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         bool        spilled_phi; /* true when the whole Phi has been spilled and
  96                                     will be replaced with a PhiM. false if only the
  97                                     value of the Phi gets spilled */
  98 };
  99
 100 struct spill_env_t {
 101         const arch_env_t *arch_env;
 102         ir_graph         *irg;
 103         struct obstack    obst;
 104         int               spill_cost;     /**< the cost of a single spill node */
 105         int               reload_cost;    /**< the cost of a reload node */
 106         set              *spills;         /**< all spill_info_t's, which must be
 107                                                placed */
 108         spill_info_t    **mem_phis;       /**< set of all spilled phis. */
 109         ir_exec_freq     *exec_freq;
 110
 111         unsigned          spill_count;
 112         unsigned          reload_count;
 113         unsigned          remat_count;
 114         unsigned          spilled_phi_count;
 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                 info.spilled_phi = false;
 146                 res = (spill_info_t*)set_insert(env->spills, &info, sizeof(info), hash);
 147         }
 148
 149         return res;
 150 }
 151
 152 spill_env_t *be_new_spill_env(ir_graph *irg)
 153 {
 154         const arch_env_t *arch_env = be_get_irg_arch_env(irg);
 155
 156         spill_env_t *env = XMALLOC(spill_env_t);
 157         env->spills         = new_set(cmp_spillinfo, 1024);
 158         env->irg            = irg;
 159         env->arch_env       = arch_env;
 160         env->mem_phis       = NEW_ARR_F(spill_info_t*, 0);
 161         env->spill_cost     = arch_env->spill_cost;
 162         env->reload_cost    = arch_env->reload_cost;
 163         env->exec_freq      = be_get_irg_exec_freq(irg);
 164         obstack_init(&env->obst);
 165
 166         env->spill_count       = 0;
 167         env->reload_count      = 0;
 168         env->remat_count       = 0;
 169         env->spilled_phi_count = 0;
 170
 171         return env;
 172 }
 173
 174 void be_delete_spill_env(spill_env_t *env)
 175 {
 176         del_set(env->spills);
 177         DEL_ARR_F(env->mem_phis);
 178         obstack_free(&env->obst, NULL);
 179         free(env);
 180 }
 181
 182 /*
 183  *  ____  _                  ____      _                 _
 184  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 185  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 186  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 187  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 188  *
 189  */
 190
 191 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 192 {
 193         spill_info_t  *spill_info = get_spillinfo(env, to_spill);
 194         spill_t       *spill;
 195         spill_t       *s;
 196         spill_t       *last;
 197
 198         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 199         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 200
 201         /* Just for safety make sure that we do not insert the spill in front of a phi */
 202         assert(!is_Phi(sched_next(after)));
 203
 204         /* spills that are dominated by others are not needed */
 205         last = NULL;
 206         s    = spill_info->spills;
 207         for ( ; s != NULL; s = s->next) {
 208                 /* no need to add this spill if it is dominated by another */
 209                 if (value_dominates(s->after, after)) {
 210                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 211                         return;
 212                 }
 213                 /* remove spills that we dominate */
 214                 if (value_dominates(after, s->after)) {
 215                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 216                         if (last != NULL) {
 217                                 last->next         = s->next;
 218                         } else {
 219                                 spill_info->spills = s->next;
 220                         }
 221                 } else {
 222                         last = s;
 223                 }
 224         }
 225
 226         spill         = OALLOC(&env->obst, spill_t);
 227         spill->after  = after;
 228         spill->next   = spill_info->spills;
 229         spill->spill  = NULL;
 230
 231         spill_info->spills = spill;
 232 }
 233
 234 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 235                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 236                 int allow_remat)
 237 {
 238         spill_info_t  *info;
 239         reloader_t    *rel;
 240
 241         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 242
 243         info = get_spillinfo(env, to_spill);
 244
 245         if (is_Phi(to_spill)) {
 246                 int i, arity;
 247
 248                 /* create spillinfos for the phi arguments */
 249                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 250                         ir_node *arg = get_irn_n(to_spill, i);
 251                         get_spillinfo(env, arg);
 252                 }
 253         }
 254
