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         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 = (spill_info_t*)set_find(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                 res = (spill_info_t*)set_insert(env->spills, &info, sizeof(info), hash);
 144         }
 145
 146         return res;
 147 }
 148
 149 spill_env_t *be_new_spill_env(ir_graph *irg)
 150 {
 151         const arch_env_t *arch_env = be_get_irg_arch_env(irg);
 152
 153         spill_env_t *env = XMALLOC(spill_env_t);
 154         env->spills         = new_set(cmp_spillinfo, 1024);
 155         env->irg            = irg;
 156         env->arch_env       = arch_env;
 157         ir_nodeset_init(&env->mem_phis);
 158         env->spill_cost     = arch_env->spill_cost;
 159         env->reload_cost    = arch_env->reload_cost;
 160         env->exec_freq      = be_get_irg_exec_freq(irg);
 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         ir_nodeset_destroy(&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                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 233                 int allow_remat)
 234 {
 235         spill_info_t  *info;
 236         reloader_t    *rel;
 237
 238         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 239
 240         info = get_spillinfo(env, to_spill);
 241
 242         if (is_Phi(to_spill)) {
 243                 int i, arity;
 244
 245                 /* create spillinfos for the phi arguments */
 246                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 247                         ir_node *arg = get_irn_n(to_spill, i);
 248                         get_spillinfo(env, arg);
 249                 }
 250         }
 251
 252         assert(!is_Proj(before) && !be_is_Keep(before));
 253
 254         /* put reload into list */
 255         rel                   = OALLOC(&env->obst, reloader_t);
 256         rel->next             = info->reloaders;
 257         rel->reloader         = before;
 258         rel->rematted_node    = NULL;
 259         rel->can_spill_after  = can_spill_after;
 260         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 261
 262         info->reloaders  = rel;
 263         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 264         info->reload_cls = reload_cls;
 265
 266         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 267                 to_spill, before, allow_remat ? "" : " not"));
 268 }
 269
 270 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 271                    const arch_register_class_t *reload_cls, int allow_remat)
 272 {
 273         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 274
 275 }
 276
 277 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 278 {
 279         ir_node *last = sched_last(block);
 280
 281         /* we might have keeps behind the jump... */
 282         while (be_is_Keep(last)) {
 283                 last = sched_prev(last);
 284                 assert(!sched_is_end(last));
 285         }
 286
 287         assert(is_cfop(last));
 288
 289         /* add the reload before the (cond-)jump */
 290         return last;
 291 }
 292
 293 /**
 294  * determine final spill position: it should be after all phis, keep nodes
 295  * and behind nodes marked as prolog
 296  */
 297 static ir_node *determine_spill_point(ir_node *node)
 298 {
 299         node = skip_Proj(node);
 300         while (true) {
 301                 ir_node *next = sched_next(node);
 302                 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
 303                         break;
 304                 node = next;
 305         }
 306         return node;
 307 }
 308
 309 /**
 310  * Returns the point at which you can insert a node that should be executed
 311  * before block @p block when coming from pred @p pos.
 312  */
 313 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 314 {
 315         ir_node *predblock;
 316
 317         /* simply add the reload to the beginning of the block if we only have 1
 318          * predecessor. We don't need to check for phis as there can't be any in a
 319          * block with only 1 pred. */
 320         if (get_Block_n_cfgpreds(block) == 1) {
 321                 assert(!is_Phi(sched_first(block)));
 322                 return sched_first(block);
 323         }
 324
 325         /* We have to reload the value in pred-block */
 326         predblock = get_Block_cfgpred_block(block, pos);
 327         return be_get_end_of_block_insertion_point(predblock);
 328 }
 329
 330 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
 331                           const ir_node *block,
 332                           const arch_register_class_t *reload_cls,
 333                           int allow_remat)
 334 {
 335         ir_node *before = be_get_end_of_block_insertion_point(block);
 336         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 337 }
 338
 339 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 340                            int pos, const arch_register_class_t *reload_cls,
 341                            int allow_remat)
 342 {
 343         ir_node *before = get_block_insertion_point(block, pos);
 344         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 345 }
 346
 347 void be_spill_phi(spill_env_t *env, ir_node *node)
 348 {
 349         ir_node *block;
 350         int i, arity;
 351
 352         assert(is_Phi(node));
 353
 354         ir_nodeset_insert(&env->mem_phis, node);
 355
 356         /* create spills for the phi arguments */
 357         block = get_nodes_block(node);
 358         for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 359                 ir_node *arg = get_irn_n(node, i);
 360                 ir_node *insert;
 361
 362                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 363                  * and simply always available. */
 364                 if (!sched_is_scheduled(arg)) {
 365                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 366                         insert = be_get_end_of_block_insertion_point(pred_block);
 367                         insert = sched_prev(insert);
 368                 } else {
 369                         insert = determine_spill_point(arg);
 370                 }
 371
 372                 be_add_spill(env, arg, insert);
 373         }
 374 }
 375
 376 /*
 377  *   ____                _         ____        _ _ _
 378  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 379  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 380  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 381  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 382  *                                      |_|
 383  */
 384
 385 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 386
 387 /**
 388  * Creates a spill.
