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

   1 /*
   2  * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  */
  25 #include "config.h"
  26
  27 #include <string.h>
  28
  29 #include "iroptimize.h"
  30 #include "irnode_t.h"
  31 #include "irgraph_t.h"
  32 #include "irmode_t.h"
  33 #include "iropt_t.h"
  34 #include "ircons_t.h"
  35 #include "irgmod.h"
  36 #include "irgwalk.h"
  37 #include "irtools.h"
  38 #include "tv_t.h"
  39 #include "dbginfo_t.h"
  40 #include "iropt_dbg.h"
  41 #include "irflag_t.h"
  42 #include "array_t.h"
  43 #include "irhooks.h"
  44 #include "iredges.h"
  45 #include "irpass.h"
  46 #include "irmemory.h"
  47 #include "irnodehashmap.h"
  48 #include "irgopt.h"
  49 #include "set.h"
  50 #include "be.h"
  51 #include "debug.h"
  52
  53 /** The debug handle. */
  54 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  55
  56 #undef IMAX
  57 #define IMAX(a,b)   ((a) > (b) ? (a) : (b))
  58
  59 #define MAX_PROJ    IMAX(IMAX((long)pn_Load_max, (long)pn_Store_max), (long)pn_Call_max)
  60
  61 enum changes_t {
  62         DF_CHANGED = 1,       /**< data flow changed */
  63         CF_CHANGED = 2,       /**< control flow changed */
  64 };
  65
  66 /**
  67  * walker environment
  68  */
  69 typedef struct walk_env_t {
  70         struct obstack obst;          /**< list of all stores */
  71         unsigned changes;             /**< a bitmask of graph changes */
  72 } walk_env_t;
  73
  74 /** A Load/Store info. */
  75 typedef struct ldst_info_t {
  76         ir_node  *projs[MAX_PROJ+1];  /**< list of Proj's of this node */
  77         ir_node  *exc_block;          /**< the exception block if available */
  78         int      exc_idx;             /**< predecessor index in the exception block */
  79         unsigned visited;             /**< visited counter for breaking loops */
  80 } ldst_info_t;
  81
  82 /**
  83  * flags for control flow.
  84  */
  85 enum block_flags_t {
  86         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  87         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  88 };
  89
  90 /**
  91  * a Block info.
  92  */
  93 typedef struct block_info_t {
  94         unsigned flags;               /**< flags for the block */
  95 } block_info_t;
  96
  97 /** the master visited flag for loop detection. */
  98 static unsigned master_visited = 0;
  99
 100 #define INC_MASTER()       ++master_visited
 101 #define MARK_NODE(info)    (info)->visited = master_visited
 102 #define NODE_VISITED(info) (info)->visited >= master_visited
 103
 104 /**
 105  * get the Load/Store info of a node
 106  */
 107 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
 108 {
 109         ldst_info_t *info = (ldst_info_t*)get_irn_link(node);
 110
 111         if (! info) {
 112                 info = OALLOCZ(obst, ldst_info_t);
 113                 set_irn_link(node, info);
 114         }
 115         return info;
 116 }  /* get_ldst_info */
 117
 118 /**
 119  * get the Block info of a node
 120  */
 121 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
 122 {
 123         block_info_t *info = (block_info_t*)get_irn_link(node);
 124
 125         if (! info) {
 126                 info = OALLOCZ(obst, block_info_t);
 127                 set_irn_link(node, info);
 128         }
 129         return info;
 130 }  /* get_block_info */
 131
 132 /**
 133  * update the projection info for a Load/Store
 134  */
 135 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 136 {
 137         long nr = get_Proj_proj(proj);
 138
 139         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 140
 141         if (info->projs[nr]) {
 142                 /* there is already one, do CSE */
 143                 exchange(proj, info->projs[nr]);
 144                 return DF_CHANGED;
 145         }
 146         else {
 147                 info->projs[nr] = proj;
 148                 return 0;
 149         }
 150 }  /* update_projs */
 151
 152 /**
 153  * update the exception block info for a Load/Store node.
 154  *
 155  * @param info   the load/store info struct
 156  * @param block  the exception handler block for this load/store
 157  * @param pos    the control flow input of the block
 158  */
 159 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 160 {
 161         assert(info->exc_block == NULL && "more than one exception block found");
 162
 163         info->exc_block = block;
 164         info->exc_idx   = pos;
 165         return 0;
 166 }  /* update_exc */
 167
 168 /**
 169  * walker, collects all Load/Store/Proj nodes
 170  *
 171  * walks from Start -> End
 172  */
 173 static void collect_nodes(ir_node *node, void *env)
 174 {
 175         walk_env_t  *wenv   = (walk_env_t *)env;
 176         unsigned     opcode = get_irn_opcode(node);
 177         ir_node     *pred, *blk, *pred_blk;
 178         ldst_info_t *ldst_info;
 179
 180         if (opcode == iro_Proj) {
 181                 pred   = get_Proj_pred(node);
 182                 opcode = get_irn_opcode(pred);
 183
 184                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 185                         ldst_info = get_ldst_info(pred, &wenv->obst);
 186
 187                         wenv->changes |= update_projs(ldst_info, node);
 188
 189                         /*
 190                          * Place the Proj's to the same block as the
 191                          * predecessor Load. This is always ok and prevents
 192                          * "non-SSA" form after optimizations if the Proj
 193                          * is in a wrong block.
 194                          */
 195                         blk      = get_nodes_block(node);
 196                         pred_blk = get_nodes_block(pred);
 197                         if (blk != pred_blk) {
 198                                 wenv->changes |= DF_CHANGED;
 199                                 set_nodes_block(node, pred_blk);
 200                         }
 201                 }
 202         } else if (opcode == iro_Block) {
 203                 int i;
 204
 205                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 206                         ir_node      *pred_block, *proj;
 207                         block_info_t *bl_info;
 208                         int          is_exc = 0;
 209
 210                         pred = proj = get_Block_cfgpred(node, i);
 211
 212                         if (is_Proj(proj)) {
 213                                 pred   = get_Proj_pred(proj);
 214                                 is_exc = is_x_except_Proj(proj);
 215                         }
 216
 217                         /* ignore Bad predecessors, they will be removed later */
 218                         if (is_Bad(pred))
 219                                 continue;
 220
 221                         pred_block = get_nodes_block(pred);
 222                         bl_info    = get_block_info(pred_block, &wenv->obst);
 223
 224                         if (is_fragile_op(pred) && is_exc)
 225                                 bl_info->flags |= BLOCK_HAS_EXC;
 226                         else if (is_irn_forking(pred))
 227                                 bl_info->flags |= BLOCK_HAS_COND;
 228
 229                         opcode = get_irn_opcode(pred);
 230                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 231                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 232
 233                                 wenv->changes |= update_exc(ldst_info, node, i);
 234                         }
 235                 }
 236         }
 237 }  /* collect_nodes */
 238
 239 /**
 240  * Returns an entity if the address ptr points to a constant one.
