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 "adt/cpset.h"
  42
  43 #include "bearch.h"
  44 #include "beuses.h"
  45 #include "besched.h"
  46 #include "belive_t.h"
  47 #include "bemodule.h"
  48 #include "benode.h"
  49 #include "beirgmod.h"
  50 #include "bessaconstr.h"
  51
  52 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  53
  54 typedef struct spill_info_t {
  55         struct spill_info_t *next;
  56         ir_node *value;
  57         ir_node *spill;
  58         ir_node **reloads;
  59 } spill_info_t;
  60
  61 typedef struct minibelady_env_t {
  62         struct obstack         obst;
  63         const arch_register_t *reg;
  64         const be_lv_t         *lv;
  65         void                  *func_env;
  66         create_reload_func     create_reload;
  67         create_spill_func      create_spill;
  68         spill_info_t          *spills;
  69         ir_nodehashmap_t       spill_infos;
  70
  71         be_uses_t             *uses;           /**< env for the next-use magic */
  72 } minibelady_env_t;
  73
  74 typedef struct block_info_t {
  75         ir_node *start_state;
  76         ir_node *end_state;
  77 } block_info_t;
  78
  79 static inline block_info_t *new_block_info(struct obstack *obst, ir_node *block)
  80 {
  81         block_info_t *res = OALLOCZ(obst, block_info_t);
  82
  83         assert(is_Block(block));
  84         set_irn_link(block, res);
  85         mark_irn_visited(block);
  86
  87         return res;
  88 }
  89
  90 static inline block_info_t *get_block_info(ir_node *block)
  91 {
  92         assert(irn_visited(block));
  93         return (block_info_t*) get_irn_link(block);
  94 }
  95
  96 static inline spill_info_t *create_spill_info(minibelady_env_t *env, ir_node *state)
  97 {
  98         spill_info_t *spill_info = OALLOCZ(&env->obst, spill_info_t);
  99         spill_info->value = state;
 100         spill_info->reloads = NEW_ARR_F(ir_node*, 0);
 101
 102         ir_nodehashmap_insert(&env->spill_infos, state, spill_info);
 103         //ir_fprintf(stderr, "Insert %+F -> %p\n", state, spill_info);
 104
 105         spill_info->next = env->spills;
 106         env->spills = spill_info;
 107
 108         return spill_info;
 109 }
 110
 111 static inline spill_info_t *get_spill_info(minibelady_env_t *env, const ir_node *node)
 112 {
 113         spill_info_t *spill_info
 114                 = (spill_info_t*) ir_nodehashmap_get(&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_Proj(next) || 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         int            i;
 222
 223         /* Create the block info for this block. */
 224         block_info = new_block_info(&env->obst, block);
 225         n_cfgpreds = get_Block_n_cfgpreds(block);
 226
 227         /* no cfgpred -> no value active */
 228         if (n_cfgpreds == 0) {
 229                 block_info->start_state = NULL;
 230                 return block_info;
 231         }
 232
 233         /* for 1 pred only: simply take the the end-state of the pred */
 234         if (n_cfgpreds == 1) {
 235                 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
 236                 block_info_t *pred_info;
 237
 238                 /* process pred block */
 239                 belady(env, pred_block);
 240
 241                 pred_info = get_block_info(pred_block);
 242
 243                 DBG((dbg, LEVEL_2, "Taking end state from %+F: %+F\n", pred_block, pred_info->end_state));
 244                 block_info->start_state = pred_info->end_state;
 245                 return block_info;
 246         }
 247
 248         /* Collect all values living at start of block */
 249         DBG((dbg, LEVEL_2, "Living at start of %+F:\n", block));
 250         first = sched_first(block);
 251         loop = get_irn_loop(block);
 252         best_starter = NULL;
 253         best_time = USES_INFINITY;
 254         outer_loop_allowed = 1;
 255
 256         /* check all Phis first */
 257         sched_foreach(block, node) {
 258                 if (!is_Phi(node))
 259                         break;
 260                 if (arch_get_irn_register(node) != env->reg)
 261                         continue;
 262
 263                 DBG((dbg, LEVEL_2, "\t...checking %+F\n", node));
 264                 next_use = be_get_next_use(env->uses, first, node, 0);
 265
 266                 if (USES_IS_INFINITE(next_use.time)) {
 267                         DBG((dbg, LEVEL_2, "\tnot taken (dead)\n"));
 268                         continue;
 269                 }
 270
 271                 if (next_use.outermost_loop >= get_loop_depth(loop)) {
 272                         if (outer_loop_allowed || next_use.time < best_time) {
 273                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 274                                      next_use.outermost_loop));
 275
 276                                 if (best_starter != NULL) {
