nsz Git - libfirm/blob - ir/opt/ldst2.c

   1 /*
   2  * Copyright (C) 1995-2007 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   parallelizing Load/Store optimisation
  23  * @author  Christoph Mallon
  24  * @version $Id$
  25  */
  26 #ifdef HAVE_CONFIG_H
  27 #include "config.h"
  28 #endif
  29
  30 #include "array.h"
  31 #include "debug.h"
  32 #include "ircons.h"
  33 #include "irgraph.h"
  34 #include "irgmod.h"
  35 #include "irgopt.h"
  36 #include "irgwalk.h"
  37 #include "irmemory.h"
  38 #include "irnode.h"
  39 #include "irnodeset.h"
  40 #include "ldst2.h"
  41 #include "obst.h"
  42 #include "return.h"
  43 #include "irdump.h"
  44
  45
  46 #define OPTIMISE_LOAD_AFTER_LOAD
  47
  48
  49 #define UNIMPLEMENTED abort();
  50
  51
  52 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  53
  54
  55 static struct obstack obst;
  56 static size_t count_addrs;
  57 static ir_node** addrs;
  58
  59
  60 static void AddressCollector(ir_node* node, void* env)
  61 {
  62         ir_nodeset_t* addrs_set = env;
  63         ir_node* addr;
  64         if (is_Load(node)) {
  65                 addr = get_Load_ptr(node);
  66         } else if (is_Store(node)) {
  67                 addr = get_Store_ptr(node);
  68         } else {
  69                 return;
  70         }
  71         ir_nodeset_insert(addrs_set, addr);
  72 }
  73
  74
  75 /* Collects all unique addresses used by load and store nodes of a graph and
  76  * puts them into an array for later use */
  77 static void CollectAddresses(ir_graph* irg)
  78 {
  79         ir_nodeset_t addrs_set;
  80
  81         ir_nodeset_init(&addrs_set);
  82         irg_walk_graph(irg, AddressCollector, NULL, &addrs_set);
  83
  84         count_addrs = ir_nodeset_size(&addrs_set);
  85         DB((dbg, LEVEL_1, "===> %+F uses %u unique addresses\n", irg, (unsigned int)count_addrs));
  86         if (count_addrs != 0) {
  87                 ir_nodeset_iterator_t addr_iter;
  88                 size_t i;
  89
  90                 addrs = NEW_ARR_D(ir_node*, &obst, count_addrs);
  91                 ir_nodeset_iterator_init(&addr_iter, &addrs_set);
  92                 for (i = 0; i < count_addrs; i++) {
  93                         ir_node* addr = ir_nodeset_iterator_next(&addr_iter);
  94                         assert(addr != NULL);
  95                         set_irn_link(addr, (void *)i);
  96                         addrs[i] = addr;
  97                         DB((dbg, LEVEL_2, "===> Collected unique symbolic address %+F\n", addr));
  98                 }
  99         }
 100 }
 101
 102
 103 static void AliasSetAdder(ir_node* block, void* env)
 104 {
 105         ir_nodeset_t* alias_set;
 106         size_t i;
 107
 108         alias_set = NEW_ARR_D(ir_nodeset_t, &obst, count_addrs);
 109         for (i = 0; i < count_addrs; i++) {
 110                 ir_nodeset_init(&alias_set[i]);
 111         }
 112         set_irn_link(block, alias_set);
 113 }
 114
 115
 116 static void SetStartAddressesTop(ir_graph* irg)
 117 {
 118         ir_node* initial_mem;
 119         ir_node* start_block;
 120         ir_nodeset_t* start_addrs;
 121         size_t i;
 122
 123         initial_mem = get_irg_initial_mem(irg);
 124         start_block = get_irg_start_block(irg);
 125         start_addrs = get_irn_link(start_block);
 126         for (i = 0; i < count_addrs; i++) {
 127                 ir_nodeset_insert(&start_addrs[i], initial_mem);
 128         }
 129         mark_Block_block_visited(start_block);
 130 }
 131
 132
 133 static void AliasSetDestroyer(ir_node* block, void* env)
 134 {
 135         ir_nodeset_t* alias_set = get_irn_link(block);
 136         size_t i;
 137
 138         for (i = 0; i < count_addrs; i++) {
 139                 ir_nodeset_destroy(&alias_set[i]);
 140         }
 141 }
 142
 143
 144 static ir_alias_relation AliasTest(ir_graph* irg, ir_node* addr, ir_mode* mode, ir_node* other)
 145 {
 146         ir_node* other_addr;
 147         ir_mode* other_mode;
