nsz Git - libfirm/blob - ir/opt/ldst2.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   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 "iroptimize.h"
  31
  32 #include "array.h"
  33 #include "debug.h"
  34 #include "ircons.h"
  35 #include "irgraph.h"
  36 #include "irgmod.h"
  37 #include "irgopt.h"
  38 #include "irgwalk.h"
  39 #include "irmemory.h"
  40 #include "irnode.h"
  41 #include "irnodeset.h"
  42 #include "obst.h"
  43 #include "irdump.h"
  44 #include "irflag_t.h"
  45 #include "irprintf.h"
  46
  47 #if +0
  48 #define OPTIMISE_LOAD_AFTER_LOAD
  49
  50
  51 #define UNIMPLEMENTED abort();
  52
  53
  54 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  55
  56
  57 static struct obstack obst;
  58 static size_t count_addrs;
  59 static ir_node** addrs;
  60
  61
  62 static void AddressCollector(ir_node* node, void* env)
  63 {
  64         ir_nodeset_t* addrs_set = env;
  65         ir_node* addr;
  66         if (is_Load(node)) {
  67                 addr = get_Load_ptr(node);
  68         } else if (is_Store(node)) {
  69                 addr = get_Store_ptr(node);
  70         } else {
  71                 return;
  72         }
  73         ir_nodeset_insert(addrs_set, addr);
  74 }
  75
  76
  77 /* Collects all unique addresses used by load and store nodes of a graph and
  78  * puts them into an array for later use */
  79 static void CollectAddresses(ir_graph* irg)
  80 {
  81         ir_nodeset_t addrs_set;
  82
  83         ir_nodeset_init(&addrs_set);
  84         irg_walk_graph(irg, AddressCollector, NULL, &addrs_set);
  85
  86         count_addrs = ir_nodeset_size(&addrs_set);
  87         DB((dbg, LEVEL_1, "===> %+F uses %u unique addresses\n", irg, (unsigned int)count_addrs));
  88         if (count_addrs != 0) {
  89                 ir_nodeset_iterator_t addr_iter;
  90                 size_t i;
  91
  92                 addrs = NEW_ARR_D(ir_node*, &obst, count_addrs);
  93                 ir_nodeset_iterator_init(&addr_iter, &addrs_set);
  94                 for (i = 0; i < count_addrs; i++) {
  95                         ir_node* addr = ir_nodeset_iterator_next(&addr_iter);
  96                         assert(addr != NULL);
  97                         set_irn_link(addr, (void *)i);
  98                         addrs[i] = addr;
  99                         DB((dbg, LEVEL_2, "===> Collected unique symbolic address %+F\n", addr));
 100                 }
 101         }
 102 }
 103
 104
 105 static void AliasSetAdder(ir_node* block, void* env)
 106 {
 107         ir_nodeset_t* alias_set;
 108         size_t i;
 109         (void) env;
 110
 111         alias_set = NEW_ARR_D(ir_nodeset_t, &obst, count_addrs);
 112         for (i = 0; i < count_addrs; i++) {
 113                 ir_nodeset_init(&alias_set[i]);
 114         }
 115         set_irn_link(block, alias_set);
 116 }
 117
 118
 119 static void SetStartAddressesTop(ir_graph* irg)
 120 {
 121         ir_node* initial_mem;
 122         ir_node* start_block;
 123         ir_nodeset_t* start_addrs;
 124         size_t i;
 125
 126         initial_mem = get_irg_initial_mem(irg);
 127         start_block = get_irg_start_block(irg);
 128         start_addrs = get_irn_link(start_block);
 129         for (i = 0; i < count_addrs; i++) {
 130                 ir_nodeset_insert(&start_addrs[i], initial_mem);
 131         }
 132         mark_Block_block_visited(start_block);
 133 }
 134
 135
 136 static void AliasSetDestroyer(ir_node* block, void* env)
 137 {
 138         ir_nodeset_t* alias_set = get_irn_link(block);
 139         size_t i;
 140         (void) env;
 141
 142         for (i = 0; i < count_addrs; i++) {
 143                 ir_nodeset_destroy(&alias_set[i]);
 144         }
 145 }
 146
 147
 148 static ir_alias_relation AliasTest(ir_graph* irg, ir_node* addr, ir_mode* mode, ir_node* other)
 149 {
 150         ir_node* other_addr;
 151         ir_mode* other_mode;
 152
 153         if (is_Proj(other)) other = get_Proj_pred(other);
 154
 155         if (is_Load(other)) {
 156                 other_addr = get_Load_ptr(other);
 157         } else if (is_Store(other)) {
 158                 other_addr = get_Store_ptr(other);
