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

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief       implementation of the spill/reload placement abstraction layer
   9  * @author      Daniel Grund, Sebastian Hack, Matthias Braun
  10  * @date        29.09.2005
  11  */
  12 #include "config.h"
  13
  14 #include <stdlib.h>
  15 #include <stdbool.h>
  16
  17 #include "pset.h"
  18 #include "irnode_t.h"
  19 #include "ircons_t.h"
  20 #include "iredges_t.h"
  21 #include "irbackedge_t.h"
  22 #include "irprintf.h"
  23 #include "ident_t.h"
  24 #include "type_t.h"
  25 #include "entity_t.h"
  26 #include "debug.h"
  27 #include "irgwalk.h"
  28 #include "array.h"
  29 #include "pdeq.h"
  30 #include "execfreq.h"
  31 #include "irnodeset.h"
  32 #include "error.h"
  33
  34 #include "bearch.h"
  35 #include "belive_t.h"
  36 #include "besched.h"
  37 #include "bespill.h"
  38 #include "bespillutil.h"
  39 #include "belive_t.h"
  40 #include "benode.h"
  41 #include "bechordal_t.h"
  42 #include "statev_t.h"
  43 #include "bessaconstr.h"
  44 #include "beirg.h"
  45 #include "beirgmod.h"
  46 #include "beintlive_t.h"
  47 #include "bemodule.h"
  48 #include "be_t.h"
  49
  50 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  51
  52 #define REMAT_COST_INFINITE  1000
  53
  54 typedef struct reloader_t reloader_t;
  55 struct reloader_t {
  56         reloader_t *next;
  57         ir_node    *reloader;
  58         ir_node    *rematted_node;
  59         int         remat_cost_delta; /** costs needed for rematerialization,
  60                                            compared to placing a reload */
  61 };
  62
  63 typedef struct spill_t spill_t;
  64 struct spill_t {
  65         spill_t *next;
  66         ir_node *after;  /**< spill has to be placed after this node (or earlier) */
  67         ir_node *spill;
  68 };
  69
  70 typedef struct spill_info_t spill_info_t;
  71 struct spill_info_t {
  72         ir_node    *to_spill;  /**< the value that should get spilled */
  73         reloader_t *reloaders; /**< list of places where the value should get
  74                                     reloaded */
  75         spill_t    *spills;    /**< list of latest places where spill must be
  76                                     placed */
  77         double      spill_costs; /**< costs needed for spilling the value */
  78         const arch_register_class_t *reload_cls; /** the register class in which the
  79                                                      reload should be placed */
  80         bool        spilled_phi; /* true when the whole Phi has been spilled and
  81                                     will be replaced with a PhiM. false if only the
  82                                     value of the Phi gets spilled */
  83 };
  84
  85 struct spill_env_t {
  86         const arch_env_t *arch_env;
  87         ir_graph         *irg;
  88         struct obstack    obst;
  89         int               spill_cost;     /**< the cost of a single spill node */
  90         int               reload_cost;    /**< the cost of a reload node */
  91         set              *spills;         /**< all spill_info_t's, which must be
  92                                                placed */
  93         spill_info_t    **mem_phis;       /**< set of all spilled phis. */
  94
  95         unsigned          spill_count;
  96         unsigned          reload_count;
  97         unsigned          remat_count;
  98         unsigned          spilled_phi_count;
  99 };
 100
 101 /**
 102  * Compare two spill infos.
