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 #include "config.h"
  27
  28 #include "iroptimize.h"
  29
  30 #include "array_t.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 "obst.h"
  41 #include "irdump.h"
  42 #include "irflag_t.h"
  43 #include "irprintf.h"
  44 #include "irtools.h"
  45
  46 #if +0
  47 #define OPTIMISE_LOAD_AFTER_LOAD
  48
  49
  50 #define UNIMPLEMENTED abort();
  51
  52
  53 DEBUG_ONLY(static firm_dbg_module_t *dbg);
  54
  55
  56 static struct obstack obst;
  57 static size_t count_addrs;
  58 static ir_node** addrs;
  59
  60
  61 static void AddressCollector(ir_node* node, void* env)
  62 {
  63         ir_nodeset_t* addrs_set = env;
  64         ir_node* addr;
  65         if (is_Load(node)) {
  66                 addr = get_Load_ptr(node);
  67         } else if (is_Store(node)) {
  68                 addr = get_Store_ptr(node);
  69         } else {
  70                 return;
  71         }
  72         ir_nodeset_insert(addrs_set, addr);
  73 }
  74
  75
  76 /* Collects all unique addresses used by load and store nodes of a graph and
  77  * puts them into an array for later use */
  78 static void CollectAddresses(ir_graph* irg)
  79 {
  80         ir_nodeset_t addrs_set;
  81
  82         ir_nodeset_init(&addrs_set);
  83         irg_walk_graph(irg, AddressCollector, NULL, &addrs_set);
  84
  85         count_addrs = ir_nodeset_size(&addrs_set);
  86         DB((dbg, LEVEL_1, "===> %+F uses %u unique addresses\n", irg, (unsigned int)count_addrs));
  87         if (count_addrs != 0) {
  88                 ir_nodeset_iterator_t addr_iter;
  89                 size_t i;
  90
  91                 addrs = NEW_ARR_D(ir_node*, &obst, count_addrs);
  92                 ir_nodeset_iterator_init(&addr_iter, &addrs_set);
  93                 for (i = 0; i < count_addrs; i++) {
  94                         ir_node* addr = ir_nodeset_iterator_next(&addr_iter);
  95                         assert(addr != NULL);
  96                         set_irn_link(addr, (void *)i);
  97                         addrs[i] = addr;
  98                         DB((dbg, LEVEL_2, "===> Collected unique symbolic address %+F\n", addr));
  99                 }
 100         }
 101 }
 102
 103
 104 static void AliasSetAdder(ir_node* block, void* env)
 105 {
 106         ir_nodeset_t* alias_set;
 107         size_t i;
 108         (void) env;
 109
 110         alias_set = NEW_ARR_D(ir_nodeset_t, &obst, count_addrs);
 111         for (i = 0; i < count_addrs; i++) {
 112                 ir_nodeset_init(&alias_set[i]);
 113         }
 114         set_irn_link(block, alias_set);
 115 }
 116
 117
 118 static void SetStartAddressesTop(ir_graph* irg)
 119 {
 120         ir_node* initial_mem;
 121         ir_node* start_block;
 122         ir_nodeset_t* start_addrs;
 123         size_t i;
 124
 125         initial_mem = get_irg_initial_mem(irg);
 126         start_block = get_irg_start_block(irg);
 127         start_addrs = get_irn_link(start_block);
 128         for (i = 0; i < count_addrs; i++) {
 129                 ir_nodeset_insert(&start_addrs[i], initial_mem);
 130         }
 131         mark_Block_block_visited(start_block);
 132 }
 133
 134
 135 static void AliasSetDestroyer(ir_node* block, void* env)
 136 {
 137         ir_nodeset_t* alias_set = get_irn_link(block);
 138         size_t i;
 139         (void) env;
 140
 141         for (i = 0; i < count_addrs; i++) {
 142                 ir_nodeset_destroy(&alias_set[i]);
 143         }
 144 }
 145
 146
 147 static ir_alias_relation AliasTest(ir_graph* irg, ir_node* addr, ir_mode* mode, ir_node* other)
 148 {
 149         ir_node* other_addr;
 150         ir_mode* other_mode;
 151
 152         if (is_Proj(other)) other = get_Proj_pred(other);
 153
 154         if (is_Load(other)) {
 155                 other_addr = get_Load_ptr(other);
 156         } else if (is_Store(other)) {
 157                 other_addr = get_Store_ptr(other);
 158         } else {
 159                 return ir_may_alias;
 160         }
 161
 162         other_mode = get_irn_mode(other);
 163         return get_alias_relation(irg, addr, mode, other_addr, other_mode);
