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 #include "config.h"
  27
  28 #include <string.h>
  29
  30 #include "iroptimize.h"
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irmode_t.h"
  34 #include "iropt_t.h"
  35 #include "ircons_t.h"
  36 #include "irgmod.h"
  37 #include "irgwalk.h"
  38 #include "irvrfy.h"
  39 #include "tv_t.h"
  40 #include "dbginfo_t.h"
  41 #include "iropt_dbg.h"
  42 #include "irflag_t.h"
  43 #include "array_t.h"
  44 #include "irhooks.h"
  45 #include "iredges.h"
  46 #include "irtools.h"
  47 #include "opt_polymorphy.h"
  48 #include "irmemory.h"
  49 #include "irphase_t.h"
  50 #include "irgopt.h"
  51 #include "set.h"
  52 #include "debug.h"
  53
  54 /** The debug handle. */
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  56
  57 #ifdef DO_CACHEOPT
  58 #include "cacheopt/cachesim.h"
  59 #endif
  60
  61 #undef IMAX
  62 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  63
  64 #define MAX_PROJ        IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
  65
  66 enum changes_t {
  67         DF_CHANGED = 1,       /**< data flow changed */
  68         CF_CHANGED = 2,       /**< control flow changed */
  69 };
  70
  71 /**
  72  * walker environment
  73  */
  74 typedef struct _walk_env_t {
  75         struct obstack obst;          /**< list of all stores */
  76         unsigned changes;             /**< a bitmask of graph changes */
  77 } walk_env_t;
  78
  79 /** A Load/Store info. */
  80 typedef struct _ldst_info_t {
  81         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  82         ir_node  *exc_block;          /**< the exception block if available */
  83         int      exc_idx;             /**< predecessor index in the exception block */
  84         unsigned visited;             /**< visited counter for breaking loops */
  85 } ldst_info_t;
  86
  87 /**
  88  * flags for control flow.
  89  */
  90 enum block_flags_t {
  91         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  92         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  93 };
  94
  95 /**
  96  * a Block info.
  97  */
  98 typedef struct _block_info_t {
  99         unsigned flags;               /**< flags for the block */
 100 } block_info_t;
 101
 102 /** the master visited flag for loop detection. */
 103 static unsigned master_visited = 0;
 104
 105 #define INC_MASTER()       ++master_visited
 106 #define MARK_NODE(info)    (info)->visited = master_visited
 107 #define NODE_VISITED(info) (info)->visited >= master_visited
 108
 109 /**
 110  * get the Load/Store info of a node
 111  */
 112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 113         ldst_info_t *info = get_irn_link(node);
 114
 115         if (! info) {
 116                 info = obstack_alloc(obst, sizeof(*info));
 117                 memset(info, 0, sizeof(*info));
 118                 set_irn_link(node, info);
 119         }
 120         return info;
 121 }  /* get_ldst_info */
 122
 123 /**
 124  * get the Block info of a node
 125  */
 126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
 127         block_info_t *info = get_irn_link(node);
 128
 129         if (! info) {
 130                 info = obstack_alloc(obst, sizeof(*info));
 131                 memset(info, 0, sizeof(*info));
 132                 set_irn_link(node, info);
 133         }
 134         return info;
 135 }  /* get_block_info */
 136
 137 /**
 138  * update the projection info for a Load/Store
 139  */
 140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 141 {
 142         long nr = get_Proj_proj(proj);
 143
 144         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 145
 146         if (info->projs[nr]) {
 147                 /* there is already one, do CSE */
 148                 exchange(proj, info->projs[nr]);
 149                 return DF_CHANGED;
 150         }
 151         else {
 152                 info->projs[nr] = proj;
 153                 return 0;
 154         }
 155 }  /* update_projs */
 156
 157 /**
 158  * update the exception block info for a Load/Store node.
 159  *
 160  * @param info   the load/store info struct
 161  * @param block  the exception handler block for this load/store
 162  * @param pos    the control flow input of the block
 163  */
 164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 165 {
 166         assert(info->exc_block == NULL && "more than one exception block found");
 167
 168         info->exc_block = block;
 169         info->exc_idx   = pos;
 170         return 0;
 171 }  /* update_exc */
 172
 173 /** Return the number of uses of an address node */
 174 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 175
 176 /**
 177  * walker, collects all Load/Store/Proj nodes
 178  *
 179  * walks from Start -> End
 180  */
 181 static void collect_nodes(ir_node *node, void *env)
 182 {
 183         ir_opcode   opcode = get_irn_opcode(node);
 184         ir_node     *pred, *blk, *pred_blk;
 185         ldst_info_t *ldst_info;
 186         walk_env_t  *wenv = env;
 187
 188         if (opcode == iro_Proj) {
 189                 pred   = get_Proj_pred(node);
 190                 opcode = get_irn_opcode(pred);
 191
 192                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 193                         ldst_info = get_ldst_info(pred, &wenv->obst);
 194
 195                         wenv->changes |= update_projs(ldst_info, node);
 196
 197                         /*
 198                          * Place the Proj's to the same block as the
 199                          * predecessor Load. This is always ok and prevents
 200                          * "non-SSA" form after optimizations if the Proj
 201                          * is in a wrong block.
 202                          */
 203                         blk      = get_nodes_block(node);
 204                         pred_blk = get_nodes_block(pred);
 205                         if (blk != pred_blk) {
 206                                 wenv->changes |= DF_CHANGED;
 207                                 set_nodes_block(node, pred_blk);
 208                         }
 209                 }
 210         } else if (opcode == iro_Block) {
 211                 int i;
 212
 213                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 214                         ir_node      *pred_block, *proj;
 215                         block_info_t *bl_info;
 216                         int          is_exc = 0;
 217
 218                         pred = proj = get_Block_cfgpred(node, i);
 219
 220                         if (is_Proj(proj)) {
 221                                 pred   = get_Proj_pred(proj);
 222                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 223                         }
 224
 225                         /* ignore Bad predecessors, they will be removed later */
 226                         if (is_Bad(pred))
 227                                 continue;
 228
 229                         pred_block = get_nodes_block(pred);
 230                         bl_info    = get_block_info(pred_block, &wenv->obst);
 231
 232                         if (is_fragile_op(pred) && is_exc)
 233                                 bl_info->flags |= BLOCK_HAS_EXC;
 234                         else if (is_irn_forking(pred))
 235                                 bl_info->flags |= BLOCK_HAS_COND;
 236
 237                         opcode = get_irn_opcode(pred);
 238                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 239                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 240
 241                                 wenv->changes |= update_exc(ldst_info, node, i);
 242                         }
 243                 }
 244         }
 245 }  /* collect_nodes */
 246
 247 /**
 248  * Returns an entity if the address ptr points to a constant one.
 249  *
 250  * @param ptr  the address
 251  *
 252  * @return an entity or NULL
 253  */
 254 static ir_entity *find_constant_entity(ir_node *ptr)
 255 {
 256         for (;;) {
 257                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 258                         return get_SymConst_entity(ptr);
 259                 } else if (is_Sel(ptr)) {
 260                         ir_entity *ent = get_Sel_entity(ptr);
 261                         ir_type   *tp  = get_entity_owner(ent);
 262
 263                         /* Do not fiddle with polymorphism. */
 264                         if (is_Class_type(get_entity_owner(ent)) &&
 265                                 ((get_entity_n_overwrites(ent)    != 0) ||
 266                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 267                                 return NULL;
 268
 269                         if (is_Array_type(tp)) {
 270                                 /* check bounds */
 271                                 int i, n;
 272
 273                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 274                                         ir_node *bound;
 275                                         tarval *tlower, *tupper;
 276                                         ir_node *index = get_Sel_index(ptr, i);
 277                                         tarval *tv     = computed_value(index);
 278
 279                                         /* check if the index is constant */
 280                                         if (tv == tarval_bad)
 281                                                 return NULL;
 282
 283                                         bound  = get_array_lower_bound(tp, i);
 284                                         tlower = computed_value(bound);
 285                                         bound  = get_array_upper_bound(tp, i);
 286                                         tupper = computed_value(bound);
 287
 288                                         if (tlower == tarval_bad || tupper == tarval_bad)
 289                                                 return NULL;
 290
 291                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 292                                                 return NULL;
 293                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 294                                                 return NULL;
 295
 296                                         /* ok, bounds check finished */
 297                                 }
 298                         }
 299
 300                         if (variability_constant == get_entity_variability(ent))
 301                                 return ent;
 302
 303                         /* try next */
 304                         ptr = get_Sel_ptr(ptr);
 305                 } else if (is_Add(ptr)) {
 306                         ir_node *l = get_Add_left(ptr);
 307                         ir_node *r = get_Add_right(ptr);
 308
 309                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 310                                 ptr = l;
 311                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 312                                 ptr = r;
 313                         else
 314                                 return NULL;
 315
 316                         /* for now, we support only one addition, reassoc should fold all others */
 317                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 318                                 return NULL;
 319                 } else if (is_Sub(ptr)) {
 320                         ir_node *l = get_Sub_left(ptr);
 321                         ir_node *r = get_Sub_right(ptr);
 322
 323                         if (get_irn_mode(l) == get_irn_mode(ptr) &&     is_Const(r))
 324                                 ptr = l;
 325                         else
 326                                 return NULL;
 327                         /* for now, we support only one substraction, reassoc should fold all others */
 328                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 329                                 return NULL;
 330                 } else
 331                         return NULL;
 332         }
 333 }  /* find_constant_entity */
 334
 335 /**
 336  * Return the Selection index of a Sel node from dimension n
 337  */
 338 static long get_Sel_array_index_long(ir_node *n, int dim) {
 339         ir_node *index = get_Sel_index(n, dim);
 340         assert(is_Const(index));
 341         return get_tarval_long(get_Const_tarval(index));
 342 }  /* get_Sel_array_index_long */
 343
 344 /**
 345  * Returns the accessed component graph path for an
 346  * node computing an address.