 103  */
 104 static int cmp_spillinfo(const void *x, const void *y, size_t size)
 105 {
 106         const spill_info_t *xx = (const spill_info_t*)x;
 107         const spill_info_t *yy = (const spill_info_t*)y;
 108         (void) size;
 109
 110         return xx->to_spill != yy->to_spill;
 111 }
 112
 113 /**
 114  * Returns spill info for a specific value (the value that is to be spilled)
 115  */
 116 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
 117 {
 118         spill_info_t info, *res;
 119         int hash = hash_irn(value);
 120
 121         info.to_spill = value;
 122         res = set_find(spill_info_t, env->spills, &info, sizeof(info), hash);
 123
 124         if (res == NULL) {
 125                 info.reloaders   = NULL;
 126                 info.spills      = NULL;
 127                 info.spill_costs = -1;
 128                 info.reload_cls  = NULL;
 129                 info.spilled_phi = false;
 130                 res = set_insert(spill_info_t, env->spills, &info, sizeof(info), hash);
 131         }
 132
 133         return res;
 134 }
 135
 136 spill_env_t *be_new_spill_env(ir_graph *irg)
 137 {
 138         const arch_env_t *arch_env = be_get_irg_arch_env(irg);
 139
 140         spill_env_t *env = XMALLOC(spill_env_t);
 141         env->spills         = new_set(cmp_spillinfo, 1024);
 142         env->irg            = irg;
 143         env->arch_env       = arch_env;
 144         env->mem_phis       = NEW_ARR_F(spill_info_t*, 0);
 145         env->spill_cost     = arch_env->spill_cost;
 146         env->reload_cost    = arch_env->reload_cost;
 147         obstack_init(&env->obst);
 148
 149         env->spill_count       = 0;
 150         env->reload_count      = 0;
 151         env->remat_count       = 0;
 152         env->spilled_phi_count = 0;
 153
 154         return env;
 155 }
 156
 157 void be_delete_spill_env(spill_env_t *env)
 158 {
 159         del_set(env->spills);
 160         DEL_ARR_F(env->mem_phis);
 161         obstack_free(&env->obst, NULL);
 162         free(env);
 163 }
 164
 165 /*
 166  *  ____  _                  ____      _                 _
 167  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 168  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 169  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 170  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 171  *
 172  */
 173
 174 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 175 {
 176         spill_info_t  *spill_info = get_spillinfo(env, to_spill);
 177         spill_t       *spill;
 178         spill_t       *s;
 179         spill_t       *last;
 180
 181         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 182         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 183
 184         /* Just for safety make sure that we do not insert the spill in front of a phi */
 185         assert(!is_Phi(sched_next(after)));
 186
 187         /* spills that are dominated by others are not needed */
 188         last = NULL;
 189         s    = spill_info->spills;
 190         for ( ; s != NULL; s = s->next) {
 191                 /* no need to add this spill if it is dominated by another */
 192                 if (value_dominates(s->after, after)) {
 193                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 194                         return;
 195                 }
 196                 /* remove spills that we dominate */
 197                 if (value_dominates(after, s->after)) {
 198                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 199                         if (last != NULL) {
 200                                 last->next         = s->next;
 201                         } else {
 202                                 spill_info->spills = s->next;
 203                         }
 204                 } else {
 205                         last = s;
 206                 }
 207         }
 208
 209         spill         = OALLOC(&env->obst, spill_t);
 210         spill->after  = after;
 211         spill->next   = spill_info->spills;
 212         spill->spill  = NULL;
 213
 214         spill_info->spills = spill;
 215 }
 216
 217 void be_add_reload(spill_env_t *env, ir_node *to_spill, ir_node *before, const arch_register_class_t *reload_cls, int allow_remat)
 218 {
 219         spill_info_t  *info;
 220         reloader_t    *rel;
 221
 222         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 223
 224         info = get_spillinfo(env, to_spill);
 225
 226         if (is_Phi(to_spill)) {
 227                 int i, arity;
 228
 229                 /* create spillinfos for the phi arguments */
 230                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 231                         ir_node *arg = get_irn_n(to_spill, i);
 232                         get_spillinfo(env, arg);
 233                 }
 234         }
 235
 236         assert(!be_is_Keep(before));
 237
 238         /* put reload into list */
 239         rel                   = OALLOC(&env->obst, reloader_t);
 240         rel->next             = info->reloaders;
 241         rel->reloader         = before;
 242         rel->rematted_node    = NULL;
 243         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 244
 245         info->reloaders  = rel;
 246         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 247         info->reload_cls = reload_cls;
 248
 249         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 250                 to_spill, before, allow_remat ? "" : " not"));
 251 }
 252
 253 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 254 {
 255         ir_node *last = sched_last(block);
 256
 257         /* we might have keeps behind the jump... */
 258         while (be_is_Keep(last)) {
 259                 last = sched_prev(last);
 260                 assert(!sched_is_end(last));
 261         }
 262
 263         assert(is_cfop(last));
 264
 265         /* add the reload before the (cond-)jump */
 266         return last;
 267 }
 268
 269 /**
 270  * determine final spill position: it should be after all phis, keep nodes
 271  * and behind nodes marked as prolog
 272  */
 273 static ir_node *determine_spill_point(ir_node *node)
 274 {
 275         node = skip_Proj(node);
 276         while (true) {
 277                 ir_node *next = sched_next(node);
 278                 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
 279                         break;
 280                 node = next;
 281         }
 282         return node;
 283 }
 284
 285 /**
 286  * Returns the point at which you can insert a node that should be executed
 287  * before block @p block when coming from pred @p pos.