 164 }
 165
 166
 167 static int in_cmp(void const* va, void const* vb)
 168 {
 169         ir_node const* const a = *(ir_node const*const*)va;
 170         ir_node const* const b = *(ir_node const*const*)vb;
 171         return get_irn_idx(a) - get_irn_idx(b);
 172 }
 173
 174
 175 static ir_node* GenerateSync(ir_graph* irg, ir_node* block, ir_nodeset_t* after_set)
 176 {
 177         size_t set_size = ir_nodeset_size(after_set);
 178         ir_nodeset_iterator_t iter;
 179
 180         assert(set_size != 0);
 181
 182         ir_nodeset_iterator_init(&iter, after_set);
 183         if (set_size == 1) {
 184                 return ir_nodeset_iterator_next(&iter);
 185         } else {
 186                 ir_node** in;
 187                 size_t i;
 188
 189                 NEW_ARR_A(ir_node*, in, set_size);
 190                 for (i = 0; i < set_size; i++) {
 191                         in[i] = ir_nodeset_iterator_next(&iter);
 192                 }
 193                 qsort(in, set_size, sizeof(*in), in_cmp);
 194                 return new_r_Sync(irg, block, set_size, in);
 195         }
 196 }
 197
 198
 199 static ir_node** unfinished_phis;
 200
 201
 202 static void PlaceMemPhis(ir_graph* irg, ir_node* block, ir_node* phi)
 203 {
 204         int unfinished = 0;
 205         size_t block_n_preds = get_Block_n_cfgpreds(block);
 206         ir_nodeset_t* thissets;
 207         ir_node** in;
 208         size_t i;
 209         size_t j;
 210
 211         thissets = get_irn_link(block);
 212         NEW_ARR_A(ir_node*, in, block_n_preds);
 213         for (j = 0; j < count_addrs; j++) {
 214                 ir_node* new_phi;
 215
 216                 for (i = 0; i < block_n_preds; i++) {
 217                         ir_node* pred_block = get_nodes_block(get_Phi_pred(phi, i)); // TODO get_Block_cfgpred_block(block, i);
 218                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 219                         size_t predset_size = ir_nodeset_size(&predsets[j]);
 220
 221                         if (predset_size == 0) {
 222                                 in[i] = new_r_Unknown(irg, mode_M);
 223                                 unfinished = 1;
 224                         } else {
 225                                 in[i] = GenerateSync(irg, pred_block, &predsets[j]);
 226                         }
 227                 }
 228                 new_phi = new_r_Phi(irg, block, block_n_preds, in, mode_M);
 229                 if (unfinished) {
 230                         set_irn_link(new_phi, unfinished_phis[j]);
 231                         unfinished_phis[j] = new_phi;
 232                 }
 233                 ir_nodeset_insert(&thissets[j], new_phi);
 234         }
 235 }
 236
 237
 238 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block);
 239
 240
 241 static void WalkMemPhi(ir_graph* irg, ir_node* block, ir_node* phi)
 242 {
 243         size_t n = get_Phi_n_preds(phi);
 244         size_t i;
 245
 246         for (i = 0; i < n; i++) {
 247                 WalkMem(irg, get_Phi_pred(phi, i), block);
 248         }
 249
 250         PlaceMemPhis(irg, block, phi);
 251         exchange(phi, new_Bad());
 252 }
 253
 254
 255 static void PlaceLoad(ir_graph* irg, ir_node* block, ir_node* load, ir_node* memory)
 256 {
 257         ir_node* addr = get_Load_ptr(load);
 258         size_t addr_idx = (size_t)get_irn_link(addr);
 259         ir_nodeset_t* interfere_sets = get_irn_link(block);
 260         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 261         size_t size = ir_nodeset_size(interfere_set);
 262         ir_nodeset_iterator_t interfere_iter;
 263         size_t i;
 264
 265         assert(size > 0);
 266         ir_nodeset_iterator_init(&interfere_iter, interfere_set);
 267         if (size == 1) {
 268                 ir_node* after = ir_nodeset_iterator_next(&interfere_iter);
 269                 assert(!is_Proj(after) || !is_Load(get_Proj_pred(after)));
 270                 DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, after));
 271                 set_Load_mem(load, after);
 272         } else {
 273                 ir_node** after_set;