 148
 149         if (is_Proj(other)) other = get_Proj_pred(other);
 150
 151         if (is_Load(other)) {
 152                 other_addr = get_Load_ptr(other);
 153         } else if (is_Store(other)) {
 154                 other_addr = get_Store_ptr(other);
 155         } else {
 156                 return may_alias;
 157         }
 158
 159         other_mode = get_irn_mode(other);
 160         return get_alias_relation(irg, addr, mode, other_addr, other_mode);
 161 }
 162
 163
 164 static ir_node* GenerateSync(ir_graph* irg, ir_node* block, ir_nodeset_t* after_set)
 165 {
 166         size_t set_size = ir_nodeset_size(after_set);
 167         ir_nodeset_iterator_t iter;
 168
 169         assert(set_size != 0);
 170
 171         ir_nodeset_iterator_init(&iter, after_set);
 172         if (set_size == 1) {
 173                 return ir_nodeset_iterator_next(&iter);
 174         } else {
 175                 ir_node** in;
 176                 size_t i;
 177
 178                 NEW_ARR_A(ir_node*, in, set_size);
 179                 for (i = 0; i < set_size; i++) {
 180                         in[i] = ir_nodeset_iterator_next(&iter);
 181                 }
 182                 return new_r_Sync(irg, block, set_size, in);
 183         }
 184 }
 185
 186
 187 static ir_node** unfinished_phis;
 188
 189
 190 static void PlaceMemPhis(ir_graph* irg, ir_node* block, ir_node* phi)
 191 {
 192         int unfinished = 0;
 193         size_t block_n_preds = get_Block_n_cfgpreds(block);
 194         ir_nodeset_t* thissets;
 195         ir_node** in;
 196         size_t i;
 197         size_t j;
 198
 199         thissets = get_irn_link(block);
 200         NEW_ARR_A(ir_node*, in, block_n_preds);
 201         for (j = 0; j < count_addrs; j++) {
 202                 ir_node* new_phi;
 203
 204                 for (i = 0; i < block_n_preds; i++) {
 205                         ir_node* pred_block = get_nodes_block(get_Phi_pred(phi, i)); // TODO get_Block_cfgpred_block(block, i);
 206                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 207                         size_t predset_size = ir_nodeset_size(&predsets[j]);
 208
 209                         if (predset_size == 0) {
 210                                 in[i] = new_r_Unknown(irg, mode_M);
 211                                 unfinished = 1;
 212                         } else {
 213                                 in[i] = GenerateSync(irg, pred_block, &predsets[j]);
 214                         }
 215                 }
 216                 new_phi = new_r_Phi(irg, block, block_n_preds, in, mode_M);
 217                 if (unfinished) {
 218                         set_irn_link(new_phi, unfinished_phis[j]);
 219                         unfinished_phis[j] = new_phi;
 220                 }
 221                 ir_nodeset_insert(&thissets[j], new_phi);
 222         }
 223 }
 224
 225
 226 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block);
 227
 228
 229 static void WalkMemPhi(ir_graph* irg, ir_node* block, ir_node* phi)
 230 {
 231         size_t n = get_Phi_n_preds(phi);
 232         size_t i;
 233
 234         for (i = 0; i < n; i++) {
 235                 WalkMem(irg, get_Phi_pred(phi, i), block);
 236         }
 237
 238         PlaceMemPhis(irg, block, phi);
 239         exchange(phi, new_Bad());
 240 }
 241
 242
 243 static void PlaceLoad(ir_graph* irg, ir_node* block, ir_node* load, ir_node* memory)
 244 {
 245         ir_node* addr = get_Load_ptr(load);
 246         size_t addr_idx = (size_t)get_irn_link(addr);
 247         ir_nodeset_t* interfere_sets = get_irn_link(block);
 248         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 249         size_t size = ir_nodeset_size(interfere_set);
 250         ir_nodeset_iterator_t interfere_iter;
 251         size_t i;
 252
 253         assert(size > 0);
 254         ir_nodeset_iterator_init(&interfere_iter, interfere_set);
 255         if (size == 1) {