 241  *
 242  * @param ptr  the address
 243  *
 244  * @return an entity or NULL
 245  */
 246 static ir_entity *find_constant_entity(ir_node *ptr)
 247 {
 248         for (;;) {
 249                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 250                         return get_SymConst_entity(ptr);
 251                 } else if (is_Sel(ptr)) {
 252                         ir_entity *ent = get_Sel_entity(ptr);
 253                         ir_type   *tp  = get_entity_owner(ent);
 254
 255                         /* Do not fiddle with polymorphism. */
 256                         if (is_Class_type(get_entity_owner(ent)) &&
 257                                 ((get_entity_n_overwrites(ent)    != 0) ||
 258                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 259                                 return NULL;
 260
 261                         if (is_Array_type(tp)) {
 262                                 /* check bounds */
 263                                 int i, n;
 264
 265                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 266                                         ir_node   *bound;
 267                                         ir_tarval *tlower, *tupper;
 268                                         ir_node   *index = get_Sel_index(ptr, i);
 269                                         ir_tarval *tv    = computed_value(index);
 270
 271                                         /* check if the index is constant */
 272                                         if (tv == tarval_bad)
 273                                                 return NULL;
 274
 275                                         bound  = get_array_lower_bound(tp, i);
 276                                         tlower = computed_value(bound);
 277                                         bound  = get_array_upper_bound(tp, i);
 278                                         tupper = computed_value(bound);
 279
 280                                         if (tlower == tarval_bad || tupper == tarval_bad)
 281                                                 return NULL;
 282
 283                                         if (tarval_cmp(tv, tlower) == ir_relation_less)
 284                                                 return NULL;
 285                                         if (tarval_cmp(tupper, tv) == ir_relation_less)
 286                                                 return NULL;
 287
 288                                         /* ok, bounds check finished */
 289                                 }
 290                         }
 291
 292                         if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
 293                                 return ent;
 294
 295                         /* try next */
 296                         ptr = get_Sel_ptr(ptr);
 297                 } else if (is_Add(ptr)) {
 298                         ir_node *l = get_Add_left(ptr);
 299                         ir_node *r = get_Add_right(ptr);
 300
 301                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 302                                 ptr = l;
 303                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 304                                 ptr = r;
 305                         else
 306                                 return NULL;
 307
 308                         /* for now, we support only one addition, reassoc should fold all others */
 309                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 310                                 return NULL;
 311                 } else if (is_Sub(ptr)) {
 312                         ir_node *l = get_Sub_left(ptr);
 313                         ir_node *r = get_Sub_right(ptr);
 314
 315                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 316                                 ptr = l;
 317                         else
 318                                 return NULL;
 319                         /* for now, we support only one substraction, reassoc should fold all others */
 320                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 321                                 return NULL;
 322                 } else
 323                         return NULL;
 324         }
 325 }  /* find_constant_entity */
 326
 327 /**
 328  * Return the Selection index of a Sel node from dimension n
 329  */
 330 static long get_Sel_array_index_long(ir_node *n, int dim)
 331 {
 332         ir_node *index = get_Sel_index(n, dim);
 333         assert(is_Const(index));
 334         return get_tarval_long(get_Const_tarval(index));
 335 }  /* get_Sel_array_index_long */
 336
 337 typedef struct path_entry {
 338         ir_entity         *ent;
 339         struct path_entry *next;
 340         size_t            index;
 341 } path_entry;
 342
 343 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
 344 {
 345         path_entry       entry, *p;
 346         ir_entity        *ent, *field;
 347         ir_initializer_t *initializer;
 348         ir_tarval        *tv;
 349         ir_type          *tp;
 350         size_t           n;
 351
 352         entry.next = next;
 353         if (is_SymConst(ptr)) {
 354                 /* found the root */
 355                 ent         = get_SymConst_entity(ptr);
 356                 initializer = get_entity_initializer(ent);
 357                 for (p = next; p != NULL;) {
 358                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 359                                 return NULL;
 360                         n  = get_initializer_compound_n_entries(initializer);
 361                         tp = get_entity_type(ent);
 362
 363                         if (is_Array_type(tp)) {
 364                                 ent = get_array_element_entity(tp);
 365                                 if (ent != p->ent) {
 366                                         /* a missing [0] */
 367                                         if (0 >= n)
 368                                                 return NULL;
 369                                         initializer = get_initializer_compound_value(initializer, 0);
 370                                         continue;
 371                                 }
 372                         }
 373                         if (p->index >= n)
 374                                 return NULL;
 375                         initializer = get_initializer_compound_value(initializer, p->index);
 376
 377                         ent = p->ent;
 378                         p   = p->next;
 379                 }
 380                 tp = get_entity_type(ent);
 381                 while (is_Array_type(tp)) {
 382                         ent = get_array_element_entity(tp);
 383                         tp = get_entity_type(ent);
 384                         /* a missing [0] */
 385                         n  = get_initializer_compound_n_entries(initializer);
 386                         if (0 >= n)
 387                                 return NULL;
 388                         initializer = get_initializer_compound_value(initializer, 0);
 389                 }
 390
 391                 switch (initializer->kind) {
 392                 case IR_INITIALIZER_CONST:
 393                         return get_initializer_const_value(initializer);
 394                 case IR_INITIALIZER_TARVAL:
 395                 case IR_INITIALIZER_NULL:
 396                 default:
 397                         return NULL;
 398                 }
 399         } else if (is_Sel(ptr)) {
 400                 entry.ent = field = get_Sel_entity(ptr);
 401                 tp = get_entity_owner(field);
 402                 if (is_Array_type(tp)) {
 403                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 404                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 405                 } else {
 406                         size_t i, n_members = get_compound_n_members(tp);
 407                         for (i = 0; i < n_members; ++i) {
 408                                 if (get_compound_member(tp, i) == field)
 409                                         break;
 410                         }
 411                         if (i >= n_members) {
 412                                 /* not found: should NOT happen */
 413                                 return NULL;
 414                         }
 415                         entry.index = i;
 416                 }
 417                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 418         }  else if (is_Add(ptr)) {
 419                 ir_mode  *mode;
 420                 unsigned pos;
 421
 422                 {
 423                         ir_node *l = get_Add_left(ptr);
 424                         ir_node *r = get_Add_right(ptr);
 425                         if (is_Const(r)) {
 426                                 ptr = l;
 427                                 tv  = get_Const_tarval(r);
 428                         } else {
 429                                 ptr = r;
 430                                 tv  = get_Const_tarval(l);
 431                         }
 432                 }
 433 ptr_arith:
 434                 mode = get_tarval_mode(tv);
 435
 436                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 437                 if (is_Sel(ptr)) {
 438                         field = get_Sel_entity(ptr);
 439                 } else {
 440                         field = get_SymConst_entity(ptr);
 441                 }
 442
 443                 /* count needed entries */
 444                 pos = 0;
 445                 for (ent = field;;) {
 446                         tp = get_entity_type(ent);
 447                         if (! is_Array_type(tp))
 448                                 break;
 449                         ent = get_array_element_entity(tp);
 450                         ++pos;
 451                 }
 452                 /* should be at least ONE entry */
 453                 if (pos == 0)
 454                         return NULL;
 455
 456                 /* allocate the right number of entries */
 457                 NEW_ARR_A(path_entry, p, pos);
 458
 459                 /* fill them up */
 460                 pos = 0;
 461                 for (ent = field;;) {
 462                         unsigned   size;
 463                         ir_tarval *sz, *tv_index, *tlower, *tupper;
 464                         long       index;
 465                         ir_node   *bound;
 466
 467                         tp = get_entity_type(ent);
 468                         if (! is_Array_type(tp))
 469                                 break;
 470                         ent = get_array_element_entity(tp);
 471                         p[pos].ent  = ent;
 472                         p[pos].next = &p[pos + 1];
 473
 474                         size = get_type_size_bytes(get_entity_type(ent));
 475                         sz   = new_tarval_from_long(size, mode);
 476
 477                         tv_index = tarval_div(tv, sz);
 478                         tv       = tarval_mod(tv, sz);
 479
 480                         if (tv_index == tarval_bad || tv == tarval_bad)
 481                                 return NULL;
 482
 483                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 484                         bound  = get_array_lower_bound(tp, 0);
 485                         tlower = computed_value(bound);
 486                         bound  = get_array_upper_bound(tp, 0);
 487                         tupper = computed_value(bound);
 488
 489                         if (tlower == tarval_bad || tupper == tarval_bad)
 490                                 return NULL;
 491
 492                         if (tarval_cmp(tv_index, tlower) == ir_relation_less)
 493                                 return NULL;
 494                         if (tarval_cmp(tupper, tv_index) == ir_relation_less)
 495                                 return NULL;
 496
 497                         /* ok, bounds check finished */
 498                         index = get_tarval_long(tv_index);
 499                         p[pos].index = index;
 500                         ++pos;
 501                 }
 502                 if (! tarval_is_null(tv)) {
 503                         /* hmm, wrong access */
 504                         return NULL;
 505                 }
 506                 p[pos - 1].next = next;
 507                 return rec_find_compound_ent_value(ptr, p);
 508         } else if (is_Sub(ptr)) {
 509                 ir_node *l = get_Sub_left(ptr);
 510                 ir_node *r = get_Sub_right(ptr);
 511
 512                 ptr = l;
 513                 tv  = get_Const_tarval(r);
 514                 tv  = tarval_neg(tv);
 515                 goto ptr_arith;
 516         }
 517         return NULL;
 518 }
 519
 520 static ir_node *find_compound_ent_value(ir_node *ptr)
 521 {
 522         return rec_find_compound_ent_value(ptr, NULL);
 523 }
 524
 525 /* forward */
 526 static void reduce_adr_usage(ir_node *ptr);
 527
 528 /**
 529  * Update a Load that may have lost its users.
 530  */
 531 static void handle_load_update(ir_node *load)
 532 {
 533         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
 534
 535         /* do NOT touch volatile loads for now */
 536         if (get_Load_volatility(load) == volatility_is_volatile)
 537                 return;
 538
 539         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 540                 ir_node *ptr = get_Load_ptr(load);
 541                 ir_node *mem = get_Load_mem(load);
 542
 543                 /* a Load whose value is neither used nor exception checked, remove it */
 544                 exchange(info->projs[pn_Load_M], mem);
 545                 if (info->projs[pn_Load_X_regular])
 546                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 547                 kill_node(load);
 548                 reduce_adr_usage(ptr);
 549         }
 550 }  /* handle_load_update */
 551
 552 /**
 553  * A use of an address node has vanished. Check if this was a Proj
 554  * node and update the counters.
 555  */
 556 static void reduce_adr_usage(ir_node *ptr)
 557 {
 558         ir_node *pred;
 559         if (!is_Proj(ptr))
 560                 return;
 561         if (get_irn_n_edges(ptr) > 0)
 562                 return;
 563
 564         /* this Proj is dead now */
 565         pred = get_Proj_pred(ptr);
 566         if (is_Load(pred)) {
 567                 ldst_info_t *info = (ldst_info_t*)get_irn_link(pred);
 568                 info->projs[get_Proj_proj(ptr)] = NULL;
 569
 570                 /* this node lost its result proj, handle that */
 571                 handle_load_update(pred);
 572         }
 573 }  /* reduce_adr_usage */
 574
 575 /**
 576  * Check, if an already existing value of mode old_mode can be converted
 577  * into the needed one new_mode without loss.