 288  */
 289 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 290 {
 291         ir_node *predblock;
 292
 293         /* simply add the reload to the beginning of the block if we only have 1
 294          * predecessor. We don't need to check for phis as there can't be any in a
 295          * block with only 1 pred. */
 296         if (get_Block_n_cfgpreds(block) == 1) {
 297                 assert(!is_Phi(sched_first(block)));
 298                 return sched_first(block);
 299         }
 300
 301         /* We have to reload the value in pred-block */
 302         predblock = get_Block_cfgpred_block(block, pos);
 303         return be_get_end_of_block_insertion_point(predblock);
 304 }
 305
 306 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 307                            int pos, const arch_register_class_t *reload_cls,
 308                            int allow_remat)
 309 {
 310         ir_node *before = get_block_insertion_point(block, pos);
 311         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 312 }
 313
 314 void be_spill_phi(spill_env_t *env, ir_node *node)
 315 {
 316         ir_node *block;
 317         int i, arity;
 318         spill_info_t *info;
 319
 320         assert(is_Phi(node));
 321
 322         info              = get_spillinfo(env, node);
 323         info->spilled_phi = true;
 324         ARR_APP1(spill_info_t*, env->mem_phis, info);
 325
 326         /* create spills for the phi arguments */
 327         block = get_nodes_block(node);
 328         for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 329                 ir_node *arg = get_irn_n(node, i);
 330                 ir_node *insert;
 331
 332                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 333                  * and simply always available. */
 334                 if (!sched_is_scheduled(arg)) {
 335                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 336                         insert = be_get_end_of_block_insertion_point(pred_block);
 337                         insert = sched_prev(insert);
 338                 } else {
 339                         insert = determine_spill_point(arg);
 340                 }
 341
 342                 be_add_spill(env, arg, insert);
 343         }
 344 }
 345
 346 /*
 347  *   ____                _         ____        _ _ _
 348  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 349  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 350  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 351  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 352  *                                      |_|
 353  */
 354
 355 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 356
 357 /**
 358  * Creates a spill.
 359  *
 360  * @param senv      the spill environment
 361  * @param irn       the node that should be spilled
 362  * @param ctx_irn   an user of the spilled node
 363  *
 364  * @return a be_Spill node
 365  */
 366 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 367 {
 368         ir_node       *to_spill = spillinfo->to_spill;
 369         const ir_node *insn     = skip_Proj_const(to_spill);
 370         spill_t *spill;
 371
 372         /* determine_spill_costs must have been run before */
 373         assert(spillinfo->spill_costs >= 0);
 374
 375         /* some backends have virtual noreg/unknown nodes that are not scheduled
 376          * and simply always available. */
 377         if (!sched_is_scheduled(insn)) {
 378                 /* override spillinfos or create a new one */
 379                 ir_graph *irg = get_irn_irg(to_spill);
 380                 spillinfo->spills->spill = get_irg_no_mem(irg);
 381                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 382                 return;
 383         }
 384
 385         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 386         spill = spillinfo->spills;
 387         for ( ; spill != NULL; spill = spill->next) {
 388                 ir_node *after = spill->after;
 389                 after = determine_spill_point(after);
 390
 391                 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
 392                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 393                 env->spill_count++;
 394         }
 395         DBG((dbg, LEVEL_1, "\n"));
 396 }
 397
 398 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 399
 400 /**
 401  * If the first usage of a Phi result would be out of memory
 402  * there is no sense in allocating a register for it.