 389  *
 390  * @param senv      the spill environment
 391  * @param irn       the node that should be spilled
 392  * @param ctx_irn   an user of the spilled node
 393  *
 394  * @return a be_Spill node
 395  */
 396 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 397 {
 398         ir_node       *to_spill = spillinfo->to_spill;
 399         const ir_node *insn     = skip_Proj_const(to_spill);
 400         spill_t *spill;
 401
 402         /* determine_spill_costs must have been run before */
 403         assert(spillinfo->spill_costs >= 0);
 404
 405         /* some backends have virtual noreg/unknown nodes that are not scheduled
 406          * and simply always available. */
 407         if (!sched_is_scheduled(insn)) {
 408                 /* override spillinfos or create a new one */
 409                 ir_graph *irg = get_irn_irg(to_spill);
 410                 spillinfo->spills->spill = get_irg_no_mem(irg);
 411                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 412                 return;
 413         }
 414
 415         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 416         spill = spillinfo->spills;
 417         for ( ; spill != NULL; spill = spill->next) {
 418                 ir_node *after = spill->after;
 419                 after = determine_spill_point(after);
 420
 421                 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
 422                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 423                 env->spill_count++;
 424         }
 425         DBG((dbg, LEVEL_1, "\n"));
 426 }
 427
 428 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 429
 430 /**
 431  * If the first usage of a Phi result would be out of memory
 432  * there is no sense in allocating a register for it.
 433  * Thus we spill it and all its operands to the same spill slot.
 434  * Therefore the phi/dataB becomes a phi/Memory
 435  *
 436  * @param senv      the spill environment
 437  * @param phi       the Phi node that should be spilled
 438  * @param ctx_irn   an user of the spilled node
 439  */
 440 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 441 {
 442         ir_graph *irg   = env->irg;
 443         ir_node  *phi   = spillinfo->to_spill;
 444         ir_node  *block = get_nodes_block(phi);
 445         ir_node  *unknown;
 446         ir_node **ins;
 447         spill_t  *spill;
 448         int       i;
 449         int       arity;
 450
 451         assert(is_Phi(phi));
 452         assert(!get_opt_cse());
 453         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 454
 455         /* build a new PhiM */
 456         arity   = get_irn_arity(phi);
 457         ins     = ALLOCAN(ir_node*, arity);
 458         unknown = new_r_Unknown(irg, mode_M);
 459         for (i = 0; i < arity; ++i) {
 460                 ins[i] = unknown;
 461         }
 462
 463         /* override or replace spills list... */
 464         spill         = OALLOC(&env->obst, spill_t);
 465         spill->after  = determine_spill_point(phi);
 466         spill->spill  = be_new_Phi(block, arity, ins, mode_M, NULL);
 467         spill->next   = NULL;
 468         sched_add_after(block, spill->spill);
 469
 470         spillinfo->spills = spill;
 471         env->spilled_phi_count++;
 472
 473         for (i = 0; i < arity; ++i) {
 474                 ir_node      *arg      = get_irn_n(phi, i);
 475                 spill_info_t *arg_info = get_spillinfo(env, arg);
 476
 477                 determine_spill_costs(env, arg_info);
 478                 spill_node(env, arg_info);
 479
 480                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 481         }
 482         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 483              spill->spill));
 484 }
 485
 486 /**
 487  * Spill a node.
