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 "statev_t.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
 110         unsigned          spill_count;
 111         unsigned          reload_count;
 112         unsigned          remat_count;
 113         unsigned          spilled_phi_count;
 114 };
 115
 116 /**
 117  * Compare two spill infos.
 118  */
 119 static int cmp_spillinfo(const void *x, const void *y, size_t size)
 120 {
 121         const spill_info_t *xx = (const spill_info_t*)x;
 122         const spill_info_t *yy = (const spill_info_t*)y;
 123         (void) size;
 124
 125         return xx->to_spill != yy->to_spill;
 126 }
 127
 128 /**
 129  * Returns spill info for a specific value (the value that is to be spilled)
 130  */
 131 static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value)
 132 {
 133         spill_info_t info, *res;
 134         int hash = hash_irn(value);
 135
 136         info.to_spill = value;
 137         res = set_find(spill_info_t, env->spills, &info, sizeof(info), hash);
 138
 139         if (res == NULL) {
 140                 info.reloaders   = NULL;
 141                 info.spills      = NULL;
 142                 info.spill_costs = -1;
 143                 info.reload_cls  = NULL;
 144                 info.spilled_phi = false;
 145                 res = set_insert(spill_info_t, env->spills, &info, sizeof(info), hash);
 146         }
 147
 148         return res;
 149 }
 150
 151 spill_env_t *be_new_spill_env(ir_graph *irg)
 152 {
 153         const arch_env_t *arch_env = be_get_irg_arch_env(irg);
 154
 155         spill_env_t *env = XMALLOC(spill_env_t);
 156         env->spills         = new_set(cmp_spillinfo, 1024);
 157         env->irg            = irg;
 158         env->arch_env       = arch_env;
 159         env->mem_phis       = NEW_ARR_F(spill_info_t*, 0);
 160         env->spill_cost     = arch_env->spill_cost;
 161         env->reload_cost    = arch_env->reload_cost;
 162         obstack_init(&env->obst);
 163
 164         env->spill_count       = 0;
 165         env->reload_count      = 0;
 166         env->remat_count       = 0;
 167         env->spilled_phi_count = 0;
 168
 169         return env;
 170 }
 171
 172 void be_delete_spill_env(spill_env_t *env)
 173 {
 174         del_set(env->spills);
 175         DEL_ARR_F(env->mem_phis);
 176         obstack_free(&env->obst, NULL);
 177         free(env);
 178 }
 179
 180 /*
 181  *  ____  _                  ____      _                 _
 182  * |  _ \| | __ _  ___ ___  |  _ \ ___| | ___   __ _  __| |___
 183  * | |_) | |/ _` |/ __/ _ \ | |_) / _ \ |/ _ \ / _` |/ _` / __|
 184  * |  __/| | (_| | (_|  __/ |  _ <  __/ | (_) | (_| | (_| \__ \
 185  * |_|   |_|\__,_|\___\___| |_| \_\___|_|\___/ \__,_|\__,_|___/
 186  *
 187  */
 188
 189 void be_add_spill(spill_env_t *env, ir_node *to_spill, ir_node *after)
 190 {
 191         spill_info_t  *spill_info = get_spillinfo(env, to_spill);
 192         spill_t       *spill;
 193         spill_t       *s;
 194         spill_t       *last;
 195
 196         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 197         DB((dbg, LEVEL_1, "Add spill of %+F after %+F\n", to_spill, after));
 198
 199         /* Just for safety make sure that we do not insert the spill in front of a phi */
 200         assert(!is_Phi(sched_next(after)));
 201
 202         /* spills that are dominated by others are not needed */
 203         last = NULL;
 204         s    = spill_info->spills;
 205         for ( ; s != NULL; s = s->next) {
 206                 /* no need to add this spill if it is dominated by another */
 207                 if (value_dominates(s->after, after)) {
 208                         DB((dbg, LEVEL_1, "...dominated by %+F, not added\n", s->after));
 209                         return;
 210                 }
 211                 /* remove spills that we dominate */
 212                 if (value_dominates(after, s->after)) {
 213                         DB((dbg, LEVEL_1, "...remove old spill at %+F\n", s->after));
 214                         if (last != NULL) {
 215                                 last->next         = s->next;
 216                         } else {
 217                                 spill_info->spills = s->next;
 218                         }
 219                 } else {
 220                         last = s;
 221                 }
 222         }
 223
 224         spill         = OALLOC(&env->obst, spill_t);
 225         spill->after  = after;
 226         spill->next   = spill_info->spills;
 227         spill->spill  = NULL;
 228
 229         spill_info->spills = spill;
 230 }
 231
 232 void be_add_reload2(spill_env_t *env, ir_node *to_spill, ir_node *before,
 233                 ir_node *can_spill_after, const arch_register_class_t *reload_cls,
 234                 int allow_remat)
 235 {
 236         spill_info_t  *info;
 237         reloader_t    *rel;
 238
 239         assert(!arch_irn_is(skip_Proj_const(to_spill), dont_spill));
 240
 241         info = get_spillinfo(env, to_spill);
 242
 243         if (is_Phi(to_spill)) {
 244                 int i, arity;
 245
 246                 /* create spillinfos for the phi arguments */
 247                 for (i = 0, arity = get_irn_arity(to_spill); i < arity; ++i) {