 347  *
 348  * @param ptr    the node computing the address
 349  * @param depth  current depth in steps upward from the root
 350  *               of the address
 351  */
 352 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 353         compound_graph_path *res = NULL;
 354         ir_entity           *root, *field, *ent;
 355         int                 path_len, pos, idx;
 356         tarval              *tv;
 357         ir_type             *tp;
 358
 359         if (is_SymConst(ptr)) {
 360                 /* a SymConst. If the depth is 0, this is an access to a global
 361                  * entity and we don't need a component path, else we know
 362                  * at least its length.
 363                  */
 364                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 365                 root = get_SymConst_entity(ptr);
 366                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 367         } else if (is_Sel(ptr)) {
 368                 /* it's a Sel, go up until we find the root */
 369                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 370                 if (res == NULL)
 371                         return NULL;
 372
 373                 /* fill up the step in the path at the current position */
 374                 field    = get_Sel_entity(ptr);
 375                 path_len = get_compound_graph_path_length(res);
 376                 pos      = path_len - depth - 1;
 377                 set_compound_graph_path_node(res, pos, field);
 378
 379                 if (is_Array_type(get_entity_owner(field))) {
 380                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 381                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 382                 }
 383         } else if (is_Add(ptr)) {
 384                 ir_node *l    = get_Add_left(ptr);
 385                 ir_node *r    = get_Add_right(ptr);
 386                 ir_mode *mode = get_irn_mode(ptr);
 387                 tarval  *tmp;
 388
 389                 if (is_Const(r) && get_irn_mode(l) == mode) {
 390                         ptr = l;
 391                         tv  = get_Const_tarval(r);
 392                 } else {
 393                         ptr = r;
 394                         tv  = get_Const_tarval(l);
 395                 }
 396 ptr_arith:
 397                 mode = get_tarval_mode(tv);
 398                 tmp  = tv;
 399
 400                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 401                 if (is_Sel(ptr)) {
 402                         field = get_Sel_entity(ptr);
 403                 } else {
 404                         field = get_SymConst_entity(ptr);
 405                 }
 406                 idx = 0;
 407                 for (ent = field;;) {
 408                         unsigned size;
 409                         tarval   *sz, *tv_index, *tlower, *tupper;
 410                         ir_node  *bound;
 411
 412                         tp = get_entity_type(ent);
 413                         if (! is_Array_type(tp))
 414                                 break;
 415                         ent = get_array_element_entity(tp);
 416                         size = get_type_size_bytes(get_entity_type(ent));
 417                         sz   = new_tarval_from_long(size, mode);
 418
 419                         tv_index = tarval_div(tmp, sz);
 420                         tmp      = tarval_mod(tmp, sz);
 421
 422                         if (tv_index == tarval_bad || tmp == tarval_bad)
 423                                 return NULL;
 424
 425                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 426                         bound  = get_array_lower_bound(tp, 0);
 427                         tlower = computed_value(bound);
 428                         bound  = get_array_upper_bound(tp, 0);
 429                         tupper = computed_value(bound);
 430
 431                         if (tlower == tarval_bad || tupper == tarval_bad)
 432                                 return NULL;
 433
 434                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 435                                 return NULL;
 436                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 437                                 return NULL;
 438
 439                         /* ok, bounds check finished */
 440                         ++idx;
 441                 }
 442                 if (! tarval_is_null(tmp)) {
 443                         /* access to some struct/union member */
 444                         return NULL;
 445                 }
 446
 447                 /* should be at least ONE array */
 448                 if (idx == 0)
 449                         return NULL;
 450
 451                 res = rec_get_accessed_path(ptr, depth + idx);
 452                 if (res == NULL)
 453                         return NULL;
 454
 455                 path_len = get_compound_graph_path_length(res);
 456                 pos      = path_len - depth - idx;
 457
 458                 for (ent = field;;) {
 459                         unsigned size;
 460                         tarval   *sz, *tv_index;
 461                         long     index;
 462
 463                         tp = get_entity_type(ent);
 464                         if (! is_Array_type(tp))
 465                                 break;
 466                         ent = get_array_element_entity(tp);
 467                         set_compound_graph_path_node(res, pos, ent);
 468
 469                         size = get_type_size_bytes(get_entity_type(ent));
 470                         sz   = new_tarval_from_long(size, mode);
 471
 472                         tv_index = tarval_div(tv, sz);
 473                         tv       = tarval_mod(tv, sz);
 474
 475                         /* worked above, should work again */
 476                         assert(tv_index != tarval_bad && tv != tarval_bad);
 477
 478                         /* bounds already checked above */
 479                         index = get_tarval_long(tv_index);
 480                         set_compound_graph_path_array_index(res, pos, index);
 481                         ++pos;
 482                 }
 483         } else if (is_Sub(ptr)) {
 484                 ir_node *l = get_Sub_left(ptr);
 485                 ir_node *r = get_Sub_right(ptr);
 486
 487                 ptr = l;
 488                 tv  = get_Const_tarval(r);
 489                 tv  = tarval_neg(tv);
 490                 goto ptr_arith;
 491         }
 492         return res;
 493 }  /* rec_get_accessed_path */
 494
 495 /**
 496  * Returns an access path or NULL.  The access path is only
 497  * valid, if the graph is in phase_high and _no_ address computation is used.
 498  */
 499 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 500         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 501         return gr;
 502 }  /* get_accessed_path */
 503
 504 typedef struct path_entry {
 505         ir_entity         *ent;
 506         struct path_entry *next;
 507         long              index;
 508 } path_entry;
 509
 510 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
 511         path_entry       entry, *p;
 512         ir_entity        *ent, *field;
 513         ir_initializer_t *initializer;
 514         tarval           *tv;
 515         ir_type          *tp;
 516         unsigned         n;
 517
 518         entry.next = next;
 519         if (is_SymConst(ptr)) {
 520                 /* found the root */
 521                 ent         = get_SymConst_entity(ptr);
 522                 initializer = get_entity_initializer(ent);
 523                 for (p = next; p != NULL;) {
 524                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 525                                 return NULL;
 526                         n  = get_initializer_compound_n_entries(initializer);
 527                         tp = get_entity_type(ent);
 528
 529                         if (is_Array_type(tp)) {
 530                                 ent = get_array_element_entity(tp);
 531                                 if (ent != p->ent) {
 532                                         /* a missing [0] */
 533                                         if (0 >= n)
 534                                                 return NULL;
 535                                         initializer = get_initializer_compound_value(initializer, 0);
 536                                         continue;
 537                                 }
 538                         }
 539                         if (p->index >= (int) n)
 540                                 return NULL;
 541                         initializer = get_initializer_compound_value(initializer, p->index);
 542
 543                         ent = p->ent;
 544                         p   = p->next;
 545                 }
 546                 tp = get_entity_type(ent);
 547                 while (is_Array_type(tp)) {
 548                         ent = get_array_element_entity(tp);
 549                         tp = get_entity_type(ent);
 550                         /* a missing [0] */
 551                         n  = get_initializer_compound_n_entries(initializer);
 552                         if (0 >= n)
 553                                 return NULL;
 554                         initializer = get_initializer_compound_value(initializer, 0);
 555                 }
 556
 557                 switch (initializer->kind) {
 558                 case IR_INITIALIZER_CONST:
 559                         return get_initializer_const_value(initializer);
 560                 case IR_INITIALIZER_TARVAL:
 561                 case IR_INITIALIZER_NULL:
 562                 default:
 563                         return NULL;
 564                 }
 565         } else if (is_Sel(ptr)) {
 566                 entry.ent = field = get_Sel_entity(ptr);
 567                 tp = get_entity_owner(field);
 568                 if (is_Array_type(tp)) {
 569                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 570                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 571                 } else {
 572                         int i, n_members = get_compound_n_members(tp);
 573                         for (i = 0; i < n_members; ++i) {
 574                                 if (get_compound_member(tp, i) == field)
 575                                         break;
 576                         }
 577                         if (i >= n_members) {
 578                                 /* not found: should NOT happen */
 579                                 return NULL;
 580                         }
 581                         entry.index = i;
 582                 }
 583                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 584         }  else if (is_Add(ptr)) {
 585                 ir_node  *l = get_Add_left(ptr);
 586                 ir_node  *r = get_Add_right(ptr);
 587                 ir_mode  *mode;
 588                 unsigned pos;
 589
 590                 if (is_Const(r)) {
 591                         ptr = l;
 592                         tv  = get_Const_tarval(r);
 593                 } else {
 594                         ptr = r;
 595                         tv  = get_Const_tarval(l);
 596                 }
 597 ptr_arith:
 598                 mode = get_tarval_mode(tv);
 599
 600                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 601                 if (is_Sel(ptr)) {
 602                         field = get_Sel_entity(ptr);
 603                 } else {
 604                         field = get_SymConst_entity(ptr);
 605                 }
 606
 607                 /* count needed entries */
 608                 pos = 0;
 609                 for (ent = field;;) {
 610                         tp = get_entity_type(ent);
 611                         if (! is_Array_type(tp))
 612                                 break;
 613                         ent = get_array_element_entity(tp);
 614                         ++pos;
 615                 }
 616                 /* should be at least ONE entry */
 617                 if (pos == 0)
 618                         return NULL;
 619
 620                 /* allocate the right number of entries */
 621                 NEW_ARR_A(path_entry, p, pos);
 622
 623                 /* fill them up */
 624                 pos = 0;
 625                 for (ent = field;;) {
 626                         unsigned size;
 627                         tarval   *sz, *tv_index, *tlower, *tupper;
 628                         long     index;
 629                         ir_node  *bound;
 630
 631                         tp = get_entity_type(ent);
 632                         if (! is_Array_type(tp))
 633                                 break;
 634                         ent = get_array_element_entity(tp);
 635                         p[pos].ent  = ent;
 636                         p[pos].next = &p[pos + 1];
 637
 638                         size = get_type_size_bytes(get_entity_type(ent));
 639                         sz   = new_tarval_from_long(size, mode);
 640
 641                         tv_index = tarval_div(tv, sz);
 642                         tv       = tarval_mod(tv, sz);
 643
 644                         if (tv_index == tarval_bad || tv == tarval_bad)
 645                                 return NULL;
 646
 647                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 648                         bound  = get_array_lower_bound(tp, 0);
 649                         tlower = computed_value(bound);
 650                         bound  = get_array_upper_bound(tp, 0);
 651                         tupper = computed_value(bound);
 652
 653                         if (tlower == tarval_bad || tupper == tarval_bad)
 654                                 return NULL;
 655
 656                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 657                                 return NULL;
 658                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 659                                 return NULL;
 660
 661                         /* ok, bounds check finished */
 662                         index = get_tarval_long(tv_index);
 663                         p[pos].index = index;
 664                         ++pos;
 665                 }
 666                 if (! tarval_is_null(tv)) {
 667                         /* hmm, wrong access */
 668                         return NULL;
 669                 }
 670                 p[pos - 1].next = next;
 671                 return rec_find_compound_ent_value(ptr, p);
 672         } else if (is_Sub(ptr)) {
 673                 ir_node *l = get_Sub_left(ptr);
 674                 ir_node *r = get_Sub_right(ptr);
 675
 676                 ptr = l;
 677                 tv  = get_Const_tarval(r);
 678                 tv  = tarval_neg(tv);
 679                 goto ptr_arith;
 680         }
 681         return NULL;
 682 }
 683
 684 static ir_node *find_compound_ent_value(ir_node *ptr) {
 685         return rec_find_compound_ent_value(ptr, NULL);
 686 }
 687
 688 /* forward */
 689 static void reduce_adr_usage(ir_node *ptr);
 690
 691 /**
 692  * Update a Load that may have lost its users.