 159         } else {
 160                 return may_alias;
 161         }
 162
 163         other_mode = get_irn_mode(other);
 164         return get_alias_relation(irg, addr, mode, other_addr, other_mode);
 165 }
 166
 167
 168 static ir_node* GenerateSync(ir_graph* irg, ir_node* block, ir_nodeset_t* after_set)
 169 {
 170         size_t set_size = ir_nodeset_size(after_set);
 171         ir_nodeset_iterator_t iter;
 172
 173         assert(set_size != 0);
 174
 175         ir_nodeset_iterator_init(&iter, after_set);
 176         if (set_size == 1) {
 177                 return ir_nodeset_iterator_next(&iter);
 178         } else {
 179                 ir_node** in;
 180                 size_t i;
 181
 182                 NEW_ARR_A(ir_node*, in, set_size);
 183                 for (i = 0; i < set_size; i++) {
 184                         in[i] = ir_nodeset_iterator_next(&iter);
 185                 }
 186                 return new_r_Sync(irg, block, set_size, in);
 187         }
 188 }
 189
 190
 191 static ir_node** unfinished_phis;
 192
 193
 194 static void PlaceMemPhis(ir_graph* irg, ir_node* block, ir_node* phi)
 195 {
 196         int unfinished = 0;
 197         size_t block_n_preds = get_Block_n_cfgpreds(block);
 198         ir_nodeset_t* thissets;
 199         ir_node** in;
 200         size_t i;
 201         size_t j;
 202
 203         thissets = get_irn_link(block);
 204         NEW_ARR_A(ir_node*, in, block_n_preds);
 205         for (j = 0; j < count_addrs; j++) {
 206                 ir_node* new_phi;
 207
 208                 for (i = 0; i < block_n_preds; i++) {
 209                         ir_node* pred_block = get_nodes_block(get_Phi_pred(phi, i)); // TODO get_Block_cfgpred_block(block, i);
 210                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 211                         size_t predset_size = ir_nodeset_size(&predsets[j]);
 212
 213                         if (predset_size == 0) {
 214                                 in[i] = new_r_Unknown(irg, mode_M);
 215                                 unfinished = 1;
 216                         } else {
 217                                 in[i] = GenerateSync(irg, pred_block, &predsets[j]);
 218                         }
 219                 }
 220                 new_phi = new_r_Phi(irg, block, block_n_preds, in, mode_M);
 221                 if (unfinished) {
 222                         set_irn_link(new_phi, unfinished_phis[j]);
 223                         unfinished_phis[j] = new_phi;
 224                 }
 225                 ir_nodeset_insert(&thissets[j], new_phi);
 226         }
 227 }
 228
 229
 230 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block);
 231
 232
 233 static void WalkMemPhi(ir_graph* irg, ir_node* block, ir_node* phi)
 234 {
 235         size_t n = get_Phi_n_preds(phi);
 236         size_t i;
 237
 238         for (i = 0; i < n; i++) {
 239                 WalkMem(irg, get_Phi_pred(phi, i), block);
 240         }
 241
 242         PlaceMemPhis(irg, block, phi);
 243         exchange(phi, new_Bad());
 244 }
 245
 246
 247 static void PlaceLoad(ir_graph* irg, ir_node* block, ir_node* load, ir_node* memory)
 248 {
 249         ir_node* addr = get_Load_ptr(load);
 250         size_t addr_idx = (size_t)get_irn_link(addr);
 251         ir_nodeset_t* interfere_sets = get_irn_link(block);
 252         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 253         size_t size = ir_nodeset_size(interfere_set);
 254         ir_nodeset_iterator_t interfere_iter;
 255         size_t i;
 256
 257         assert(size > 0);
 258         ir_nodeset_iterator_init(&interfere_iter, interfere_set);
 259         if (size == 1) {
 260                 ir_node* after = ir_nodeset_iterator_next(&interfere_iter);
 261                 assert(!is_Proj(after) || !is_Load(get_Proj_pred(after)));
 262                 DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, after));
 263                 set_Load_mem(load, after);
 264         } else {
 265                 ir_node** after_set;
 266                 ir_node* after;
 267                 ir_node* mem;
 268                 size_t i;
 269
 270                 NEW_ARR_A(ir_node*, after_set, size);