 403  * Thus we spill it and all its operands to the same spill slot.
 404  * Therefore the phi/dataB becomes a phi/Memory
 405  *
 406  * @param senv      the spill environment
 407  * @param phi       the Phi node that should be spilled
 408  * @param ctx_irn   an user of the spilled node
 409  */
 410 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 411 {
 412         ir_graph *irg   = env->irg;
 413         ir_node  *phi   = spillinfo->to_spill;
 414         ir_node  *block = get_nodes_block(phi);
 415         spill_t  *spill;
 416         int       i;
 417
 418         assert(!get_opt_cse());
 419         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 420
 421         /* build a new PhiM */
 422         int       const arity   = get_Phi_n_preds(phi);
 423         ir_node **const ins     = ALLOCAN(ir_node*, arity);
 424         ir_node  *const unknown = new_r_Unknown(irg, mode_M);
 425         for (i = 0; i < arity; ++i) {
 426                 ins[i] = unknown;
 427         }
 428
 429         /* override or replace spills list... */
 430         spill         = OALLOC(&env->obst, spill_t);
 431         spill->after  = determine_spill_point(phi);
 432         spill->spill  = be_new_Phi(block, arity, ins, mode_M, arch_no_register_req);
 433         spill->next   = NULL;
 434         sched_add_after(block, spill->spill);
 435
 436         spillinfo->spills = spill;
 437         env->spilled_phi_count++;
 438
 439         for (i = 0; i < arity; ++i) {
 440                 ir_node      *arg      = get_irn_n(phi, i);
 441                 spill_info_t *arg_info = get_spillinfo(env, arg);
 442
 443                 determine_spill_costs(env, arg_info);
 444                 spill_node(env, arg_info);
 445
 446                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 447         }
 448         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 449              spill->spill));
 450 }
 451
 452 /**
 453  * Spill a node.
 454  *
 455  * @param senv      the spill environment
 456  * @param to_spill  the node that should be spilled
 457  */
 458 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 459 {
 460         /* node is already spilled */
 461         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 462                 return;
 463
 464         if (spillinfo->spilled_phi) {
 465                 spill_phi(env, spillinfo);
 466         } else {
 467                 spill_irn(env, spillinfo);
 468         }
 469 }
 470
 471 /*
 472  *
 473  *  ____                      _            _       _ _
 474  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 475  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 476  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 477  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 478  *
 479  */
 480
 481 /**
 482  * Tests whether value @p arg is available before node @p reloader
 483  * @returns 1 if value is available, 0 otherwise
 484  */
 485 static int is_value_available(spill_env_t *env, const ir_node *arg,
 486                               const ir_node *reloader)
 487 {
 488         if (is_Unknown(arg) || is_NoMem(arg))
 489                 return 1;
 490
 491         if (arch_irn_is(skip_Proj_const(arg), spill))
 492                 return 1;
 493
 494         if (arg == get_irg_frame(env->irg))
 495                 return 1;
 496
 497         (void)reloader;
 498
 499         if (get_irn_mode(arg) == mode_T)
 500                 return 0;
 501
 502         /*
 503          * Ignore registers are always available
 504          */
 505         if (arch_irn_is_ignore(arg))
 506                 return 1;
 507
 508         return 0;
 509 }
 510
 511 /**
 512  * Check if a node is rematerializable. This tests for the following conditions:
 513  *
 514  * - The node itself is rematerializable
 515  * - All arguments of the node are available or also rematerialisable
 516  * - The costs for the rematerialisation operation is less or equal a limit
 517  *
 518  * Returns the costs needed for rematerialisation or something
 519  * >= REMAT_COST_INFINITE if remat is not possible.
