the costs for the spill if we can remat
 924                                    all reloaders */
 925                                 all_remat_costs -= si->spill_costs;
 926
 927                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 928                                      env->spill_cost, si->spill_costs));
 929                         }
 930
 931                         if (all_remat_costs < 0) {
 932                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 933                                      all_remat_costs));
 934                                 force_remat = 1;
 935                         }
 936                 }
 937
 938                 /* go through all reloads for this spill */
 939                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 940                         ir_node *copy; /* a reload is a "copy" of the original value */
 941
 942                         if (rld->rematted_node != NULL) {
 943                                 copy = rld->rematted_node;
 944                                 sched_add_before(rld->reloader, copy);
 945                         } else if (be_do_remats &&
 946                                         (force_remat || rld->remat_cost_delta < 0)) {
 947                                 copy = do_remat(env, to_spill, rld->reloader);
 948                         } else {
 949                                 /* make sure we have a spill */
 950                                 spill_node(env, si);
 951
 952                                 /* create a reload, use the first spill for now SSA
 953                                  * reconstruction for memory comes below */
 954                                 assert(si->spills != NULL);
 955                                 copy = be_reload(si->reload_cls, rld->reloader, mode,
 956                                                  si->spills->spill);
 957 #ifdef FIRM_STATISTICS
 958                                 env->reload_count++;
 959 #endif
 960                         }
 961
 962                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 963                              copy, to_spill, rld->reloader));
 964                         ARR_APP1(ir_node*, copies, copy);
 965                 }
 966
 967                 /* if we had any reloads or remats, then we need to reconstruct the
 968                  * SSA form for the spilled value */
 969                 if (ARR_LEN(copies) > 0) {
 970                         be_ssa_construction_env_t senv;
 971                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 972
 973                         be_ssa_construction_init(&senv, env->irg);
 974                         be_ssa_construction_add_copy(&senv, to_spill);
 975                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 976                         be_ssa_construction_fix_users(&senv, to_spill);
 977
 978 #if 0
 979                         /* no need to enable this as long as we invalidate liveness
 980                            after this function... */
 981                         be_ssa_construction_update_liveness_phis(&senv);
 982                         be_liveness_update(to_spill);
 983                         len = ARR_LEN(copies);
 984                         for (i = 0; i < len; ++i) {
 985                                 be_liveness_update(lv, copies[i]);
 986                         }
 987 #endif
 988                         be_ssa_construction_destroy(&senv);
 989                 }
 990                 /* need to reconstruct SSA form if we had multiple spills */
 991                 if (si->spills != NULL && si->spills->next != NULL) {
 992                         spill_t *spill;
 993                         int      spill_count = 0;
 994
 995                         be_ssa_construction_env_t senv;
 996
 997                         be_ssa_construction_init(&senv, env->irg);
 998                         spill = si->spills;
 999                         for ( ; spill != NULL; spill = spill->next) {
1000                                 /* maybe we rematerialized the value and need no spill */
1001                                 if (spill->spill == NULL)
1002                                         continue;
1003                                 be_ssa_construction_add_copy(&senv, spill->spill);
1004                                 spill_count++;
1005                         }
1006                         if (spill_count > 1) {
1007                                 /* all reloads are attached to the first spill, fix them now */
1008                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1009                         }
1010
1011                         be_ssa_construction_destroy(&senv);
1012                 }
1013
1014                 DEL_ARR_F(copies);
1015                 si->reloaders = NULL;
1016         }
1017
1018         stat_ev_dbl("spill_spills", env->spill_count);
1019         stat_ev_dbl("spill_reloads", env->reload_count);
1020         stat_ev_dbl("spill_remats", env->remat_count);
1021         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1022
1023         /* Matze: In theory be_ssa_construction should take care of the liveness...
1024          * try to disable this again in the future */
1025         be_liveness_invalidate(be_get_irg_liveness(env->irg));
1026
1027         be_remove_dead_nodes_from_schedule(env->irg);
1028
1029         be_timer_pop(T_RA_SPILL_APPLY);
1030 }
1031
1032 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill);
1033 void be_init_spill(void)
1034 {
1035         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1036 }