 271                 i = 0;
 272                 while ((mem = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 273                         if (is_Proj(mem)) {
 274                                 ir_node* pred = get_Proj_pred(mem);
 275                                 if (is_Load(pred)) {
 276 #ifdef OPTIMISE_LOAD_AFTER_LOAD
 277                                         if (get_Load_ptr(pred) == addr && get_Load_mode(pred) == get_Load_mode(load)) {
 278                                                 exchange(load, pred);
 279                                                 return;
 280                                         }
 281 #endif
 282                                         continue;
 283                                 }
 284                         }
 285                         DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, mem));
 286                         after_set[i++] = mem;
 287                 }
 288                 assert(i != 0);
 289                 if (i == 1) {
 290                         after = after_set[0];
 291                 } else {
 292                         after = new_r_Sync(irg, block, i, after_set);
 293                 }
 294                 set_Load_mem(load, after);
 295         }
 296
 297         for (i = 0; i < count_addrs; i++) {
 298                 ir_mode* mode = get_Load_mode(load);
 299                 ir_node* other_addr = addrs[i];
 300                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 301                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 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_insert(&interfere_sets[i], memory);
 310         }
 311 }
 312
 313
 314 static void PlaceStore(ir_graph* irg, ir_node* block, ir_node* store, ir_node* memory)
 315 {
 316         ir_node* addr = get_Store_ptr(store);
 317         size_t addr_idx = (size_t)get_irn_link(addr);
 318         ir_nodeset_t* interfere_sets = get_irn_link(block);
 319         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 320         ir_node* after;
 321         size_t i;
 322
 323         after = GenerateSync(irg, block, interfere_set);
 324         set_Store_mem(store, after);
 325
 326         for (i = 0; i < count_addrs; i++) {
 327                 ir_nodeset_iterator_t interfere_iter;
 328                 ir_mode* mode = get_irn_mode(get_Store_value(store));
 329                 ir_node* other_addr = addrs[i];
 330                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 331                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 332                 ir_node* other_node;
 333
 334                 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
 335                 if (rel == no_alias) {
 336                         continue;
 337                 }
 338                 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", store, other_addr));
 339
 340                 ir_nodeset_iterator_init(&interfere_iter, &interfere_sets[i]);
 341                 while ((other_node = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 342                         if (AliasTest(irg, addr, mode, other_node) != no_alias) {
 343                                 DB((dbg, LEVEL_3, "===> Removing %+F from execute-after set of %+F due to %+F\n", other_node, addrs[i], store));
 344                                 ir_nodeset_remove_iterator(&interfere_sets[i], &interfere_iter);
 345                         }
 346                 }
 347
 348                 ir_nodeset_insert(&interfere_sets[i], memory);
 349         }
 350 }
 351
 352
 353 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block)
 354 {
 355         int block_change = 0;
 356         ir_node* block = get_nodes_block(node);
 357         ir_node* pred;
 358         ir_node* memory = node;
 359         ir_nodeset_t* addr_sets;
 360
 361         if (block != last_block) {
 362                 DB((dbg, LEVEL_3, "===> Changing block from %+F to %+F\n", last_block, block));
 363                 block_change = 1;
 364                 if (Block_not_block_visited(block)) {
 365                         mark_Block_block_visited(block);
 366                 } else {