 578  */
 579 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
 580 {
 581         unsigned old_size;
 582         unsigned new_size;
 583         if (old_mode == new_mode)
 584                 return true;
 585
 586         old_size = get_mode_size_bits(old_mode);
 587         new_size = get_mode_size_bits(new_mode);
 588
 589         /* if both modes are two-complement ones, we can always convert the
 590            Stored value into the needed one. (on big endian machines we currently
 591            only support this for modes of same size) */
 592         if (old_size >= new_size &&
 593                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 594                   get_mode_arithmetic(new_mode) == irma_twos_complement &&
 595                   (!be_get_backend_param()->byte_order_big_endian
 596                 || old_size == new_size)) {
 597                 return true;
 598         }
 599         return false;
 600 }
 601
 602 /**
 603  * Check whether a Call is at least pure, i.e. does only read memory.
 604  */
 605 static unsigned is_Call_pure(ir_node *call)
 606 {
 607         ir_type *call_tp = get_Call_type(call);
 608         unsigned prop = get_method_additional_properties(call_tp);
 609
 610         /* check first the call type */
 611         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 612                 /* try the called entity */
 613                 ir_node *ptr = get_Call_ptr(call);
 614
 615                 if (is_SymConst_addr_ent(ptr)) {
 616                         ir_entity *ent = get_SymConst_entity(ptr);
 617
 618                         prop = get_entity_additional_properties(ent);
 619                 }
 620         }
 621         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 622 }  /* is_Call_pure */
 623
 624 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 625 {
 626         ir_mode *mode  = get_irn_mode(ptr);
 627         long    offset = 0;
 628
 629         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 630         for (;;) {
 631                 if (is_Add(ptr)) {
 632                         ir_node *l = get_Add_left(ptr);
 633                         ir_node *r = get_Add_right(ptr);
 634
 635                         if (get_irn_mode(l) != mode || !is_Const(r))
 636                                 break;
 637
 638                         offset += get_tarval_long(get_Const_tarval(r));
 639                         ptr     = l;
 640                 } else if (is_Sub(ptr)) {
 641                         ir_node *l = get_Sub_left(ptr);
 642                         ir_node *r = get_Sub_right(ptr);
 643
 644                         if (get_irn_mode(l) != mode || !is_Const(r))
 645                                 break;
 646
 647                         offset -= get_tarval_long(get_Const_tarval(r));
 648                         ptr     = l;
 649                 } else if (is_Sel(ptr)) {
 650                         ir_entity *ent = get_Sel_entity(ptr);
 651                         ir_type   *tp  = get_entity_owner(ent);
 652
 653                         if (is_Array_type(tp)) {
 654                                 int     size;
 655                                 ir_node *index;
 656
 657                                 /* only one dimensional arrays yet */
 658                                 if (get_Sel_n_indexs(ptr) != 1)
 659                                         break;
 660                                 index = get_Sel_index(ptr, 0);
 661                                 if (! is_Const(index))
 662                                         break;
 663
 664                                 tp = get_entity_type(ent);
 665                                 if (get_type_state(tp) != layout_fixed)
 666                                         break;
 667
 668                                 size    = get_type_size_bytes(tp);
 669                                 offset += size * get_tarval_long(get_Const_tarval(index));
 670                         } else {
 671                                 if (get_type_state(tp) != layout_fixed)
 672                                         break;
 673                                 offset += get_entity_offset(ent);
 674                         }
 675                         ptr = get_Sel_ptr(ptr);
 676                 } else
 677                         break;
 678         }
 679
 680         *pOffset = offset;
 681         return ptr;
 682 }
 683
 684 static int try_load_after_store(ir_node *load,
 685                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 686 {
 687         ldst_info_t *info;
 688         ir_node *store_ptr      = get_Store_ptr(store);
 689         long     store_offset;
 690         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 691         ir_node *store_value;
 692         ir_mode *store_mode;
 693         ir_node *load_ptr;
 694         ir_mode *load_mode;
 695         long     load_mode_len;
 696         long     store_mode_len;
 697         long     delta;
 698         int      res;
 699
 700         if (load_base_ptr != store_base_ptr)
 701                 return 0;
 702
 703         load_mode      = get_Load_mode(load);
 704         load_mode_len  = get_mode_size_bytes(load_mode);
 705         store_mode     = get_irn_mode(get_Store_value(store));
 706         store_mode_len = get_mode_size_bytes(store_mode);
 707         delta          = load_offset - store_offset;
 708         store_value    = get_Store_value(store);
 709
 710         if (delta != 0 || store_mode != load_mode) {
 711                 /* TODO: implement for big-endian */
 712                 if (delta < 0 || delta + load_mode_len > store_mode_len
 713                                 || (be_get_backend_param()->byte_order_big_endian
 714                                     && load_mode_len != store_mode_len))
 715                         return 0;
 716
 717                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 718                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 719                         return 0;
 720
 721
 722                 /* produce a shift to adjust offset delta */
 723                 if (delta > 0) {
 724                         ir_node *cnst;
 725                         ir_graph *irg = get_irn_irg(load);
 726
 727                         cnst        = new_r_Const_long(irg, mode_Iu, delta * 8);
 728                         store_value = new_r_Shr(get_nodes_block(load),
 729                                                                         store_value, cnst, store_mode);
 730                 }
 731
 732                 /* add an convert if needed */
 733                 if (store_mode != load_mode) {
 734                         store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
 735                 }
 736         }
 737
 738         DBG_OPT_RAW(load, store_value);
 739
 740         info = (ldst_info_t*)get_irn_link(load);
 741         if (info->projs[pn_Load_M])
 742                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 743
 744         res = 0;
 745         /* no exception */
 746         if (info->projs[pn_Load_X_except]) {
 747                 ir_graph *irg = get_irn_irg(load);
 748                 exchange( info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
 749                 res |= CF_CHANGED;
 750         }
 751         if (info->projs[pn_Load_X_regular]) {
 752                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 753                 res |= CF_CHANGED;
 754         }
 755
 756         if (info->projs[pn_Load_res])
 757                 exchange(info->projs[pn_Load_res], store_value);
 758
 759         load_ptr = get_Load_ptr(load);
 760         kill_node(load);
 761         reduce_adr_usage(load_ptr);
 762         return res | DF_CHANGED;
 763 }
 764
 765 /**
 766  * Follow the memory chain as long as there are only Loads,
 767  * alias free Stores, and constant Calls and try to replace the
 768  * current Load by a previous ones.
 769  * Note that in unreachable loops it might happen that we reach
 770  * load again, as well as we can fall into a cycle.
 771  * We break such cycles using a special visited flag.
 772  *
 773  * INC_MASTER() must be called before dive into
 774  */
 775 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
 776 {
 777         unsigned    res = 0;
 778         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
 779         ir_node     *pred;
 780         ir_node     *ptr       = get_Load_ptr(load);
 781         ir_node     *mem       = get_Load_mem(load);
 782         ir_mode     *load_mode = get_Load_mode(load);
 783
 784         for (pred = curr; load != pred; ) {
 785                 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
 786
 787                 /*
 788                  * a Load immediately after a Store -- a read after write.
 789                  * We may remove the Load, if both Load & Store does not have an
 790                  * exception handler OR they are in the same Block. In the latter
 791                  * case the Load cannot throw an exception when the previous Store was
 792                  * quiet.
 793                  *
 794                  * Why we need to check for Store Exception? If the Store cannot
 795                  * be executed (ROM) the exception handler might simply jump into
 796                  * the load Block :-(
 797                  * We could make it a little bit better if we would know that the
 798                  * exception handler of the Store jumps directly to the end...
 799                  */
 800                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 801                                 && info->projs[pn_Load_X_except] == NULL)
 802                                 || get_nodes_block(load) == get_nodes_block(pred)))
 803                 {
 804                         long    load_offset;
 805                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 806                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 807
 808                         if (changes != 0)
 809                                 return res | changes;
 810                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 811                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 812                         /*
 813                          * a Load after a Load -- a read after read.
 814                          * We may remove the second Load, if it does not have an exception
 815                          * handler OR they are in the same Block. In the later case
 816                          * the Load cannot throw an exception when the previous Load was
 817                          * quiet.
 818                          *
 819                          * Here, there is no need to check if the previous Load has an
 820                          * exception hander because they would have exact the same
 821                          * exception...