 248                         ir_node *arg = get_irn_n(to_spill, i);
 249                         get_spillinfo(env, arg);
 250                 }
 251         }
 252
 253         assert(!is_Proj(before) && !be_is_Keep(before));
 254
 255         /* put reload into list */
 256         rel                   = OALLOC(&env->obst, reloader_t);
 257         rel->next             = info->reloaders;
 258         rel->reloader         = before;
 259         rel->rematted_node    = NULL;
 260         rel->can_spill_after  = can_spill_after;
 261         rel->remat_cost_delta = allow_remat ? 0 : REMAT_COST_INFINITE;
 262
 263         info->reloaders  = rel;
 264         assert(info->reload_cls == NULL || info->reload_cls == reload_cls);
 265         info->reload_cls = reload_cls;
 266
 267         DBG((dbg, LEVEL_1, "creating spillinfo for %+F, will be reloaded before %+F, may%s be rematerialized\n",
 268                 to_spill, before, allow_remat ? "" : " not"));
 269 }
 270
 271 void be_add_reload(spill_env_t *senv, ir_node *to_spill, ir_node *before,
 272                    const arch_register_class_t *reload_cls, int allow_remat)
 273 {
 274         be_add_reload2(senv, to_spill, before, to_spill, reload_cls, allow_remat);
 275
 276 }
 277
 278 ir_node *be_get_end_of_block_insertion_point(const ir_node *block)
 279 {
 280         ir_node *last = sched_last(block);
 281
 282         /* we might have keeps behind the jump... */
 283         while (be_is_Keep(last)) {
 284                 last = sched_prev(last);
 285                 assert(!sched_is_end(last));
 286         }
 287
 288         assert(is_cfop(last));
 289
 290         /* add the reload before the (cond-)jump */
 291         return last;
 292 }
 293
 294 /**
 295  * determine final spill position: it should be after all phis, keep nodes
 296  * and behind nodes marked as prolog
 297  */
 298 static ir_node *determine_spill_point(ir_node *node)
 299 {
 300         node = skip_Proj(node);
 301         while (true) {
 302                 ir_node *next = sched_next(node);
 303                 if (!is_Phi(next) && !be_is_Keep(next) && !be_is_CopyKeep(next))
 304                         break;
 305                 node = next;
 306         }
 307         return node;
 308 }
 309
 310 /**
 311  * Returns the point at which you can insert a node that should be executed
 312  * before block @p block when coming from pred @p pos.
 313  */
 314 static ir_node *get_block_insertion_point(ir_node *block, int pos)
 315 {
 316         ir_node *predblock;
 317
 318         /* simply add the reload to the beginning of the block if we only have 1
 319          * predecessor. We don't need to check for phis as there can't be any in a
 320          * block with only 1 pred. */
 321         if (get_Block_n_cfgpreds(block) == 1) {
 322                 assert(!is_Phi(sched_first(block)));
 323                 return sched_first(block);
 324         }
 325
 326         /* We have to reload the value in pred-block */
 327         predblock = get_Block_cfgpred_block(block, pos);
 328         return be_get_end_of_block_insertion_point(predblock);
 329 }
 330
 331 void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block,
 332                            int pos, const arch_register_class_t *reload_cls,
 333                            int allow_remat)
 334 {
 335         ir_node *before = get_block_insertion_point(block, pos);
 336         be_add_reload(env, to_spill, before, reload_cls, allow_remat);
 337 }
 338
 339 void be_spill_phi(spill_env_t *env, ir_node *node)
 340 {
 341         ir_node *block;
 342         int i, arity;
 343         spill_info_t *info;
 344
 345         assert(is_Phi(node));
 346
 347         info              = get_spillinfo(env, node);
 348         info->spilled_phi = true;
 349         ARR_APP1(spill_info_t*, env->mem_phis, info);
 350
 351         /* create spills for the phi arguments */
 352         block = get_nodes_block(node);
 353         for (i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 354                 ir_node *arg = get_irn_n(node, i);
 355                 ir_node *insert;
 356
 357                 /* some backends have virtual noreg/unknown nodes that are not scheduled
 358                  * and simply always available. */
 359                 if (!sched_is_scheduled(arg)) {
 360                         ir_node *pred_block = get_Block_cfgpred_block(block, i);
 361                         insert = be_get_end_of_block_insertion_point(pred_block);
 362                         insert = sched_prev(insert);
 363                 } else {
 364                         insert = determine_spill_point(arg);
 365                 }
 366
 367                 be_add_spill(env, arg, insert);
 368         }
 369 }
 370
 371 /*
 372  *   ____                _         ____        _ _ _
 373  *  / ___|_ __ ___  __ _| |_ ___  / ___| _ __ (_) | |___
 374  * | |   | '__/ _ \/ _` | __/ _ \ \___ \| '_ \| | | / __|
 375  * | |___| | |  __/ (_| | ||  __/  ___) | |_) | | | \__ \
 376  *  \____|_|  \___|\__,_|\__\___| |____/| .__/|_|_|_|___/
 377  *                                      |_|
 378  */
 379
 380 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo);
 381
 382 /**
 383  * Creates a spill.