 693  */
 694 static void handle_load_update(ir_node *load) {
 695         ldst_info_t *info = get_irn_link(load);
 696
 697         /* do NOT touch volatile loads for now */
 698         if (get_Load_volatility(load) == volatility_is_volatile)
 699                 return;
 700
 701         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 702                 ir_node *ptr = get_Load_ptr(load);
 703                 ir_node *mem = get_Load_mem(load);
 704
 705                 /* a Load whose value is neither used nor exception checked, remove it */
 706                 exchange(info->projs[pn_Load_M], mem);
 707                 if (info->projs[pn_Load_X_regular])
 708                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 709                 kill_node(load);
 710                 reduce_adr_usage(ptr);
 711         }
 712 }  /* handle_load_update */
 713
 714 /**
 715  * A use of an address node has vanished. Check if this was a Proj
 716  * node and update the counters.
 717  */
 718 static void reduce_adr_usage(ir_node *ptr) {
 719         if (is_Proj(ptr)) {
 720                 if (get_irn_n_edges(ptr) <= 0) {
 721                         /* this Proj is dead now */
 722                         ir_node *pred = get_Proj_pred(ptr);
 723
 724                         if (is_Load(pred)) {
 725                                 ldst_info_t *info = get_irn_link(pred);
 726                                 info->projs[get_Proj_proj(ptr)] = NULL;
 727
 728                                 /* this node lost its result proj, handle that */
 729                                 handle_load_update(pred);
 730                         }
 731                 }
 732         }
 733 }  /* reduce_adr_usage */
 734
 735 /**
 736  * Check, if an already existing value of mode old_mode can be converted
 737  * into the needed one new_mode without loss.
 738  */
 739 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 740         if (old_mode == new_mode)
 741                 return 1;
 742
 743         /* if both modes are two-complement ones, we can always convert the
 744            Stored value into the needed one. */
 745         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 746                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 747                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 748                 return 1;
 749         return 0;
 750 }  /* can_use_stored_value */
 751
 752 /**
 753  * Check whether a Call is at least pure, ie. does only read memory.
 754  */
 755 static unsigned is_Call_pure(ir_node *call) {
 756         ir_type *call_tp = get_Call_type(call);
 757         unsigned prop = get_method_additional_properties(call_tp);
 758
 759         /* check first the call type */
 760         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 761                 /* try the called entity */
 762                 ir_node *ptr = get_Call_ptr(call);
 763
 764                 if (is_Global(ptr)) {
 765                         ir_entity *ent = get_Global_entity(ptr);
 766
 767                         prop = get_entity_additional_properties(ent);
 768                 }
 769         }
 770         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 771 }  /* is_Call_pure */
 772
 773 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 774 {
 775         ir_mode *mode  = get_irn_mode(ptr);
 776         long    offset = 0;
 777
 778         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 779         for (;;) {
 780                 if (is_Add(ptr)) {
 781                         ir_node *l = get_Add_left(ptr);
 782                         ir_node *r = get_Add_right(ptr);
 783
 784                         if (get_irn_mode(l) != mode || !is_Const(r))
 785                                 break;
 786
 787                         offset += get_tarval_long(get_Const_tarval(r));
 788                         ptr     = l;
 789                 } else if (is_Sub(ptr)) {
 790                         ir_node *l = get_Sub_left(ptr);
 791                         ir_node *r = get_Sub_right(ptr);
 792
 793                         if (get_irn_mode(l) != mode || !is_Const(r))
 794                                 break;
 795
 796                         offset -= get_tarval_long(get_Const_tarval(r));
 797                         ptr     = l;
 798                 } else if (is_Sel(ptr)) {
 799                         ir_entity *ent = get_Sel_entity(ptr);
 800                         ir_type   *tp  = get_entity_owner(ent);
 801
 802                         if (is_Array_type(tp)) {
 803                                 int     size;
 804                                 ir_node *index;
 805
 806                                 /* only one dimensional arrays yet */
 807                                 if (get_Sel_n_indexs(ptr) != 1)
 808                                         break;
 809                                 index = get_Sel_index(ptr, 0);
 810                                 if (! is_Const(index))
 811                                         break;
 812
 813                                 tp = get_entity_type(ent);
 814                                 if (get_type_state(tp) != layout_fixed)
 815                                         break;
 816
 817                                 size    = get_type_size_bytes(tp);
 818                                 offset += size * get_tarval_long(get_Const_tarval(index));
 819                         } else {
 820                                 if (get_type_state(tp) != layout_fixed)
 821                                         break;
 822                                 offset += get_entity_offset(ent);
 823                         }
 824                         ptr = get_Sel_ptr(ptr);
 825                 } else
 826                         break;
 827         }
 828
 829         *pOffset = offset;
 830         return ptr;
 831 }
 832
 833 static int try_load_after_store(ir_node *load,
 834                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 835 {
 836         ldst_info_t *info;
 837         ir_node *store_ptr      = get_Store_ptr(store);
 838         long     store_offset;
 839         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 840         ir_node *store_value;
 841         ir_mode *store_mode;
 842         ir_node *load_ptr;
 843         ir_mode *load_mode;
 844         long     load_mode_len;
 845         long     store_mode_len;
 846         long     delta;
 847         int      res;
 848
 849         if (load_base_ptr != store_base_ptr)
 850                 return 0;
 851
 852         load_mode      = get_Load_mode(load);
 853         load_mode_len  = get_mode_size_bytes(load_mode);
 854         store_mode     = get_irn_mode(get_Store_value(store));
 855         store_mode_len = get_mode_size_bytes(store_mode);
 856         delta          = load_offset - store_offset;
 857         store_value    = get_Store_value(store);
 858
 859         if (delta != 0 || store_mode != load_mode) {
 860                 if (delta < 0 || delta + load_mode_len > store_mode_len)
 861                         return 0;
 862
 863                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 864                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 865                         return 0;
 866
 867
 868                 /* produce a shift to adjust offset delta */
 869                 if (delta > 0) {
 870                         ir_node *cnst;
 871
 872                         /* FIXME: only true for little endian */
 873                         cnst        = new_Const_long(mode_Iu, delta * 8);
 874                         store_value = new_r_Shr(get_nodes_block(load),
 875                                                                         store_value, cnst, store_mode);
 876                 }
 877
 878                 /* add an convert if needed */
 879                 if (store_mode != load_mode) {
 880                         store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
 881                 }
 882         }
 883
 884         DBG_OPT_RAW(load, store_value);
 885
 886         info = get_irn_link(load);
 887         if (info->projs[pn_Load_M])
 888                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 889
 890         res = 0;
 891         /* no exception */
 892         if (info->projs[pn_Load_X_except]) {
 893                 exchange( info->projs[pn_Load_X_except], new_Bad());
 894                 res |= CF_CHANGED;
 895         }
 896         if (info->projs[pn_Load_X_regular]) {
 897                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 898                 res |= CF_CHANGED;
 899         }
 900
 901         if (info->projs[pn_Load_res])
 902                 exchange(info->projs[pn_Load_res], store_value);
 903
 904         load_ptr = get_Load_ptr(load);
 905         kill_node(load);
 906         reduce_adr_usage(load_ptr);
 907         return res | DF_CHANGED;
 908 }
 909
 910 /**
 911  * Follow the memory chain as long as there are only Loads,
 912  * alias free Stores, and constant Calls and try to replace the
 913  * current Load by a previous ones.