 255         assert(!is_Proj(before) && !be_is_Keep(before));
 256
 257         /* put reload into list */
 258         rel                   = OALLOC(&env->obst, reloader_t);
 259         rel->next             = info->reloaders;
 260         rel->reloader         = before;
 261         rel->rematted_node    = NULL;
 262         rel->can_spill_after  = can_spill_after;
 263         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 264
 265         info->reloaders  = rel;
 266         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 267         info->reload_cls = reload_cls;
 268
 269         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 270                 to_spill, before, allow_remat ? "" : " not"));
 271 }
 272
 273 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 274                    const arch_register_class_t *reload_cls, int allow_remat)
 275 {
 276         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 277
 278 }
 279
 280 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 281 {
 282         ir_node *last = sched_last(block);
 283
 284         /* we might have keeps behind the jump... */
 285         while (be_is_Keep(last)) {
 286                 last = sched_prev(last);
 287                 assert(!sched_is_end(last));
 288         }
 289
 290         assert(is_cfop(last));
 291
 292         /* add the reload before the (cond-)jump */
 293         return last;
 294 }
 295
 296 /**
 297  * determine final spill position: it should be after all phis, keep nodes
 298  * and behind nodes marked as prolog
 299  */
 300 static ir_node *determine_spill_point(ir_node *node)
 301 {
 302         node = skip_Proj(node);
 303         while (true) {
 304                 ir_node *next = sched_next(node);
 305                 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
 306                         break;
 307                 node = next;
 308         }
 309         return node;
 310 }
 311
 312 /**
 313  * Returns the point at which you can insert a node that should be executed
 314  * before block @p block when coming from pred @p pos.
 315  */
 316 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 317 {
 318         ir_node *predblock;
 319
 320         /* simply add the reload to the beginning of the block if we only have 1
 321          * predecessor. We don't need to check for phis as there can't be any in a
 322          * block with only 1 pred. */
 323         if (get_Block_n_cfgpreds(block) == 1) {
 324                 assert(!is_Phi(sched_first(block)));
 325                 return sched_first(block);
 326         }
 327
 328         /* We have to reload the value in pred-block */
 329         predblock = get_Block_cfgpred_block(block, pos);
 330         return be_get_end_of_block_insertion_point(predblock);
 331 }
 332
 333 void be_add_reload_at_end(spill_env_t *env, ir_node *to_spill,
 334                           const ir_node *block,
 335                           const arch_register_class_t *reload_cls,
 336                           int allow_remat)
 337 {
 338         ir_node *before = be_get_end_of_block_insertion_point(block);
 339         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 340 }
 341
 342 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 343                            int pos, const arch_register_class_t *reload_cls,
 344                            int allow_remat)
 345 {
 346         ir_node *before = get_block_insertion_point(block, pos);
 347         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 348 }
 349
 350 void be_spill_phi(spill_env_t *env, ir_node *node)
 351 {
 352         ir_node *block;
 353         int i, arity;
 354         spill_info_t *info;
 355
 356         assert(is_Phi(node));
 357
 358         info              = get_spillinfo(env, node);
 359         info->spilled_phi = true;
 360         ARR_APP1(spill_info_t*, env->mem_phis, info);
 361
 362         /* create spills for the phi arguments */
 363         block = get_nodes_block(node);
 364         for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 365                 ir_node *arg = get_irn_n(node, i);
 366                 ir_node *insert;
 367
 368                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 369                  * and simply always available. */
 370                 if (!sched_is_scheduled(arg)) {
 371                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 372                         insert = be_get_end_of_block_insertion_point(pred_block);
 373                         insert = sched_prev(insert);
 374                 } else {
 375                         insert = determine_spill_point(arg);
 376                 }
 377
 378                 be_add_spill(env, arg, insert);
 379         }
 380 }
 381
 382 /*
 383  *   ____                _         ____        _ _ _
 384  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 385  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 386  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 387  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 388  *                                      |_|
 389  */
 390
 391 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 392
 393 /**
 394  * Creates a spill.