 384  *
 385  * @param senv      the spill environment
 386  * @param irn       the node that should be spilled
 387  * @param ctx_irn   an user of the spilled node
 388  *
 389  * @return a be_Spill node
 390  */
 391 static void spill_irn(spill_env_t *env, spill_info_t *spillinfo)
 392 {
 393         ir_node       *to_spill = spillinfo->to_spill;
 394         const ir_node *insn     = skip_Proj_const(to_spill);
 395         spill_t *spill;
 396
 397         /* determine_spill_costs must have been run before */
 398         assert(spillinfo->spill_costs >= 0);
 399
 400         /* some backends have virtual noreg/unknown nodes that are not scheduled
 401          * and simply always available. */
 402         if (!sched_is_scheduled(insn)) {
 403                 /* override spillinfos or create a new one */
 404                 ir_graph *irg = get_irn_irg(to_spill);
 405                 spillinfo->spills->spill = get_irg_no_mem(irg);
 406                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 407                 return;
 408         }
 409
 410         DBG((dbg, LEVEL_1, "spilling %+F ... \n", to_spill));
 411         spill = spillinfo->spills;
 412         for ( ; spill != NULL; spill = spill->next) {
 413                 ir_node *after = spill->after;
 414                 after = determine_spill_point(after);
 415
 416                 spill->spill = arch_env_new_spill(env->arch_env, to_spill, after);
 417                 DB((dbg, LEVEL_1, "\t%+F after %+F\n", spill->spill, after));
 418                 env->spill_count++;
 419         }
 420         DBG((dbg, LEVEL_1, "\n"));
 421 }
 422
 423 static void spill_node(spill_env_t *env, spill_info_t *spillinfo);
 424
 425 /**
 426  * If the first usage of a Phi result would be out of memory
 427  * there is no sense in allocating a register for it.
 428  * Thus we spill it and all its operands to the same spill slot.
 429  * Therefore the phi/dataB becomes a phi/Memory
 430  *
 431  * @param senv      the spill environment
 432  * @param phi       the Phi node that should be spilled
 433  * @param ctx_irn   an user of the spilled node
 434  */
 435 static void spill_phi(spill_env_t *env, spill_info_t *spillinfo)
 436 {
 437         ir_graph *irg   = env->irg;
 438         ir_node  *phi   = spillinfo->to_spill;
 439         ir_node  *block = get_nodes_block(phi);
 440         ir_node  *unknown;
 441         ir_node **ins;
 442         spill_t  *spill;
 443         int       i;
 444         int       arity;
 445
 446         assert(is_Phi(phi));
 447         assert(!get_opt_cse());
 448         DBG((dbg, LEVEL_1, "spilling Phi %+F:\n", phi));
 449
 450         /* build a new PhiM */
 451         arity   = get_irn_arity(phi);
 452         ins     = ALLOCAN(ir_node*, arity);
 453         unknown = new_r_Unknown(irg, mode_M);
 454         for (i = 0; i < arity; ++i) {
 455                 ins[i] = unknown;
 456         }
 457
 458         /* override or replace spills list... */
 459         spill         = OALLOC(&env->obst, spill_t);
 460         spill->after  = determine_spill_point(phi);
 461         spill->spill  = be_new_Phi(block, arity, ins, mode_M, arch_no_register_req);
 462         spill->next   = NULL;
 463         sched_add_after(block, spill->spill);
 464
 465         spillinfo->spills = spill;
 466         env->spilled_phi_count++;
 467
 468         for (i = 0; i < arity; ++i) {
 469                 ir_node      *arg      = get_irn_n(phi, i);
 470                 spill_info_t *arg_info = get_spillinfo(env, arg);
 471
 472                 determine_spill_costs(env, arg_info);
 473                 spill_node(env, arg_info);
 474
 475                 set_irn_n(spill->spill, i, arg_info->spills->spill);
 476         }
 477         DBG((dbg, LEVEL_1, "... done spilling Phi %+F, created PhiM %+F\n", phi,
 478              spill->spill));
 479 }
 480
 481 /**
 482  * Spill a node.
