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

   1 /*
   2  * Copyright (C) 1995-2011 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       Handles state switching. This is basically the belady spill
  23  *              algorithm optimized for the 1-register case.
  24  * @author      Matthias Braun
  25  * @date        26.03.2007
  26  */
  27 #include "config.h"
  28
  29 #include "bestate.h"
  30
  31 #include "obst.h"
  32 #include "irgraph_t.h"
  33 #include "irnode_t.h"
  34 #include "irgwalk.h"
  35 #include "irloop.h"
  36 #include "iredges_t.h"
  37 #include "ircons_t.h"
  38 #include "irgmod.h"
  39 #include "irnodeset.h"
  40 #include "irnodehashmap.h"
  41 #include "cpset.h"
  42
  43 #include "bearch.h"
  44 #include "beirg.h"
  45 #include "beuses.h"
  46 #include "besched.h"
  47 #include "belive_t.h"
  48 #include "bemodule.h"
  49 #include "benode.h"
  50 #include "beirgmod.h"
  51 #include "bessaconstr.h"
  52
  53 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  54
  55 typedef struct spill_info_t {
  56         struct spill_info_t *next;
  57         ir_node *value;
  58         ir_node *spill;
  59         ir_node **reloads;
  60 } spill_info_t;
  61
  62 typedef struct minibelady_env_t {
  63         struct obstack         obst;
  64         const arch_register_t *reg;
  65         const be_lv_t         *lv;
  66         void                  *func_env;
  67         create_reload_func     create_reload;
  68         create_spill_func      create_spill;
  69         spill_info_t          *spills;
  70         ir_nodehashmap_t       spill_infos;
  71
  72         be_uses_t             *uses;           /**< env for the next-use magic */
  73 } minibelady_env_t;
  74
  75 typedef struct block_info_t {
  76         ir_node *start_state;
  77         ir_node *end_state;
  78 } block_info_t;
  79
  80 static inline block_info_t *new_block_info(struct obstack *obst, ir_node *block)
  81 {
  82         block_info_t *res = OALLOCZ(obst, block_info_t);
  83
  84         assert(is_Block(block));
  85         set_irn_link(block, res);
  86         mark_irn_visited(block);
  87
  88         return res;
  89 }
  90
  91 static inline block_info_t *get_block_info(ir_node *block)
  92 {
  93         assert(irn_visited(block));
  94         return (block_info_t*) get_irn_link(block);
  95 }
  96
  97 static inline spill_info_t *create_spill_info(minibelady_env_t *env, ir_node *state)
  98 {
  99         spill_info_t *spill_info = OALLOCZ(&env->obst, spill_info_t);
 100         spill_info->value = state;
 101         spill_info->reloads = NEW_ARR_F(ir_node*, 0);
 102
 103         ir_nodehashmap_insert(&env->spill_infos, state, spill_info);
 104         //ir_fprintf(stderr, "Insert %+F -> %p\n", state, spill_info);
 105
 106         spill_info->next = env->spills;
 107         env->spills = spill_info;
 108
 109         return spill_info;
 110 }
 111
 112 static inline spill_info_t *get_spill_info(minibelady_env_t *env, const ir_node *node)
 113 {
 114         spill_info_t *spill_info = ir_nodehashmap_get(spill_info_t, &env->spill_infos, node);
 115         //ir_fprintf(stderr, "Get %+F -> %p\n", node, spill_info);
 116         return spill_info;
 117 }
 118
 119 static spill_info_t *create_spill(minibelady_env_t *env, ir_node *state, int force)
 120 {
 121         spill_info_t *spill_info;
 122         ir_node *next;
 123         ir_node *after;
 124
 125         spill_info = get_spill_info(env, state);
 126         if (spill_info == NULL) {
 127                 spill_info = create_spill_info(env, state);
 128         } else if (spill_info->spill != NULL) {
 129                 return spill_info;
 130         }
 131
 132         if (sched_is_scheduled(state)) {
 133                 next = state;
 134                 do {
 135                         after = next;
 136                         next = sched_next(after);
 137                 } while (is_Phi(next) || be_is_Keep(next));
 138         } else {
 139                 after = state;
 140         }
 141         spill_info->spill = env->create_spill(env->func_env, state, force, after);
 142
 143         return spill_info;
 144 }
 145
 146 static void create_reload(minibelady_env_t *env, ir_node *state,
 147                           ir_node *before, ir_node *last_state)
 148 {
 149         spill_info_t *spill_info = create_spill(env, state, 0);
 150         ir_node *spill = spill_info->spill;
 151         ir_node *reload;
 152
 153         reload = env->create_reload(env->func_env, state, spill, before,
 154                                     last_state);
 155         ARR_APP1(ir_node*, spill_info->reloads, reload);
 156 }
 157
 158 static void spill_phi(minibelady_env_t *env, ir_node *phi)
 159 {
 160         ir_graph     *irg           = get_irn_irg(phi);