 822                          *
 823                          * TODO: implement load-after-load with different mode for big
 824                          *       endian
 825                          */
 826                         if (info->projs[pn_Load_X_except] == NULL
 827                                         || get_nodes_block(load) == get_nodes_block(pred)) {
 828                                 ir_node *value;
 829
 830                                 DBG_OPT_RAR(load, pred);
 831
 832                                 /* the result is used */
 833                                 if (info->projs[pn_Load_res]) {
 834                                         if (pred_info->projs[pn_Load_res] == NULL) {
 835                                                 /* create a new Proj again */
 836                                                 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
 837                                         }
 838                                         value = pred_info->projs[pn_Load_res];
 839
 840                                         /* add an convert if needed */
 841                                         if (get_Load_mode(pred) != load_mode) {
 842                                                 value = new_r_Conv(get_nodes_block(load), value, load_mode);
 843                                         }
 844
 845                                         exchange(info->projs[pn_Load_res], value);
 846                                 }
 847
 848                                 if (info->projs[pn_Load_M])
 849                                         exchange(info->projs[pn_Load_M], mem);
 850
 851                                 /* no exception */
 852                                 if (info->projs[pn_Load_X_except]) {
 853                                         ir_graph *irg = get_irn_irg(load);
 854                                         exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
 855                                         res |= CF_CHANGED;
 856                                 }
 857                                 if (info->projs[pn_Load_X_regular]) {
 858                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 859                                         res |= CF_CHANGED;
 860                                 }
 861
 862                                 kill_node(load);
 863                                 reduce_adr_usage(ptr);
 864                                 return res |= DF_CHANGED;
 865                         }
 866                 }
 867
 868                 if (is_Store(pred)) {
 869                         /* check if we can pass through this store */
 870                         ir_alias_relation rel = get_alias_relation(
 871                                 get_Store_ptr(pred),
 872                                 get_irn_mode(get_Store_value(pred)),
 873                                 ptr, load_mode);
 874                         /* if the might be an alias, we cannot pass this Store */
 875                         if (rel != ir_no_alias)
 876                                 break;
 877                         pred = skip_Proj(get_Store_mem(pred));
 878                 } else if (is_Load(pred)) {
 879                         pred = skip_Proj(get_Load_mem(pred));
 880                 } else if (is_Call(pred)) {
 881                         if (is_Call_pure(pred)) {
 882                                 /* The called graph is at least pure, so there are no Store's
 883                                    in it. We can handle it like a Load and skip it. */
 884                                 pred = skip_Proj(get_Call_mem(pred));
 885                         } else {
 886                                 /* there might be Store's in the graph, stop here */
 887                                 break;
 888                         }
 889                 } else {
 890                         /* follow only Load chains */
 891                         break;
 892                 }
 893
 894                 /* check for cycles */
 895                 if (NODE_VISITED(pred_info))
 896                         break;
 897                 MARK_NODE(pred_info);
 898         }
 899
 900         if (is_Sync(pred)) {
 901                 int i;
 902
 903                 /* handle all Sync predecessors */
 904                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 905                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 906                         if (res)
 907                                 return res;
 908                 }
 909         }
 910
 911         return res;
 912 }  /* follow_Mem_chain */
 913
 914 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
 915 {
 916         ir_mode  *c_mode = get_irn_mode(c);
 917         ir_mode  *l_mode = get_Load_mode(load);
 918         ir_node  *block  = get_nodes_block(load);
 919         dbg_info *dbgi   = get_irn_dbg_info(load);
 920         ir_node  *res    = copy_const_value(dbgi, c, block);
 921
 922         if (c_mode != l_mode) {
 923                 /* check, if the mode matches OR can be easily converted info */
 924                 if (is_reinterpret_cast(c_mode, l_mode)) {
 925                         /* copy the value from the const code irg and cast it */
 926                         res = new_rd_Conv(dbgi, block, res, l_mode);
 927                 } else {
 928                         return NULL;
 929                 }
 930         }
 931         return res;
 932 }
 933
 934 /**
 935  * optimize a Load
 936  *
 937  * @param load  the Load node
 938  */
 939 static unsigned optimize_load(ir_node *load)
 940 {
 941         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
 942         ir_node     *mem, *ptr, *value;
 943         ir_entity   *ent;
 944         long        dummy;
 945         unsigned    res = 0;
 946
 947         /* do NOT touch volatile loads for now */
 948         if (get_Load_volatility(load) == volatility_is_volatile)
 949                 return 0;
 950
 951         /* the address of the load to be optimized */
 952         ptr = get_Load_ptr(load);
 953
 954         /* The mem of the Load. Must still be returned after optimization. */
 955         mem = get_Load_mem(load);
 956
 957         if (info->projs[pn_Load_res] == NULL
 958                         && info->projs[pn_Load_X_except] == NULL) {
 959                 /* the value is never used and we don't care about exceptions, remove */
 960                 exchange(info->projs[pn_Load_M], mem);
 961
 962                 if (info->projs[pn_Load_X_regular]) {
 963                         /* should not happen, but if it does, remove it */
 964                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 965                         res |= CF_CHANGED;
 966                 }
 967                 kill_node(load);
 968                 reduce_adr_usage(ptr);
 969                 return res | DF_CHANGED;
 970         }
 971
 972         value = NULL;
 973         /* check if we can determine the entity that will be loaded */
 974         ent = find_constant_entity(ptr);
 975         if (ent != NULL
 976                         && get_entity_visibility(ent) != ir_visibility_external) {
 977                 /* a static allocation that is not external: there should be NO
 978                  * exception when loading even if we cannot replace the load itself.
 979                  */
 980
 981                 /* no exception, clear the info field as it might be checked later again */
 982                 if (info->projs[pn_Load_X_except]) {
 983                         ir_graph *irg = get_irn_irg(load);
 984                         exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
 985                         info->projs[pn_Load_X_except] = NULL;
 986                         res |= CF_CHANGED;
 987                 }
 988                 if (info->projs[pn_Load_X_regular]) {
 989                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 990                         info->projs[pn_Load_X_regular] = NULL;
 991                         res |= CF_CHANGED;
 992                 }
 993
 994                 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
 995                         if (has_entity_initializer(ent)) {
 996                                 /* new style initializer */
 997                                 value = find_compound_ent_value(ptr);
 998                         }
 999                         if (value != NULL) {
1000                                 ir_graph *irg = get_irn_irg(load);
1001                                 value = can_replace_load_by_const(load, value);
1002                                 if (value != NULL && is_Sel(ptr)) {
1003                                         /* frontend has inserted masking operations after bitfield accesses,
1004                                          * so we might have to shift the const. */
1005                                         unsigned char bit_offset = get_entity_offset_bits_remainder(get_Sel_entity(ptr));
1006                                         if (bit_offset != 0) {
1007                                                 if (is_Const(value)) {
1008                                                         ir_tarval *tv_old = get_Const_tarval(value);
1009                                                         ir_tarval *tv_offset = new_tarval_from_long(bit_offset, mode_Bu);
1010                                                         ir_tarval *tv_new = tarval_shl(tv_old, tv_offset);
1011                                                         value = new_r_Const(irg, tv_new);
1012                                                 } else {
1013                                                         value = NULL;
1014                                                 }
1015                                         }
1016                                 }
1017                         }
1018                 }
1019         }
1020         if (value != NULL) {
1021                 /* we completely replace the load by this value */
1022                 if (info->projs[pn_Load_X_except]) {
1023                         ir_graph *irg = get_irn_irg(load);
1024                         exchange(info->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
1025                         info->projs[pn_Load_X_except] = NULL;
1026                         res |= CF_CHANGED;
1027                 }
1028                 if (info->projs[pn_Load_X_regular]) {
1029                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1030                         info->projs[pn_Load_X_regular] = NULL;
1031                         res |= CF_CHANGED;
1032                 }
1033                 if (info->projs[pn_Load_M]) {
1034                         exchange(info->projs[pn_Load_M], mem);
1035                         res |= DF_CHANGED;
1036                 }
1037                 if (info->projs[pn_Load_res]) {
1038                         exchange(info->projs[pn_Load_res], value);
1039                         res |= DF_CHANGED;
1040                 }
1041                 kill_node(load);
1042                 reduce_adr_usage(ptr);
1043                 return res;
1044         }
1045
1046         /* Check, if the address of this load is used more than once.
1047          * If not, more load cannot be removed in any case. */
1048         if (get_irn_n_edges(ptr) <= 1 && get_irn_n_edges(get_base_and_offset(ptr, &dummy)) <= 1)
1049                 return res;
1050
1051         /*
1052          * follow the memory chain as long as there are only Loads
1053          * and try to replace current Load or Store by a previous one.
1054          * Note that in unreachable loops it might happen that we reach
1055          * load again, as well as we can fall into a cycle.
1056          * We break such cycles using a special visited flag.
1057          */
1058         INC_MASTER();
1059         res = follow_Mem_chain(load, skip_Proj(mem));
1060         return res;
1061 }  /* optimize_load */
1062
1063 /**
1064  * Check whether a value of mode new_mode would completely overwrite a value
1065  * of mode old_mode in memory.
1066  */
1067 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1068 {
1069         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1070 }  /* is_completely_overwritten */
1071
1072 /**
1073  * Check whether small is a part of large (starting at same address).
1074  */
1075 static int is_partially_same(ir_node *small, ir_node *large)
1076 {
1077         ir_mode *sm = get_irn_mode(small);
1078         ir_mode *lm = get_irn_mode(large);
1079
1080         /* FIXME: Check endianness */
1081         return is_Conv(small) && get_Conv_op(small) == large
1082             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1083             && get_mode_arithmetic(sm) == irma_twos_complement
1084             && get_mode_arithmetic(lm) == irma_twos_complement;
1085 }  /* is_partially_same */
1086
1087 /**
1088  * follow the memory chain as long as there are only Loads and alias free Stores.