 483  *
 484  * @param senv      the spill environment
 485  * @param to_spill  the node that should be spilled
 486  */
 487 static void spill_node(spill_env_t *env, spill_info_t *spillinfo)
 488 {
 489         /* node is already spilled */
 490         if (spillinfo->spills != NULL && spillinfo->spills->spill != NULL)
 491                 return;
 492
 493         if (spillinfo->spilled_phi) {
 494                 spill_phi(env, spillinfo);
 495         } else {
 496                 spill_irn(env, spillinfo);
 497         }
 498 }
 499
 500 /*
 501  *
 502  *  ____                      _            _       _ _
 503  * |  _ \ ___ _ __ ___   __ _| |_ ___ _ __(_) __ _| (_)_______
 504  * | |_) / _ \ '_ ` _ \ / _` | __/ _ \ '__| |/ _` | | |_  / _ \
 505  * |  _ <  __/ | | | | | (_| | ||  __/ |  | | (_| | | |/ /  __/
 506  * |_| \_\___|_| |_| |_|\__,_|\__\___|_|  |_|\__,_|_|_/___\___|
 507  *
 508  */
 509
 510 /**
 511  * Tests whether value @p arg is available before node @p reloader
 512  * @returns 1 if value is available, 0 otherwise
 513  */
 514 static int is_value_available(spill_env_t *env, const ir_node *arg,
 515                               const ir_node *reloader)
 516 {
 517         if (is_Unknown(arg) || is_NoMem(arg))
 518                 return 1;
 519
 520         if (be_is_Spill(skip_Proj_const(arg)))
 521                 return 1;
 522
 523         if (arg == get_irg_frame(env->irg))
 524                 return 1;
 525
 526         (void)reloader;
 527
 528         if (get_irn_mode(arg) == mode_T)
 529                 return 0;
 530
 531         /*
 532          * Ignore registers are always available
 533          */
 534         if (arch_irn_is_ignore(arg))
 535                 return 1;
 536
 537         return 0;
 538 }
 539
 540 /**
 541  * Check if a node is rematerializable. This tests for the following conditions:
 542  *
 543  * - The node itself is rematerializable
 544  * - All arguments of the node are available or also rematerialisable
 545  * - The costs for the rematerialisation operation is less or equal a limit
 546  *
 547  * Returns the costs needed for rematerialisation or something
 548  * >= REMAT_COST_INFINITE if remat is not possible.
 549  */
 550 static int check_remat_conditions_costs(spill_env_t *env,
 551                 const ir_node *spilled, const ir_node *reloader, int parentcosts)
 552 {
 553         int i, arity;
 554         int argremats;
 555         int costs = 0;
 556         const ir_node *insn = skip_Proj_const(spilled);
 557
 558         assert(!be_is_Spill(insn));
 559         if (!arch_irn_is(insn, rematerializable))
 560                 return REMAT_COST_INFINITE;
 561
 562         if (be_is_Reload(insn)) {
 563                 costs += 2;
 564         } else {
 565                 costs += arch_get_op_estimated_cost(insn);
 566         }
 567         if (parentcosts + costs >= env->reload_cost + env->spill_cost) {
 568                 return REMAT_COST_INFINITE;
 569         }
 570         /* never rematerialize a node which modifies the flags.
 571          * (would be better to test whether the flags are actually live at point
 572          * reloader...)
 573          */
 574         if (arch_irn_is(insn, modify_flags)) {
 575                 return REMAT_COST_INFINITE;
 576         }
 577
 578         argremats = 0;
 579         for (i = 0, arity = get_irn_arity(insn); i < arity; ++i) {
 580                 ir_node *arg = get_irn_n(insn, i);
 581
 582                 if (is_value_available(env, arg, reloader))
 583                         continue;
 584
 585                 /* we have to rematerialize the argument as well */
 586                 ++argremats;
 587                 if (argremats > 1) {
 588                         /* we only support rematerializing 1 argument at the moment,
 589                          * as multiple arguments could increase register pressure */
 590                         return REMAT_COST_INFINITE;
 591                 }
 592
 593                 costs += check_remat_conditions_costs(env, arg, reloader,
 594                                                       parentcosts + costs);
 595                 if (parentcosts + costs >= env->reload_cost + env->spill_cost)
 596                         return REMAT_COST_INFINITE;
 597         }
 598
 599         return costs;
 600 }
 601
 602 /**
 603  * Re-materialize a node.