 367                         DB((dbg, LEVEL_2, "===> Hit already visited block at %+F\n", node));
 368                         return block_change;
 369                 }
 370         }
 371
 372         // Skip projs
 373         if (is_Proj(node)) node = get_Proj_pred(node);
 374
 375         if (is_Phi(node)) {
 376                 WalkMemPhi(irg, block, node);
 377                 return block_change;
 378         } else if (is_Sync(node)) {
 379                 UNIMPLEMENTED
 380         } else if (is_Return(node)) {
 381                 pred = get_Return_mem(node);
 382         } else {
 383                 pred = get_fragile_op_mem(node);
 384         }
 385
 386         if (WalkMem(irg, pred, block)) {
 387                 // There was a block change
 388                 size_t block_arity = get_Block_n_cfgpreds(block);
 389
 390                 DB((dbg, LEVEL_3, "===> There is a block change before %+F\n", node));
 391                 if (block_arity == 1) {
 392                         // Just one predecessor, inherit its alias sets
 393                         ir_node* pred_block = get_nodes_block(pred);
 394                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 395                         ir_nodeset_t* thissets = get_irn_link(block);
 396                         size_t i;
 397
 398                         DB((dbg, LEVEL_3, "===> Copying the only predecessor's address sets\n"));
 399
 400                         if (ir_nodeset_size(&predsets[0]) == 0) {
 401                                 ir_node* unknown;
 402
 403                                 DB((dbg, LEVEL_3, "===> The predecessor was not finished yet\n"));
 404                                 assert(!Block_not_block_visited(pred_block));
 405
 406                                 unknown = new_r_Unknown(irg, mode_M);
 407                                 for (i = 0; i < count_addrs; i++) {
 408                                         ir_node* phi_unk = new_r_Phi(irg, block, 1, &unknown, mode_M);
 409                                         DB((dbg, LEVEL_3, "===> Placing unfinished %+F for %+F in %+F\n", phi_unk, addrs[i], block));
 410                                         set_irn_link(phi_unk, unfinished_phis[i]);
 411                                         unfinished_phis[i] = phi_unk;
 412                                         ir_nodeset_insert(&thissets[i], phi_unk);
 413                                 }
 414                         } else {
 415                                 for (i = 0; i < count_addrs; i++) {
 416                                         ir_nodeset_iterator_t prediter;
 417                                         ir_node* addr;
 418
 419                                         ir_nodeset_iterator_init(&prediter, &predsets[i]);
 420                                         while ((addr = ir_nodeset_iterator_next(&prediter)) != NULL) {
 421                                                 ir_nodeset_insert(&thissets[i], addr);
 422                                         }
 423                                 }
 424                         }
 425                 }
 426         }
 427
 428         DB((dbg, LEVEL_3, "===> Detotalising %+F\n", node));
 429
 430         addr_sets = get_irn_link(block);
 431
 432         if (is_Load(node)) {
 433                 PlaceLoad(irg, block, node, memory);
 434         } else if (is_Store(node)) {
 435                 PlaceStore(irg, block, node, memory);
 436         } else {
 437                 ir_nodeset_t sync_set;
 438                 size_t i;
 439                 ir_node* after;
 440
 441                 DB((dbg, LEVEL_3, "===> Fallback: %+F aliases everything\n", node));
 442
 443                 ir_nodeset_init(&sync_set);
 444                 for (i = 0; i < count_addrs; i++) {
 445                         ir_nodeset_iterator_t iter;
 446                         ir_node* mem;
 447
 448                         ir_nodeset_iterator_init(&iter, &addr_sets[i]);
 449                         while ((mem = ir_nodeset_iterator_next(&iter)) != NULL) {
 450                                 ir_nodeset_insert(&sync_set, mem);
 451                         }
 452                 }
 453