1089  *
1090  * INC_MASTER() must be called before dive into
1091  */
1092 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1093 {
1094         unsigned res = 0;
1095         ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1096         ir_node *pred;
1097         ir_node *ptr = get_Store_ptr(store);
1098         ir_node *mem = get_Store_mem(store);
1099         ir_node *value = get_Store_value(store);
1100         ir_mode *mode  = get_irn_mode(value);
1101         ir_node *block = get_nodes_block(store);
1102
1103         for (pred = curr; pred != store;) {
1104                 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
1105
1106                 /*
1107                  * BEWARE: one might think that checking the modes is useless, because
1108                  * if the pointers are identical, they refer to the same object.
1109                  * This is only true in strong typed languages, not is C were the following
1110                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1111                  * However, if the size of the mode that is written is bigger or equal the
1112                  * size of the old one, the old value is completely overwritten and can be
1113                  * killed ...
1114                  */
1115                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1116             get_nodes_block(pred) == block) {
1117                         /*
1118                          * a Store after a Store in the same Block -- a write after write.
1119                          */
1120
1121                         /*
1122                          * We may remove the first Store, if the old value is completely
1123                          * overwritten or the old value is a part of the new value,
1124                          * and if it does not have an exception handler.
1125                          *
1126                          * TODO: What, if both have the same exception handler ???
1127                          */
1128                         if (get_Store_volatility(pred) != volatility_is_volatile
1129                                 && !pred_info->projs[pn_Store_X_except]) {
1130                                 ir_node *predvalue = get_Store_value(pred);
1131                                 ir_mode *predmode  = get_irn_mode(predvalue);
1132
1133                                 if (is_completely_overwritten(predmode, mode)
1134                                         || is_partially_same(predvalue, value)) {
1135                                         DBG_OPT_WAW(pred, store);
1136                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1137                                         kill_node(pred);
1138                                         reduce_adr_usage(ptr);
1139                                         return DF_CHANGED;
1140                                 }
1141                         }
1142
1143                         /*
1144                          * We may remove the Store, if the old value already contains
1145                          * the new value, and if it does not have an exception handler.
1146                          *
1147                          * TODO: What, if both have the same exception handler ???
1148                          */
1149                         if (get_Store_volatility(store) != volatility_is_volatile
1150                                 && !info->projs[pn_Store_X_except]) {
1151                                 ir_node *predvalue = get_Store_value(pred);
1152
1153                                 if (is_partially_same(value, predvalue)) {
1154                                         DBG_OPT_WAW(pred, store);
1155                                         exchange(info->projs[pn_Store_M], mem);
1156                                         kill_node(store);
1157                                         reduce_adr_usage(ptr);
1158                                         return DF_CHANGED;
1159                                 }
1160                         }
1161                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1162                            value == pred_info->projs[pn_Load_res]) {
1163                         /*
1164                          * a Store of a value just loaded from the same address
1165                          * -- a write after read.
1166                          * We may remove the Store, if it does not have an exception
1167                          * handler.
1168                          */
1169                         if (! info->projs[pn_Store_X_except]) {
1170                                 DBG_OPT_WAR(store, pred);
1171                                 exchange(info->projs[pn_Store_M], mem);
1172                                 kill_node(store);
1173                                 reduce_adr_usage(ptr);
1174                                 return DF_CHANGED;
1175                         }
1176                 }
1177
1178                 if (is_Store(pred)) {
1179                         /* check if we can pass through this store */
1180                         ir_alias_relation rel = get_alias_relation(
1181                                 get_Store_ptr(pred),
1182                                 get_irn_mode(get_Store_value(pred)),
1183                                 ptr, mode);
1184                         /* if the might be an alias, we cannot pass this Store */
1185                         if (rel != ir_no_alias)
1186                                 break;
1187                         pred = skip_Proj(get_Store_mem(pred));
1188                 } else if (is_Load(pred)) {
1189                         ir_alias_relation rel = get_alias_relation(
1190                                 get_Load_ptr(pred), get_Load_mode(pred),
1191                                 ptr, mode);
1192                         if (rel != ir_no_alias)
1193                                 break;
1194
1195                         pred = skip_Proj(get_Load_mem(pred));
1196                 } else {
1197                         /* follow only Load chains */
1198                         break;
1199                 }
1200
1201                 /* check for cycles */
1202                 if (NODE_VISITED(pred_info))
1203                         break;
1204                 MARK_NODE(pred_info);
1205         }
1206
1207         if (is_Sync(pred)) {
1208                 int i;
1209
1210                 /* handle all Sync predecessors */
1211                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1212                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1213                         if (res)
1214                                 break;
1215                 }
1216         }
1217         return res;
1218 }  /* follow_Mem_chain_for_Store */
1219
1220 /** find entity used as base for an address calculation */
1221 static ir_entity *find_entity(ir_node *ptr)
1222 {
1223         switch (get_irn_opcode(ptr)) {
1224         case iro_SymConst:
1225                 return get_SymConst_entity(ptr);
1226         case iro_Sel: {
1227                 ir_node *pred = get_Sel_ptr(ptr);
1228                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1229                         return get_Sel_entity(ptr);
1230
1231                 return find_entity(pred);
1232         }
1233         case iro_Sub:
1234         case iro_Add: {
1235                 ir_node *left = get_binop_left(ptr);
1236                 ir_node *right;
1237                 if (mode_is_reference(get_irn_mode(left)))
1238                         return find_entity(left);
1239                 right = get_binop_right(ptr);
1240                 if (mode_is_reference(get_irn_mode(right)))
1241                         return find_entity(right);
1242                 return NULL;
1243         }
1244         default:
1245                 return NULL;
1246         }
1247 }
1248
1249 /**
1250  * optimize a Store
1251  *
1252  * @param store  the Store node
1253  */
1254 static unsigned optimize_store(ir_node *store)
1255 {
1256         ir_node   *ptr;
1257         ir_node   *mem;
1258         ir_entity *entity;
1259
1260         if (get_Store_volatility(store) == volatility_is_volatile)
1261                 return 0;
1262
1263         ptr    = get_Store_ptr(store);
1264         entity = find_entity(ptr);
1265
1266         /* a store to an entity which is never read is unnecessary */
1267         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1268                 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1269                 if (info->projs[pn_Store_X_except] == NULL) {
1270                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1271                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1272                         kill_node(store);
1273                         reduce_adr_usage(ptr);
1274                         return DF_CHANGED;
1275                 }
1276         }
1277
1278         /* Check, if the address of this Store is used more than once.
1279          * If not, this Store cannot be removed in any case. */
1280         if (get_irn_n_edges(ptr) <= 1)
1281                 return 0;
1282
1283         mem = get_Store_mem(store);
1284
1285         /* follow the memory chain as long as there are only Loads */
1286         INC_MASTER();
1287
1288         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1289 }  /* optimize_store */
1290
1291 /* check if a node has more than one real user. Keepalive edges do not count as
1292  * real users */
1293 static bool has_multiple_users(const ir_node *node)
1294 {
1295         unsigned real_users = 0;
1296         foreach_out_edge(node, edge) {
1297                 ir_node *user = get_edge_src_irn(edge);
1298                 if (is_End(user))
1299                         continue;
1300                 ++real_users;
1301                 if (real_users > 1)
1302                         return true;
1303         }
1304         return false;
1305 }
1306
1307 /**
1308  * walker, optimizes Phi after Stores to identical places:
1309  * Does the following optimization:
1310  * @verbatim
1311  *
1312  *   val1   val2   val3          val1  val2  val3
1313  *    |      |      |               \    |    /
1314  *  Store  Store  Store              \   |   /
1315  *      \    |    /                   PhiData
1316  *       \   |   /                       |
1317  *        \  |  /                      Store
1318  *          PhiM
1319  *
1320  * @endverbatim
1321  * This reduces the number of stores and allows for predicated execution.
1322  * Moves Stores back to the end of a function which may be bad.
1323  *
1324  * This is only possible if the predecessor blocks have only one successor.