 256                 ir_node* after = ir_nodeset_iterator_next(&interfere_iter);
 257                 if (is_Proj(after)) {
 258                         ir_node* pred = get_Proj_pred(after);
 259                         if (is_Load(pred)) {
 260 #ifdef OPTIMISE_LOAD_AFTER_LOAD
 261                                 if (get_Load_ptr(pred) == addr && get_Load_mode(pred) == get_Load_mode(load)) {
 262                                         exchange(load, pred);
 263                                         return;
 264                                 }
 265 #endif
 266                                 after = get_Load_mem(pred);
 267                         }
 268                 }
 269                 DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, after));
 270                 set_Load_mem(load, after);
 271         } else {
 272                 ir_node** after_set;
 273                 ir_node* sync;
 274
 275                 NEW_ARR_A(ir_node*, after_set, size);
 276                 for (i = 0; i < size; i++) {
 277                         ir_node* mem = ir_nodeset_iterator_next(&interfere_iter);
 278                         if (is_Proj(mem)) {
 279                                 ir_node* pred = get_Proj_pred(mem);
 280                                 if (is_Load(pred)) {
 281 #ifdef OPTIMISE_LOAD_AFTER_LOAD
 282                                         if (get_Load_ptr(pred) == addr && get_Load_mode(pred) == get_Load_mode(load)) {
 283                                                 exchange(load, pred);
 284                                                 return;
 285                                         }
 286 #endif
 287                                         mem = get_Load_mem(pred);
 288                                 }
 289                         }
 290                         after_set[i] = mem;
 291                         sync = new_r_Sync(irg, block, size, after_set);
 292                 }
 293                 set_Load_mem(load, sync);
 294         }
 295
 296         for (i = 0; i < count_addrs; i++) {
 297                 ir_mode* mode = get_Load_mode(load);
 298                 ir_node* other_addr = addrs[i];
 299                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 300                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 301                 ir_node* other_node;
 302
 303                 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
 304                 if (rel == no_alias) {
 305                         continue;
 306                 }
 307                 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", load, other_addr));
 308
 309                 ir_nodeset_iterator_init(&interfere_iter, &interfere_sets[i]);
 310                 while ((other_node = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 311                         if (is_Proj(other_node) && is_Load(get_Proj_pred(other_node))) continue;
 312                         if (AliasTest(irg, addr, mode, other_node) != no_alias) {
 313                                 DB((dbg, LEVEL_3, "===> Removing %+F from execute-after set of %+F due to %+F\n", other_node, addrs[i], load));
 314                                 ir_nodeset_remove_iterator(&interfere_sets[i], &interfere_iter);
 315                         }
 316                 }
 317
 318                 ir_nodeset_insert(&interfere_sets[i], memory);
 319         }
 320 }
 321
 322
 323 static void PlaceStore(ir_graph* irg, ir_node* block, ir_node* store, ir_node* memory)
 324 {
 325         ir_node* addr = get_Store_ptr(store);
 326         size_t addr_idx = (size_t)get_irn_link(addr);
 327         ir_nodeset_t* interfere_sets = get_irn_link(block);
 328         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 329         ir_node* after;
 330         size_t i;
 331
 332         after = GenerateSync(irg, block, interfere_set);
 333         set_Store_mem(store, after);
 334
 335         for (i = 0; i < count_addrs; i++) {
 336                 ir_nodeset_iterator_t interfere_iter;