 274                 ir_node* after;
 275                 ir_node* mem;
 276                 size_t i;
 277
 278                 NEW_ARR_A(ir_node*, after_set, size);
 279                 i = 0;
 280                 while ((mem = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 281                         if (is_Proj(mem)) {
 282                                 ir_node* pred = get_Proj_pred(mem);
 283                                 if (is_Load(pred)) {
 284 #ifdef OPTIMISE_LOAD_AFTER_LOAD
 285                                         if (get_Load_ptr(pred) == addr && get_Load_mode(pred) == get_Load_mode(load)) {
 286                                                 exchange(load, pred);
 287                                                 return;
 288                                         }
 289 #endif
 290                                         continue;
 291                                 }
 292                         }
 293                         DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, mem));
 294                         after_set[i++] = mem;
 295                 }
 296                 assert(i != 0);
 297                 if (i == 1) {
 298                         after = after_set[0];
 299                 } else {
 300                         after = new_r_Sync(irg, block, i, after_set);
 301                 }
 302                 set_Load_mem(load, after);
 303         }
 304
 305         for (i = 0; i < count_addrs; i++) {
 306                 ir_mode* mode = get_Load_mode(load);
 307                 ir_node* other_addr = addrs[i];
 308                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 309                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 310
 311                 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
 312                 if (rel == ir_no_alias) {
 313                         continue;
 314                 }
 315                 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", load, other_addr));
 316
 317                 ir_nodeset_insert(&interfere_sets[i], memory);
 318         }
 319 }
 320
 321
 322 static void PlaceStore(ir_graph* irg, ir_node* block, ir_node* store, ir_node* memory)
 323 {
 324         ir_node* addr = get_Store_ptr(store);
 325         size_t addr_idx = (size_t)get_irn_link(addr);
 326         ir_nodeset_t* interfere_sets = get_irn_link(block);
 327         ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
 328         ir_node* after;
 329         size_t i;
 330
 331         after = GenerateSync(irg, block, interfere_set);
 332         set_Store_mem(store, after);
 333
 334         for (i = 0; i < count_addrs; i++) {
 335                 ir_nodeset_iterator_t interfere_iter;
 336                 ir_mode* mode = get_irn_mode(get_Store_value(store));
 337                 ir_node* other_addr = addrs[i];
 338                 ir_mode* other_mode = mode; // XXX second mode is nonsense
 339                 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
 340                 ir_node* other_node;
 341
 342                 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
 343                 if (rel == ir_no_alias) {
 344                         continue;
 345                 }
 346                 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", store, other_addr));
 347
 348                 ir_nodeset_iterator_init(&interfere_iter, &interfere_sets[i]);
 349                 while ((other_node = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
 350                         if (AliasTest(irg, addr, mode, other_node) != ir_no_alias) {
 351                                 DB((dbg, LEVEL_3, "===> Removing %+F from execute-after set of %+F due to %+F\n", other_node, addrs[i], store));
 352                                 ir_nodeset_remove_iterator(&interfere_sets[i], &interfere_iter);
 353                         }
 354                 }
 355
 356                 ir_nodeset_insert(&interfere_sets[i], memory);
 357         }
 358 }
 359
 360
 361 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block)
 362 {
 363         int block_change = 0;
