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

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