 604  *
 605  * @param env       the spill environment
 606  * @param spilled   the node that was spilled
 607  * @param reloader  a irn that requires a reload
 608  */
 609 static ir_node *do_remat(spill_env_t *env, ir_node *spilled, ir_node *reloader)
 610 {
 611         int i, arity;
 612         ir_node *res;
 613         ir_node *bl;
 614         ir_node **ins;
 615
 616         if (is_Block(reloader)) {
 617                 bl = reloader;
 618         } else {
 619                 bl = get_nodes_block(reloader);
 620         }
 621
 622         ins = ALLOCAN(ir_node*, get_irn_arity(spilled));
 623         for (i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
 624                 ir_node *arg = get_irn_n(spilled, i);
 625
 626                 if (is_value_available(env, arg, reloader)) {
 627                         ins[i] = arg;
 628                 } else {
 629                         ins[i] = do_remat(env, arg, reloader);
 630                         /* don't count the argument rematerialization as an extra remat */
 631                         --env->remat_count;
 632                 }
 633         }
 634
 635         /* create a copy of the node */
 636         res = new_ir_node(get_irn_dbg_info(spilled), env->irg, bl,
 637                           get_irn_op(spilled), get_irn_mode(spilled),
 638                           get_irn_arity(spilled), ins);
 639         copy_node_attr(env->irg, spilled, res);
 640         arch_env_mark_remat(env->arch_env, res);
 641
 642         DBG((dbg, LEVEL_1, "Insert remat %+F of %+F before reloader %+F\n", res, spilled, reloader));
 643
 644         if (! is_Proj(res)) {
 645                 /* insert in schedule */
 646                 sched_reset(res);
 647                 sched_add_before(reloader, res);
 648                 ++env->remat_count;
 649         }
 650
 651         return res;
 652 }
 653
 654 double be_get_spill_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 655 {
 656         ir_node *block = get_nodes_block(before);
 657         double   freq  = get_block_execfreq(block);
 658         (void) to_spill;
 659
 660         return env->spill_cost * freq;
 661 }
 662
 663 unsigned be_get_reload_costs_no_weight(spill_env_t *env, const ir_node *to_spill,
 664                                        const ir_node *before)
 665 {
 666         if (be_do_remats) {
 667                 /* is the node rematerializable? */
 668                 unsigned costs = check_remat_conditions_costs(env, to_spill, before, 0);
 669                 if (costs < (unsigned) env->reload_cost)
 670                         return costs;
 671         }
 672
 673         return env->reload_cost;
 674 }
 675
 676 double be_get_reload_costs(spill_env_t *env, ir_node *to_spill, ir_node *before)
 677 {
 678         ir_node *block = get_nodes_block(before);
 679         double   freq  = get_block_execfreq(block);
 680
 681         if (be_do_remats) {
 682                 /* is the node rematerializable? */
 683                 int costs = check_remat_conditions_costs(env, to_spill, before, 0);
 684                 if (costs < env->reload_cost)
 685                         return costs * freq;
 686         }
 687
 688         return env->reload_cost * freq;
 689 }
 690
 691 int be_is_rematerializable(spill_env_t *env, const ir_node *to_remat,
 692                            const ir_node *before)
 693 {
 694         return check_remat_conditions_costs(env, to_remat, before, 0) < REMAT_COST_INFINITE;
 695 }
 696
 697 double be_get_reload_costs_on_edge(spill_env_t *env, ir_node *to_spill,
 698                                    ir_node *block, int pos)
 699 {
 700         ir_node *before = get_block_insertion_point(block, pos);
 701         return be_get_reload_costs(env, to_spill, before);
 702 }
 703
 704 ir_node *be_new_spill(ir_node *value, ir_node *after)
 705 {
 706         ir_graph                    *irg       = get_irn_irg(value);
 707         ir_node                     *frame     = get_irg_frame(irg);
 708         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 709         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 710         ir_node                     *block     = get_block(after);
 711         ir_node                     *spill
 712                 = be_new_Spill(cls, cls_frame, block, frame, value);
 713
 714         sched_add_after(after, spill);
 715         return spill;
 716 }
 717
 718 ir_node *be_new_reload(ir_node *value, ir_node *spill, ir_node *before)
 719 {
 720         ir_graph *irg   = get_irn_irg(value);
 721         ir_node  *frame = get_irg_frame(irg);
 722         ir_node  *block = get_block(before);
 723         const arch_register_class_t *cls       = arch_get_irn_reg_class(value);
 724         const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame);
 725         ir_mode                     *mode      = get_irn_mode(value);
 726         ir_node  *reload;
 727
 728         assert(be_is_Spill(spill) || is_Phi(spill));
 729         assert(get_irn_mode(spill) == mode_M);
 730
 731         reload = be_new_Reload(cls, cls_frame, block, frame, spill, mode);