 364         ir_node* block = get_nodes_block(node);
 365         ir_node* pred;
 366         ir_node* memory = node;
 367         ir_nodeset_t* addr_sets;
 368
 369         if (block != last_block) {
 370                 DB((dbg, LEVEL_3, "===> Changing block from %+F to %+F\n", last_block, block));
 371                 block_change = 1;
 372                 if (!Block_block_visited(block)) {
 373                         mark_Block_block_visited(block);
 374                 } else {
 375                         DB((dbg, LEVEL_2, "===> Hit already visited block at %+F\n", node));
 376                         return block_change;
 377                 }
 378         }
 379
 380         // Skip projs
 381         if (is_Proj(node)) node = get_Proj_pred(node);
 382
 383         if (is_Phi(node)) {
 384                 WalkMemPhi(irg, block, node);
 385                 return block_change;
 386         } else if (is_Sync(node)) {
 387                 UNIMPLEMENTED
 388         } else if (is_Return(node)) {
 389                 pred = get_Return_mem(node);
 390         } else {
 391                 pred = get_fragile_op_mem(node);
 392         }
 393
 394         if (WalkMem(irg, pred, block)) {
 395                 // There was a block change
 396                 size_t block_arity = get_Block_n_cfgpreds(block);
 397
 398                 DB((dbg, LEVEL_3, "===> There is a block change before %+F\n", node));
 399                 if (block_arity == 1) {
 400                         // Just one predecessor, inherit its alias sets
 401                         ir_node* pred_block = get_nodes_block(pred);
 402                         ir_nodeset_t* predsets = get_irn_link(pred_block);
 403                         ir_nodeset_t* thissets = get_irn_link(block);
 404                         size_t i;
 405
 406                         DB((dbg, LEVEL_3, "===> Copying the only predecessor's address sets\n"));
 407
 408                         if (ir_nodeset_size(&predsets[0]) == 0) {
 409                                 ir_node* unknown;
 410
 411                                 DB((dbg, LEVEL_3, "===> The predecessor was not finished yet\n"));
 412                                 assert(Block_block_visited(pred_block));
 413
 414                                 unknown = new_r_Unknown(irg, mode_M);
 415                                 for (i = 0; i < count_addrs; i++) {
 416                                         ir_node* phi_unk = new_r_Phi(irg, block, 1, &unknown, mode_M);
 417                                         DB((dbg, LEVEL_3, "===> Placing unfinished %+F for %+F in %+F\n", phi_unk, addrs[i], block));
 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 = XMALLOCN(ir_node, 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 #endif
 543
 544
 545 #if 0
 546 static void AddSyncPreds(ir_nodeset_t* preds, ir_node* sync)
 547 {
 548         size_t n = get_Sync_n_preds(sync);
 549         size_t i;
 550
 551         for (i = 0; i < n; i++) {
 552                 ir_node* pred = get_Sync_pred(sync, i);
 553                 if (is_Sync(pred)) {
 554                         AddSyncPreds(preds, pred);
 555                 } else {
 556                         ir_nodeset_insert(preds, pred);
 557                 }
 558         }
 559 }
 560
 561 static void NormaliseSync(ir_node* node, void* env)
 562 {
 563         ir_nodeset_t preds;
 564         ir_nodeset_iterator_t iter;
 565         ir_node** in;
 566         size_t count_preds;
 567         size_t i;
 568         (void) env;
 569
 570         if (!is_Sync(node)) return;
 571
 572         ir_nodeset_init(&preds);
 573         AddSyncPreds(&preds, node);
 574
 575         count_preds = ir_nodeset_size(&preds);
 576         if (count_preds != (unsigned)get_Sync_n_preds(node)) {
 577                 NEW_ARR_A(ir_node*, in, count_preds);
 578                 ir_nodeset_iterator_init(&iter, &preds);
 579                 for (i = 0; i < count_preds; i++) {
 580                         ir_node* pred = ir_nodeset_iterator_next(&iter);
 581                         assert(pred != NULL);
 582                         in[i] = pred;
 583                 }