 914  * Note that in unreachable loops it might happen that we reach
 915  * load again, as well as we can fall into a cycle.
 916  * We break such cycles using a special visited flag.
 917  *
 918  * INC_MASTER() must be called before dive into
 919  */
 920 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 921         unsigned    res = 0;
 922         ldst_info_t *info = get_irn_link(load);
 923         ir_node     *pred;
 924         ir_node     *ptr       = get_Load_ptr(load);
 925         ir_node     *mem       = get_Load_mem(load);
 926         ir_mode     *load_mode = get_Load_mode(load);
 927
 928         for (pred = curr; load != pred; ) {
 929                 ldst_info_t *pred_info = get_irn_link(pred);
 930
 931                 /*
 932                  * a Load immediately after a Store -- a read after write.
 933                  * We may remove the Load, if both Load & Store does not have an
 934                  * exception handler OR they are in the same MacroBlock. In the latter
 935                  * case the Load cannot throw an exception when the previous Store was
 936                  * quiet.
 937                  *
 938                  * Why we need to check for Store Exception? If the Store cannot
 939                  * be executed (ROM) the exception handler might simply jump into
 940                  * the load MacroBlock :-(
 941                  * We could make it a little bit better if we would know that the
 942                  * exception handler of the Store jumps directly to the end...
 943                  */
 944                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 945                                 && info->projs[pn_Load_X_except] == NULL)
 946                                 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
 947                 {
 948                         long    load_offset;
 949                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 950                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 951
 952                         if (changes != 0)
 953                                 return res | changes;
 954                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 955                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 956                         /*
 957                          * a Load after a Load -- a read after read.
 958                          * We may remove the second Load, if it does not have an exception handler
 959                          * OR they are in the same MacroBlock. In the later case the Load cannot
 960                          * throw an exception when the previous Load was quiet.
 961                          *
 962                          * Here, there is no need to check if the previous Load has an exception
 963                          * hander because they would have exact the same exception...
 964                          */
 965                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 966                                 ir_node *value;
 967
 968                                 DBG_OPT_RAR(load, pred);
 969
 970                                 /* the result is used */
 971                                 if (info->projs[pn_Load_res]) {
 972                                         if (pred_info->projs[pn_Load_res] == NULL) {
 973                                                 /* create a new Proj again */
 974                                                 pred_info->projs[pn_Load_res] = new_r_Proj(get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 975                                         }
 976                                         value = pred_info->projs[pn_Load_res];
 977
 978                                         /* add an convert if needed */
 979                                         if (get_Load_mode(pred) != load_mode) {
 980                                                 value = new_r_Conv(get_nodes_block(load), value, load_mode);
 981                                         }
 982
 983                                         exchange(info->projs[pn_Load_res], value);
 984                                 }
 985
 986                                 if (info->projs[pn_Load_M])
 987                                         exchange(info->projs[pn_Load_M], mem);
 988
 989                                 /* no exception */
 990                                 if (info->projs[pn_Load_X_except]) {
 991                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 992                                         res |= CF_CHANGED;
 993                                 }
 994                                 if (info->projs[pn_Load_X_regular]) {
 995                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 996                                         res |= CF_CHANGED;
 997                                 }
 998
 999                                 kill_node(load);
1000                                 reduce_adr_usage(ptr);
1001                                 return res |= DF_CHANGED;
1002                         }
1003                 }
1004
1005                 if (is_Store(pred)) {
1006                         /* check if we can pass through this store */
1007                         ir_alias_relation rel = get_alias_relation(
1008                                 current_ir_graph,
1009                                 get_Store_ptr(pred),
1010                                 get_irn_mode(get_Store_value(pred)),
1011                                 ptr, load_mode);
1012                         /* if the might be an alias, we cannot pass this Store */
1013                         if (rel != ir_no_alias)
1014                                 break;
1015                         pred = skip_Proj(get_Store_mem(pred));
1016                 } else if (is_Load(pred)) {
1017                         pred = skip_Proj(get_Load_mem(pred));
1018                 } else if (is_Call(pred)) {
1019                         if (is_Call_pure(pred)) {
1020                                 /* The called graph is at least pure, so there are no Store's
1021                                    in it. We can handle it like a Load and skip it. */
1022                                 pred = skip_Proj(get_Call_mem(pred));
1023                         } else {
1024                                 /* there might be Store's in the graph, stop here */
1025                                 break;
1026                         }
1027                 } else {
1028                         /* follow only Load chains */
1029                         break;
1030                 }
1031
1032                 /* check for cycles */
1033                 if (NODE_VISITED(pred_info))
1034                         break;
1035                 MARK_NODE(pred_info);
1036         }
1037
1038         if (is_Sync(pred)) {
1039                 int i;
1040
1041                 /* handle all Sync predecessors */
1042                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1043                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1044                         if (res)
1045                                 return res;
1046                 }
1047         }
1048
1049         return res;
1050 }  /* follow_Mem_chain */
1051
1052 /*
1053  * Check if we can replace the load by a given const from
1054  * the const code irg.
1055  */
1056 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1057         ir_mode *c_mode = get_irn_mode(c);
1058         ir_mode *l_mode = get_Load_mode(load);
1059         ir_node *res    = NULL;
1060
1061         if (c_mode != l_mode) {
1062                 /* check, if the mode matches OR can be easily converted info */
1063                 if (is_reinterpret_cast(c_mode, l_mode)) {
1064                         /* we can safely cast */
1065                         dbg_info *dbg   = get_irn_dbg_info(load);
1066                         ir_node  *block = get_nodes_block(load);
1067
1068                         /* copy the value from the const code irg and cast it */
1069                         res = copy_const_value(dbg, c);
1070                         res = new_rd_Conv(dbg, block, res, l_mode);
1071                 }
1072         } else {
1073                 /* copy the value from the const code irg */
1074                 res = copy_const_value(get_irn_dbg_info(load), c);
1075         }
1076         return res;
1077 }  /* can_replace_load_by_const */
1078
1079 /**
1080  * optimize a Load
1081  *
1082  * @param load  the Load node
1083  */
1084 static unsigned optimize_load(ir_node *load)
1085 {
1086         ldst_info_t *info = get_irn_link(load);
1087         ir_node     *mem, *ptr, *value;
1088         ir_entity   *ent;
1089         long        dummy;
1090         unsigned    res = 0;
1091
1092         /* do NOT touch volatile loads for now */
1093         if (get_Load_volatility(load) == volatility_is_volatile)
1094                 return 0;
1095
1096         /* the address of the load to be optimized */
1097         ptr = get_Load_ptr(load);
1098
1099         /*
1100          * Check if we can remove the exception from a Load:
1101          * This can be done, if the address is from an Sel(Alloc) and
1102          * the Sel type is a subtype of the allocated type.
1103          *
1104          * This optimizes some often used OO constructs,
1105          * like x = new O; x->t;
1106          */
1107         if (info->projs[pn_Load_X_except]) {
1108                 ir_node *addr = ptr;
1109
1110                 /* find base address */
1111                 while (is_Sel(addr))
1112                         addr = get_Sel_ptr(addr);
1113                 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1114                         /* simple case: a direct load after an Alloc. Firm Alloc throw
1115                          * an exception in case of out-of-memory. So, there is no way for an
1116                          * exception in this load.
1117                          * This code is constructed by the "exception lowering" in the Jack compiler.
