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

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