 277                                         /* spill the phi as it is not used */
 278                                         spill_phi(env, best_starter);
 279                                 }
 280                                 best_starter = node;
 281                                 best_time = next_use.time;
 282                                 outer_loop_allowed = 0;
 283                         }
 284                 } else {
 285                         if (outer_loop_allowed && next_use.time < best_time) {
 286                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 287                                      next_use.outermost_loop));
 288                                 if (best_starter != NULL) {
 289                                         /* spill the phi as it is not used */
 290                                         spill_phi(env, best_starter);
 291                                 }
 292                                 best_starter = node;
 293                                 best_time = next_use.time;
 294                         }
 295                 }
 296
 297                 if (best_starter != node) {
 298                         /* spill the phi as it is not used */
 299                         spill_phi(env, best_starter);
 300                 }
 301         }
 302
 303         /* check all Live-Ins */
 304         be_lv_foreach(env->lv, block, be_lv_state_in, i) {
 305                 ir_node *const node = be_lv_get_irn(env->lv, block, i);
 306
 307                 if (!mode_is_data(get_irn_mode(node)))
 308                         continue;
 309
 310                 if (arch_get_irn_register(node) != env->reg)
 311                         continue;
 312
 313                 DBG((dbg, LEVEL_2, "\t...checking %+F\n", node));
 314                 next_use = be_get_next_use(env->uses, first, node, 0);
 315
 316                 if (USES_IS_INFINITE(next_use.time)) {
 317                         DBG((dbg, LEVEL_2, "\tnot taken (dead)\n"));
 318                         continue;
 319                 }
 320
 321                 if (next_use.outermost_loop >= get_loop_depth(loop)) {
 322                         if (outer_loop_allowed || next_use.time < best_time) {
 323                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 324                                      next_use.outermost_loop));
 325
 326                                 if (best_starter != NULL && is_Phi(best_starter)) {
 327                                         /* spill the phi as it is not used */
 328                                         spill_phi(env, best_starter);
 329                                 }
 330                                 best_starter = node;
 331                                 best_time = next_use.time;
 332                                 outer_loop_allowed = 0;
 333                         }
 334                 } else {
 335                         if (outer_loop_allowed && next_use.time < best_time) {
 336                                 DBG((dbg, LEVEL_2, "\ttaken (%u, loop %d)\n", next_use.time,
 337                                      next_use.outermost_loop));
 338                                 if (best_starter != NULL && is_Phi(best_starter)) {
 339                                         /* spill the phi as it is not used */
 340                                         spill_phi(env, best_starter);
 341                                 }
 342                                 best_starter = node;
 343                                 best_time = next_use.time;
 344                         }
 345                 }
 346         }
 347
 348         block_info->start_state = best_starter;
 349
 350         return block_info;
 351 }
 352
 353 /**
 354  * For the given block @p block, decide for each values
 355  * whether it is used from a register or is reloaded
 356  * before the use.
 357  */
 358 static void belady(minibelady_env_t *env, ir_node *block)
 359 {
 360         ir_node *current_state;
 361         block_info_t *block_info;
 362
 363         /* Don't do a block twice */
 364         if (irn_visited(block))
 365                 return;
 366
 367         /* compute value to start with */
 368         block_info = compute_block_start_state(env, block);
 369
 370         /* get the starting workset for this block */
 371         DBG((dbg, LEVEL_3, "\n"));
 372         DBG((dbg, LEVEL_3, "Decide for %+F\n", block));
 373
 374         current_state = block_info->start_state;
 375         DBG((dbg, LEVEL_3, "Start value: %+F\n", current_state));
 376
 377         /* process the block from start to end */
 378         DBG((dbg, LEVEL_3, "Processing...\n"));
 379
 380         sched_foreach(block, node) {
 381                 int i, arity;
 382                 ir_node *need_val = NULL;
 383
 384                 /* projs are handled with the tuple value.