 454                 after = GenerateSync(irg, block, &sync_set);
 455                 set_irn_n(node, 0, after); // XXX unnice way to set the memory input
 456
 457                 for (i = 0; i < count_addrs; i++) {
 458                         ir_nodeset_iterator_t iter;
 459                         ir_nodeset_iterator_init(&iter, &addr_sets[i]);
 460                         while (ir_nodeset_iterator_next(&iter) != NULL) {
 461                                 ir_nodeset_remove_iterator(&addr_sets[i], &iter);
 462                         }
 463                         ir_nodeset_insert(&addr_sets[i], memory);
 464                 }
 465         }
 466
 467         return block_change;
 468 }
 469
 470
 471 static void FinalisePhis(ir_graph* irg)
 472 {
 473         size_t i;
 474
 475         for (i = 0; i < count_addrs; i++) {
 476                 ir_node* next_phi;
 477                 ir_node* phi;
 478
 479                 for (phi = unfinished_phis[i]; phi != NULL; phi = next_phi) {
 480                         ir_node* block = get_nodes_block(phi);
 481                         size_t block_n_preds = get_Block_n_cfgpreds(block);
 482
 483                         next_phi = get_irn_link(phi);
 484
 485                         DB((dbg, LEVEL_4, "===> Finialising phi %+F in %+F\n", phi, block));
 486
 487                         if (block_n_preds == 1) {
 488                                 ir_node* pred_block = get_Block_cfgpred_block(block, 0);
 489                                 ir_nodeset_t* pred_sets = get_irn_link(pred_block);
 490                                 ir_node* after = GenerateSync(irg, pred_block, &pred_sets[i]);
 491
 492                                 assert(is_Unknown(get_Phi_pred(phi, 0)));
 493                                 exchange(phi, after);
 494                         } else {
 495                                 ir_node** in;
 496                                 size_t j;
 497
 498                                 NEW_ARR_A(ir_node*, in, block_n_preds);
 499                                 for (j = 0; j < block_n_preds; j++) {
 500                                         ir_node* pred_block = get_Block_cfgpred_block(block, j);
 501                                         ir_nodeset_t* pred_sets = get_irn_link(pred_block);
 502
 503                                         if (is_Unknown(get_Phi_pred(phi, j))) {
 504                                                 set_Phi_pred(phi, j, GenerateSync(irg, pred_block, &pred_sets[i]));
 505                                         }
 506                                 }
 507                         }
 508                 }
 509         }
 510 }
 511
 512
 513 static void Detotalise(ir_graph* irg)
 514 {
 515         ir_node* end_block = get_irg_end_block(irg);
 516         size_t npreds = get_Block_n_cfgpreds(end_block);
 517         size_t i;
 518
 519         unfinished_phis = xmalloc(sizeof(*unfinished_phis) * count_addrs);
 520         for (i = 0; i < count_addrs; i++) {
 521                 unfinished_phis[i] = NULL;
 522         }
 523
 524         for (i = 0; i < npreds; i++) {
 525                 ir_node* pred = get_Block_cfgpred(end_block, i);
 526                 assert(is_Return(pred));
 527                 DB((dbg, LEVEL_2, "===> Starting memory walk at %+F\n", pred));
 528                 WalkMem(irg, pred, NULL);
 529         }
 530
 531         FinalisePhis(irg);
 532         xfree(unfinished_phis);
 533 }
 534 #endif
 535
 536
 537 static void AddSyncPreds(ir_nodeset_t* preds, ir_node* sync)
 538 {
 539         size_t n = get_Sync_n_preds(sync);
 540         size_t i;
 541
 542         for (i = 0; i < n; i++) {
 543                 ir_node* pred = get_Sync_pred(sync, i);
 544                 if (is_Sync(pred)) {
 545                         AddSyncPreds(preds, pred);
 546                 } else {
 547                         ir_nodeset_insert(preds, pred);
 548                 }
 549         }
 550 }
 551
 552 #if 0
 553 static void NormaliseSync(ir_node* node, void* env)
 554 {
 555         ir_nodeset_t preds;
 556         ir_nodeset_iterator_t iter;
 557         ir_node** in;
 558         size_t count_preds;
 559         size_t i;
 560         (void) env;
 561