1325  */
1326 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1327 {
1328         int i, n;
1329         ir_node *store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1330 #ifdef DO_CACHEOPT
1331         ir_node *old_store;
1332 #endif
1333         ir_mode *mode;
1334         ir_node **inM, **inD, **projMs;
1335         int *idx;
1336         dbg_info *db = NULL;
1337         ldst_info_t *info;
1338         block_info_t *bl_info;
1339         unsigned res = 0;
1340
1341         /* Must be a memory Phi */
1342         if (get_irn_mode(phi) != mode_M)
1343                 return 0;
1344
1345         n = get_Phi_n_preds(phi);
1346         if (n <= 0)
1347                 return 0;
1348
1349         /* must be only one user */
1350         projM = get_Phi_pred(phi, 0);
1351         if (has_multiple_users(projM))
1352                 return 0;
1353
1354         store = skip_Proj(projM);
1355 #ifdef DO_CACHEOPT
1356         old_store = store;
1357 #endif
1358         if (!is_Store(store))
1359                 return 0;
1360
1361         block = get_nodes_block(store);
1362
1363         /* check if the block is post dominated by Phi-block
1364            and has no exception exit */
1365         bl_info = (block_info_t*)get_irn_link(block);
1366         if (bl_info->flags & BLOCK_HAS_EXC)
1367                 return 0;
1368
1369         phi_block = get_nodes_block(phi);
1370         if (! block_strictly_postdominates(phi_block, block))
1371                 return 0;
1372
1373         /* this is the address of the store */
1374         ptr  = get_Store_ptr(store);
1375         mode = get_irn_mode(get_Store_value(store));
1376         info = (ldst_info_t*)get_irn_link(store);
1377         exc  = info->exc_block;
1378
1379         for (i = 1; i < n; ++i) {
1380                 ir_node *pred = get_Phi_pred(phi, i);
1381
1382                 if (has_multiple_users(pred))
1383                         return 0;
1384
1385                 pred = skip_Proj(pred);
1386                 if (!is_Store(pred))
1387                         return 0;
1388
1389                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1390                         return 0;
1391
1392                 info = (ldst_info_t*)get_irn_link(pred);
1393
1394                 /* check, if all stores have the same exception flow */
1395                 if (exc != info->exc_block)
1396                         return 0;
1397
1398                 block = get_nodes_block(pred);
1399
1400                 /* check if the block is post dominated by Phi-block
1401                    and has no exception exit. Note that block must be different from
1402                    Phi-block, else we would move a Store from end End of a block to its
1403                    Start... */
1404                 bl_info = (block_info_t*)get_irn_link(block);
1405                 if (bl_info->flags & BLOCK_HAS_EXC)
1406                         return 0;
1407                 if (block == phi_block || ! block_postdominates(phi_block, block))
1408                         return 0;
1409         }
1410
1411         /*
1412          * ok, when we are here, we found all predecessors of a Phi that
1413          * are Stores to the same address and size. That means whatever
1414          * we do before we enter the block of the Phi, we do a Store.
1415          * So, we can move the Store to the current block:
1416          *
1417          *   val1    val2    val3          val1  val2  val3
1418          *    |       |       |               \    |    /
1419          * | Str | | Str | | Str |             \   |   /
1420          *      \     |     /                   PhiData
1421          *       \    |    /                       |
1422          *        \   |   /                       Str
1423          *           PhiM
1424          *
1425          * Is only allowed if the predecessor blocks have only one successor.
1426          */
1427
1428         NEW_ARR_A(ir_node *, projMs, n);
1429         NEW_ARR_A(ir_node *, inM, n);
1430         NEW_ARR_A(ir_node *, inD, n);
1431         NEW_ARR_A(int, idx, n);
1432
1433         /* Prepare: Collect all Store nodes.  We must do this
1434            first because we otherwise may loose a store when exchanging its
1435            memory Proj.
1436          */
1437         for (i = n - 1; i >= 0; --i) {
1438                 ir_node *store;
1439
1440                 projMs[i] = get_Phi_pred(phi, i);
1441                 assert(is_Proj(projMs[i]));
1442
1443                 store = get_Proj_pred(projMs[i]);
1444                 info  = (ldst_info_t*)get_irn_link(store);
1445
1446                 inM[i] = get_Store_mem(store);
1447                 inD[i] = get_Store_value(store);
1448                 idx[i] = info->exc_idx;
1449         }
1450         block = get_nodes_block(phi);
1451
1452         /* second step: create a new memory Phi */
1453         phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1454
1455         /* third step: create a new data Phi */
1456         phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1457
1458         /* rewire memory and kill the node */
1459         for (i = n - 1; i >= 0; --i) {
1460                 ir_node *proj  = projMs[i];
1461
1462                 if (is_Proj(proj)) {
1463                         ir_node *store = get_Proj_pred(proj);
1464                         exchange(proj, inM[i]);
1465                         kill_node(store);
1466                 }
1467         }
1468
1469         /* fourth step: create the Store */
1470         store = new_rd_Store(db, block, phiM, ptr, phiD, cons_none);
1471 #ifdef DO_CACHEOPT
1472         co_set_irn_name(store, co_get_irn_ident(old_store));
1473 #endif
1474
1475         projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1476
1477         info = get_ldst_info(store, &wenv->obst);
1478         info->projs[pn_Store_M] = projM;
1479
1480         /* fifths step: repair exception flow */
1481         if (exc) {
1482                 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1483
1484                 info->projs[pn_Store_X_except] = projX;
1485                 info->exc_block                = exc;
1486                 info->exc_idx                  = idx[0];
1487
1488                 for (i = 0; i < n; ++i) {
1489                         set_Block_cfgpred(exc, idx[i], projX);
1490                 }
1491
1492                 if (n > 1) {
1493                         /* the exception block should be optimized as some inputs are identical now */
1494                 }
1495
1496                 res |= CF_CHANGED;
1497         }
1498
1499         /* sixth step: replace old Phi */
1500         exchange(phi, projM);
1501
1502         return res | DF_CHANGED;
1503 }  /* optimize_phi */
1504
1505 static int optimize_conv_load(ir_node *conv)
1506 {
1507         ir_node *op = get_Conv_op(conv);
1508         if (!is_Proj(op))
1509                 return 0;
1510         if (has_multiple_users(op))
1511                 return 0;
1512         /* shrink mode of load if possible. */
1513         ir_node *load = get_Proj_pred(op);
1514         if (!is_Load(load))
1515                 return 0;
1516
1517         /* only do it if we are the only user (otherwise the risk is too
1518          * great that we end up with 2 loads instead of one). */
1519         ir_mode *mode      = get_irn_mode(conv);
1520         ir_mode *load_mode = get_Load_mode(load);
1521         int      bits_diff
1522                 = get_mode_size_bits(load_mode) - get_mode_size_bits(mode);
1523         if (mode_is_float(load_mode) || mode_is_float(mode) || bits_diff < 0)
1524             return 0;
1525
1526         if (be_get_backend_param()->byte_order_big_endian) {
1527                 if (bits_diff % 8 != 0)
1528                         return 0;
1529                 ir_graph *irg   = get_irn_irg(conv);
1530                 ir_node  *ptr   = get_Load_ptr(load);
1531                 ir_mode  *mode  = get_irn_mode(ptr);
1532                 ir_node  *delta = new_r_Const_long(irg, mode, bits_diff/8);
1533                 ir_node  *block = get_nodes_block(load);
1534                 ir_node  *add   = new_r_Add(block, ptr, delta, mode);
1535                 set_Load_ptr(load, add);
1536         }
1537         set_Load_mode(load, mode);
1538         set_irn_mode(op, mode);
1539         exchange(conv, op);
1540         return DF_CHANGED;
1541 }
1542
1543 /**
1544  * walker, do the optimizations
1545  */
1546 static void do_load_store_optimize(ir_node *n, void *env)
1547 {
1548         walk_env_t *wenv = (walk_env_t*)env;
1549
1550         switch (get_irn_opcode(n)) {
1551
1552         case iro_Load:
1553                 wenv->changes |= optimize_load(n);
1554                 break;
1555
1556         case iro_Store:
1557                 wenv->changes |= optimize_store(n);
1558                 break;
1559
1560         case iro_Phi:
1561                 wenv->changes |= optimize_phi(n, wenv);
1562                 break;
1563
1564         case iro_Conv:
1565                 wenv->changes |= optimize_conv_load(n);
1566                 break;
1567
1568         default:
1569                 break;
1570         }
1571 }  /* do_load_store_optimize */
1572
1573 /** A scc. */
1574 typedef struct scc {
1575         ir_node *head;      /**< the head of the list */
1576 } scc;
1577
1578 /** A node entry. */
1579 typedef struct node_entry {
1580         unsigned DFSnum;    /**< the DFS number of this node */
1581         unsigned low;       /**< the low number of this node */
1582         int      in_stack;  /**< flag, set if the node is on the stack */
1583         ir_node  *next;     /**< link to the next node the the same scc */
1584         scc      *pscc;     /**< the scc of this node */
1585         unsigned POnum;     /**< the post order number for blocks */
1586 } node_entry;
1587
1588 /** A loop entry. */
1589 typedef struct loop_env {
1590         ir_nodehashmap_t map;
1591         struct obstack   obst;
1592         ir_node          **stack;      /**< the node stack */
1593         size_t           tos;          /**< tos index */
1594         unsigned         nextDFSnum;   /**< the current DFS number */
1595         unsigned         POnum;        /**< current post order number */
1596
1597         unsigned         changes;      /**< a bitmask of graph changes */
1598 } loop_env;
1599
1600 /**
1601 * Gets the node_entry of a node
1602 */
1603 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1604 {
1605         node_entry *e = ir_nodehashmap_get(node_entry, &env->map, irn);
1606
1607         if (e == NULL) {
1608                 e = OALLOC(&env->obst, node_entry);
1609                 memset(e, 0, sizeof(*e));
1610                 ir_nodehashmap_insert(&env->map, irn, e);
1611         }
1612         return e;
1613 }  /* get_irn_ne */
1614
1615 /**
1616  * Push a node onto the stack.