 520  */
 521 static int check_remat_conditions_costs(spill_env_t *env,
 522                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 523 {
 524         int i, arity;
 525         int argremats;
 526         int costs = 0;
 527         const ir_node *insn = skip_Proj_const(spilled);
 528
 529         assert(!be_is_Spill(insn));
 530         if (!arch_irn_is(insn, rematerializable))
 531                 return REMAT_COST_INFINITE;
 532
 533         if (be_is_Reload(insn)) {
 534                 costs += 2;
 535         } else {
 536                 costs += arch_get_op_estimated_cost(insn);
 537         }
 538         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 539                 return REMAT_COST_INFINITE;
 540         }
 541         /* never rematerialize a node which modifies the flags.
 542          * (would be better to test whether the flags are actually live at point
 543          * reloader...)
 544          */
 545         if (arch_irn_is(insn, modify_flags)) {
 546                 return REMAT_COST_INFINITE;
 547         }
 548
 549         argremats = 0;
 550         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 551                 ir_node *arg = get_irn_n(insn, i);
 552
 553                 if (is_value_available(env, arg, reloader))
 554                         continue;
 555
 556                 /* we have to rematerialize the argument as well */
 557                 ++argremats;
 558                 if (argremats > 1) {
 559                         /* we only support rematerializing 1 argument at the moment,
 560                          * as multiple arguments could increase register pressure */
 561                         return REMAT_COST_INFINITE;
 562                 }
 563
 564                 costs += check_remat_conditions_costs(env, arg, reloader,
 565                                                       parentcosts + costs);
 566                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 567                         return REMAT_COST_INFINITE;
 568         }
 569
 570         return costs;
 571 }
 572
 573 /**
 574  * Re-materialize a node.
 575  *
 576  * @param env       the spill environment
 577  * @param spilled   the node that was spilled
 578  * @param reloader  a irn that requires a reload
 579  */
 580 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 581 {
 582         int i, arity;
 583         ir_node *res;
 584         ir_node **ins;
 585
 586         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 587         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 588                 ir_node *arg = get_irn_n(spilled, i);
 589
 590                 if (is_value_available(env, arg, reloader)) {
 591                         ins[i] = arg;
 592                 } else {
 593                         ins[i] = do_remat(env, arg, reloader);
 594                         /* don't count the argument rematerialization as an extra remat */
 595                         --env->remat_count;
 596                 }
 597         }
 598
 599         /* create a copy of the node */
 600         ir_node *const bl = get_nodes_block(reloader);
 601         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 602                           get_irn_op(spilled), get_irn_mode(spilled),
 603                           get_irn_arity(spilled), ins);
 604         copy_node_attr(env->irg, spilled, res);
 605         arch_env_mark_remat(env->arch_env, res);
 606
 607         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 608
 609         if (! is_Proj(res)) {
 610                 /* insert in schedule */
 611                 sched_reset(res);
 612                 sched_add_before(reloader, res);
 613                 ++env->remat_count;
 614         }
 615
 616         return res;
 617 }
 618
 619 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 620 {
 621         ir_node *block = get_nodes_block(before);
 622         double   freq  = get_block_execfreq(block);
 623         (void) to_spill;
 624
 625         return env->spill_cost * freq;
 626 }
 627
 628 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 629                                        const ir_node *before)
 630 {
 631         if (be_do_remats) {
 632                 /* is the node rematerializable? */
 633                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 634                 if (costs < (unsigned) env->reload_cost)
 635                         return costs;
 636         }
 637