 562         if (!is_Sync(node)) return;
 563
 564         ir_nodeset_init(&preds);
 565         AddSyncPreds(&preds, node);
 566
 567         count_preds = ir_nodeset_size(&preds);
 568         if (count_preds != (unsigned)get_Sync_n_preds(node)) {
 569                 NEW_ARR_A(ir_node*, in, count_preds);
 570                 ir_nodeset_iterator_init(&iter, &preds);
 571                 for (i = 0; i < count_preds; i++) {
 572                         ir_node* pred = ir_nodeset_iterator_next(&iter);
 573                         assert(pred != NULL);
 574                         in[i] = pred;
 575                 }
 576                 set_irn_in(node, count_preds, in);
 577         }
 578
 579         ir_nodeset_destroy(&preds);
 580 }
 581
 582 void opt_ldst2(ir_graph* irg)
 583 {
 584         FIRM_DBG_REGISTER(dbg, "firm.opt.ldst2");
 585         DB((dbg, LEVEL_1, "===> Performing load/store optimisation on %+F\n", irg));
 586
 587         normalize_one_return(irg);
 588         dump_ir_block_graph(irg, "-prefluffig");
 589
 590         obstack_init(&obst);
 591
 592         if (1 /* XXX */ || get_opt_alias_analysis()) {
 593                 assure_irg_address_taken_computed(irg);
 594                 assure_irp_globals_address_taken_computed();
 595         }
 596
 597
 598         CollectAddresses(irg);
 599         if (count_addrs == 0) return;
 600
 601         irg_block_walk_graph(irg, AliasSetAdder, NULL, NULL);
 602         inc_irg_block_visited(irg);
 603         SetStartAddressesTop(irg);
 604         Detotalise(irg);
 605         dump_ir_block_graph(irg, "-fluffig");
 606
 607         irg_block_walk_graph(irg, AliasSetDestroyer, NULL, NULL);
 608         obstack_free(&obst, NULL);
 609
 610         normalize_proj_nodes(irg);
 611         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 612   optimize_graph_df(irg);
 613         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 614         dump_ir_block_graph(irg, "-postfluffig");
 615 }
 616 #endif
 617
 618
 619 typedef struct parallelise_info
 620 {
 621         ir_node      *origin_block;
 622         ir_node      *origin_ptr;
 623         ir_mode      *origin_mode;
 624         ir_nodeset_t  this_mem;
 625         ir_nodeset_t  user_mem;
 626 } parallelise_info;
 627
 628
 629 static void parallelise_load(parallelise_info *pi, ir_node *irn)
 630 {
 631         //ir_fprintf(stderr, "considering %+F\n", irn);
 632         if (get_nodes_block(irn) == pi->origin_block) {
 633                 if (is_Proj(irn)) {
 634                         ir_node *pred = get_Proj_pred(irn);
 635                         if (is_Load(pred) &&
 636                                         get_Load_volatility(pred) == volatility_non_volatile) {
 637                                 ir_node *mem = get_Load_mem(pred);
 638                                 //ir_nodeset_insert(&pi->this_mem, mem);
 639                                 ir_nodeset_insert(&pi->user_mem, irn);
 640                                 //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 641                                 parallelise_load(pi, mem);
 642                                 return;
 643                         } else if (is_Store(pred) &&
 644                                         get_Store_volatility(pred) == volatility_non_volatile) {
 645                                 ir_mode *org_mode   = pi->origin_mode;
 646                                 ir_node *org_ptr    = pi->origin_ptr;
 647                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 648                                 ir_node *store_ptr  = get_Store_ptr(pred);
 649                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, store_ptr, store_mode) == no_alias) {
 650                                         ir_node *mem = get_Store_mem(pred);
 651                                         ir_fprintf(stderr, "Ld after St: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, store_ptr, store_mode);
 652                                         ir_nodeset_insert(&pi->user_mem, irn);
 653                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 654                                         parallelise_load(pi, mem);