 395  *
 396  * @param senv      the spill environment
 397  * @param irn       the node that should be spilled
 398  * @param ctx_irn   an user of the spilled node
 399  *
 400  * @return a be_Spill node
 401  */
 402 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 403 {
 404         ir_node       *to_spill = spillinfo->to_spill;
 405         const ir_node *insn     = skip_Proj_const(to_spill);
 406         spill_t *spill;
 407
 408         /* determine_spill_costs must have been run before */
 409         assert(spillinfo->spill_costs >= 0);
 410
 411         /* some backends have virtual noreg/unknown nodes that are not scheduled
 412          * and simply always available. */
 413         if (!sched_is_scheduled(insn)) {
 414                 /* override spillinfos or create a new one */
 415                 ir_graph *irg = get_irn_irg(to_spill);
 416                 spillinfo->spills->spill = get_irg_no_mem(irg);
 417                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 418                 return;
 419         }
 420
 421         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 422         spill = spillinfo->spills;
 423         for ( ; spill != NULL; spill = spill->next) {
 424                 ir_node *after = spill->after;
 425                 after = determine_spill_point(after);
 426
 427                 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
 428                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 429                 env->spill_count++;
 430         }
 431         DBG((dbg, LEVEL_1, "\n"));
 432 }
 433
 434 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 435
 436 /**
 437  * If the first usage of a Phi result would be out of memory
 438  * there is no sense in allocating a register for it.
 439  * Thus we spill it and all its operands to the same spill slot.
 440  * Therefore the phi/dataB becomes a phi/Memory
 441  *
 442  * @param senv      the spill environment
 443  * @param phi       the Phi node that should be spilled
 444  * @param ctx_irn   an user of the spilled node
 445  */
 446 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 447 {
 448         ir_graph *irg   = env->irg;
 449         ir_node  *phi   = spillinfo->to_spill;
 450         ir_node  *block = get_nodes_block(phi);
 451         ir_node  *unknown;
 452         ir_node **ins;
 453         spill_t  *spill;
 454         int       i;
 455         int       arity;
 456
 457         assert(is_Phi(phi));
 458         assert(!get_opt_cse());
 459         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 460
 461         /* build a new PhiM */
 462         arity   = get_irn_arity(phi);
 463         ins     = ALLOCAN(ir_node*, arity);
 464         unknown = new_r_Unknown(irg, mode_M);
 465         for (i = 0; i < arity; ++i) {
 466                 ins[i] = unknown;
 467         }
 468
 469         /* override or replace spills list... */
 470         spill         = OALLOC(&env->obst, spill_t);
 471         spill->after  = determine_spill_point(phi);
 472         spill->spill  = be_new_Phi(block, arity, ins, mode_M, arch_no_register_req);
 473         spill->next   = NULL;
 474         sched_add_after(block, spill->spill);
 475
 476         spillinfo->spills = spill;
 477         env->spilled_phi_count++;
 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         /* node is already spilled */
 501         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 502                 return;
 503
 504         if (spillinfo->spilled_phi) {
 505                 spill_phi(env, spillinfo);
 506         } else {
 507                 spill_irn(env, spillinfo);
 508         }
 509 }
 510
 511 /*
 512  *
 513  *  ____                      _            _       _ _
 514  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 515  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 516  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 517  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 518  *
 519  */
 520
 521 /**
 522  * Tests whether value @p arg is available before node @p reloader
 523  * @returns 1 if value is available, 0 otherwise
 524  */
 525 static int is_value_available(spill_env_t *env, const ir_node *arg,
 526                               const ir_node *reloader)
 527 {
 528         if (is_Unknown(arg) || is_NoMem(arg))
 529                 return 1;
 530
 531         if (be_is_Spill(skip_Proj_const(arg)))
 532                 return 1;
 533
 534         if (arg == get_irg_frame(env->irg))
 535                 return 1;
 536
 537         (void)reloader;
 538
 539         if (get_irn_mode(arg) == mode_T)
 540                 return 0;
 541
 542         /*
 543          * Ignore registers are always available
 544          */
 545         if (arch_irn_is_ignore(arg))
 546                 return 1;
 547
 548         return 0;
 549 }
 550
 551 /**
 552  * Check if a node is rematerializable. This tests for the following conditions:
 553  *
 554  * - The node itself is rematerializable
 555  * - All arguments of the node are available or also rematerialisable
 556  * - The costs for the rematerialisation operation is less or equal a limit
 557  *
 558  * Returns the costs needed for rematerialisation or something
 559  * >= REMAT_COST_INFINITE if remat is not possible.