1617  *
1618  * @param env   the loop environment
1619  * @param n     the node to push
1620  */
1621 static void push(loop_env *env, ir_node *n)
1622 {
1623         node_entry *e;
1624
1625         if (env->tos == ARR_LEN(env->stack)) {
1626                 size_t nlen = ARR_LEN(env->stack) * 2;
1627                 ARR_RESIZE(ir_node *, env->stack, nlen);
1628         }
1629         env->stack[env->tos++] = n;
1630         e = get_irn_ne(n, env);
1631         e->in_stack = 1;
1632 }  /* push */
1633
1634 /**
1635  * pop a node from the stack
1636  *
1637  * @param env   the loop environment
1638  *
1639  * @return  The topmost node
1640  */
1641 static ir_node *pop(loop_env *env)
1642 {
1643         ir_node *n = env->stack[--env->tos];
1644         node_entry *e = get_irn_ne(n, env);
1645
1646         e->in_stack = 0;
1647         return n;
1648 }  /* pop */
1649
1650 /**
1651  * Check if irn is a region constant.
1652  * The block or irn must strictly dominate the header block.
1653  *
1654  * @param irn           the node to check
1655  * @param header_block  the header block of the induction variable
1656  */
1657 static int is_rc(ir_node *irn, ir_node *header_block)
1658 {
1659         ir_node *block = get_nodes_block(irn);
1660
1661         return (block != header_block) && block_dominates(block, header_block);
1662 }  /* is_rc */
1663
1664 typedef struct phi_entry phi_entry;
1665 struct phi_entry {
1666         ir_node   *phi;    /**< A phi with a region const memory. */
1667         int       pos;     /**< The position of the region const memory */
1668         ir_node   *load;   /**< the newly created load for this phi */
1669         phi_entry *next;
1670 };
1671
1672 /**
1673  * An entry in the avail set.
1674  */
1675 typedef struct avail_entry_t {
1676         ir_node *ptr;   /**< the address pointer */
1677         ir_mode *mode;  /**< the load mode */
1678         ir_node *load;  /**< the associated Load */
1679 } avail_entry_t;
1680
1681 /**
1682  * Compare two avail entries.
1683  */
1684 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1685 {
1686         const avail_entry_t *a = (const avail_entry_t*)elt;
1687         const avail_entry_t *b = (const avail_entry_t*)key;
1688         (void) size;
1689
1690         return a->ptr != b->ptr || a->mode != b->mode;
1691 }  /* cmp_avail_entry */
1692
1693 /**
1694  * Calculate the hash value of an avail entry.
1695  */
1696 static unsigned hash_cache_entry(const avail_entry_t *entry)
1697 {
1698         return get_irn_idx(entry->ptr) * 9 + hash_ptr(entry->mode);
1699 }  /* hash_cache_entry */
1700
1701 /**
1702  * Move loops out of loops if possible.
1703  *
1704  * @param pscc   the loop described by an SCC
1705  * @param env    the loop environment
1706  */
1707 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1708 {
1709         ir_node   *phi, *load, *next, *other, *next_other;
1710         int       j;
1711         phi_entry *phi_list = NULL;
1712         set       *avail;
1713
1714         /* collect all outer memories */
1715         for (phi = pscc->head; phi != NULL; phi = next) {
1716                 node_entry *ne = get_irn_ne(phi, env);
1717                 next = ne->next;
1718
1719                 /* check all memory Phi's */
1720                 if (! is_Phi(phi))
1721                         continue;
1722
1723                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1724
1725                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1726                         ir_node    *pred = get_irn_n(phi, j);
1727                         node_entry *pe   = get_irn_ne(pred, env);
1728
1729                         if (pe->pscc != ne->pscc) {
1730                                 /* not in the same SCC, is region const */
1731                                 phi_entry *pe = OALLOC(&env->obst, phi_entry);
1732
1733                                 pe->phi  = phi;
1734                                 pe->pos  = j;
1735                                 pe->next = phi_list;
1736                                 phi_list = pe;
1737                         }
1738                 }
1739         }
1740         /* no Phis no fun */
1741         assert(phi_list != NULL && "DFS found a loop without Phi");
1742
1743         /* for now, we cannot handle more than one input (only reducible cf) */
1744         if (phi_list->next != NULL)
1745                 return;
1746
1747         avail = new_set(cmp_avail_entry, 8);
1748
1749         for (load = pscc->head; load; load = next) {
1750                 ir_mode *load_mode;
1751                 node_entry *ne = get_irn_ne(load, env);
1752                 next = ne->next;
1753
1754                 if (is_Load(load)) {
1755                         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1756                         ir_node     *ptr = get_Load_ptr(load);
1757
1758                         /* for now, we cannot handle Loads with exceptions */
1759                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1760                                 continue;
1761
1762                         /* for now, we can only move Load(Global) */
1763                         if (! is_SymConst_addr_ent(ptr))
1764                                 continue;
1765                         load_mode = get_Load_mode(load);
1766                         for (other = pscc->head; other != NULL; other = next_other) {
1767                                 node_entry *ne = get_irn_ne(other, env);
1768                                 next_other = ne->next;
1769
1770                                 if (is_Store(other)) {
1771                                         ir_alias_relation rel = get_alias_relation(
1772                                                 get_Store_ptr(other),
1773                                                 get_irn_mode(get_Store_value(other)),
1774                                                 ptr, load_mode);
1775                                         /* if the might be an alias, we cannot pass this Store */
1776                                         if (rel != ir_no_alias)
1777                                                 break;
1778                                 }
1779                                 /* only Phis and pure Calls are allowed here, so ignore them */
1780                         }
1781                         if (other == NULL) {
1782                                 ldst_info_t *ninfo = NULL;
1783                                 phi_entry   *pe;
1784                                 dbg_info    *db;
1785
1786                                 /* yep, no aliasing Store found, Load can be moved */
1787                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1788
1789                                 db   = get_irn_dbg_info(load);
1790                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1791                                         int     pos   = pe->pos;
1792                                         ir_node *phi  = pe->phi;
1793                                         ir_node *blk  = get_nodes_block(phi);
1794                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1795                                         ir_node *irn, *mem;
1796                                         avail_entry_t entry, *res;
1797
1798                                         entry.ptr  = ptr;
1799                                         entry.mode = load_mode;
1800                                         res = set_find(avail_entry_t, avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1801                                         if (res != NULL) {
1802                                                 irn = res->load;
1803                                         } else {
1804                                                 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, cons_none);
1805                                                 entry.load = irn;
1806                                                 (void)set_insert(avail_entry_t, avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1807                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1808                                         }
1809                                         pe->load = irn;
1810                                         ninfo = get_ldst_info(irn, &env->obst);
1811
1812                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1813                                         if (res == NULL) {
1814                                                 /* irn is from cache, so do not set phi pred again.
1815                                                  * There might be other Loads between phi and irn already.
1816                                                  */
1817                                                 set_Phi_pred(phi, pos, mem);
1818                                         }
1819
1820                                         ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1821                                 }
1822
1823                                 /* now kill the old Load */
1824                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1825                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1826
1827                                 env->changes |= DF_CHANGED;
1828                         }
1829                 }
1830         }
1831         del_set(avail);
1832 }  /* move_loads_out_of_loops */
1833
1834 /**
1835  * Process a loop SCC.
1836  *
1837  * @param pscc  the SCC
1838  * @param env   the loop environment
1839  */
1840 static void process_loop(scc *pscc, loop_env *env)
1841 {
1842         ir_node *irn, *next, *header = NULL;
1843         node_entry *b, *h = NULL;
1844         int j, only_phi, num_outside, process = 0;
1845         ir_node *out_rc;
1846
1847         /* find the header block for this scc */
1848         for (irn = pscc->head; irn; irn = next) {
1849                 node_entry *e = get_irn_ne(irn, env);
1850                 ir_node *block = get_nodes_block(irn);
1851
1852                 next = e->next;
1853                 b = get_irn_ne(block, env);
1854
1855                 if (header != NULL) {
1856                         if (h->POnum < b->POnum) {
1857                                 header = block;
1858                                 h      = b;
1859                         }
1860                 } else {
1861                         header = block;
1862                         h      = b;
1863                 }
1864         }
1865
1866         /* check if this scc contains only Phi, Loads or Stores nodes */
1867         only_phi    = 1;
1868         num_outside = 0;
1869         out_rc      = NULL;
1870         for (irn = pscc->head; irn; irn = next) {
1871                 node_entry *e = get_irn_ne(irn, env);
1872
1873                 next = e->next;
1874                 switch (get_irn_opcode(irn)) {
1875                 case iro_Call:
1876                         if (is_Call_pure(irn)) {
1877                                 /* pure calls can be treated like loads */
1878                                 only_phi = 0;
1879                                 break;
1880                         }
1881                         /* non-pure calls must be handle like may-alias Stores */
1882                         goto fail;
1883                 case iro_CopyB:
1884                         /* cannot handle CopyB yet */
1885                         goto fail;
1886                 case iro_Load:
1887                         process = 1;
1888                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1889                                 /* cannot handle loops with volatile Loads */
1890                                 goto fail;
1891                         }
1892                         only_phi = 0;
1893                         break;
1894                 case iro_Store:
1895                         if (get_Store_volatility(irn) == volatility_is_volatile) {
1896                                 /* cannot handle loops with volatile Stores */
1897                                 goto fail;
1898                         }
1899                         only_phi = 0;
1900                         break;
1901                 default:
1902                         only_phi = 0;
1903                         break;
1904                 case iro_Phi:
1905                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1906                                 ir_node *pred  = get_irn_n(irn, j);
1907                                 node_entry *pe = get_irn_ne(pred, env);
1908
1909                                 if (pe->pscc != e->pscc) {
1910                                         /* not in the same SCC, must be a region const */
1911                                         if (! is_rc(pred, header)) {
1912                                                 /* not a memory loop */
1913                                                 goto fail;
1914                                         }
1915                                         if (out_rc == NULL) {
1916                                                 /* first region constant */
1917                                                 out_rc = pred;
1918                                                 ++num_outside;
1919                                         } else if (out_rc != pred) {
1920                                                 /* another region constant */
1921                                                 ++num_outside;
1922                                         }
1923                                 }
1924                         }
1925                         break;
1926                 }
1927         }
1928         if (! process)
1929                 goto fail;
1930
1931         /* found a memory loop */
1932         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
1933         if (only_phi && num_outside == 1) {
1934                 /* a phi cycle with only one real predecessor can be collapsed */
1935                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
1936
1937                 for (irn = pscc->head; irn; irn = next) {
1938                         node_entry *e = get_irn_ne(irn, env);
1939                         next = e->next;
1940                         exchange(irn, out_rc);
1941                 }
1942                 env->changes |= DF_CHANGED;
1943                 return;
1944         }
1945
1946 #ifdef DEBUG_libfirm
1947         for (irn = pscc->head; irn; irn = next) {
1948                 node_entry *e = get_irn_ne(irn, env);
1949                 next = e->next;
1950                 DB((dbg, LEVEL_2, " %+F,", irn));
1951         }
1952         DB((dbg, LEVEL_2, "\n"));
1953 #endif
1954         move_loads_out_of_loops(pscc, env);
1955
1956 fail:
1957         ;
1958 }  /* process_loop */
1959
1960 /**
1961  * Process a SCC.