 655                                         return;
 656                                 }
 657                         }
 658                 } else if (is_Sync(irn)) {
 659                         int n = get_Sync_n_preds(irn);
 660                         int i;
 661
 662                         for (i = 0; i < n; ++i) {
 663                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 664                                 parallelise_load(pi, sync_pred);
 665                         }
 666                         return;
 667                 }
 668         }
 669         ir_nodeset_insert(&pi->this_mem, irn);
 670         //ir_fprintf(stderr, "adding %+F to this set\n", irn);
 671 }
 672
 673
 674 static void parallelise_store(parallelise_info *pi, ir_node *irn)
 675 {
 676         //ir_fprintf(stderr, "considering %+F\n", irn);
 677         if (get_nodes_block(irn) == pi->origin_block) {
 678                 if (is_Proj(irn)) {
 679                         ir_node *pred = get_Proj_pred(irn);
 680                         if (is_Load(pred) &&
 681                                         get_Load_volatility(pred) == volatility_non_volatile) {
 682                                 ir_mode *org_mode  = pi->origin_mode;
 683                                 ir_node *org_ptr   = pi->origin_ptr;
 684                                 ir_mode *load_mode = get_Load_mode(pred);
 685                                 ir_node *load_ptr  = get_Load_ptr(pred);
 686                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, load_ptr, load_mode) == no_alias) {
 687                                         ir_node *mem = get_Load_mem(pred);
 688                                         ir_fprintf(stderr, "St after Ld: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, load_ptr, load_mode);
 689                                         ir_nodeset_insert(&pi->user_mem, irn);
 690                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 691                                         parallelise_store(pi, mem);
 692                                         return;
 693                                 }
 694                         } else if (is_Store(pred) &&
 695                                         get_Store_volatility(pred) == volatility_non_volatile) {
 696                                 ir_mode *org_mode   = pi->origin_mode;
 697                                 ir_node *org_ptr    = pi->origin_ptr;
 698                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 699                                 ir_node *store_ptr  = get_Store_ptr(pred);
 700                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, store_ptr, store_mode) == no_alias) {
 701                                         ir_node *mem;
 702
 703                                         ir_fprintf(stderr, "St after St: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, store_ptr, store_mode);
 704                                         ir_nodeset_insert(&pi->user_mem, irn);
 705                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 706                                         mem = get_Store_mem(pred);
 707                                         parallelise_store(pi, mem);
 708                                         return;
 709                                 }
 710                         }
 711                 } else if (is_Sync(irn)) {
 712                         int n = get_Sync_n_preds(irn);
 713                         int i;
 714
 715                         for (i = 0; i < n; ++i) {
 716                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 717                                 parallelise_store(pi, sync_pred);
 718                         }
 719                         return;
 720                 }
 721         }
 722         ir_nodeset_insert(&pi->this_mem, irn);
 723         //ir_fprintf(stderr, "adding %+F to this set\n", irn);
 724 }
 725
 726
 727 static void walker(ir_node *proj, void *env)
 728 {
 729         ir_node          *mem_op;
 730         ir_node          *pred;
 731         ir_node          *block;
 732         int               n;
 733         parallelise_info  pi;
 734