 584                 set_irn_in(node, count_preds, in);
 585         }
 586
 587         ir_nodeset_destroy(&preds);
 588 }
 589
 590 void opt_ldst2(ir_graph* irg)
 591 {
 592         FIRM_DBG_REGISTER(dbg, "firm.opt.ldst2");
 593         DB((dbg, LEVEL_1, "===> Performing load/store optimisation on %+F\n", irg));
 594
 595         normalize_one_return(irg);
 596         dump_ir_block_graph(irg, "-prefluffig");
 597
 598         obstack_init(&obst);
 599
 600         if (1 /* XXX */ || get_opt_alias_analysis()) {
 601                 assure_irg_address_taken_computed(irg);
 602                 assure_irp_globals_address_taken_computed();
 603         }
 604
 605
 606         CollectAddresses(irg);
 607         if (count_addrs == 0) return;
 608
 609         irg_block_walk_graph(irg, AliasSetAdder, NULL, NULL);
 610         inc_irg_block_visited(irg);
 611         SetStartAddressesTop(irg);
 612         Detotalise(irg);
 613         dump_ir_block_graph(irg, "-fluffig");
 614
 615         irg_block_walk_graph(irg, AliasSetDestroyer, NULL, NULL);
 616         obstack_free(&obst, NULL);
 617
 618         normalize_proj_nodes(irg);
 619         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 620   optimize_graph_df(irg);
 621         irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 622         dump_ir_block_graph(irg, "-postfluffig");
 623 }
 624 #endif
 625
 626
 627 typedef struct parallelise_info
 628 {
 629         ir_node      *origin_block;
 630         ir_node      *origin_ptr;
 631         ir_mode      *origin_mode;
 632         ir_nodeset_t  this_mem;
 633         ir_nodeset_t  user_mem;
 634 } parallelise_info;
 635
 636
 637 static void parallelise_load(parallelise_info *pi, ir_node *irn)
 638 {
 639         /* There is no point in investigating the same subgraph twice */
 640         if (ir_nodeset_contains(&pi->user_mem, irn))
 641                 return;
 642
 643         //ir_fprintf(stderr, "considering %+F\n", irn);
 644         if (get_nodes_block(irn) == pi->origin_block) {
 645                 if (is_Proj(irn)) {
 646                         ir_node *pred = get_Proj_pred(irn);
 647                         if (is_Load(pred) &&
 648                                         get_Load_volatility(pred) == volatility_non_volatile) {
 649                                 ir_node *mem = get_Load_mem(pred);
 650                                 //ir_nodeset_insert(&pi->this_mem, mem);
 651                                 ir_nodeset_insert(&pi->user_mem, irn);
 652                                 //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 653                                 parallelise_load(pi, mem);
 654                                 return;
 655                         } else if (is_Store(pred) &&
 656                                         get_Store_volatility(pred) == volatility_non_volatile) {
 657                                 ir_mode *org_mode   = pi->origin_mode;
 658                                 ir_node *org_ptr    = pi->origin_ptr;
 659                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 660                                 ir_node *store_ptr  = get_Store_ptr(pred);
 661                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
 662                                         ir_node *mem = get_Store_mem(pred);
 663                                         //ir_fprintf(stderr, "Ld after St: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, store_ptr, store_mode);
 664                                         ir_nodeset_insert(&pi->user_mem, irn);
 665                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 666                                         parallelise_load(pi, mem);
 667                                         return;
 668                                 }
 669                         }
 670                 } else if (is_Sync(irn)) {
 671                         int n = get_Sync_n_preds(irn);
 672                         int i;
 673
 674                         for (i = 0; i < n; ++i) {
 675                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 676                                 parallelise_load(pi, sync_pred);