 732         sched_add_before(before, reload);
 733
 734         return reload;
 735 }
 736
 737 /*
 738  *  ___                     _     ____      _                 _
 739  * |_ _|_ __  ___  ___ _ __| |_  |  _ \ ___| | ___   __ _  __| |___
 740  *  | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ |/ _ \ / _` |/ _` / __|
 741  *  | || | | \__ \  __/ |  | |_  |  _ <  __/ | (_) | (_| | (_| \__ \
 742  * |___|_| |_|___/\___|_|   \__| |_| \_\___|_|\___/ \__,_|\__,_|___/
 743  *
 744  */
 745
 746 /**
 747  * analyzes how to best spill a node and determine costs for that
 748  */
 749 static void determine_spill_costs(spill_env_t *env, spill_info_t *spillinfo)
 750 {
 751         ir_node       *to_spill = spillinfo->to_spill;
 752         const ir_node *insn     = skip_Proj_const(to_spill);
 753         ir_node       *spill_block;
 754         spill_t       *spill;
 755         double         spill_execfreq;
 756
 757         /* already calculated? */
 758         if (spillinfo->spill_costs >= 0)
 759                 return;
 760
 761         assert(!arch_irn_is(insn, dont_spill));
 762         assert(!be_is_Reload(insn));
 763
 764         /* some backends have virtual noreg/unknown nodes that are not scheduled
 765          * and simply always available.
 766          * TODO: this is kinda hairy, the NoMem is correct for an Unknown as Phi
 767          * predecessor (of a PhiM) but this test might match other things too...
 768          */
 769         if (!sched_is_scheduled(insn)) {
 770                 ir_graph *irg = get_irn_irg(to_spill);
 771                 /* override spillinfos or create a new one */
 772                 spill_t *spill = OALLOC(&env->obst, spill_t);
 773                 spill->after = NULL;
 774                 spill->next  = NULL;
 775                 spill->spill = get_irg_no_mem(irg);
 776
 777                 spillinfo->spills      = spill;
 778                 spillinfo->spill_costs = 0;
 779
 780                 DB((dbg, LEVEL_1, "don't spill %+F use NoMem\n", to_spill));
 781                 return;
 782         }
 783
 784         spill_block    = get_nodes_block(insn);
 785         spill_execfreq = get_block_execfreq(spill_block);
 786
 787         if (spillinfo->spilled_phi) {
 788                 /* TODO calculate correct costs...
 789                  * (though we can't remat this node anyway so no big problem) */
 790                 spillinfo->spill_costs = env->spill_cost * spill_execfreq;
 791                 return;
 792         }
 793
 794         if (spillinfo->spills != NULL) {
 795                 spill_t *s;
 796                 double   spills_execfreq;
 797
 798                 /* calculate sum of execution frequencies of individual spills */
 799                 spills_execfreq = 0;
 800                 s               = spillinfo->spills;
 801                 for ( ; s != NULL; s = s->next) {
 802                         ir_node *spill_block = get_block(s->after);
 803                         double   freq = get_block_execfreq(spill_block);
 804
 805                         spills_execfreq += freq;
 806                 }
 807
 808                 DB((dbg, LEVEL_1, "%+F: latespillcosts %f after def: %f\n", to_spill,
 809                     spills_execfreq * env->spill_cost,
 810                     spill_execfreq * env->spill_cost));
 811
 812                 /* multi-/latespill is advantageous -> return*/
 813                 if (spills_execfreq < spill_execfreq) {
 814                         DB((dbg, LEVEL_1, "use latespills for %+F\n", to_spill));
 815                         spillinfo->spill_costs = spills_execfreq * env->spill_cost;
 816                         return;
 817                 }
 818         }
 819
 820         /* override spillinfos or create a new one */
 821         spill        = OALLOC(&env->obst, spill_t);
 822         spill->after = determine_spill_point(to_spill);
 823         spill->next  = NULL;
 824         spill->spill = NULL;
 825
 826         spillinfo->spills      = spill;
 827         spillinfo->spill_costs = spill_execfreq * env->spill_cost;
 828         DB((dbg, LEVEL_1, "spill %+F after definition\n", to_spill));
 829 }
 830
 831 void make_spill_locations_dominate_irn(spill_env_t *env, ir_node *irn)
 832 {
 833         const spill_info_t *si = get_spillinfo(env, irn);
 834         ir_node *start_block   = get_irg_start_block(get_irn_irg(irn));
 835         int n_blocks           = get_Block_dom_max_subtree_pre_num(start_block);
 836         bitset_t *reloads      = bitset_alloca(n_blocks);
 837         reloader_t *r;
 838         spill_t *s;
 839
 840         if (si == NULL)
 841                 return;
 842
 843         /* Fill the bitset with the dominance pre-order numbers
 844          * of the blocks the reloads are located in. */
 845         for (r = si->reloaders; r != NULL; r = r->next) {
 846                 ir_node *bl = get_nodes_block(r->reloader);
 847                 bitset_set(reloads, get_Block_dom_tree_pre_num(bl));
 848         }
 849
 850         /* Now, cancel out all the blocks that are dominated by each spill.