 488  *
 489  * @param senv      the spill environment
 490  * @param to_spill  the node that should be spilled
 491  */
 492 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 493 {
 494         ir_node *to_spill;
 495
 496         /* node is already spilled */
 497         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 498                 return;
 499
 500         to_spill = spillinfo->to_spill;
 501
 502         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 503                 spill_phi(env, spillinfo);
 504         } else {
 505                 spill_irn(env, spillinfo);
 506         }
 507 }
 508
 509 /*
 510  *
 511  *  ____                      _            _       _ _
 512  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 513  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 514  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 515  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 516  *
 517  */
 518
 519 /**
 520  * Tests whether value @p arg is available before node @p reloader
 521  * @returns 1 if value is available, 0 otherwise
 522  */
 523 static int is_value_available(spill_env_t *env, const ir_node *arg,
 524                               const ir_node *reloader)
 525 {
 526         if (is_Unknown(arg) || is_NoMem(arg))
 527                 return 1;
 528
 529         if (be_is_Spill(skip_Proj_const(arg)))
 530                 return 1;
 531
 532         if (arg == get_irg_frame(env->irg))
 533                 return 1;
 534
 535         (void)reloader;
 536
 537         if (get_irn_mode(arg) == mode_T)
 538                 return 0;
 539
 540         /*
 541          * Ignore registers are always available
 542          */
 543         if (arch_irn_is_ignore(arg))
 544                 return 1;
 545
 546         return 0;
 547 }
 548
 549 /**
 550  * Check if a node is rematerializable. This tests for the following conditions:
 551  *
 552  * - The node itself is rematerializable
 553  * - All arguments of the node are available or also rematerialisable
 554  * - The costs for the rematerialisation operation is less or equal a limit
 555  *
 556  * Returns the costs needed for rematerialisation or something
 557  * >= REMAT_COST_INFINITE if remat is not possible.
 558  */
 559 static int check_remat_conditions_costs(spill_env_t *env,
 560                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 561 {
 562         int i, arity;
 563         int argremats;
 564         int costs = 0;
 565         const ir_node *insn = skip_Proj_const(spilled);
 566
 567         assert(!be_is_Spill(insn));
 568         if (!arch_irn_is(insn, rematerializable))
 569                 return REMAT_COST_INFINITE;
 570
 571         if (be_is_Reload(insn)) {
 572                 costs += 2;
 573         } else {
 574                 costs += arch_get_op_estimated_cost(insn);
 575         }
 576         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 577                 return REMAT_COST_INFINITE;
 578         }
 579         /* never rematerialize a node which modifies the flags.
 580          * (would be better to test whether the flags are actually live at point
 581          * reloader...)
 582          */
 583         if (arch_irn_is(insn, modify_flags)) {
 584                 return REMAT_COST_INFINITE;
 585         }
 586
 587         argremats = 0;
 588         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 589                 ir_node *arg = get_irn_n(insn, i);
 590
 591                 if (is_value_available(env, arg, reloader))
 592                         continue;
 593
 594                 /* we have to rematerialize the argument as well */
 595                 ++argremats;
 596                 if (argremats > 1) {
 597                         /* we only support rematerializing 1 argument at the moment,
 598                          * as multiple arguments could increase register pressure */
 599                         return REMAT_COST_INFINITE;
 600                 }
 601
 602                 costs += check_remat_conditions_costs(env, arg, reloader,
 603                                                       parentcosts + costs);
 604                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 605                         return REMAT_COST_INFINITE;
 606         }
 607
 608         return costs;
 609 }
 610
 611 /**
 612  * Re-materialize a node.