 677                         }
 678                         return;
 679                 }
 680         }
 681         ir_nodeset_insert(&pi->this_mem, irn);
 682         //ir_fprintf(stderr, "adding %+F to this set\n", irn);
 683 }
 684
 685
 686 static void parallelise_store(parallelise_info *pi, ir_node *irn)
 687 {
 688         /* There is no point in investigating the same subgraph twice */
 689         if (ir_nodeset_contains(&pi->user_mem, irn))
 690                 return;
 691
 692         //ir_fprintf(stderr, "considering %+F\n", irn);
 693         if (get_nodes_block(irn) == pi->origin_block) {
 694                 if (is_Proj(irn)) {
 695                         ir_node *pred = get_Proj_pred(irn);
 696                         if (is_Load(pred) &&
 697                                         get_Load_volatility(pred) == volatility_non_volatile) {
 698                                 ir_mode *org_mode  = pi->origin_mode;
 699                                 ir_node *org_ptr   = pi->origin_ptr;
 700                                 ir_mode *load_mode = get_Load_mode(pred);
 701                                 ir_node *load_ptr  = get_Load_ptr(pred);
 702                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, load_ptr, load_mode) == ir_no_alias) {
 703                                         ir_node *mem = get_Load_mem(pred);
 704                                         //ir_fprintf(stderr, "St after Ld: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, load_ptr, load_mode);
 705                                         ir_nodeset_insert(&pi->user_mem, irn);
 706                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 707                                         parallelise_store(pi, mem);
 708                                         return;
 709                                 }
 710                         } else if (is_Store(pred) &&
 711                                         get_Store_volatility(pred) == volatility_non_volatile) {
 712                                 ir_mode *org_mode   = pi->origin_mode;
 713                                 ir_node *org_ptr    = pi->origin_ptr;
 714                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 715                                 ir_node *store_ptr  = get_Store_ptr(pred);
 716                                 if (get_alias_relation(current_ir_graph, org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
 717                                         ir_node *mem;
 718
 719                                         //ir_fprintf(stderr, "St after St: %+F (%+F) does not alias %+F (%+F)\n", org_ptr, org_mode, store_ptr, store_mode);
 720                                         ir_nodeset_insert(&pi->user_mem, irn);
 721                                         //ir_fprintf(stderr, "adding %+F to user set\n", irn);
 722                                         mem = get_Store_mem(pred);
 723                                         parallelise_store(pi, mem);
 724                                         return;
 725                                 }
 726                         }
 727                 } else if (is_Sync(irn)) {
 728                         int n = get_Sync_n_preds(irn);
 729                         int i;
 730
 731                         for (i = 0; i < n; ++i) {
 732                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 733                                 parallelise_store(pi, sync_pred);
 734                         }
 735                         return;
 736                 }
 737         }
 738         ir_nodeset_insert(&pi->this_mem, irn);
 739         //ir_fprintf(stderr, "adding %+F to this set\n", irn);
 740 }
 741
 742
 743 static void walker(ir_node *proj, void *env)
 744 {
 745         ir_node          *mem_op;
 746         ir_node          *pred;
 747         ir_node          *block;
 748         int               n;
 749         parallelise_info  pi;
 750
 751         (void)env;
 752
 753         if (!is_Proj(proj)) return;
 754         if (get_irn_mode(proj) != mode_M) return;
 755
 756         mem_op = get_Proj_pred(proj);
 757         if (is_Load(mem_op)) {