 638         return env->reload_cost;
 639 }
 640
 641 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 642 {
 643         ir_node *block = get_nodes_block(before);
 644         double   freq  = get_block_execfreq(block);
 645
 646         if (be_do_remats) {
 647                 /* is the node rematerializable? */
 648                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 649                 if (costs < env->reload_cost)
 650                         return costs * freq;
 651         }
 652
 653         return env->reload_cost * freq;
 654 }
 655
 656 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 657                            const ir_node *before)
 658 {
 659         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 660 }
 661
 662 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 663                                    ir_node *block, int pos)
 664 {
 665         ir_node *before = get_block_insertion_point(block, pos);
 666         return be_get_reload_costs(env, to_spill, before);
 667 }
 668
 669 ir_node *be_new_spill(ir_node *value, ir_node *after)
 670 {
 671         ir_graph                    *irg       = get_irn_irg(value);
 672         ir_node                     *frame     = get_irg_frame(irg);
 673         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 674         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 675         ir_node                     *block     = get_block(after);
 676         ir_node                     *spill
 677                 = be_new_Spill(cls, cls_frame, block, frame, value);
 678
 679         sched_add_after(after, spill);
 680         return spill;
 681 }
 682
 683 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
 684 {
 685         ir_graph *irg   = get_irn_irg(value);
 686         ir_node  *frame = get_irg_frame(irg);
 687         ir_node  *block = get_block(before);
 688         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 689         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 690         ir_mode                     *mode      = get_irn_mode(value);
 691         ir_node  *reload;
 692
 693         assert(be_is_Spill(spill) || is_Phi(spill));
 694         assert(get_irn_mode(spill) == mode_M);
 695
 696         reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
 697         sched_add_before(before, reload);
 698
 699         return reload;
 700 }
 701
 702 /*
 703  *  ___                     _     ____      _                 _
 704  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 705  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 706  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 707  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 708  *
 709  */
 710
 711 /**
 712  * analyzes how to best spill a node and determine costs for that
 713  */
 714 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 715 {
 716         ir_node       *to_spill = spillinfo->to_spill;
 717         const ir_node *insn     = skip_Proj_const(to_spill);
 718         ir_node       *spill_block;
 719         spill_t       *spill;
 720         double         spill_execfreq;
 721
 722         /* already calculated? */
 723         if (spillinfo->spill_costs >= 0)
 724                 return;
 725
 726         assert(!arch_irn_is(insn, dont_spill));
 727         assert(!be_is_Reload(insn));
 728
 729         /* some backends have virtual noreg/unknown nodes that are not scheduled
 730          * and simply always available.
 731          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 732          * predecessor (of a PhiM) but this test might match other things too...
 733          */
 734         if (!sched_is_scheduled(insn)) {
 735                 ir_graph *irg = get_irn_irg(to_spill);
 736                 /* override spillinfos or create a new one */
 737                 spill_t *spill = OALLOC(&env->obst, spill_t);
 738                 spill->after = NULL;
 739                 spill->next  = NULL;
 740                 spill->spill = get_irg_no_mem(irg);
 741
 742                 spillinfo->spills      = spill;
 743                 spillinfo->spill_costs = 0;
 744
 745                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 746                 return;
 747         }
 748
 749         spill_block    = get_nodes_block(insn);
 750         spill_execfreq = get_block_execfreq(spill_block);
 751
 752         if (spillinfo->spilled_phi) {
 753                 /* TODO calculate correct costs...