 385                  * Phis are no real instr (see insert_starters()) */
 386                 if (is_Proj(node) || is_Phi(node)) {
 387                         continue;
 388                 }
 389
 390                 /* check which state is desired for the node */
 391                 arity = get_irn_arity(node);
 392                 for (i = 0; i < arity; ++i) {
 393                         const arch_register_t *reg;
 394                         ir_node *in = get_irn_n(node, i);
 395
 396                         if (!mode_is_data(get_irn_mode(in)))
 397                                 continue;
 398
 399                         reg = arch_get_irn_register(in);
 400                         if (reg == env->reg) {
 401                                 assert(need_val == NULL);
 402                                 need_val = in;
 403                                 DBG((dbg, LEVEL_3, "\t... need state %+F\n", need_val));
 404                         }
 405                 }
 406                 /* create a reload to match state if necessary */
 407                 if (need_val != NULL && need_val != current_state) {
 408                         ir_node *before = node;
 409                         DBG((dbg, LEVEL_3, "\t... reloading %+F\n", need_val));
 410                         create_reload(env, need_val, before, current_state);
 411                         current_state = need_val;
 412                 }
 413
 414                 DBG((dbg, LEVEL_3, "  ...%+F\n", node));
 415
 416                 /* record state changes by the node */
 417                 if (get_irn_mode(node) == mode_T) {
 418                         foreach_out_edge(node, edge) {
 419                                 const arch_register_t *reg;
 420                                 ir_node *proj = get_edge_src_irn(edge);
 421
 422                                 if (!mode_is_data(get_irn_mode(proj)))
 423                                         continue;
 424
 425                                 reg = arch_get_irn_register(proj);
 426                                 if (reg == env->reg) {
 427                                         current_state = proj;
 428                                         DBG((dbg, LEVEL_3, "\t... current_state <- %+F\n", current_state));
 429                                 }
 430                         }
 431                 } else {
 432                         if (mode_is_data(get_irn_mode(node))) {
 433                                 const arch_register_t *reg = arch_get_irn_register(node);
 434                                 if (reg == env->reg) {
 435                                         current_state = node;
 436                                         DBG((dbg, LEVEL_3, "\t... current_state <- %+F\n", current_state));
 437                                 }
 438                         }
 439                 }
 440         }
 441
 442         /* Remember end-workset for this block */
 443         block_info->end_state = current_state;
 444         DBG((dbg, LEVEL_3, "End value for %+F: %+F\n", block, current_state));
 445 }
 446
 447 static void belady_walker(ir_node *block, void *data)
 448 {
 449         belady((minibelady_env_t*) data, block);
 450 }
 451
 452 static ir_node *get_end_of_block_insertion_point(ir_node *block)
 453 {
 454         ir_node *last = sched_last(block);
 455
 456         /* skip Projs and Keep-alikes behind the jump... */
 457         while (is_Proj(last) || be_is_Keep(last)) {
 458                 last = sched_prev(last);
 459         }
 460
 461         if (!is_cfop(last)) {
 462                 last = sched_next(last);
 463                 /* last node must be a cfop, only exception is the start block */
 464                 assert(last == get_irg_start_block(get_irn_irg(block)));
 465         }
 466
 467         return last;
 468 }
 469
 470 /**
 471  * We must adapt the live-outs to the live-ins at each block-border.