 161         ir_node      *block         = get_nodes_block(phi);
 162         int           arity         = get_irn_arity(phi);
 163         ir_node     **phi_in        = ALLOCAN(ir_node*, arity);
 164         ir_node      *dummy         = new_r_Dummy(irg, mode_M);
 165         ir_node      *spill_to_kill = NULL;
 166         spill_info_t *spill_info;
 167         int           i;
 168
 169         /* does a spill exist for the phis value? */
 170         spill_info = get_spill_info(env, phi);
 171         if (spill_info != NULL) {
 172                 spill_to_kill = spill_info->spill;
 173         } else {
 174                 spill_info = create_spill_info(env, phi);
 175         }
 176
 177         /* create a new phi-M with bad preds */
 178         for (i = 0; i < arity; ++i) {
 179                 phi_in[i] = dummy;
 180         }
 181
 182         DBG((dbg, LEVEL_2, "\tcreate Phi-M for %+F\n", phi));
 183
 184         /* create a Phi-M */
 185         spill_info->spill = be_new_Phi(block, arity, phi_in, mode_M,
 186                                        arch_no_register_req);
 187         sched_add_after(block, spill_info->spill);
 188
 189         if (spill_to_kill != NULL) {
 190                 exchange(spill_to_kill, spill_info->spill);
 191                 sched_remove(spill_to_kill);
 192         }
 193
 194         /* create spills for the phi values */
 195         for (i = 0; i < arity; ++i) {
 196                 ir_node *in = get_irn_n(phi, i);
 197                 spill_info_t *pred_spill = create_spill(env, in, 1);
 198                 set_irn_n(spill_info->spill, i, pred_spill->spill);
 199         }
 200 }
 201
 202 static void belady(minibelady_env_t *env, ir_node *block);
 203
 204 /**
 205  * Collects all values live-in at block @p block and all phi results in this
 206  * block.
 207  * Then it adds the best values (at most n_regs) to the blocks start_workset.
 208  * The phis among the remaining values get spilled: Introduce pseudo-copies of
 209  * their args to break interference and make it possible to spill them to the
 210  * same spill slot.
 211  */
 212 static block_info_t *compute_block_start_state(minibelady_env_t *env, ir_node *block)
 213 {
 214         block_info_t  *block_info;
 215         be_next_use_t  next_use;
 216         ir_loop       *loop;
 217         ir_node       *best_starter, *first;
 218         int            n_cfgpreds;
 219         unsigned       best_time;
 220         int            outer_loop_allowed;
 221
 222         /* Create the block info for this block. */
 223         block_info = new_block_info(&env->obst, block);
 224         n_cfgpreds = get_Block_n_cfgpreds(block);
 225
 226         /* no cfgpred -> no value active */
 227         if (n_cfgpreds == 0) {
 228                 block_info->start_state = NULL;
 229                 return block_info;
 230         }
 231
 232         /* for 1 pred only: simply take the the end-state of the pred */
 233         if (n_cfgpreds == 1) {
 234                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 235                 block_info_t *pred_info;
 236
 237                 /* process pred block */
 238                 belady(env, pred_block);
 239
 240                 pred_info = get_block_info(pred_block);
 241
 242                 DBG((dbg, LEVEL_2, "Taking end state from %+F: %+F\n", pred_block, pred_info->end_state));
 243                 block_info->start_state = pred_info->end_state;
 244                 return block_info;
 245         }
 246
 247         /* Collect all values living at start of block */
 248         DBG((dbg, LEVEL_2, "Living at start of %+F:\n", block));
 249         first = sched_first(block);
 250         loop = get_irn_loop(block);
 251         best_starter = NULL;
 252         best_time = USES_INFINITY;
 253         outer_loop_allowed = 1;
 254
 255         /* check all Phis first */
 256         sched_foreach(block, node) {
 257                 if (!is_Phi(node))
 258                         break;
 259                 if (arch_get_irn_register(node) != env->reg)
 260                         continue;
 261
 262                 DBG((dbg, LEVEL_2, "\t...checking %+F\n", node));
 263                 next_use = be_get_next_use(env->uses, first, node, 0);
 264
 265                 if (USES_IS_INFINITE(next_use.time)) {
 266                         DBG((dbg, LEVEL_2, "\tnot taken (dead)\n"));
 267                         continue;
 268                 }
 269
 270                 if (next_use.outermost_loop >= get_loop_depth(loop)) {
 271                         if (outer_loop_allowed || next_use.time < best_time) {
 272                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 273                                      next_use.outermost_loop));