 754                  * (though we can't remat this node anyway so no big problem) */
 755                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 756                 return;
 757         }
 758
 759         if (spillinfo->spills != NULL) {
 760                 spill_t *s;
 761                 double   spills_execfreq;
 762
 763                 /* calculate sum of execution frequencies of individual spills */
 764                 spills_execfreq = 0;
 765                 s               = spillinfo->spills;
 766                 for ( ; s != NULL; s = s->next) {
 767                         ir_node *spill_block = get_block(s->after);
 768                         double   freq = get_block_execfreq(spill_block);
 769
 770                         spills_execfreq += freq;
 771                 }
 772
 773                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 774                     spills_execfreq * env->spill_cost,
 775                     spill_execfreq * env->spill_cost));
 776
 777                 /* multi-/latespill is advantageous -> return*/
 778                 if (spills_execfreq < spill_execfreq) {
 779                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 780                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 781                         return;
 782                 }
 783         }
 784
 785         /* override spillinfos or create a new one */
 786         spill        = OALLOC(&env->obst, spill_t);
 787         spill->after = determine_spill_point(to_spill);
 788         spill->next  = NULL;
 789         spill->spill = NULL;
 790
 791         spillinfo->spills      = spill;
 792         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 793         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 794 }
 795
 796 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 797 {
 798         const spill_info_t *si = get_spillinfo(env, irn);
 799         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 800         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 801         bitset_t *reloads      = bitset_alloca(n_blocks);
 802         reloader_t *r;
 803         spill_t *s;
 804
 805         if (si == NULL)
 806                 return;
 807
 808         /* Fill the bitset with the dominance pre-order numbers
 809          * of the blocks the reloads are located in. */
 810         for (r = si->reloaders; r != NULL; r = r->next) {
 811                 ir_node *bl = get_nodes_block(r->reloader);
 812                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 813         }
 814
 815         /* Now, cancel out all the blocks that are dominated by each spill.
 816          * If the bitset is not empty after that, we have reloads that are
 817          * not dominated by any spill. */
 818         for (s = si->spills; s != NULL; s = s->next) {
 819                 ir_node *bl = get_nodes_block(s->after);
 820                 int start   = get_Block_dom_tree_pre_num(bl);
 821                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 822
 823                 bitset_clear_range(reloads, start, end);
 824         }
 825
 826         if (!bitset_is_empty(reloads))
 827                 be_add_spill(env, si->to_spill, si->to_spill);
 828 }
 829
 830 void be_insert_spills_reloads(spill_env_t *env)
 831 {
 832         size_t n_mem_phis = ARR_LEN(env->mem_phis);
 833         size_t i;
 834
 835         be_timer_push(T_RA_SPILL_APPLY);
 836
 837         /* create all phi-ms first, this is needed so, that phis, hanging on
 838            spilled phis work correctly */
 839         for (i = 0; i < n_mem_phis; ++i) {
 840                 spill_info_t *info = env->mem_phis[i];
 841                 spill_node(env, info);
 842         }
 843
 844         /* process each spilled node */
 845         foreach_set(env->spills, spill_info_t, si) {
 846                 ir_node  *to_spill        = si->to_spill;
 847                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 848                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 849                 bool      force_remat     = false;
 850                 reloader_t *rld;
 851
 852                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 853
 854                 determine_spill_costs(env, si);
 855
 856                 /* determine possibility of rematerialisations */
 857                 if (be_do_remats) {
 858                         /* calculate cost savings for each indivial value when it would
 859                            be rematted instead of reloaded */
 860                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 861                                 double   freq;
 862                                 int      remat_cost;
 863                                 int      remat_cost_delta;
 864                                 ir_node *block;
 865                                 ir_node *reloader = rld->reloader;
 866
 867                                 if (rld->rematted_node != NULL) {
 868                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 869                                              rld->rematted_node, reloader));
 870                                         continue;
 871                                 }
 872                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 873                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 874                                              reloader));
 875                                         all_remat_costs = REMAT_COST_INFINITE;
 876                                         continue;
 877                                 }
 878
 879                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 880                                                                            reloader, 0);
 881                                 if (remat_cost >= REMAT_COST_INFINITE) {
 882                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 883                                              reloader));