 560  */
 561 static int check_remat_conditions_costs(spill_env_t *env,
 562                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 563 {
 564         int i, arity;
 565         int argremats;
 566         int costs = 0;
 567         const ir_node *insn = skip_Proj_const(spilled);
 568
 569         assert(!be_is_Spill(insn));
 570         if (!arch_irn_is(insn, rematerializable))
 571                 return REMAT_COST_INFINITE;
 572
 573         if (be_is_Reload(insn)) {
 574                 costs += 2;
 575         } else {
 576                 costs += arch_get_op_estimated_cost(insn);
 577         }
 578         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 579                 return REMAT_COST_INFINITE;
 580         }
 581         /* never rematerialize a node which modifies the flags.
 582          * (would be better to test whether the flags are actually live at point
 583          * reloader...)
 584          */
 585         if (arch_irn_is(insn, modify_flags)) {
 586                 return REMAT_COST_INFINITE;
 587         }
 588
 589         argremats = 0;
 590         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 591                 ir_node *arg = get_irn_n(insn, i);
 592
 593                 if (is_value_available(env, arg, reloader))
 594                         continue;
 595
 596                 /* we have to rematerialize the argument as well */
 597                 ++argremats;
 598                 if (argremats > 1) {
 599                         /* we only support rematerializing 1 argument at the moment,
 600                          * as multiple arguments could increase register pressure */
 601                         return REMAT_COST_INFINITE;
 602                 }
 603
 604                 costs += check_remat_conditions_costs(env, arg, reloader,
 605                                                       parentcosts + costs);
 606                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 607                         return REMAT_COST_INFINITE;
 608         }
 609
 610         return costs;
 611 }
 612
 613 /**
 614  * Re-materialize a node.
 615  *
 616  * @param env       the spill environment
 617  * @param spilled   the node that was spilled
 618  * @param reloader  a irn that requires a reload
 619  */
 620 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 621 {
 622         int i, arity;
 623         ir_node *res;
 624         ir_node *bl;
 625         ir_node **ins;
 626
 627         if (is_Block(reloader)) {
 628                 bl = reloader;
 629         } else {
 630                 bl = get_nodes_block(reloader);
 631         }
 632
 633         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 634         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 635                 ir_node *arg = get_irn_n(spilled, i);
 636
 637                 if (is_value_available(env, arg, reloader)) {
 638                         ins[i] = arg;
 639                 } else {
 640                         ins[i] = do_remat(env, arg, reloader);
 641                         /* don't count the argument rematerialization as an extra remat */
 642                         --env->remat_count;
 643                 }
 644         }
 645
 646         /* create a copy of the node */
 647         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 648                           get_irn_op(spilled), get_irn_mode(spilled),
 649                           get_irn_arity(spilled), ins);
 650         copy_node_attr(env->irg, spilled, res);
 651         arch_env_mark_remat(env->arch_env, res);
 652
 653         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 654
 655         if (! is_Proj(res)) {
 656                 /* insert in schedule */
 657                 sched_reset(res);
 658                 sched_add_before(reloader, res);
 659                 ++env->remat_count;
 660         }
 661
 662         return res;
 663 }
 664
 665 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 666 {
 667         ir_node *block = get_nodes_block(before);
 668         double   freq  = get_block_execfreq(env->exec_freq, block);
 669         (void) to_spill;
 670
 671         return env->spill_cost * freq;
 672 }
 673
 674 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 675                                        const ir_node *before)
 676 {
 677         if (be_do_remats) {
 678                 /* is the node rematerializable? */
 679                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 680                 if (costs < (unsigned) env->reload_cost)