 851          * If the bitset is not empty after that, we have reloads that are
 852          * not dominated by any spill. */
 853         for (s = si->spills; s != NULL; s = s->next) {
 854                 ir_node *bl = get_nodes_block(s->after);
 855                 int start   = get_Block_dom_tree_pre_num(bl);
 856                 int end     = get_Block_dom_max_subtree_pre_num(bl);
 857
 858                 bitset_clear_range(reloads, start, end);
 859         }
 860
 861         if (!bitset_is_empty(reloads))
 862                 be_add_spill(env, si->to_spill, si->to_spill);
 863 }
 864
 865 void be_insert_spills_reloads(spill_env_t *env)
 866 {
 867         size_t n_mem_phis = ARR_LEN(env->mem_phis);
 868         size_t i;
 869
 870         be_timer_push(T_RA_SPILL_APPLY);
 871
 872         /* create all phi-ms first, this is needed so, that phis, hanging on
 873            spilled phis work correctly */
 874         for (i = 0; i < n_mem_phis; ++i) {
 875                 spill_info_t *info = env->mem_phis[i];
 876                 spill_node(env, info);
 877         }
 878
 879         /* process each spilled node */
 880         foreach_set(env->spills, spill_info_t, si) {
 881                 ir_node  *to_spill        = si->to_spill;
 882                 ir_node **copies          = NEW_ARR_F(ir_node*, 0);
 883                 double    all_remat_costs = 0; /** costs when we would remat all nodes */
 884                 bool      force_remat     = false;
 885                 reloader_t *rld;
 886
 887                 DBG((dbg, LEVEL_1, "\nhandling all reloaders of %+F:\n", to_spill));
 888
 889                 determine_spill_costs(env, si);
 890
 891                 /* determine possibility of rematerialisations */
 892                 if (be_do_remats) {
 893                         /* calculate cost savings for each indivial value when it would
 894                            be rematted instead of reloaded */
 895                         for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 896                                 double   freq;
 897                                 int      remat_cost;
 898                                 int      remat_cost_delta;
 899                                 ir_node *block;
 900                                 ir_node *reloader = rld->reloader;
 901
 902                                 if (rld->rematted_node != NULL) {
 903                                         DBG((dbg, LEVEL_2, "\tforced remat %+F before %+F\n",
 904                                              rld->rematted_node, reloader));
 905                                         continue;
 906                                 }
 907                                 if (rld->remat_cost_delta >= REMAT_COST_INFINITE) {
 908                                         DBG((dbg, LEVEL_2, "\treload before %+F is forbidden\n",
 909                                              reloader));
 910                                         all_remat_costs = REMAT_COST_INFINITE;
 911                                         continue;
 912                                 }
 913
 914                                 remat_cost  = check_remat_conditions_costs(env, to_spill,
 915                                                                            reloader, 0);
 916                                 if (remat_cost >= REMAT_COST_INFINITE) {
 917                                         DBG((dbg, LEVEL_2, "\tremat before %+F not possible\n",
 918                                              reloader));
 919                                         rld->remat_cost_delta = REMAT_COST_INFINITE;
 920                                         all_remat_costs       = REMAT_COST_INFINITE;
 921                                         continue;
 922                                 }
 923
 924                                 remat_cost_delta      = remat_cost - env->reload_cost;
 925                                 rld->remat_cost_delta = remat_cost_delta;
 926                                 block                 = is_Block(reloader) ? reloader : get_nodes_block(reloader);
 927                                 freq                  = get_block_execfreq(block);
 928                                 all_remat_costs      += remat_cost_delta * freq;
 929                                 DBG((dbg, LEVEL_2, "\tremat costs delta before %+F: "
 930                                      "%d (rel %f)\n", reloader, remat_cost_delta,
 931                                      remat_cost_delta * freq));
 932                         }
 933                         if (all_remat_costs < REMAT_COST_INFINITE) {
 934                                 /* we don't need the costs for the spill if we can remat
 935                                    all reloaders */
 936                                 all_remat_costs -= si->spill_costs;
 937
 938                                 DBG((dbg, LEVEL_2, "\tspill costs %d (rel %f)\n",
 939                                      env->spill_cost, si->spill_costs));
 940                         }
 941
 942                         if (all_remat_costs < 0) {
 943                                 DBG((dbg, LEVEL_1, "\nforcing remats of all reloaders (%f)\n",