 613  *
 614  * @param env       the spill environment
 615  * @param spilled   the node that was spilled
 616  * @param reloader  a irn that requires a reload
 617  */
 618 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 619 {
 620         int i, arity;
 621         ir_node *res;
 622         ir_node *bl;
 623         ir_node **ins;
 624
 625         if (is_Block(reloader)) {
 626                 bl = reloader;
 627         } else {
 628                 bl = get_nodes_block(reloader);
 629         }
 630
 631         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 632         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 633                 ir_node *arg = get_irn_n(spilled, i);
 634
 635                 if (is_value_available(env, arg, reloader)) {
 636                         ins[i] = arg;
 637                 } else {
 638                         ins[i] = do_remat(env, arg, reloader);
 639                         /* don't count the argument rematerialization as an extra remat */
 640                         --env->remat_count;
 641                 }
 642         }
 643
 644         /* create a copy of the node */
 645         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 646                           get_irn_op(spilled), get_irn_mode(spilled),
 647                           get_irn_arity(spilled), ins);
 648         copy_node_attr(env->irg, spilled, res);
 649         arch_env_mark_remat(env->arch_env, res);
 650
 651         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 652
 653         if (! is_Proj(res)) {
 654                 /* insert in schedule */
 655                 sched_reset(res);
 656                 sched_add_before(reloader, res);
 657                 ++env->remat_count;
 658         }
 659
 660         return res;
 661 }
 662
 663 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 664 {
 665         ir_node *block = get_nodes_block(before);
 666         double   freq  = get_block_execfreq(env->exec_freq, block);
 667         (void) to_spill;
 668
 669         return env->spill_cost * freq;
 670 }
 671
 672 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 673                                        const ir_node *before)
 674 {
 675         if (be_do_remats) {
 676                 /* is the node rematerializable? */
 677                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 678                 if (costs < (unsigned) env->reload_cost)
 679                         return costs;
 680         }
 681
 682         return env->reload_cost;
 683 }
 684
 685 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 686 {
 687         ir_node      *block = get_nodes_block(before);
 688         double        freq  = get_block_execfreq(env->exec_freq, block);
 689
 690         if (be_do_remats) {
 691                 /* is the node rematerializable? */
 692                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 693                 if (costs < env->reload_cost)
 694                         return costs * freq;
 695         }
 696
 697         return env->reload_cost * freq;
 698 }
 699
 700 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 701                            const ir_node *before)
 702 {
 703         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 704 }
 705
 706 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 707                                    ir_node *block, int pos)
 708 {
 709         ir_node *before = get_block_insertion_point(block, pos);
 710         return be_get_reload_costs(env, to_spill, before);
 711 }
 712
 713 ir_node *be_new_spill(ir_node *value, ir_node *after)
 714 {
 715         ir_graph                    *irg       = get_irn_irg(value);
 716         ir_node                     *frame     = get_irg_frame(irg);
 717         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 718         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 719         ir_node                     *block     = get_block(after);
 720         ir_node                     *spill
 721                 = be_new_Spill(cls, cls_frame, block, frame, value);
 722
 723         sched_add_after(after, spill);
 724         return spill;
 725 }
 726
 727 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
 728 {
 729         ir_graph *irg   = get_irn_irg(value);
 730         ir_node  *frame = get_irg_frame(irg);
 731         ir_node  *block = get_block(before);
 732         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 733         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 734         ir_mode                     *mode      = get_irn_mode(value);
 735         ir_node  *reload;
 736
 737         assert(be_is_Spill(spill) || is_Phi(spill));
 738         assert(get_irn_mode(spill) == mode_M);
 739
 740         reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
 741         sched_add_before(before, reload);
 742
 743         return reload;
 744 }
 745
 746 /*
 747  *  ___                     _     ____      _                 _
 748  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 749  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 750  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 751  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 752  *
 753  */
 754
 755 /**
 756  * analyzes how to best spill a node and determine costs for that
 757  */
 758 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 759 {
 760         ir_node       *to_spill = spillinfo->to_spill;
 761         const ir_node *insn     = skip_Proj_const(to_spill);
 762         ir_node       *spill_block;
 763         spill_t       *spill;
 764         double         spill_execfreq;
 765
 766         /* already calculated? */
 767         if (spillinfo->spill_costs >= 0)
 768                 return;
 769
 770         assert(!arch_irn_is(insn, dont_spill));
 771         assert(!be_is_Reload(insn));
 772
 773         /* some backends have virtual noreg/unknown nodes that are not scheduled
 774          * and simply always available.
 775          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 776          * predecessor (of a PhiM) but this test might match other things too...
 777          */
 778         if (!sched_is_scheduled(insn)) {
 779                 ir_graph *irg = get_irn_irg(to_spill);
 780                 /* override spillinfos or create a new one */
 781                 spill_t *spill = OALLOC(&env->obst, spill_t);
 782                 spill->after = NULL;
 783                 spill->next  = NULL;
 784                 spill->spill = get_irg_no_mem(irg);
 785
 786                 spillinfo->spills      = spill;
 787                 spillinfo->spill_costs = 0;
 788
 789                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 790                 return;
 791         }
 792
 793         spill_block    = get_nodes_block(insn);
 794         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 795
 796         if (is_Phi(to_spill) && ir_nodeset_contains(&env->mem_phis, to_spill)) {
 797                 /* TODO calculate correct costs...