 274
 275                                 if (best_starter != NULL) {
 276                                         /* spill the phi as it is not used */
 277                                         spill_phi(env, best_starter);
 278                                 }
 279                                 best_starter = node;
 280                                 best_time = next_use.time;
 281                                 outer_loop_allowed = 0;
 282                         }
 283                 } else {
 284                         if (outer_loop_allowed && next_use.time < best_time) {
 285                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 286                                      next_use.outermost_loop));
 287                                 if (best_starter != NULL) {
 288                                         /* spill the phi as it is not used */
 289                                         spill_phi(env, best_starter);
 290                                 }
 291                                 best_starter = node;
 292                                 best_time = next_use.time;
 293                         }
 294                 }
 295
 296                 if (best_starter != node) {
 297                         /* spill the phi as it is not used */
 298                         spill_phi(env, best_starter);
 299                 }
 300         }
 301
 302         /* check all Live-Ins */
 303         be_lv_foreach_cls(env->lv, block, be_lv_state_in, env->reg->reg_class, node) {
 304                 if (arch_get_irn_register(node) != env->reg)
 305                         continue;
 306
 307                 DBG((dbg, LEVEL_2, "\t...checking %+F\n", node));
 308                 next_use = be_get_next_use(env->uses, first, node, 0);
 309
 310                 if (USES_IS_INFINITE(next_use.time)) {
 311                         DBG((dbg, LEVEL_2, "\tnot taken (dead)\n"));
 312                         continue;
 313                 }
 314
 315                 if (next_use.outermost_loop >= get_loop_depth(loop)) {
 316                         if (outer_loop_allowed || next_use.time < best_time) {
 317                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 318                                      next_use.outermost_loop));
 319
 320                                 if (best_starter != NULL && is_Phi(best_starter)) {
 321                                         /* spill the phi as it is not used */
 322                                         spill_phi(env, best_starter);
 323                                 }
 324                                 best_starter = node;
 325                                 best_time = next_use.time;
 326                                 outer_loop_allowed = 0;
 327                         }
 328                 } else {
 329                         if (outer_loop_allowed && next_use.time < best_time) {
 330                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 331                                      next_use.outermost_loop));
 332                                 if (best_starter != NULL && is_Phi(best_starter)) {
 333                                         /* spill the phi as it is not used */
 334                                         spill_phi(env, best_starter);
 335                                 }
 336                                 best_starter = node;
 337                                 best_time = next_use.time;
 338                         }
 339                 }
 340         }
 341
 342         block_info->start_state = best_starter;
 343
 344         return block_info;
 345 }
 346
 347 /**
 348  * For the given block @p block, decide for each values
 349  * whether it is used from a register or is reloaded
 350  * before the use.
 351  */
 352 static void belady(minibelady_env_t *env, ir_node *block)
 353 {
 354         ir_node *current_state;
 355         block_info_t *block_info;
 356
 357         /* Don't do a block twice */
 358         if (irn_visited(block))
 359                 return;
 360
 361         /* compute value to start with */
 362         block_info = compute_block_start_state(env, block);
 363
 364         /* get the starting workset for this block */
 365         DBG((dbg, LEVEL_3, "\n"));
 366         DBG((dbg, LEVEL_3, "Decide for %+F\n", block));
 367
 368         current_state = block_info->start_state;
 369         DBG((dbg, LEVEL_3, "Start value: %+F\n", current_state));
 370
 371         /* process the block from start to end */
 372         DBG((dbg, LEVEL_3, "Processing...\n"));
 373
 374         sched_foreach(block, node) {
 375                 int i, arity;
 376                 ir_node *need_val = NULL;
 377
 378                 /* projs are handled with the tuple value.