 944                                      all_remat_costs));
 945                                 force_remat = true;
 946                         }
 947                 }
 948
 949                 /* go through all reloads for this spill */
 950                 for (rld = si->reloaders; rld != NULL; rld = rld->next) {
 951                         ir_node *copy; /* a reload is a "copy" of the original value */
 952
 953                         if (rld->rematted_node != NULL) {
 954                                 copy = rld->rematted_node;
 955                                 sched_add_before(rld->reloader, copy);
 956                         } else if (be_do_remats &&
 957                                         (force_remat || rld->remat_cost_delta < 0)) {
 958                                 copy = do_remat(env, to_spill, rld->reloader);
 959                         } else {
 960                                 /* make sure we have a spill */
 961                                 spill_node(env, si);
 962
 963                                 /* create a reload, use the first spill for now SSA
 964                                  * reconstruction for memory comes below */
 965                                 assert(si->spills != NULL);
 966                                 copy = arch_env_new_reload(env->arch_env, si->to_spill,
 967                                                            si->spills->spill, rld->reloader);
 968                                 env->reload_count++;
 969                         }
 970
 971                         DBG((dbg, LEVEL_1, " %+F of %+F before %+F\n",
 972                              copy, to_spill, rld->reloader));
 973                         ARR_APP1(ir_node*, copies, copy);
 974                 }
 975
 976                 /* if we had any reloads or remats, then we need to reconstruct the
 977                  * SSA form for the spilled value */
 978                 if (ARR_LEN(copies) > 0) {
 979                         be_ssa_construction_env_t senv;
 980                         /* be_lv_t *lv = be_get_irg_liveness(env->irg); */
 981
 982                         be_ssa_construction_init(&senv, env->irg);
 983                         be_ssa_construction_add_copy(&senv, to_spill);
 984                         be_ssa_construction_add_copies(&senv, copies, ARR_LEN(copies));
 985                         be_ssa_construction_fix_users(&senv, to_spill);
 986
 987 #if 0
 988                         /* no need to enable this as long as we invalidate liveness
 989                            after this function... */
 990                         be_ssa_construction_update_liveness_phis(&senv);
 991                         be_liveness_update(to_spill);
 992                         len = ARR_LEN(copies);
 993                         for (i = 0; i < len; ++i) {
 994                                 be_liveness_update(lv, copies[i]);
 995                         }
 996 #endif
 997                         be_ssa_construction_destroy(&senv);
 998                 }
 999                 /* need to reconstruct SSA form if we had multiple spills */
1000                 if (si->spills != NULL && si->spills->next != NULL) {
1001                         spill_t *spill;
1002                         int      spill_count = 0;
1003
1004                         be_ssa_construction_env_t senv;
1005
1006                         be_ssa_construction_init(&senv, env->irg);
1007                         spill = si->spills;
1008                         for ( ; spill != NULL; spill = spill->next) {
1009                                 /* maybe we rematerialized the value and need no spill */
1010                                 if (spill->spill == NULL)
1011                                         continue;
1012                                 be_ssa_construction_add_copy(&senv, spill->spill);
1013                                 spill_count++;
1014                         }
1015                         if (spill_count > 1) {
1016                                 /* all reloads are attached to the first spill, fix them now */
1017                                 be_ssa_construction_fix_users(&senv, si->spills->spill);
1018                         }
1019
1020                         be_ssa_construction_destroy(&senv);
1021                 }
1022
1023                 DEL_ARR_F(copies);
1024                 si->reloaders = NULL;
1025         }
1026
1027         stat_ev_dbl("spill_spills", env->spill_count);
1028         stat_ev_dbl("spill_reloads", env->reload_count);
1029         stat_ev_dbl("spill_remats", env->remat_count);
1030         stat_ev_dbl("spill_spilled_phis", env->spilled_phi_count);
1031
1032         /* Matze: In theory be_ssa_construction should take care of the liveness...
1033          * try to disable this again in the future */
1034         be_invalidate_live_sets(env->irg);
1035
1036         be_remove_dead_nodes_from_schedule(env->irg);
1037
1038         be_timer_pop(T_RA_SPILL_APPLY);
1039 }
1040
1041 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spill)
1042 void be_init_spill(void)
1043 {
1044         FIRM_DBG_REGISTER(dbg, "firm.be.spill");
1045 }