 884                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 885                                         all_remat_costs       = REMAT_COST_INFINITE;
 886                                         continue;
 887                                 }
 888
 889                                 remat_cost_delta      = remat_cost - env->reload_cost;
 890                                 rld->remat_cost_delta = remat_cost_delta;
 891                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 892                                 freq                  = get_block_execfreq(block);
 893                                 all_remat_costs      += remat_cost_delta * freq;
 894                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 895                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 896                                      remat_cost_delta * freq));
 897                         }
 898                         if (all_remat_costs < REMAT_COST_INFINITE) {
 899                                 /* we don't need the costs for the spill if we can remat
 900                                    all reloaders */
 901                                 all_remat_costs -= si->spill_costs;
 902
 903                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 904                                      env->spill_cost, si->spill_costs));
 905                         }
 906
 907                         if (all_remat_costs < 0) {
 908                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 909                                      all_remat_costs));
 910                                 force_remat = true;
 911                         }
 912                 }
 913
 914                 /* go through all reloads for this spill */
 915                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 916                         ir_node *copy; /* a reload is a "copy" of the original value */
 917
 918                         if (rld->rematted_node != NULL) {
 919                                 copy = rld->rematted_node;
 920                                 sched_add_before(rld->reloader, copy);
 921                         } else if (be_do_remats &&
 922                                         (force_remat || rld->remat_cost_delta < 0)) {
 923                                 copy = do_remat(env, to_spill, rld->reloader);
 924                         } else {
 925                                 /* make sure we have a spill */
 926                                 spill_node(env, si);
 927
 928                                 /* create a reload, use the first spill for now SSA
 929                                  * reconstruction for memory comes below */
 930                                 assert(si->spills != NULL);
 931                                 copy = arch_env_new_reload(env->arch_env, si->to_spill,
 932                                                            si->spills->spill, rld->reloader);
 933                                 env->reload_count++;
 934                         }
 935
 936                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 937                              copy, to_spill, rld->reloader));
 938                         ARR_APP1(ir_node*, copies, copy);
 939                 }
 940
 941                 /* if we had any reloads or remats, then we need to reconstruct the
 942                  * SSA form for the spilled value */
 943                 if (ARR_LEN(copies) > 0) {
 944                         be_ssa_construction_env_t senv;
 945                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 946
 947                         be_ssa_construction_init(&senv, env->irg);
 948                         be_ssa_construction_add_copy(&senv, to_spill);
 949                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 950                         be_ssa_construction_fix_users(&senv, to_spill);
 951
 952                         be_ssa_construction_destroy(&senv);
 953                 }
 954                 /* need to reconstruct SSA form if we had multiple spills */
 955                 if (si->spills != NULL && si->spills->next != NULL) {
 956                         spill_t *spill;
 957                         int      spill_count = 0;
 958
 959                         be_ssa_construction_env_t senv;
 960
 961                         be_ssa_construction_init(&senv, env->irg);
 962                         spill = si->spills;
 963                         for ( ; spill != NULL; spill = spill->next) {
 964                                 /* maybe we rematerialized the value and need no spill */
 965                                 if (spill->spill == NULL)
 966                                         continue;
 967                                 be_ssa_construction_add_copy(&senv, spill->spill);
 968                                 spill_count++;
 969                         }
 970                         if (spill_count > 1) {
 971                                 /* all reloads are attached to the first spill, fix them now */
 972                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
 973                         }
 974
 975                         be_ssa_construction_destroy(&senv);
 976                 }
 977
 978                 DEL_ARR_F(copies);
 979                 si->reloaders = NULL;
 980         }
 981
 982         stat_ev_dbl("spill_spills", env->spill_count);
 983         stat_ev_dbl("spill_reloads", env->reload_count);
 984         stat_ev_dbl("spill_remats", env->remat_count);
 985         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
 986
 987         /* Matze: In theory be_ssa_construction should take care of the liveness...
 988          * try to disable this again in the future */
 989         be_invalidate_live_sets(env->irg);
 990
 991         be_remove_dead_nodes_from_schedule(env->irg);
 992
 993         be_timer_pop(T_RA_SPILL_APPLY);
 994 }
 995
 996 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
 997 void be_init_spill(void)
 998 {
 999         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1000 }