1118                          */
1119                         exchange(info->projs[pn_Load_X_except], new_Bad());
1120                         info->projs[pn_Load_X_except] = NULL;
1121                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1122                         info->projs[pn_Load_X_regular] = NULL;
1123                         res |= CF_CHANGED;
1124                 }
1125         }
1126
1127         /* The mem of the Load. Must still be returned after optimization. */
1128         mem = get_Load_mem(load);
1129
1130         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1131                 /* a Load which value is neither used nor exception checked, remove it */
1132                 exchange(info->projs[pn_Load_M], mem);
1133
1134                 if (info->projs[pn_Load_X_regular]) {
1135                         /* should not happen, but if it does, remove it */
1136                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1137                         res |= CF_CHANGED;
1138                 }
1139                 kill_node(load);
1140                 reduce_adr_usage(ptr);
1141                 return res | DF_CHANGED;
1142         }
1143
1144         /* Load from a constant polymorphic field, where we can resolve
1145            polymorphism. */
1146         value = transform_polymorph_Load(load);
1147         if (value == load) {
1148                 value = NULL;
1149                 /* check if we can determine the entity that will be loaded */
1150                 ent = find_constant_entity(ptr);
1151                 if (ent != NULL                                     &&
1152                     allocation_static == get_entity_allocation(ent) &&
1153                     visibility_external_allocated != get_entity_visibility(ent)) {
1154                         /* a static allocation that is not external: there should be NO exception
1155                          * when loading even if we cannot replace the load itself. */
1156
1157                         /* no exception, clear the info field as it might be checked later again */
1158                         if (info->projs[pn_Load_X_except]) {
1159                                 exchange(info->projs[pn_Load_X_except], new_Bad());
1160                                 info->projs[pn_Load_X_except] = NULL;
1161                                 res |= CF_CHANGED;
1162                         }
1163                         if (info->projs[pn_Load_X_regular]) {
1164                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1165                                 info->projs[pn_Load_X_regular] = NULL;
1166                                 res |= CF_CHANGED;
1167                         }
1168
1169                         if (variability_constant == get_entity_variability(ent)) {
1170                                 if (is_atomic_entity(ent)) {
1171                                         /* Might not be atomic after lowering of Sels.  In this case we
1172                                          * could also load, but it's more complicated. */
1173                                         /* more simpler case: we load the content of a constant value:
1174                                          * replace it by the constant itself */
1175                                         value = get_atomic_ent_value(ent);
1176                                 } else if (ent->has_initializer) {
1177                                         /* new style initializer */
1178                                         value = find_compound_ent_value(ptr);
1179                                 } else {
1180                                         /* old style initializer */
1181                                         compound_graph_path *path = get_accessed_path(ptr);
1182
1183                                         if (path != NULL) {
1184                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1185
1186                                                 value = get_compound_ent_value_by_path(ent, path);
1187                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1188                                                 free_compound_graph_path(path);
1189                                         }
1190                                 }
1191                                 if (value != NULL)
1192                                         value = can_replace_load_by_const(load, value);
1193                         }
1194                 }
1195         }
1196         if (value != NULL) {
1197                 /* we completely replace the load by this value */
1198                 if (info->projs[pn_Load_X_except]) {
1199                         exchange(info->projs[pn_Load_X_except], new_Bad());
1200                         info->projs[pn_Load_X_except] = NULL;
1201                         res |= CF_CHANGED;
1202                 }
1203                 if (info->projs[pn_Load_X_regular]) {
1204                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1205                         info->projs[pn_Load_X_regular] = NULL;
1206                         res |= CF_CHANGED;
1207                 }
1208                 if (info->projs[pn_Load_M]) {
1209                         exchange(info->projs[pn_Load_M], mem);
1210                         res |= DF_CHANGED;
1211                 }
1212                 if (info->projs[pn_Load_res]) {
1213                         exchange(info->projs[pn_Load_res], value);
1214                         res |= DF_CHANGED;
1215                 }
1216                 kill_node(load);
1217                 reduce_adr_usage(ptr);
1218                 return res;
1219         }
1220
1221         /* Check, if the address of this load is used more than once.
1222          * If not, more load cannot be removed in any case. */
1223         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1224                 return res;
1225
1226         /*
1227          * follow the memory chain as long as there are only Loads
1228          * and try to replace current Load or Store by a previous one.
1229          * Note that in unreachable loops it might happen that we reach
1230          * load again, as well as we can fall into a cycle.
1231          * We break such cycles using a special visited flag.
1232          */
1233         INC_MASTER();
1234         res = follow_Mem_chain(load, skip_Proj(mem));
1235         return res;
1236 }  /* optimize_load */
1237
1238 /**
1239  * Check whether a value of mode new_mode would completely overwrite a value
1240  * of mode old_mode in memory.
1241  */
1242 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1243 {
1244         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1245 }  /* is_completely_overwritten */
1246
1247 /**
1248  * Check whether small is a part of large (starting at same address).
1249  */
1250 static int is_partially_same(ir_node *small, ir_node *large)
1251 {
1252         ir_mode *sm = get_irn_mode(small);
1253         ir_mode *lm = get_irn_mode(large);
1254
1255         /* FIXME: Check endianness */
1256         return is_Conv(small) && get_Conv_op(small) == large
1257             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1258             && get_mode_arithmetic(sm) == irma_twos_complement
1259             && get_mode_arithmetic(lm) == irma_twos_complement;
1260 }  /* is_partially_same */
1261
1262 /**
1263  * follow the memory chain as long as there are only Loads and alias free Stores.
1264  *
1265  * INC_MASTER() must be called before dive into
1266  */
1267 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1268         unsigned res = 0;
1269         ldst_info_t *info = get_irn_link(store);
1270         ir_node *pred;
1271         ir_node *ptr = get_Store_ptr(store);
1272         ir_node *mem = get_Store_mem(store);
1273         ir_node *value = get_Store_value(store);
1274         ir_mode *mode  = get_irn_mode(value);
1275         ir_node *block = get_nodes_block(store);
1276         ir_node *mblk  = get_Block_MacroBlock(block);
1277
1278         for (pred = curr; pred != store;) {
1279                 ldst_info_t *pred_info = get_irn_link(pred);
1280
1281                 /*
1282                  * BEWARE: one might think that checking the modes is useless, because
1283                  * if the pointers are identical, they refer to the same object.
1284                  * This is only true in strong typed languages, not is C were the following
1285                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1286                  * However, if the size of the mode that is written is bigger or equal the
1287                  * size of the old one, the old value is completely overwritten and can be
1288                  * killed ...
1289                  */
1290                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1291             get_nodes_MacroBlock(pred) == mblk) {
1292                         /*
1293                          * a Store after a Store in the same MacroBlock -- a write after write.
1294                          */
1295
1296                         /*
1297                          * We may remove the first Store, if the old value is completely
1298                          * overwritten or the old value is a part of the new value,
1299                          * and if it does not have an exception handler.
1300                          *
1301                          * TODO: What, if both have the same exception handler ???
1302                          */
1303                         if (get_Store_volatility(pred) != volatility_is_volatile
1304                                 && !pred_info->projs[pn_Store_X_except]) {
1305                                 ir_node *predvalue = get_Store_value(pred);
1306                                 ir_mode *predmode  = get_irn_mode(predvalue);
1307
1308                                 if(is_completely_overwritten(predmode, mode)
1309                                         || is_partially_same(predvalue, value)) {
1310                                         DBG_OPT_WAW(pred, store);
1311                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1312                                         kill_node(pred);
1313                                         reduce_adr_usage(ptr);
1314                                         return DF_CHANGED;
1315                                 }
1316                         }
1317
1318                         /*
1319                          * We may remove the Store, if the old value already contains
1320                          * the new value, and if it does not have an exception handler.
1321                          *
1322                          * TODO: What, if both have the same exception handler ???
1323                          */
1324                         if (get_Store_volatility(store) != volatility_is_volatile
1325                                 && !info->projs[pn_Store_X_except]) {
1326                                 ir_node *predvalue = get_Store_value(pred);
1327
1328                                 if(is_partially_same(value, predvalue)) {
1329                                         DBG_OPT_WAW(pred, store);
1330                                         exchange(info->projs[pn_Store_M], mem);
1331                                         kill_node(store);
1332                                         reduce_adr_usage(ptr);
1333                                         return DF_CHANGED;
1334                                 }
1335                         }
1336                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1337                            value == pred_info->projs[pn_Load_res]) {
1338                         /*
1339                          * a Store of a value just loaded from the same address
1340                          * -- a write after read.
1341                          * We may remove the Store, if it does not have an exception
1342                          * handler.
1343                          */
1344                         if (! info->projs[pn_Store_X_except]) {
1345                                 DBG_OPT_WAR(store, pred);
1346                                 exchange(info->projs[pn_Store_M], mem);
1347                                 kill_node(store);
1348                                 reduce_adr_usage(ptr);
1349                                 return DF_CHANGED;
1350                         }
1351                 }
1352
1353                 if (is_Store(pred)) {
1354                         /* check if we can pass through this store */
1355                         ir_alias_relation rel = get_alias_relation(
1356                                 current_ir_graph,
1357                                 get_Store_ptr(pred),
1358                                 get_irn_mode(get_Store_value(pred)),
1359                                 ptr, mode);
1360                         /* if the might be an alias, we cannot pass this Store */
1361                         if (rel != ir_no_alias)
1362                                 break;
1363                         pred = skip_Proj(get_Store_mem(pred));
1364                 } else if (is_Load(pred)) {
1365                         ir_alias_relation rel = get_alias_relation(
1366                                 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1367                                 ptr, mode);
1368                         if (rel != ir_no_alias)
1369                                 break;
1370
1371                         pred = skip_Proj(get_Load_mem(pred));
1372                 } else {
1373                         /* follow only Load chains */
1374                         break;
1375                 }
1376
1377                 /* check for cycles */
1378                 if (NODE_VISITED(pred_info))
1379                         break;
1380                 MARK_NODE(pred_info);
1381         }
1382
1383         if (is_Sync(pred)) {
1384                 int i;
1385
1386                 /* handle all Sync predecessors */
1387                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1388                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1389                         if (res)
1390                                 break;
1391                 }
1392         }
1393         return res;
1394 }  /* follow_Mem_chain_for_Store */
1395
1396 /** find entity used as base for an address calculation */
1397 static ir_entity *find_entity(ir_node *ptr)
1398 {
1399         switch(get_irn_opcode(ptr)) {
1400         case iro_SymConst:
1401                 return get_SymConst_entity(ptr);
1402         case iro_Sel: {
1403                 ir_node *pred = get_Sel_ptr(ptr);
1404                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1405                         return get_Sel_entity(ptr);
1406
1407                 return find_entity(pred);
1408         }
1409         case iro_Sub:
1410         case iro_Add: {
1411                 ir_node *left = get_binop_left(ptr);
1412                 ir_node *right;
1413                 if (mode_is_reference(get_irn_mode(left)))
1414                         return find_entity(left);
1415                 right = get_binop_right(ptr);
1416                 if (mode_is_reference(get_irn_mode(right)))
1417                         return find_entity(right);
1418                 return NULL;
1419         }
1420         default:
1421                 return NULL;
1422         }
1423 }
1424
1425 /**
1426  * optimize a Store
1427  *
1428  * @param store  the Store node
1429  */
1430 static unsigned optimize_store(ir_node *store) {
1431         ir_node   *ptr;
1432         ir_node   *mem;
1433         ir_entity *entity;
1434
1435         if (get_Store_volatility(store) == volatility_is_volatile)
1436                 return 0;
1437
1438         ptr    = get_Store_ptr(store);
1439         entity = find_entity(ptr);
1440
1441         /* a store to an entity which is never read is unnecessary */
1442         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1443                 ldst_info_t *info = get_irn_link(store);
1444                 if (info->projs[pn_Store_X_except] == NULL) {
1445                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1446                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1447                         kill_node(store);
1448                         reduce_adr_usage(ptr);
1449                         return DF_CHANGED;
1450                 }
1451         }
1452
1453         /* Check, if the address of this Store is used more than once.