 379                  * Phis are no real instr (see insert_starters()) */
 380                 if (is_Proj(node) || is_Phi(node)) {
 381                         continue;
 382                 }
 383
 384                 /* check which state is desired for the node */
 385                 arity = get_irn_arity(node);
 386                 for (i = 0; i < arity; ++i) {
 387                         const arch_register_t *reg;
 388                         ir_node *in = get_irn_n(node, i);
 389
 390                         if (!mode_is_data(get_irn_mode(in)))
 391                                 continue;
 392
 393                         reg = arch_get_irn_register(in);
 394                         if (reg == env->reg) {
 395                                 assert(need_val == NULL);
 396                                 need_val = in;
 397                                 DBG((dbg, LEVEL_3, "\t... need state %+F\n", need_val));
 398                         }
 399                 }
 400                 /* create a reload to match state if necessary */
 401                 if (need_val != NULL && need_val != current_state) {
 402                         ir_node *before = node;
 403                         DBG((dbg, LEVEL_3, "\t... reloading %+F\n", need_val));
 404                         create_reload(env, need_val, before, current_state);
 405                         current_state = need_val;
 406                 }
 407
 408                 DBG((dbg, LEVEL_3, "  ...%+F\n", node));
 409
 410                 /* record state changes by the node */
 411                 if (get_irn_mode(node) == mode_T) {
 412                         foreach_out_edge(node, edge) {
 413                                 const arch_register_t *reg;
 414                                 ir_node *proj = get_edge_src_irn(edge);
 415
 416                                 if (!mode_is_data(get_irn_mode(proj)))
 417                                         continue;
 418
 419                                 reg = arch_get_irn_register(proj);
 420                                 if (reg == env->reg) {
 421                                         current_state = proj;
 422                                         DBG((dbg, LEVEL_3, "\t... current_state <- %+F\n", current_state));
 423                                 }
 424                         }
 425                 } else {
 426                         if (mode_is_data(get_irn_mode(node))) {
 427                                 const arch_register_t *reg = arch_get_irn_register(node);
 428                                 if (reg == env->reg) {
 429                                         current_state = node;
 430                                         DBG((dbg, LEVEL_3, "\t... current_state <- %+F\n", current_state));
 431                                 }
 432                         }
 433                 }
 434         }
 435
 436         /* Remember end-workset for this block */
 437         block_info->end_state = current_state;
 438         DBG((dbg, LEVEL_3, "End value for %+F: %+F\n", block, current_state));
 439 }
 440
 441 static void belady_walker(ir_node *block, void *data)
 442 {
 443         belady((minibelady_env_t*) data, block);
 444 }
 445
 446 static ir_node *get_end_of_block_insertion_point(ir_node *block)
 447 {
 448         ir_node *last = sched_last(block);
 449
 450         /* skip Projs and Keep-alikes behind the jump... */
 451         while (is_Proj(last) || be_is_Keep(last)) {
 452                 last = sched_prev(last);
 453         }
 454
 455         if (!is_cfop(last)) {
 456                 last = sched_next(last);
 457                 /* last node must be a cfop, only exception is the start block */
 458                 assert(last == get_irg_start_block(get_irn_irg(block)));
 459         }
 460
 461         return last;
 462 }
 463
 464 /**
 465  * We must adapt the live-outs to the live-ins at each block-border.