 681                         return costs;
 682         }
 683
 684         return env->reload_cost;
 685 }
 686
 687 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 688 {
 689         ir_node      *block = get_nodes_block(before);
 690         double        freq  = get_block_execfreq(env->exec_freq, block);
 691
 692         if (be_do_remats) {
 693                 /* is the node rematerializable? */
 694                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 695                 if (costs < env->reload_cost)
 696                         return costs * freq;
 697         }
 698
 699         return env->reload_cost * freq;
 700 }
 701
 702 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 703                            const ir_node *before)
 704 {
 705         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 706 }
 707
 708 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 709                                    ir_node *block, int pos)
 710 {
 711         ir_node *before = get_block_insertion_point(block, pos);
 712         return be_get_reload_costs(env, to_spill, before);
 713 }
 714
 715 ir_node *be_new_spill(ir_node *value, ir_node *after)
 716 {
 717         ir_graph                    *irg       = get_irn_irg(value);
 718         ir_node                     *frame     = get_irg_frame(irg);
 719         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 720         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 721         ir_node                     *block     = get_block(after);
 722         ir_node                     *spill
 723                 = be_new_Spill(cls, cls_frame, block, frame, value);
 724
 725         sched_add_after(after, spill);
 726         return spill;
 727 }
 728
 729 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
 730 {
 731         ir_graph *irg   = get_irn_irg(value);
 732         ir_node  *frame = get_irg_frame(irg);
 733         ir_node  *block = get_block(before);
 734         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 735         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 736         ir_mode                     *mode      = get_irn_mode(value);
 737         ir_node  *reload;
 738
 739         assert(be_is_Spill(spill) || is_Phi(spill));
 740         assert(get_irn_mode(spill) == mode_M);
 741
 742         reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
 743         sched_add_before(before, reload);
 744
 745         return reload;
 746 }
 747
 748 /*
 749  *  ___                     _     ____      _                 _
 750  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 751  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 752  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 753  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 754  *
 755  */
 756
 757 /**
 758  * analyzes how to best spill a node and determine costs for that
 759  */
 760 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 761 {
 762         ir_node       *to_spill = spillinfo->to_spill;
 763         const ir_node *insn     = skip_Proj_const(to_spill);
 764         ir_node       *spill_block;
 765         spill_t       *spill;
 766         double         spill_execfreq;
 767
 768         /* already calculated? */
 769         if (spillinfo->spill_costs >= 0)
 770                 return;
 771
 772         assert(!arch_irn_is(insn, dont_spill));
 773         assert(!be_is_Reload(insn));
 774
 775         /* some backends have virtual noreg/unknown nodes that are not scheduled
 776          * and simply always available.
 777          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 778          * predecessor (of a PhiM) but this test might match other things too...
 779          */
 780         if (!sched_is_scheduled(insn)) {
 781                 ir_graph *irg = get_irn_irg(to_spill);
 782                 /* override spillinfos or create a new one */
 783                 spill_t *spill = OALLOC(&env->obst, spill_t);
 784                 spill->after = NULL;
 785                 spill->next  = NULL;
 786                 spill->spill = get_irg_no_mem(irg);
 787
 788                 spillinfo->spills      = spill;
 789                 spillinfo->spill_costs = 0;
 790
 791                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 792                 return;
 793         }
 794
 795         spill_block    = get_nodes_block(insn);
 796         spill_execfreq = get_block_execfreq(env->exec_freq, spill_block);
 797
 798         if (spillinfo->spilled_phi) {
 799                 /* TODO calculate correct costs...