1454          * If not, this Store cannot be removed in any case. */
1455         if (get_irn_n_uses(ptr) <= 1)
1456                 return 0;
1457
1458         mem = get_Store_mem(store);
1459
1460         /* follow the memory chain as long as there are only Loads */
1461         INC_MASTER();
1462
1463         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1464 }  /* optimize_store */
1465
1466 /**
1467  * walker, optimizes Phi after Stores to identical places:
1468  * Does the following optimization:
1469  * @verbatim
1470  *
1471  *   val1   val2   val3          val1  val2  val3
1472  *    |      |      |               \    |    /
1473  *  Store  Store  Store              \   |   /
1474  *      \    |    /                   PhiData
1475  *       \   |   /                       |
1476  *        \  |  /                      Store
1477  *          PhiM
1478  *
1479  * @endverbatim
1480  * This reduces the number of stores and allows for predicated execution.
1481  * Moves Stores back to the end of a function which may be bad.
1482  *
1483  * This is only possible if the predecessor blocks have only one successor.
1484  */
1485 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1486 {
1487         int i, n;
1488         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1489         ir_mode *mode;
1490         ir_node **inM, **inD, **projMs;
1491         int *idx;
1492         dbg_info *db = NULL;
1493         ldst_info_t *info;
1494         block_info_t *bl_info;
1495         unsigned res = 0;
1496
1497         /* Must be a memory Phi */
1498         if (get_irn_mode(phi) != mode_M)
1499                 return 0;
1500
1501         n = get_Phi_n_preds(phi);
1502         if (n <= 0)
1503                 return 0;
1504
1505         /* must be only one user */
1506         projM = get_Phi_pred(phi, 0);
1507         if (get_irn_n_edges(projM) != 1)
1508                 return 0;
1509
1510         store = skip_Proj(projM);
1511         old_store = store;
1512         if (!is_Store(store))
1513                 return 0;
1514
1515         block = get_nodes_block(store);
1516
1517         /* abort on dead blocks */
1518         if (is_Block_dead(block))
1519                 return 0;
1520
1521         /* check if the block is post dominated by Phi-block
1522            and has no exception exit */
1523         bl_info = get_irn_link(block);
1524         if (bl_info->flags & BLOCK_HAS_EXC)
1525                 return 0;
1526
1527         phi_block = get_nodes_block(phi);
1528         if (! block_strictly_postdominates(phi_block, block))
1529                 return 0;
1530
1531         /* this is the address of the store */
1532         ptr  = get_Store_ptr(store);
1533         mode = get_irn_mode(get_Store_value(store));
1534         info = get_irn_link(store);
1535         exc  = info->exc_block;
1536
1537         for (i = 1; i < n; ++i) {
1538                 ir_node *pred = get_Phi_pred(phi, i);
1539
1540                 if (get_irn_n_edges(pred) != 1)
1541                         return 0;
1542
1543                 pred = skip_Proj(pred);
1544                 if (!is_Store(pred))
1545                         return 0;
1546
1547                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1548                         return 0;
1549
1550                 info = get_irn_link(pred);
1551
1552                 /* check, if all stores have the same exception flow */
1553                 if (exc != info->exc_block)
1554                         return 0;
1555
1556                 /* abort on dead blocks */
1557                 block = get_nodes_block(pred);
1558                 if (is_Block_dead(block))
1559                         return 0;
1560
1561                 /* check if the block is post dominated by Phi-block
1562                    and has no exception exit. Note that block must be different from
1563                    Phi-block, else we would move a Store from end End of a block to its
1564                    Start... */
1565                 bl_info = get_irn_link(block);
1566                 if (bl_info->flags & BLOCK_HAS_EXC)
1567                         return 0;
1568                 if (block == phi_block || ! block_postdominates(phi_block, block))
1569                         return 0;
1570         }
1571
1572         /*
1573          * ok, when we are here, we found all predecessors of a Phi that
1574          * are Stores to the same address and size. That means whatever
1575          * we do before we enter the block of the Phi, we do a Store.
1576          * So, we can move the Store to the current block:
1577          *
1578          *   val1    val2    val3          val1  val2  val3
1579          *    |       |       |               \    |    /
1580          * | Str | | Str | | Str |             \   |   /
1581          *      \     |     /                   PhiData
1582          *       \    |    /                       |
1583          *        \   |   /                       Str
1584          *           PhiM
1585          *
1586          * Is only allowed if the predecessor blocks have only one successor.
1587          */
1588
1589         NEW_ARR_A(ir_node *, projMs, n);
1590         NEW_ARR_A(ir_node *, inM, n);
1591         NEW_ARR_A(ir_node *, inD, n);
1592         NEW_ARR_A(int, idx, n);
1593
1594         /* Prepare: Collect all Store nodes.  We must do this
1595            first because we otherwise may loose a store when exchanging its
1596            memory Proj.
1597          */
1598         for (i = n - 1; i >= 0; --i) {
1599                 ir_node *store;
1600
1601                 projMs[i] = get_Phi_pred(phi, i);
1602                 assert(is_Proj(projMs[i]));
1603
1604                 store = get_Proj_pred(projMs[i]);
1605                 info  = get_irn_link(store);
1606
1607                 inM[i] = get_Store_mem(store);
1608                 inD[i] = get_Store_value(store);
1609                 idx[i] = info->exc_idx;
1610         }
1611         block = get_nodes_block(phi);
1612
1613         /* second step: create a new memory Phi */
1614         phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1615
1616         /* third step: create a new data Phi */
1617         phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1618
1619         /* rewire memory and kill the node */
1620         for (i = n - 1; i >= 0; --i) {
1621                 ir_node *proj  = projMs[i];
1622
1623                 if(is_Proj(proj)) {
1624                         ir_node *store = get_Proj_pred(proj);
1625                         exchange(proj, inM[i]);
1626                         kill_node(store);
1627                 }
1628         }
1629
1630         /* fourth step: create the Store */
1631         store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1632 #ifdef DO_CACHEOPT
1633         co_set_irn_name(store, co_get_irn_ident(old_store));
1634 #endif
1635
1636         projM = new_rd_Proj(NULL, block, store, mode_M, pn_Store_M);
1637
1638         info = get_ldst_info(store, &wenv->obst);
1639         info->projs[pn_Store_M] = projM;
1640
1641         /* fifths step: repair exception flow */
1642         if (exc) {
1643                 ir_node *projX = new_rd_Proj(NULL, block, store, mode_X, pn_Store_X_except);
1644
1645                 info->projs[pn_Store_X_except] = projX;
1646                 info->exc_block                = exc;
1647                 info->exc_idx                  = idx[0];
1648
1649                 for (i = 0; i < n; ++i) {
1650                         set_Block_cfgpred(exc, idx[i], projX);
1651                 }
1652
1653                 if (n > 1) {
1654                         /* the exception block should be optimized as some inputs are identical now */
1655                 }
1656
1657                 res |= CF_CHANGED;
1658         }
1659
1660         /* sixth step: replace old Phi */
1661         exchange(phi, projM);
1662
1663         return res | DF_CHANGED;
1664 }  /* optimize_phi */
1665
1666 /**
1667  * walker, do the optimizations
1668  */
1669 static void do_load_store_optimize(ir_node *n, void *env) {
1670         walk_env_t *wenv = env;
1671
1672         switch (get_irn_opcode(n)) {
1673
1674         case iro_Load:
1675                 wenv->changes |= optimize_load(n);
1676                 break;
1677
1678         case iro_Store:
1679                 wenv->changes |= optimize_store(n);
1680                 break;
1681
1682         case iro_Phi:
1683                 wenv->changes |= optimize_phi(n, wenv);
1684                 break;
1685
1686         default:
1687                 ;
1688         }
1689 }  /* do_load_store_optimize */
1690
1691 /** A scc. */
1692 typedef struct scc {
1693         ir_node *head;          /**< the head of the list */
1694 } scc;
1695
1696 /** A node entry. */
1697 typedef struct node_entry {
1698         unsigned DFSnum;    /**< the DFS number of this node */
1699         unsigned low;       /**< the low number of this node */
1700         int      in_stack;  /**< flag, set if the node is on the stack */
1701         ir_node  *next;     /**< link to the next node the the same scc */
1702         scc      *pscc;     /**< the scc of this node */
1703         unsigned POnum;     /**< the post order number for blocks */
1704 } node_entry;
1705
1706 /** A loop entry. */
1707 typedef struct loop_env {
1708         ir_phase ph;           /**< the phase object */
1709         ir_node  **stack;      /**< the node stack */
1710         int      tos;          /**< tos index */
1711         unsigned nextDFSnum;   /**< the current DFS number */
1712         unsigned POnum;        /**< current post order number */
1713
1714         unsigned changes;      /**< a bitmask of graph changes */
1715 } loop_env;
1716
1717 /**
1718 * Gets the node_entry of a node
1719 */
1720 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1721         ir_phase   *ph = &env->ph;
1722         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1723
1724         if (! e) {
1725                 e = phase_alloc(ph, sizeof(*e));
1726                 memset(e, 0, sizeof(*e));
1727                 phase_set_irn_data(ph, irn, e);
1728         }
1729         return e;
1730 }  /* get_irn_ne */
1731
1732 /**
1733  * Push a node onto the stack.