 735         (void)env;
 736
 737         if (!is_Proj(proj)) return;
 738         if (get_irn_mode(proj) != mode_M) return;
 739
 740         mem_op = get_Proj_pred(proj);
 741         if (is_Load(mem_op)) {
 742                 if (get_Load_volatility(mem_op) != volatility_non_volatile) return;
 743
 744                 block = get_nodes_block(mem_op);
 745                 pred  = get_Load_mem(mem_op);
 746                 //ir_fprintf(stderr, "starting parallelise at %+F for %+F\n", pred, proj);
 747
 748                 pi.origin_block = block,
 749                 pi.origin_ptr   = get_Load_ptr(mem_op);
 750                 pi.origin_mode  = get_Load_mode(mem_op);
 751                 ir_nodeset_init(&pi.this_mem);
 752                 ir_nodeset_init(&pi.user_mem);
 753
 754                 parallelise_load(&pi, pred);
 755         } else if (is_Store(mem_op)) {
 756                 if (get_Store_volatility(mem_op) != volatility_non_volatile) return;
 757
 758                 block = get_nodes_block(mem_op);
 759                 pred  = get_Store_mem(mem_op);
 760                 //ir_fprintf(stderr, "starting parallelise at %+F for %+F\n", pred, proj);
 761
 762                 pi.origin_block = block,
 763                 pi.origin_ptr   = get_Store_ptr(mem_op);
 764                 pi.origin_mode  = get_irn_mode(get_Store_value(mem_op));
 765                 ir_nodeset_init(&pi.this_mem);
 766                 ir_nodeset_init(&pi.user_mem);
 767
 768                 parallelise_store(&pi, pred);
 769         } else {
 770                 return;
 771         }
 772
 773         n = ir_nodeset_size(&pi.user_mem);
 774         if (n != 0) { /* nothing happend otherwise */
 775                 ir_graph               *irg  = current_ir_graph;
 776                 ir_node                *sync;
 777                 ir_node               **in;
 778                 ir_nodeset_iterator_t   iter;
 779                 int                     i;
 780
 781                 ++n;
 782                 //ir_fprintf(stderr, "creating sync for users of %+F with %d inputs\n", proj, n);
 783                 NEW_ARR_A(ir_node*, in, n);
 784                 i = 0;
 785                 in[i++] = new_r_Unknown(irg, mode_M);
 786                 ir_nodeset_iterator_init(&iter, &pi.user_mem);
 787                 for (;;) {
 788                         ir_node* p = ir_nodeset_iterator_next(&iter);
 789                         if (p == NULL) break;
 790                         in[i++] = p;
 791                 }
 792                 assert(i == n);
 793                 sync = new_r_Sync(irg, block, n, in);
 794                 exchange(proj, sync);
 795
 796                 assert(pn_Load_M == pn_Store_M);
 797                 proj = new_r_Proj(irg, block, mem_op, mode_M, pn_Load_M);
 798                 set_Sync_pred(sync, 0, proj);
 799
 800                 n = ir_nodeset_size(&pi.this_mem);
 801                 //ir_fprintf(stderr, "creating sync for %+F with %d inputs\n", mem_op, n);
 802                 ir_nodeset_iterator_init(&iter, &pi.this_mem);
 803                 if (n == 1) {
 804                         sync = ir_nodeset_iterator_next(&iter);
 805                 } else {
 806                         NEW_ARR_A(ir_node*, in, n);
 807                         i = 0;
 808                         for (;;) {
 809                                 ir_node* p = ir_nodeset_iterator_next(&iter);
 810                                 if (p == NULL) break;
 811                                 in[i++] = p;
 812                         }
 813                         assert(i == n);
 814                         sync = new_r_Sync(irg, block, n, in);
 815                 }
 816                 set_memop_mem(mem_op, sync);
 817         }
 818
 819         ir_nodeset_destroy(&pi.this_mem);
 820         ir_nodeset_destroy(&pi.user_mem);
 821 }
 822
 823
 824 void opt_ldst2(ir_graph *irg)
 825 {
 826         assure_irg_address_taken_computed(irg);
 827         assure_irp_globals_address_taken_computed();
 828
 829         irg_walk_graph(irg, NULL, walker, NULL);
 830   //optimize_graph_df(irg);
 831         //irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 832 }