 800                  * (though we can't remat this node anyway so no big problem) */
 801                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 802                 return;
 803         }
 804
 805         if (spillinfo->spills != NULL) {
 806                 spill_t *s;
 807                 double   spills_execfreq;
 808
 809                 /* calculate sum of execution frequencies of individual spills */
 810                 spills_execfreq = 0;
 811                 s               = spillinfo->spills;
 812                 for ( ; s != NULL; s = s->next) {
 813                         ir_node *spill_block = get_block(s->after);
 814                         double   freq = get_block_execfreq(env->exec_freq, spill_block);
 815
 816                         spills_execfreq += freq;
 817                 }
 818
 819                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 820                     spills_execfreq * env->spill_cost,
 821                     spill_execfreq * env->spill_cost));
 822
 823                 /* multi-/latespill is advantageous -> return*/
 824                 if (spills_execfreq < spill_execfreq) {
 825                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 826                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 827                         return;
 828                 }
 829         }
 830
 831         /* override spillinfos or create a new one */
 832         spill        = OALLOC(&env->obst, spill_t);
 833         spill->after = determine_spill_point(to_spill);
 834         spill->next  = NULL;
 835         spill->spill = NULL;
 836
 837         spillinfo->spills      = spill;
 838         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 839         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 840 }
 841
 842 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 843 {
 844         const spill_info_t *si = get_spillinfo(env, irn);
 845         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 846         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 847         bitset_t *reloads      = bitset_alloca(n_blocks);
 848         reloader_t *r;
 849         spill_t *s;
 850
 851         if (si == NULL)
 852                 return;
 853
 854         /* Fill the bitset with the dominance pre-order numbers
 855          * of the blocks the reloads are located in. */
 856         for (r = si->reloaders; r != NULL; r = r->next) {
 857                 ir_node *bl = get_nodes_block(r->reloader);
 858                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 859         }
 860
 861         /* Now, cancel out all the blocks that are dominated by each spill.
 862          * If the bitset is not empty after that, we have reloads that are
 863          * not dominated by any spill. */
 864         for (s = si->spills; s != NULL; s = s->next) {
 865                 ir_node *bl = get_nodes_block(s->after);
 866                 int start   = get_Block_dom_tree_pre_num(bl);
 867                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 868
 869                 bitset_clear_range(reloads, start, end);
 870         }
 871
 872         if (!bitset_is_empty(reloads))
 873                 be_add_spill(env, si->to_spill, si->to_spill);
 874 }
 875
 876 void be_insert_spills_reloads(spill_env_t *env)
 877 {
 878         const ir_exec_freq    *exec_freq  = env->exec_freq;
 879         size_t                 n_mem_phis = ARR_LEN(env->mem_phis);
 880         spill_info_t          *si;
 881         size_t                 i;
 882
 883         be_timer_push(T_RA_SPILL_APPLY);
 884
 885         /* create all phi-ms first, this is needed so, that phis, hanging on
 886            spilled phis work correctly */
 887         for (i = 0; i < n_mem_phis; ++i) {
 888                 spill_info_t *info = env->mem_phis[i];
 889                 spill_node(env, info);
 890         }
 891
 892         /* process each spilled node */
 893         foreach_set(env->spills, spill_info_t*, si) {
 894                 ir_node  *to_spill        = si->to_spill;
 895                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 896                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 897                 bool      force_remat     = false;
 898                 reloader_t *rld;
 899
 900                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 901
 902                 determine_spill_costs(env, si);
 903
 904                 /* determine possibility of rematerialisations */
 905                 if (be_do_remats) {
 906                         /* calculate cost savings for each indivial value when it would
 907                            be rematted instead of reloaded */
 908                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 909                                 double   freq;
 910                                 int      remat_cost;
 911                                 int      remat_cost_delta;
 912                                 ir_node *block;
 913                                 ir_node *reloader = rld->reloader;
 914
 915                                 if (rld->rematted_node != NULL) {
 916                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 917                                              rld->rematted_node, reloader));
 918                                         continue;
 919                                 }
 920                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 921                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 922                                              reloader));
 923                                         all_remat_costs = REMAT_COST_INFINITE;
 924                                         continue;
 925                                 }
 926
 927                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 928                                                                            reloader, 0);
 929                                 if (remat_cost >= REMAT_COST_INFINITE) {
 930                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 931                                              reloader));
 932                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 933                                         all_remat_costs       = REMAT_COST_INFINITE;
 934                                         continue;