1734  *
1735  * @param env   the loop environment
1736  * @param n     the node to push
1737  */
1738 static void push(loop_env *env, ir_node *n) {
1739         node_entry *e;
1740
1741         if (env->tos == ARR_LEN(env->stack)) {
1742                 int nlen = ARR_LEN(env->stack) * 2;
1743                 ARR_RESIZE(ir_node *, env->stack, nlen);
1744         }
1745         env->stack[env->tos++] = n;
1746         e = get_irn_ne(n, env);
1747         e->in_stack = 1;
1748 }  /* push */
1749
1750 /**
1751  * pop a node from the stack
1752  *
1753  * @param env   the loop environment
1754  *
1755  * @return  The topmost node
1756  */
1757 static ir_node *pop(loop_env *env) {
1758         ir_node *n = env->stack[--env->tos];
1759         node_entry *e = get_irn_ne(n, env);
1760
1761         e->in_stack = 0;
1762         return n;
1763 }  /* pop */
1764
1765 /**
1766  * Check if irn is a region constant.
1767  * The block or irn must strictly dominate the header block.
1768  *
1769  * @param irn           the node to check
1770  * @param header_block  the header block of the induction variable
1771  */
1772 static int is_rc(ir_node *irn, ir_node *header_block) {
1773         ir_node *block = get_nodes_block(irn);
1774
1775         return (block != header_block) && block_dominates(block, header_block);
1776 }  /* is_rc */
1777
1778 typedef struct phi_entry phi_entry;
1779 struct phi_entry {
1780         ir_node   *phi;    /**< A phi with a region const memory. */
1781         int       pos;     /**< The position of the region const memory */
1782         ir_node   *load;   /**< the newly created load for this phi */
1783         phi_entry *next;
1784 };
1785
1786 /**
1787  * An entry in the avail set.
1788  */
1789 typedef struct avail_entry_t {
1790         ir_node *ptr;   /**< the address pointer */
1791         ir_mode *mode;  /**< the load mode */
1792         ir_node *load;  /**< the associated Load */
1793 } avail_entry_t;
1794
1795 /**
1796  * Compare two avail entries.
1797  */
1798 static int cmp_avail_entry(const void *elt, const void *key, size_t size) {
1799         const avail_entry_t *a = elt;
1800         const avail_entry_t *b = key;
1801         (void) size;
1802
1803         return a->ptr != b->ptr || a->mode != b->mode;
1804 }  /* cmp_avail_entry */
1805
1806 /**
1807  * Calculate the hash value of an avail entry.
1808  */
1809 static unsigned hash_cache_entry(const avail_entry_t *entry) {
1810         return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1811 }  /* hash_cache_entry */
1812
1813 /**
1814  * Move loops out of loops if possible.
1815  *
1816  * @param pscc   the loop described by an SCC
1817  * @param env    the loop environment
1818  */
1819 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1820         ir_node   *phi, *load, *next, *other, *next_other;
1821         ir_entity *ent;
1822         int       j;
1823         phi_entry *phi_list = NULL;
1824         set       *avail;
1825
1826         avail = new_set(cmp_avail_entry, 8);
1827
1828         /* collect all outer memories */
1829         for (phi = pscc->head; phi != NULL; phi = next) {
1830                 node_entry *ne = get_irn_ne(phi, env);
1831                 next = ne->next;
1832
1833                 /* check all memory Phi's */
1834                 if (! is_Phi(phi))
1835                         continue;
1836
1837                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1838
1839                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1840                         ir_node    *pred = get_irn_n(phi, j);
1841                         node_entry *pe   = get_irn_ne(pred, env);
1842
1843                         if (pe->pscc != ne->pscc) {
1844                                 /* not in the same SCC, is region const */
1845                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1846
1847                                 pe->phi  = phi;
1848                                 pe->pos  = j;
1849                                 pe->next = phi_list;
1850                                 phi_list = pe;
1851                         }
1852                 }
1853         }
1854         /* no Phis no fun */
1855         assert(phi_list != NULL && "DFS found a loop without Phi");
1856
1857         /* for now, we cannot handle more than one input (only reducible cf) */
1858         if (phi_list->next != NULL)
1859                 return;
1860
1861         for (load = pscc->head; load; load = next) {
1862                 ir_mode *load_mode;
1863                 node_entry *ne = get_irn_ne(load, env);
1864                 next = ne->next;
1865
1866                 if (is_Load(load)) {
1867                         ldst_info_t *info = get_irn_link(load);
1868                         ir_node     *ptr = get_Load_ptr(load);
1869
1870                         /* for now, we cannot handle Loads with exceptions */
1871                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1872                                 continue;
1873
1874                         /* for now, we can only move Load(Global) */
1875                         if (! is_Global(ptr))
1876                                 continue;
1877                         ent       = get_Global_entity(ptr);
1878                         load_mode = get_Load_mode(load);
1879                         for (other = pscc->head; other != NULL; other = next_other) {
1880                                 node_entry *ne = get_irn_ne(other, env);
1881                                 next_other = ne->next;
1882
1883                                 if (is_Store(other)) {
1884                                         ir_alias_relation rel = get_alias_relation(
1885                                                 current_ir_graph,
1886                                                 get_Store_ptr(other),
1887                                                 get_irn_mode(get_Store_value(other)),
1888                                                 ptr, load_mode);
1889                                         /* if the might be an alias, we cannot pass this Store */
1890                                         if (rel != ir_no_alias)
1891                                                 break;
1892                                 }
1893                                 /* only Phis and pure Calls are allowed here, so ignore them */
1894                         }
1895                         if (other == NULL) {
1896                                 ldst_info_t *ninfo;
1897                                 phi_entry   *pe;
1898                                 dbg_info    *db;
1899
1900                                 /* yep, no aliasing Store found, Load can be moved */
1901                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1902
1903                                 db   = get_irn_dbg_info(load);
1904                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1905                                         int     pos   = pe->pos;
1906                                         ir_node *phi  = pe->phi;
1907                                         ir_node *blk  = get_nodes_block(phi);
1908                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1909                                         ir_node *irn, *mem;
1910                                         avail_entry_t entry, *res;
1911
1912                                         entry.ptr  = ptr;
1913                                         entry.mode = load_mode;
1914                                         res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1915                                         if (res != NULL) {
1916                                                 irn = res->load;
1917                                         } else {
1918                                                 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1919                                                 entry.load = irn;
1920                                                 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1921                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1922                                         }
1923                                         pe->load = irn;
1924                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1925
1926                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(pred, irn, mode_M, pn_Load_M);
1927                                         set_Phi_pred(phi, pos, mem);
1928
1929                                         ninfo->projs[pn_Load_res] = new_r_Proj(pred, irn, load_mode, pn_Load_res);
1930                                 }
1931
1932                                 /* now kill the old Load */
1933                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1934                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1935
1936                                 env->changes |= DF_CHANGED;
1937                         }
1938                 }
1939         }
1940         del_set(avail);
1941 }  /* move_loads_out_of_loops */
1942
1943 /**
1944  * Process a loop SCC.
1945  *
1946  * @param pscc  the SCC
1947  * @param env   the loop environment
1948  */
1949 static void process_loop(scc *pscc, loop_env *env) {
1950         ir_node *irn, *next, *header = NULL;
1951         node_entry *b, *h = NULL;
1952         int j, only_phi, num_outside, process = 0;
1953         ir_node *out_rc;
1954
1955         /* find the header block for this scc */
1956         for (irn = pscc->head; irn; irn = next) {
1957                 node_entry *e = get_irn_ne(irn, env);
1958                 ir_node *block = get_nodes_block(irn);
1959
1960                 next = e->next;
1961                 b = get_irn_ne(block, env);
1962
1963                 if (header != NULL) {
1964                         if (h->POnum < b->POnum) {
1965                                 header = block;
1966                                 h      = b;
1967                         }
1968                 } else {
1969                         header = block;
1970                         h      = b;
1971                 }
1972         }
1973
1974         /* check if this scc contains only Phi, Loads or Stores nodes */
1975         only_phi    = 1;
1976         num_outside = 0;
1977         out_rc      = NULL;
1978         for (irn = pscc->head; irn; irn = next) {
1979                 node_entry *e = get_irn_ne(irn, env);
1980
1981                 next = e->next;
1982                 switch (get_irn_opcode(irn)) {
1983                 case iro_Call:
1984                         if (is_Call_pure(irn)) {
1985                                 /* pure calls can be treated like loads */
1986                                 only_phi = 0;
1987                                 break;
1988                         }
1989                         /* non-pure calls must be handle like may-alias Stores */
1990                         goto fail;
1991                 case iro_CopyB:
1992                         /* cannot handle CopyB yet */
1993                         goto fail;
1994                 case iro_Load:
1995                         process = 1;
1996                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1997                                 /* cannot handle loops with volatile Loads */
1998                                 goto fail;
1999                         }
2000                         only_phi = 0;
2001                         break;
2002                 case iro_Store:
2003                         if (get_Store_volatility(irn) == volatility_is_volatile) {
2004                                 /* cannot handle loops with volatile Stores */
2005                                 goto fail;
2006                         }
2007                         only_phi = 0;
2008                         break;
2009                 default:
2010                         only_phi = 0;
2011                         break;
2012                 case iro_Phi:
2013                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2014                                 ir_node *pred  = get_irn_n(irn, j);
2015                                 node_entry *pe = get_irn_ne(pred, env);
2016
2017                                 if (pe->pscc != e->pscc) {
2018                                         /* not in the same SCC, must be a region const */
2019                                         if (! is_rc(pred, header)) {
2020                                                 /* not a memory loop */
2021                                                 goto fail;
2022                                         }
2023                                         if (out_rc == NULL) {
2024                                                 /* first region constant */
2025                                                 out_rc = pred;
2026                                                 ++num_outside;
2027                                         } else if (out_rc != pred) {
2028                                                 /* another region constant */
2029                                                 ++num_outside;
2030                                         }
2031                                 }
2032                         }
2033                         break;
2034                 }
2035         }
2036         if (! process)
2037                 goto fail;
2038
2039         /* found a memory loop */
2040         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
2041         if (only_phi && num_outside == 1) {
2042                 /* a phi cycle with only one real predecessor can be collapsed */
2043                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
2044
2045                 for (irn = pscc->head; irn; irn = next) {
2046                         node_entry *e = get_irn_ne(irn, env);
2047                         next = e->next;
2048                         exchange(irn, out_rc);
2049                 }
2050                 env->changes |= DF_CHANGED;
2051                 return;
2052         }
2053
2054 #ifdef DEBUG_libfirm
2055         for (irn = pscc->head; irn; irn = next) {
2056                 node_entry *e = get_irn_ne(irn, env);
2057                 next = e->next;
2058                 DB((dbg, LEVEL_2, " %+F,", irn));
2059         }
2060         DB((dbg, LEVEL_2, "\n"));
2061 #endif
2062         move_loads_out_of_loops(pscc, env);
2063
2064 fail:
2065         ;
2066 }  /* process_loop */
2067
2068 /**
2069  * Process a SCC.