 337                 ir_mode* mode = get_irn_mode(get_Store_value(store));
 338                 ir_node* other_addr = addrs[i];
 339                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 340                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 341                 ir_node* other_node;
 342
 343                 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
 344                 if (rel == no_alias) {
 345                         continue;
 346                 }
 347                 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", store, other_addr));
 348
 349                 ir_nodeset_iterator_init(&interfere_iter, &interfere_sets[i]);
 350                 while ((other_node = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 351                         if (AliasTest(irg, addr, mode, other_node) != no_alias) {
 352                                 DB((dbg, LEVEL_3, "===> Removing %+F from execute-after set of %+F due to %+F\n", other_node, addrs[i], store));
 353                                 ir_nodeset_remove_iterator(&interfere_sets[i], &interfere_iter);
 354                         }
 355                 }
 356
 357                 ir_nodeset_insert(&interfere_sets[i], memory);
 358         }
 359 }
 360
 361
 362 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block)
 363 {
 364         int block_change = 0;
 365         ir_node* block = get_nodes_block(node);
 366         ir_node* pred;
 367         ir_node* memory = node;
 368         ir_nodeset_t* addr_sets;
 369
 370         if (block != last_block) {
 371                 DB((dbg, LEVEL_3, "===> Changing block from %+F to %+F\n", last_block, block));
 372                 block_change = 1;
 373                 if (Block_not_block_visited(block)) {
 374                         mark_Block_block_visited(block);
 375                 } else {
 376                         DB((dbg, LEVEL_2, "===> Hit already visited block at %+F\n", node));
 377                         return block_change;
 378                 }
 379         }
 380
 381         // Skip projs
 382         if (is_Proj(node)) node = get_Proj_pred(node);
 383
 384         if (is_Phi(node)) {
 385                 WalkMemPhi(irg, block, node);
 386                 return block_change;
 387         } else if (is_Sync(node)) {
 388                 UNIMPLEMENTED
 389         } else if (is_Return(node)) {
 390                 pred = get_Return_mem(node);
 391         } else {
 392                 pred = get_fragile_op_mem(node);
 393         }
 394
 395         if (WalkMem(irg, pred, block)) {
 396                 // There was a block change
 397                 size_t block_arity = get_Block_n_cfgpreds(block);
 398
 399                 DB((dbg, LEVEL_3, "===> There is a block change before %+F\n", node));
 400                 if (block_arity == 1) {
 401                         // Just one predecessor, inherit its alias sets
 402                         ir_node* pred_block = get_nodes_block(pred);
 403                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 404                         ir_nodeset_t* thissets = get_irn_link(block);
 405                         size_t i;
 406
 407                         DB((dbg, LEVEL_3, "===> Copying the only predecessor's address sets\n"));
 408
 409                         if (ir_nodeset_size(&predsets[0]) == 0) {
 410                                 ir_node* unknown;
 411
 412                                 DB((dbg, LEVEL_3, "===> The predecessor was not finished yet\n"));
 413                                 assert(!Block_not_block_visited(pred_block));
 414
 415                                 unknown = new_r_Unknown(irg, mode_M);
 416                                 for (i = 0; i < count_addrs; i++) {
 417                                         ir_node* phi_unk = new_r_Phi(irg, block, 1, &unknown, mode_M);
 418                                         set_irn_link(phi_unk, unfinished_phis[i]);