 935                                 }
 936
 937                                 remat_cost_delta      = remat_cost - env->reload_cost;
 938                                 rld->remat_cost_delta = remat_cost_delta;
 939                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 940                                 freq                  = get_block_execfreq(exec_freq, block);
 941                                 all_remat_costs      += remat_cost_delta * freq;
 942                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 943                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 944                                      remat_cost_delta * freq));
 945                         }
 946                         if (all_remat_costs < REMAT_COST_INFINITE) {
 947                                 /* we don't need the costs for the spill if we can remat
 948                                    all reloaders */
 949                                 all_remat_costs -= si->spill_costs;
 950
 951                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 952                                      env->spill_cost, si->spill_costs));
 953                         }
 954
 955                         if (all_remat_costs < 0) {
 956                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 957                                      all_remat_costs));
 958                                 force_remat = true;
 959                         }
 960                 }
 961
 962                 /* go through all reloads for this spill */
 963                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 964                         ir_node *copy; /* a reload is a "copy" of the original value */
 965
 966                         if (rld->rematted_node != NULL) {
 967                                 copy = rld->rematted_node;
 968                                 sched_add_before(rld->reloader, copy);
 969                         } else if (be_do_remats &&
 970                                         (force_remat || rld->remat_cost_delta < 0)) {
 971                                 copy = do_remat(env, to_spill, rld->reloader);
 972                         } else {
 973                                 /* make sure we have a spill */
 974                                 spill_node(env, si);
 975
 976                                 /* create a reload, use the first spill for now SSA
 977                                  * reconstruction for memory comes below */
 978                                 assert(si->spills != NULL);
 979                                 copy = arch_env_new_reload(env->arch_env, si->to_spill,
 980                                                            si->spills->spill, rld->reloader);
 981                                 env->reload_count++;
 982                         }
 983
 984                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 985                              copy, to_spill, rld->reloader));
 986                         ARR_APP1(ir_node*, copies, copy);
 987                 }
 988
 989                 /* if we had any reloads or remats, then we need to reconstruct the
 990                  * SSA form for the spilled value */
 991                 if (ARR_LEN(copies) > 0) {
 992                         be_ssa_construction_env_t senv;
 993                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 994
 995                         be_ssa_construction_init(&senv, env->irg);
 996                         be_ssa_construction_add_copy(&senv, to_spill);
 997                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 998                         be_ssa_construction_fix_users(&senv, to_spill);
 999
1000 #if 0
1001                         /* no need to enable this as long as we invalidate liveness
1002                            after this function... */
1003                         be_ssa_construction_update_liveness_phis(&senv);
1004                         be_liveness_update(to_spill);
1005                         len = ARR_LEN(copies);
1006                         for (i = 0; i < len; ++i) {
1007                                 be_liveness_update(lv, copies[i]);
1008                         }
1009 #endif
1010                         be_ssa_construction_destroy(&senv);
1011                 }
1012                 /* need to reconstruct SSA form if we had multiple spills */
1013                 if (si->spills != NULL && si->spills->next != NULL) {
1014                         spill_t *spill;
1015                         int      spill_count = 0;
1016
1017                         be_ssa_construction_env_t senv;
1018
1019                         be_ssa_construction_init(&senv, env->irg);
1020                         spill = si->spills;
1021                         for ( ; spill != NULL; spill = spill->next) {
1022                                 /* maybe we rematerialized the value and need no spill */
1023                                 if (spill->spill == NULL)
1024                                         continue;
1025                                 be_ssa_construction_add_copy(&senv, spill->spill);
1026                                 spill_count++;
1027                         }
1028                         if (spill_count > 1) {
1029                                 /* all reloads are attached to the first spill, fix them now */
1030                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1031                         }
1032
1033                         be_ssa_construction_destroy(&senv);
1034                 }
1035
1036                 DEL_ARR_F(copies);
1037                 si->reloaders = NULL;
1038         }
1039
1040         stat_ev_dbl("spill_spills", env->spill_count);
1041         stat_ev_dbl("spill_reloads", env->reload_count);
1042         stat_ev_dbl("spill_remats", env->remat_count);
1043         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1044
1045         /* Matze: In theory be_ssa_construction should take care of the liveness...
1046          * try to disable this again in the future */
1047         be_invalidate_live_sets(env->irg);
1048
1049         be_remove_dead_nodes_from_schedule(env->irg);
1050
1051         be_timer_pop(T_RA_SPILL_APPLY);
1052 }
1053
1054 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
1055 void be_init_spill(void)
1056 {
1057         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1058 }