1962  *
1963  * @param pscc  the SCC
1964  * @param env   the loop environment
1965  */
1966 static void process_scc(scc *pscc, loop_env *env)
1967 {
1968         ir_node *head = pscc->head;
1969         node_entry *e = get_irn_ne(head, env);
1970
1971 #ifdef DEBUG_libfirm
1972         {
1973                 ir_node *irn, *next;
1974
1975                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
1976                 for (irn = pscc->head; irn; irn = next) {
1977                         node_entry *e = get_irn_ne(irn, env);
1978
1979                         next = e->next;
1980
1981                         DB((dbg, LEVEL_4, " %+F,", irn));
1982                 }
1983                 DB((dbg, LEVEL_4, "\n"));
1984         }
1985 #endif
1986
1987         if (e->next != NULL) {
1988                 /* this SCC has more than one member */
1989                 process_loop(pscc, env);
1990         }
1991 }  /* process_scc */
1992
1993 /**
1994  * Do Tarjan's SCC algorithm and drive load/store optimization.
1995  *
1996  * @param irn  start at this node
1997  * @param env  the loop environment
1998  */
1999 static void dfs(ir_node *irn, loop_env *env)
2000 {
2001         int i, n;
2002         node_entry *node = get_irn_ne(irn, env);
2003
2004         mark_irn_visited(irn);
2005
2006         node->DFSnum = env->nextDFSnum++;
2007         node->low    = node->DFSnum;
2008         push(env, irn);
2009
2010         /* handle preds */
2011         if (is_Phi(irn) || is_Sync(irn)) {
2012                 n = get_irn_arity(irn);
2013                 for (i = 0; i < n; ++i) {
2014                         ir_node *pred = get_irn_n(irn, i);
2015                         node_entry *o = get_irn_ne(pred, env);
2016
2017                         if (!irn_visited(pred)) {
2018                                 dfs(pred, env);
2019                                 node->low = MIN(node->low, o->low);
2020                         }
2021                         if (o->DFSnum < node->DFSnum && o->in_stack)
2022                                 node->low = MIN(o->DFSnum, node->low);
2023                 }
2024         } else if (is_fragile_op(irn)) {
2025                 ir_node *pred = get_memop_mem(irn);
2026                 node_entry *o = get_irn_ne(pred, env);
2027
2028                 if (!irn_visited(pred)) {
2029                         dfs(pred, env);
2030                         node->low = MIN(node->low, o->low);
2031                 }
2032                 if (o->DFSnum < node->DFSnum && o->in_stack)
2033                         node->low = MIN(o->DFSnum, node->low);
2034         } else if (is_Proj(irn)) {
2035                 ir_node *pred = get_Proj_pred(irn);
2036                 node_entry *o = get_irn_ne(pred, env);
2037
2038                 if (!irn_visited(pred)) {
2039                         dfs(pred, env);
2040                         node->low = MIN(node->low, o->low);
2041                 }
2042                 if (o->DFSnum < node->DFSnum && o->in_stack)
2043                         node->low = MIN(o->DFSnum, node->low);
2044         }
2045         else {
2046                  /* IGNORE predecessors */
2047         }
2048
2049         if (node->low == node->DFSnum) {
2050                 scc *pscc = OALLOC(&env->obst, scc);
2051                 ir_node *x;
2052
2053                 pscc->head = NULL;
2054                 do {
2055                         node_entry *e;
2056
2057                         x = pop(env);
2058                         e = get_irn_ne(x, env);
2059                         e->pscc    = pscc;
2060                         e->next    = pscc->head;
2061                         pscc->head = x;
2062                 } while (x != irn);
2063
2064                 process_scc(pscc, env);
2065         }
2066 }  /* dfs */
2067
2068 /**
2069  * Do the DFS on the memory edges a graph.
2070  *
2071  * @param irg  the graph to process
2072  * @param env  the loop environment
2073  */
2074 static void do_dfs(ir_graph *irg, loop_env *env)
2075 {
2076         ir_node  *endblk, *end;
2077         int      i;
2078
2079         inc_irg_visited(irg);
2080
2081         /* visit all memory nodes */
2082         endblk = get_irg_end_block(irg);
2083         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2084                 ir_node *pred = get_Block_cfgpred(endblk, i);
2085
2086                 pred = skip_Proj(pred);
2087                 if (is_Return(pred)) {
2088                         dfs(get_Return_mem(pred), env);
2089                 } else if (is_Raise(pred)) {
2090                         dfs(get_Raise_mem(pred), env);
2091                 } else if (is_fragile_op(pred)) {
2092                         dfs(get_memop_mem(pred), env);
2093                 } else if (is_Bad(pred)) {
2094                         /* ignore non-optimized block predecessor */
2095                 } else {
2096                         assert(0 && "Unknown EndBlock predecessor");
2097                 }
2098         }
2099
2100         /* visit the keep-alives */
2101         end = get_irg_end(irg);
2102         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2103                 ir_node *ka = get_End_keepalive(end, i);
2104
2105                 if (is_Phi(ka) && !irn_visited(ka))
2106                         dfs(ka, env);
2107         }
2108 }  /* do_dfs */
2109
2110 /**
2111  * Optimize Loads/Stores in loops.
2112  *
2113  * @param irg  the graph
2114  */
2115 static int optimize_loops(ir_graph *irg)
2116 {
2117         loop_env env;
2118
2119         env.stack         = NEW_ARR_F(ir_node *, 128);
2120         env.tos           = 0;
2121         env.nextDFSnum    = 0;
2122         env.POnum         = 0;
2123         env.changes       = 0;
2124         ir_nodehashmap_init(&env.map);
2125         obstack_init(&env.obst);
2126
2127         /* calculate the SCC's and drive loop optimization. */
2128         do_dfs(irg, &env);
2129
2130         DEL_ARR_F(env.stack);
2131         obstack_free(&env.obst, NULL);
2132         ir_nodehashmap_destroy(&env.map);
2133
2134         return env.changes;
2135 }  /* optimize_loops */
2136
2137 void optimize_load_store(ir_graph *irg)
2138 {
2139         walk_env_t env;
2140
2141         assure_irg_properties(irg,
2142                 IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
2143                 | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
2144                 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES
2145                 | IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
2146                 | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE);
2147
2148         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2149
2150         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2151                 "LoadStore optimization needs pinned graph");
2152
2153         if (get_opt_alias_analysis()) {
2154                 assure_irp_globals_entity_usage_computed();
2155         }
2156
2157         obstack_init(&env.obst);
2158         env.changes = 0;
2159
2160         /* init the links, then collect Loads/Stores/Proj's in lists */
2161         master_visited = 0;
2162         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2163
2164         /* now we have collected enough information, optimize */
2165         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2166
2167         env.changes |= optimize_loops(irg);
2168
2169         obstack_free(&env.obst, NULL);
2170
2171         confirm_irg_properties(irg,
2172                 env.changes
2173                 ? env.changes & CF_CHANGED
2174                         ? IR_GRAPH_PROPERTIES_NONE
2175                         : IR_GRAPH_PROPERTIES_CONTROL_FLOW
2176                 : IR_GRAPH_PROPERTIES_ALL);
2177 }
2178
2179 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2180 {
2181         return def_graph_pass(name ? name : "ldst", optimize_load_store);
2182 }  /* optimize_load_store_pass */