 419                                         unfinished_phis[i] = phi_unk;
 420                                         ir_nodeset_insert(&thissets[i], phi_unk);
 421                                 }
 422                         } else {
 423                                 for (i = 0; i < count_addrs; i++) {
 424                                         ir_nodeset_iterator_t prediter;
 425                                         ir_node* addr;
 426
 427                                         ir_nodeset_iterator_init(&prediter, &predsets[i]);
 428                                         while ((addr = ir_nodeset_iterator_next(&prediter)) != NULL) {
 429                                                 ir_nodeset_insert(&thissets[i], addr);
 430                                         }
 431                                 }
 432                         }
 433                 }
 434         }
 435
 436         DB((dbg, LEVEL_3, "===> Detotalising %+F\n", node));
 437
 438         addr_sets = get_irn_link(block);
 439
 440         if (is_Load(node)) {
 441                 PlaceLoad(irg, block, node, memory);
 442         } else if (is_Store(node)) {
 443                 PlaceStore(irg, block, node, memory);
 444         } else {
 445                 ir_nodeset_t sync_set;
 446                 size_t i;
 447                 ir_node* after;
 448
 449                 DB((dbg, LEVEL_3, "===> Fallback: %+F aliases everything\n", node));
 450
 451                 ir_nodeset_init(&sync_set);
 452                 for (i = 0; i < count_addrs; i++) {
 453                         ir_nodeset_iterator_t iter;
 454                         ir_node* mem;
 455
 456                         ir_nodeset_iterator_init(&iter, &addr_sets[i]);
 457                         while ((mem = ir_nodeset_iterator_next(&iter)) != NULL) {
 458                                 ir_nodeset_insert(&sync_set, mem);
 459                         }
 460                 }
 461
 462                 after = GenerateSync(irg, block, &sync_set);
 463                 set_irn_n(node, 0, after); // XXX unnice way to set the memory input
 464
 465                 for (i = 0; i < count_addrs; i++) {
 466                         ir_nodeset_iterator_t iter;
 467                         ir_nodeset_iterator_init(&iter, &addr_sets[i]);
 468                         while (ir_nodeset_iterator_next(&iter) != NULL) {
 469                                 ir_nodeset_remove_iterator(&addr_sets[i], &iter);
 470                         }
 471                         ir_nodeset_insert(&addr_sets[i], memory);
 472                 }
 473         }
 474
 475         return block_change;
 476 }
 477
 478
 479 static void FinalisePhis(ir_graph* irg)
 480 {
 481         size_t i;
 482
 483         for (i = 0; i < count_addrs; i++) {
 484                 ir_node* next_phi;
 485                 ir_node* phi;
 486
 487                 for (phi = unfinished_phis[i]; phi != NULL; phi = next_phi) {
 488                         ir_node* block = get_nodes_block(phi);
 489                         size_t block_n_preds = get_Block_n_cfgpreds(block);
 490
 491                         next_phi = get_irn_link(phi);
 492
 493                         DB((dbg, LEVEL_4, "===> Finialising phi %+F in %+F\n", phi, block));
 494
 495                         if (block_n_preds == 1) {
 496                                 ir_node* pred_block = get_Block_cfgpred_block(block, 0);
 497                                 ir_nodeset_t* pred_sets = get_irn_link(pred_block);
 498                                 ir_node* after = GenerateSync(irg, pred_block, &pred_sets[i]);
 499
 500                                 assert(is_Unknown(get_Phi_pred(phi, 0)));
 501                                 exchange(phi, after);
 502                         } else {
 503                                 ir_node** in;
 504                                 size_t j;
 505
 506                                 NEW_ARR_A(ir_node*, in, block_n_preds);
 507                                 for (j = 0; j < block_n_preds; j++) {