 798                  * (though we can't remat this node anyway so no big problem) */
 799                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 800                 return;
 801         }
 802
 803         if (spillinfo->spills != NULL) {
 804                 spill_t *s;
 805                 double   spills_execfreq;
 806
 807                 /* calculate sum of execution frequencies of individual spills */
 808                 spills_execfreq = 0;
 809                 s               = spillinfo->spills;
 810                 for ( ; s != NULL; s = s->next) {
 811                         ir_node *spill_block = get_block(s->after);
 812                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 813
 814                         spills_execfreq += freq;
 815                 }
 816
 817                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 818                     spills_execfreq * env->spill_cost,
 819                     spill_execfreq * env->spill_cost));
 820
 821                 /* multi-/latespill is advantageous -> return*/
 822                 if (spills_execfreq < spill_execfreq) {
 823                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 824                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 825                         return;
 826                 }
 827         }
 828
 829         /* override spillinfos or create a new one */
 830         spill        = OALLOC(&env->obst, spill_t);
 831         spill->after = determine_spill_point(to_spill);
 832         spill->next  = NULL;
 833         spill->spill = NULL;
 834
 835         spillinfo->spills      = spill;
 836         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 837         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 838 }
 839
 840 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 841 {
 842         const spill_info_t *si = get_spillinfo(env, irn);
 843         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 844         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 845         bitset_t *reloads      = bitset_alloca(n_blocks);
 846         reloader_t *r;
 847         spill_t *s;
 848
 849         if (si == NULL)
 850                 return;
 851
 852         /* Fill the bitset with the dominance pre-order numbers
 853          * of the blocks the reloads are located in. */
 854         for (r = si->reloaders; r != NULL; r = r->next) {
 855                 ir_node *bl = get_nodes_block(r->reloader);
 856                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 857         }
 858
 859         /* Now, cancel out all the blocks that are dominated by each spill.
 860          * If the bitset is not empty after that, we have reloads that are
 861          * not dominated by any spill. */
 862         for (s = si->spills; s != NULL; s = s->next) {
 863                 ir_node *bl = get_nodes_block(s->after);
 864                 int start   = get_Block_dom_tree_pre_num(bl);
 865                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 866
 867                 bitset_clear_range(reloads, start, end);
 868         }
 869
 870         if (!bitset_is_empty(reloads))
 871                 be_add_spill(env, si->to_spill, si->to_spill);
 872 }
 873
 874 void be_insert_spills_reloads(spill_env_t *env)
 875 {
 876         const ir_exec_freq    *exec_freq = env->exec_freq;
 877         spill_info_t          *si;
 878         ir_nodeset_iterator_t  iter;
 879         ir_node               *node;
 880
 881         be_timer_push(T_RA_SPILL_APPLY);
 882
 883         /* create all phi-ms first, this is needed so, that phis, hanging on
 884            spilled phis work correctly */
 885         foreach_ir_nodeset(&env->mem_phis, node, iter) {
 886                 spill_info_t *info = get_spillinfo(env, node);
 887                 spill_node(env, info);
 888         }
 889
 890         /* process each spilled node */
 891         foreach_set(env->spills, spill_info_t*, si) {
 892                 ir_node  *to_spill        = si->to_spill;
 893                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 894                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 895                 bool      force_remat     = false;
 896                 reloader_t *rld;
 897
 898                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 899
 900                 determine_spill_costs(env, si);
 901
 902                 /* determine possibility of rematerialisations */
 903                 if (be_do_remats) {
 904                         /* calculate cost savings for each indivial value when it would
 905                            be rematted instead of reloaded */
 906                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 907                                 double   freq;
 908                                 int      remat_cost;
 909                                 int      remat_cost_delta;
 910                                 ir_node *block;
 911                                 ir_node *reloader = rld->reloader;
 912
 913                                 if (rld->rematted_node != NULL) {
 914                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 915                                              rld->rematted_node, reloader));
 916                                         continue;
 917                                 }
 918                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 919                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 920                                              reloader));
 921                                         all_remat_costs = REMAT_COST_INFINITE;
 922                                         continue;
 923                                 }
 924
 925                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 926                                                                            reloader, 0);
 927                                 if (remat_cost >= REMAT_COST_INFINITE) {
 928                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 929                                              reloader));
 930                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 931                                         all_remat_costs       = REMAT_COST_INFINITE;
 932                                         continue;
 933                                 }