 466  */
 467 static void fix_block_borders(ir_node *block, void *data)
 468 {
 469         minibelady_env_t *env = (minibelady_env_t*)data;
 470         ir_graph *irg = get_irn_irg(block);
 471         ir_node *startblock = get_irg_start_block(irg);
 472         int i;
 473         int arity;
 474         block_info_t *block_info;
 475
 476         if (block == startblock)
 477                 return;
 478
 479         DBG((dbg, LEVEL_3, "\n"));
 480
 481         block_info = get_block_info(block);
 482
 483         DBG((dbg, LEVEL_3, "Fixing %+F (needs %+F)\n", block,
 484              block_info->start_state));
 485
 486         /* process all pred blocks */
 487         arity = get_irn_arity(block);
 488         for (i = 0; i < arity; ++i) {
 489                 ir_node      *pred       = get_Block_cfgpred_block(block, i);
 490                 block_info_t *pred_info  = get_block_info(pred);
 491                 ir_node      *need_state = block_info->start_state;
 492
 493                 if (need_state == NULL)
 494                         continue;
 495
 496                 if (is_Phi(need_state) && get_nodes_block(need_state) == block) {
 497                         need_state = get_irn_n(need_state, i);
 498                 }
 499
 500                 DBG((dbg, LEVEL_3, "  Pred %+F (ends in %+F, we need %+F)\n", pred,
 501                      pred_info->end_state, need_state));
 502
 503                 if (pred_info->end_state != need_state) {
 504                         ir_node *insert_point = get_end_of_block_insertion_point(pred);
 505
 506
 507                         DBG((dbg, LEVEL_3, "  Creating reload for %+F\n", need_state));
 508                         create_reload(env, need_state, insert_point, pred_info->end_state);
 509                 }
 510         }
 511 }
 512
 513 void be_assure_state(ir_graph *irg, const arch_register_t *reg, void *func_env,
 514                      create_spill_func create_spill,
 515                      create_reload_func create_reload)
 516 {
 517         minibelady_env_t env;
 518         spill_info_t *info;
 519         be_lv_t *lv = be_get_irg_liveness(irg);
 520
 521         be_assure_live_sets(irg);
 522         assure_loopinfo(irg);
 523
 524         obstack_init(&env.obst);
 525         env.reg           = reg;
 526         env.func_env      = func_env;
 527         env.create_spill  = create_spill;
 528         env.create_reload = create_reload;
 529         env.lv            = be_get_irg_liveness(irg);
 530         env.uses          = be_begin_uses(irg, env.lv);
 531         env.spills        = NULL;
 532         ir_nodehashmap_init(&env.spill_infos);
 533
 534         assure_doms(irg);
 535         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 536         inc_irg_visited(irg);
 537
 538         /* process blocks */
 539         irg_block_walk_graph(irg, NULL, belady_walker, &env);
 540
 541         /* fix block end_states that don't match the next blocks start_state */
 542         irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
 543
 544         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 545
 546         /* reconstruct ssa-form */
 547         info = env.spills;
 548         while (info != NULL) {
 549                 be_ssa_construction_env_t senv;
 550                 size_t i, len;
 551                 ir_node **phis;
 552
 553                 be_ssa_construction_init(&senv, irg);
 554                 if (sched_is_scheduled(info->value))
 555                         be_ssa_construction_add_copy(&senv, info->value);
 556                 be_ssa_construction_add_copies(&senv,
 557                                                info->reloads, ARR_LEN(info->reloads));
 558                 be_ssa_construction_fix_users(&senv, info->value);
 559
 560                 if (lv != NULL) {
 561                         be_ssa_construction_update_liveness_phis(&senv, lv);
 562
 563                         be_liveness_update(lv, info->value);
 564                         len = ARR_LEN(info->reloads);
 565                         for (i = 0; i < len; ++i) {
 566                                 ir_node *reload = info->reloads[i];
 567                                 be_liveness_update(lv, reload);
 568                         }
 569                 }
 570
 571                 phis = be_ssa_construction_get_new_phis(&senv);
 572
 573                 /* set register requirements for phis */
 574                 len = ARR_LEN(phis);
 575                 for (i = 0; i < len; ++i) {
 576                         ir_node *phi = phis[i];
 577                         arch_set_irn_register(phi, env.reg);
 578                 }
 579                 be_ssa_construction_destroy(&senv);
 580
 581                 info = info->next;
 582         }
 583
 584         /* some nodes might be dead now. */
 585         be_remove_dead_nodes_from_schedule(irg);
 586
 587         ir_nodehashmap_destroy(&env.spill_infos);
 588         be_end_uses(env.uses);
 589         obstack_free(&env.obst, NULL);
 590 }
 591
 592 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_state)
 593 void be_init_state(void)
 594 {
 595         FIRM_DBG_REGISTER(dbg, "firm.be.state");
 596 }