2070  *
2071  * @param pscc  the SCC
2072  * @param env   the loop environment
2073  */
2074 static void process_scc(scc *pscc, loop_env *env) {
2075         ir_node *head = pscc->head;
2076         node_entry *e = get_irn_ne(head, env);
2077
2078 #ifdef DEBUG_libfirm
2079         {
2080                 ir_node *irn, *next;
2081
2082                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2083                 for (irn = pscc->head; irn; irn = next) {
2084                         node_entry *e = get_irn_ne(irn, env);
2085
2086                         next = e->next;
2087
2088                         DB((dbg, LEVEL_4, " %+F,", irn));
2089                 }
2090                 DB((dbg, LEVEL_4, "\n"));
2091         }
2092 #endif
2093
2094         if (e->next != NULL) {
2095                 /* this SCC has more than one member */
2096                 process_loop(pscc, env);
2097         }
2098 }  /* process_scc */
2099
2100 /**
2101  * Do Tarjan's SCC algorithm and drive load/store optimization.
2102  *
2103  * @param irn  start at this node
2104  * @param env  the loop environment
2105  */
2106 static void dfs(ir_node *irn, loop_env *env)
2107 {
2108         int i, n;
2109         node_entry *node = get_irn_ne(irn, env);
2110
2111         mark_irn_visited(irn);
2112
2113         node->DFSnum = env->nextDFSnum++;
2114         node->low    = node->DFSnum;
2115         push(env, irn);
2116
2117         /* handle preds */
2118         if (is_Phi(irn) || is_Sync(irn)) {
2119                 n = get_irn_arity(irn);
2120                 for (i = 0; i < n; ++i) {
2121                         ir_node *pred = get_irn_n(irn, i);
2122                         node_entry *o = get_irn_ne(pred, env);
2123
2124                         if (!irn_visited(pred)) {
2125                                 dfs(pred, env);
2126                                 node->low = MIN(node->low, o->low);
2127                         }
2128                         if (o->DFSnum < node->DFSnum && o->in_stack)
2129                                 node->low = MIN(o->DFSnum, node->low);
2130                 }
2131         } else if (is_fragile_op(irn)) {
2132                 ir_node *pred = get_fragile_op_mem(irn);
2133                 node_entry *o = get_irn_ne(pred, env);
2134
2135                 if (!irn_visited(pred)) {
2136                         dfs(pred, env);
2137                         node->low = MIN(node->low, o->low);
2138                 }
2139                 if (o->DFSnum < node->DFSnum && o->in_stack)
2140                         node->low = MIN(o->DFSnum, node->low);
2141         } else if (is_Proj(irn)) {
2142                 ir_node *pred = get_Proj_pred(irn);
2143                 node_entry *o = get_irn_ne(pred, env);
2144
2145                 if (!irn_visited(pred)) {
2146                         dfs(pred, env);
2147                         node->low = MIN(node->low, o->low);
2148                 }
2149                 if (o->DFSnum < node->DFSnum && o->in_stack)
2150                         node->low = MIN(o->DFSnum, node->low);
2151         }
2152         else {
2153                  /* IGNORE predecessors */
2154         }
2155
2156         if (node->low == node->DFSnum) {
2157                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2158                 ir_node *x;
2159
2160                 pscc->head = NULL;
2161                 do {
2162                         node_entry *e;
2163
2164                         x = pop(env);
2165                         e = get_irn_ne(x, env);
2166                         e->pscc    = pscc;
2167                         e->next    = pscc->head;
2168                         pscc->head = x;
2169                 } while (x != irn);
2170
2171                 process_scc(pscc, env);
2172         }
2173 }  /* dfs */
2174
2175 /**
2176  * Do the DFS on the memory edges a graph.
2177  *
2178  * @param irg  the graph to process
2179  * @param env  the loop environment
2180  */
2181 static void do_dfs(ir_graph *irg, loop_env *env) {
2182         ir_graph *rem = current_ir_graph;
2183         ir_node  *endblk, *end;
2184         int      i;
2185
2186         current_ir_graph = irg;
2187         inc_irg_visited(irg);
2188
2189         /* visit all memory nodes */
2190         endblk = get_irg_end_block(irg);
2191         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2192                 ir_node *pred = get_Block_cfgpred(endblk, i);
2193
2194                 pred = skip_Proj(pred);
2195                 if (is_Return(pred))
2196                         dfs(get_Return_mem(pred), env);
2197                 else if (is_Raise(pred))
2198                         dfs(get_Raise_mem(pred), env);
2199                 else if (is_fragile_op(pred))
2200                         dfs(get_fragile_op_mem(pred), env);
2201                 else {
2202                         assert(0 && "Unknown EndBlock predecessor");
2203                 }
2204         }
2205
2206         /* visit the keep-alives */
2207         end = get_irg_end(irg);
2208         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2209                 ir_node *ka = get_End_keepalive(end, i);
2210
2211                 if (is_Phi(ka) && !irn_visited(ka))
2212                         dfs(ka, env);
2213         }
2214         current_ir_graph = rem;
2215 }  /* do_dfs */
2216
2217 /**
2218  * Initialize new phase data. We do this always explicit, so return NULL here
2219  */
2220 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2221         (void)ph;
2222         (void)irn;
2223         (void)data;
2224         return NULL;
2225 }  /* init_loop_data */
2226
2227 /**
2228  * Optimize Loads/Stores in loops.
2229  *
2230  * @param irg  the graph
2231  */
2232 static int optimize_loops(ir_graph *irg) {
2233         loop_env env;
2234
2235         env.stack         = NEW_ARR_F(ir_node *, 128);
2236         env.tos           = 0;
2237         env.nextDFSnum    = 0;
2238         env.POnum         = 0;
2239         env.changes       = 0;
2240         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2241
2242         /* calculate the SCC's and drive loop optimization. */
2243         do_dfs(irg, &env);
2244
2245         DEL_ARR_F(env.stack);
2246         phase_free(&env.ph);
2247
2248         return env.changes;
2249 }  /* optimize_loops */
2250
2251 /*
2252  * do the load store optimization
2253  */
2254 int optimize_load_store(ir_graph *irg) {
2255         walk_env_t env;
2256
2257         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2258
2259         assert(get_irg_phase_state(irg) != phase_building);
2260         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2261                 "LoadStore optimization needs pinned graph");
2262
2263         /* we need landing pads */
2264         remove_critical_cf_edges(irg);
2265
2266         edges_assure(irg);
2267
2268         /* for Phi optimization post-dominators are needed ... */
2269         assure_postdoms(irg);
2270
2271         if (get_opt_alias_analysis()) {
2272                 assure_irg_entity_usage_computed(irg);
2273                 assure_irp_globals_entity_usage_computed();
2274         }
2275
2276         obstack_init(&env.obst);
2277         env.changes = 0;
2278
2279         /* init the links, then collect Loads/Stores/Proj's in lists */
2280         master_visited = 0;
2281         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2282
2283         /* now we have collected enough information, optimize */
2284         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2285
2286         env.changes |= optimize_loops(irg);
2287
2288         obstack_free(&env.obst, NULL);
2289
2290         /* Handle graph state */
2291         if (env.changes) {
2292                 set_irg_outs_inconsistent(irg);
2293                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2294         }
2295
2296         if (env.changes & CF_CHANGED) {
2297                 /* is this really needed: Yes, control flow changed, block might
2298                 have Bad() predecessors. */
2299                 set_irg_doms_inconsistent(irg);
2300         }
2301         return env.changes != 0;
2302 }  /* optimize_load_store */
2303
2304 ir_graph_pass_t *optimize_load_store_pass(const char *name, int verify, int dump)
2305 {
2306         return def_graph_pass(name ? name : "ldst", verify, dump, optimize_load_store);
2307 }  /* optimize_load_store_pass */