 472  */
 473 static void fix_block_borders(ir_node *block, void *data)
 474 {
 475         minibelady_env_t *env = (minibelady_env_t*)data;
 476         ir_graph *irg = get_irn_irg(block);
 477         ir_node *startblock = get_irg_start_block(irg);
 478         int i;
 479         int arity;
 480         block_info_t *block_info;
 481
 482         if (block == startblock)
 483                 return;
 484
 485         DBG((dbg, LEVEL_3, "\n"));
 486
 487         block_info = get_block_info(block);
 488
 489         DBG((dbg, LEVEL_3, "Fixing %+F (needs %+F)\n", block,
 490              block_info->start_state));
 491
 492         /* process all pred blocks */
 493         arity = get_irn_arity(block);
 494         for (i = 0; i < arity; ++i) {
 495                 ir_node      *pred       = get_Block_cfgpred_block(block, i);
 496                 block_info_t *pred_info  = get_block_info(pred);
 497                 ir_node      *need_state = block_info->start_state;
 498
 499                 if (need_state == NULL)
 500                         continue;
 501
 502                 if (is_Phi(need_state) && get_nodes_block(need_state) == block) {
 503                         need_state = get_irn_n(need_state, i);
 504                 }
 505
 506                 DBG((dbg, LEVEL_3, "  Pred %+F (ends in %+F, we need %+F)\n", pred,
 507                      pred_info->end_state, need_state));
 508
 509                 if (pred_info->end_state != need_state) {
 510                         ir_node *insert_point = get_end_of_block_insertion_point(pred);
 511
 512
 513                         DBG((dbg, LEVEL_3, "  Creating reload for %+F\n", need_state));
 514                         create_reload(env, need_state, insert_point, pred_info->end_state);
 515                 }
 516         }
 517 }
 518
 519 void be_assure_state(ir_graph *irg, const arch_register_t *reg, void *func_env,
 520                      create_spill_func create_spill,
 521                      create_reload_func create_reload)
 522 {
 523         minibelady_env_t env;
 524         spill_info_t *info;
 525         be_lv_t *lv = be_get_irg_liveness(irg);
 526
 527         be_assure_live_sets(irg);
 528         assure_loopinfo(irg);
 529
 530         obstack_init(&env.obst);
 531         env.reg           = reg;
 532         env.func_env      = func_env;
 533         env.create_spill  = create_spill;
 534         env.create_reload = create_reload;
 535         env.lv            = be_get_irg_liveness(irg);
 536         env.uses          = be_begin_uses(irg, env.lv);
 537         env.spills        = NULL;
 538         ir_nodehashmap_init(&env.spill_infos);
 539
 540         assure_doms(irg);
 541         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 542         inc_irg_visited(irg);
 543
 544         /* process blocks */
 545         irg_block_walk_graph(irg, NULL, belady_walker, &env);
 546
 547         /* fix block end_states that don't match the next blocks start_state */
 548         irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
 549
 550         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK);
 551
 552         /* reconstruct ssa-form */
 553         info = env.spills;
 554         while (info != NULL) {
 555                 be_ssa_construction_env_t senv;
 556                 size_t i, len;
 557                 ir_node **phis;
 558
 559                 be_ssa_construction_init(&senv, irg);
 560                 if (sched_is_scheduled(info->value))
 561                         be_ssa_construction_add_copy(&senv, info->value);
 562                 be_ssa_construction_add_copies(&senv,
 563                                                info->reloads, ARR_LEN(info->reloads));
 564                 be_ssa_construction_fix_users(&senv, info->value);
 565
 566                 if (lv != NULL) {
 567                         be_ssa_construction_update_liveness_phis(&senv, lv);
 568
 569                         be_liveness_update(lv, info->value);
 570                         len = ARR_LEN(info->reloads);
 571                         for (i = 0; i < len; ++i) {
 572                                 ir_node *reload = info->reloads[i];
 573                                 be_liveness_update(lv, reload);
 574                         }
 575                 }
 576
 577                 phis = be_ssa_construction_get_new_phis(&senv);
 578
 579                 /* set register requirements for phis */
 580                 len = ARR_LEN(phis);
 581                 for (i = 0; i < len; ++i) {
 582                         ir_node *phi = phis[i];
 583                         arch_set_irn_register(phi, env.reg);
 584                 }
 585                 be_ssa_construction_destroy(&senv);
 586
 587                 info = info->next;
 588         }
 589
 590         /* some nodes might be dead now. */
 591         be_remove_dead_nodes_from_schedule(irg);
 592
 593         ir_nodehashmap_destroy(&env.spill_infos);
 594         be_end_uses(env.uses);
 595         obstack_free(&env.obst, NULL);
 596 }
 597
 598 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_state)
 599 void be_init_state(void)
 600 {
 601         FIRM_DBG_REGISTER(dbg, "firm.be.state");
 602 }