 758                 if (get_Load_volatility(mem_op) != volatility_non_volatile) return;
 759
 760                 block = get_nodes_block(mem_op);
 761                 pred  = get_Load_mem(mem_op);
 762                 //ir_fprintf(stderr, "starting parallelise at %+F for %+F\n", pred, proj);
 763
 764                 pi.origin_block = block,
 765                 pi.origin_ptr   = get_Load_ptr(mem_op);
 766                 pi.origin_mode  = get_Load_mode(mem_op);
 767                 ir_nodeset_init(&pi.this_mem);
 768                 ir_nodeset_init(&pi.user_mem);
 769
 770                 parallelise_load(&pi, pred);
 771         } else if (is_Store(mem_op)) {
 772                 if (get_Store_volatility(mem_op) != volatility_non_volatile) return;
 773
 774                 block = get_nodes_block(mem_op);
 775                 pred  = get_Store_mem(mem_op);
 776                 //ir_fprintf(stderr, "starting parallelise at %+F for %+F\n", pred, proj);
 777
 778                 pi.origin_block = block,
 779                 pi.origin_ptr   = get_Store_ptr(mem_op);
 780                 pi.origin_mode  = get_irn_mode(get_Store_value(mem_op));
 781                 ir_nodeset_init(&pi.this_mem);
 782                 ir_nodeset_init(&pi.user_mem);
 783
 784                 parallelise_store(&pi, pred);
 785         } else {
 786                 return;
 787         }
 788
 789         n = ir_nodeset_size(&pi.user_mem);
 790         if (n != 0) { /* nothing happened otherwise */
 791                 ir_graph               *irg  = current_ir_graph;
 792                 ir_node                *sync;
 793                 ir_node               **in;
 794                 ir_nodeset_iterator_t   iter;
 795                 int                     i;
 796
 797                 ++n;
 798                 //ir_fprintf(stderr, "creating sync for users of %+F with %d inputs\n", proj, n);
 799                 NEW_ARR_A(ir_node*, in, n);
 800                 i = 0;
 801                 in[i++] = new_r_Unknown(irg, mode_M);
 802                 ir_nodeset_iterator_init(&iter, &pi.user_mem);
 803                 for (;;) {
 804                         ir_node* p = ir_nodeset_iterator_next(&iter);
 805                         if (p == NULL) break;
 806                         in[i++] = p;
 807                 }
 808                 assert(i == n);
 809                 sync = new_r_Sync(block, n, in);
 810                 exchange(proj, sync);
 811
 812                 assert(pn_Load_M == pn_Store_M);
 813                 proj = new_r_Proj(block, mem_op, mode_M, pn_Load_M);
 814                 set_Sync_pred(sync, 0, proj);
 815
 816                 n = ir_nodeset_size(&pi.this_mem);
 817                 //ir_fprintf(stderr, "creating sync for %+F with %d inputs\n", mem_op, n);
 818                 ir_nodeset_iterator_init(&iter, &pi.this_mem);
 819                 if (n == 1) {
 820                         sync = ir_nodeset_iterator_next(&iter);
 821                 } else {
 822                         NEW_ARR_A(ir_node*, in, n);
 823                         i = 0;
 824                         for (;;) {
 825                                 ir_node* p = ir_nodeset_iterator_next(&iter);
 826                                 if (p == NULL) break;
 827                                 in[i++] = p;
 828                         }
 829                         assert(i == n);
 830                         sync = new_r_Sync(block, n, in);
 831                 }
 832                 set_memop_mem(mem_op, sync);
 833         }
 834
 835         ir_nodeset_destroy(&pi.this_mem);
 836         ir_nodeset_destroy(&pi.user_mem);
 837 }
 838
 839
 840 void opt_sync(ir_graph *irg)
 841 {
 842         //assure_irg_entity_usage_computed(irg);
 843         //assure_irp_globals_entity_usage_computed();
 844
 845         irg_walk_graph(irg, NULL, walker, NULL);
 846         //optimize_graph_df(irg);
 847         //irg_walk_graph(irg, NormaliseSync, NULL, NULL);
 848 }
 849
 850 ir_graph_pass_t *opt_sync_pass(const char *name, int verify, int dump)
 851 {
 852         return def_graph_pass(name ? name : "opt_sync", verify, dump, opt_sync);
 853 }