 508                                         ir_node* pred_block = get_Block_cfgpred_block(block, j);
 509                                         ir_nodeset_t* pred_sets = get_irn_link(pred_block);
 510
 511                                         if (is_Unknown(get_Phi_pred(phi, j))) {
 512                                                 set_Phi_pred(phi, j, GenerateSync(irg, pred_block, &pred_sets[i]));
 513                                         }
 514                                 }
 515                         }
 516                 }
 517         }
 518 }
 519
 520
 521 static void Detotalise(ir_graph* irg)
 522 {
 523         ir_node* end_block = get_irg_end_block(irg);
 524         size_t npreds = get_Block_n_cfgpreds(end_block);
 525         size_t i;
 526
 527         unfinished_phis = xmalloc(sizeof(*unfinished_phis) * count_addrs);
 528         for (i = 0; i < count_addrs; i++) {
 529                 unfinished_phis[i] = NULL;
 530         }
 531
 532         for (i = 0; i < npreds; i++) {
 533                 ir_node* pred = get_Block_cfgpred(end_block, i);
 534                 assert(is_Return(pred));
 535                 DB((dbg, LEVEL_2, "===> Starting memory walk at %+F\n", pred));
 536                 WalkMem(irg, pred, NULL);
 537         }
 538
 539         FinalisePhis(irg);
 540         xfree(unfinished_phis);
 541 }
 542
 543
 544 static void AddSyncPreds(ir_nodeset_t* preds, ir_node* sync)
 545 {
 546         size_t n = get_Sync_n_preds(sync);
 547         size_t i;
 548
 549         for (i = 0; i < n; i++) {
 550                 ir_node* pred = get_Sync_pred(sync, i);
 551                 if (is_Sync(pred)) {
 552                         AddSyncPreds(preds, pred);
 553                 } else {
 554                         ir_nodeset_insert(preds, pred);
 555                 }
 556         }
 557 }
 558
 559
 560 static void NormaliseSync(ir_node* node, void* env)
 561 {
 562         ir_nodeset_t preds;
 563         ir_nodeset_iterator_t iter;
 564         ir_node** in;
 565         size_t count_preds;
 566         size_t i;
 567
 568         if (!is_Sync(node)) return;
 569
 570         ir_nodeset_init(&preds);
 571         AddSyncPreds(&preds, node);
 572
 573         count_preds = ir_nodeset_size(&preds);
 574         if (count_preds != get_Sync_n_preds(node)) {
 575                 NEW_ARR_A(ir_node*, in, count_preds);
 576                 ir_nodeset_iterator_init(&iter, &preds);
 577                 for (i = 0; i < count_preds; i++) {
 578                         ir_node* pred = ir_nodeset_iterator_next(&iter);
 579                         assert(pred != NULL);
 580                         in[i] = pred;
 581                 }
 582                 set_irn_in(node, count_preds, in);
 583         }
 584
 585         ir_nodeset_destroy(&preds);
 586 }
 587
 588
 589 void opt_ldst2(ir_graph* irg)
 590 {
 591         FIRM_DBG_REGISTER(dbg, "firm.opt.ldst2");
 592         DB((dbg, LEVEL_1, "===> Performing load/store optimisation on %+F\n", irg));
 593
 594         normalize_one_return(irg);
 595
 596         obstack_init(&obst);
 597
 598         if (1 /* XXX */ || get_opt_alias_analysis()) {
 599                 assure_irg_address_taken_computed(irg);
 600                 assure_irp_globals_address_taken_computed();
 601         }
 602
 603
 604         CollectAddresses(irg);
 605         if (count_addrs == 0) return;
 606
 607         irg_block_walk_graph(irg, AliasSetAdder, NULL, NULL);
 608         inc_irg_block_visited(irg);
 609         SetStartAddressesTop(irg);
 610         Detotalise(irg);
 611
 612         dump_ir_block_graph(irg, "-fluffig");
 613
 614         irg_block_walk_graph(irg, AliasSetDestroyer, NULL, NULL);
 615         obstack_free(&obst, NULL);
 616
 617         normalize_proj_nodes(irg);
 618         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 619   optimize_graph_df(irg);
 620         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 621 }