 934
 935                                 remat_cost_delta      = remat_cost - env->reload_cost;
 936                                 rld->remat_cost_delta = remat_cost_delta;
 937                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 938                                 freq                  = get_block_execfreq(exec_freq, block);
 939                                 all_remat_costs      += remat_cost_delta * freq;
 940                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 941                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 942                                      remat_cost_delta * freq));
 943                         }
 944                         if (all_remat_costs < REMAT_COST_INFINITE) {
 945                                 /* we don't need the costs for the spill if we can remat
 946                                    all reloaders */
 947                                 all_remat_costs -= si->spill_costs;
 948
 949                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 950                                      env->spill_cost, si->spill_costs));
 951                         }
 952
 953                         if (all_remat_costs < 0) {
 954                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 955                                      all_remat_costs));
 956                                 force_remat = true;
 957                         }
 958                 }
 959
 960                 /* go through all reloads for this spill */
 961                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 962                         ir_node *copy; /* a reload is a "copy" of the original value */
 963
 964                         if (rld->rematted_node != NULL) {
 965                                 copy = rld->rematted_node;
 966                                 sched_add_before(rld->reloader, copy);
 967                         } else if (be_do_remats &&
 968                                         (force_remat || rld->remat_cost_delta < 0)) {
 969                                 copy = do_remat(env, to_spill, rld->reloader);
 970                         } else {
 971                                 /* make sure we have a spill */
 972                                 spill_node(env, si);
 973
 974                                 /* create a reload, use the first spill for now SSA
 975                                  * reconstruction for memory comes below */
 976                                 assert(si->spills != NULL);
 977                                 copy = arch_env_new_reload(env->arch_env, si->to_spill,
 978                                                            si->spills->spill, rld->reloader);
 979                                 env->reload_count++;
 980                         }
 981
 982                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 983                              copy, to_spill, rld->reloader));
 984                         ARR_APP1(ir_node*, copies, copy);
 985                 }
 986
 987                 /* if we had any reloads or remats, then we need to reconstruct the
 988                  * SSA form for the spilled value */
 989                 if (ARR_LEN(copies) > 0) {
 990                         be_ssa_construction_env_t senv;
 991                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 992
 993                         be_ssa_construction_init(&senv, env->irg);
 994                         be_ssa_construction_add_copy(&senv, to_spill);
 995                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 996                         be_ssa_construction_fix_users(&senv, to_spill);
 997
 998 #if 0
 999                         /* no need to enable this as long as we invalidate liveness
1000                            after this function... */
1001                         be_ssa_construction_update_liveness_phis(&senv);
1002                         be_liveness_update(to_spill);
1003                         len = ARR_LEN(copies);
1004                         for (i = 0; i < len; ++i) {
1005                                 be_liveness_update(lv, copies[i]);
1006                         }
1007 #endif
1008                         be_ssa_construction_destroy(&senv);
1009                 }
1010                 /* need to reconstruct SSA form if we had multiple spills */
1011                 if (si->spills != NULL && si->spills->next != NULL) {
1012                         spill_t *spill;
1013                         int      spill_count = 0;
1014
1015                         be_ssa_construction_env_t senv;
1016
1017                         be_ssa_construction_init(&senv, env->irg);
1018                         spill = si->spills;
1019                         for ( ; spill != NULL; spill = spill->next) {
1020                                 /* maybe we rematerialized the value and need no spill */
1021                                 if (spill->spill == NULL)
1022                                         continue;
1023                                 be_ssa_construction_add_copy(&senv, spill->spill);
1024                                 spill_count++;
1025                         }
1026                         if (spill_count > 1) {
1027                                 /* all reloads are attached to the first spill, fix them now */
1028                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1029                         }
1030
1031                         be_ssa_construction_destroy(&senv);
1032                 }
1033
1034                 DEL_ARR_F(copies);
1035                 si->reloaders = NULL;
1036         }
1037
1038         stat_ev_dbl("spill_spills", env->spill_count);
1039         stat_ev_dbl("spill_reloads", env->reload_count);
1040         stat_ev_dbl("spill_remats", env->remat_count);
1041         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1042
1043         /* Matze: In theory be_ssa_construction should take care of the liveness...
1044          * try to disable this again in the future */
1045         be_liveness_invalidate(be_get_irg_liveness(env->irg));
1046
1047         be_remove_dead_nodes_from_schedule(env->irg);
1048
1049         be_timer_pop(T_RA_SPILL_APPLY);
1050 }
1051
1052 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
1053 void be_init_spill(void)
1054 {
1055         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1056 }