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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  * @version $Id$
  25  */
  26 #include "config.h"
  27
  28 #include <string.h>
  29
  30 #include "iroptimize.h"
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irmode_t.h"
  34 #include "iropt_t.h"
  35 #include "ircons_t.h"
  36 #include "irgmod.h"
  37 #include "irgwalk.h"
  38 #include "irvrfy.h"
  39 #include "tv_t.h"
  40 #include "dbginfo_t.h"
  41 #include "iropt_dbg.h"
  42 #include "irflag_t.h"
  43 #include "array_t.h"
  44 #include "irhooks.h"
  45 #include "iredges.h"
  46 #include "irpass.h"
  47 #include "opt_polymorphy.h"
  48 #include "irmemory.h"
  49 #include "irphase_t.h"
  50 #include "irgopt.h"
  51 #include "set.h"
  52 #include "debug.h"
  53
  54 /** The debug handle. */
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  56
  57 #ifdef DO_CACHEOPT
  58 #include "cacheopt/cachesim.h"
  59 #endif
  60
  61 #undef IMAX
  62 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  63
  64 #define MAX_PROJ        IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
  65
  66 enum changes_t {
  67         DF_CHANGED = 1,       /**< data flow changed */
  68         CF_CHANGED = 2,       /**< control flow changed */
  69 };
  70
  71 /**
  72  * walker environment
  73  */
  74 typedef struct _walk_env_t {
  75         struct obstack obst;          /**< list of all stores */
  76         unsigned changes;             /**< a bitmask of graph changes */
  77 } walk_env_t;
  78
  79 /** A Load/Store info. */
  80 typedef struct _ldst_info_t {
  81         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  82         ir_node  *exc_block;          /**< the exception block if available */
  83         int      exc_idx;             /**< predecessor index in the exception block */
  84         unsigned visited;             /**< visited counter for breaking loops */
  85 } ldst_info_t;
  86
  87 /**
  88  * flags for control flow.
  89  */
  90 enum block_flags_t {
  91         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  92         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  93 };
  94
  95 /**
  96  * a Block info.
  97  */
  98 typedef struct _block_info_t {
  99         unsigned flags;               /**< flags for the block */
 100 } block_info_t;
 101
 102 /** the master visited flag for loop detection. */
 103 static unsigned master_visited = 0;
 104
 105 #define INC_MASTER()       ++master_visited
 106 #define MARK_NODE(info)    (info)->visited = master_visited
 107 #define NODE_VISITED(info) (info)->visited >= master_visited
 108
 109 /**
 110  * get the Load/Store info of a node
 111  */
 112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
 113 {
 114         ldst_info_t *info = get_irn_link(node);
 115
 116         if (! info) {
 117                 info = OALLOCZ(obst, ldst_info_t);
 118                 set_irn_link(node, info);
 119         }
 120         return info;
 121 }  /* get_ldst_info */
 122
 123 /**
 124  * get the Block info of a node
 125  */
 126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
 127 {
 128         block_info_t *info = get_irn_link(node);
 129
 130         if (! info) {
 131                 info = OALLOCZ(obst, block_info_t);
 132                 set_irn_link(node, info);
 133         }
 134         return info;
 135 }  /* get_block_info */
 136
 137 /**
 138  * update the projection info for a Load/Store
 139  */
 140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 141 {
 142         long nr = get_Proj_proj(proj);
 143
 144         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 145
 146         if (info->projs[nr]) {
 147                 /* there is already one, do CSE */
 148                 exchange(proj, info->projs[nr]);
 149                 return DF_CHANGED;
 150         }
 151         else {
 152                 info->projs[nr] = proj;
 153                 return 0;
 154         }
 155 }  /* update_projs */
 156
 157 /**
 158  * update the exception block info for a Load/Store node.
 159  *
 160  * @param info   the load/store info struct
 161  * @param block  the exception handler block for this load/store
 162  * @param pos    the control flow input of the block
 163  */
 164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 165 {
 166         assert(info->exc_block == NULL && "more than one exception block found");
 167
 168         info->exc_block = block;
 169         info->exc_idx   = pos;
 170         return 0;
 171 }  /* update_exc */
 172
 173 /** Return the number of uses of an address node */
 174 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 175
 176 /**
 177  * walker, collects all Load/Store/Proj nodes
 178  *
 179  * walks from Start -> End
 180  */
 181 static void collect_nodes(ir_node *node, void *env)
 182 {
 183         ir_opcode   opcode = get_irn_opcode(node);
 184         ir_node     *pred, *blk, *pred_blk;
 185         ldst_info_t *ldst_info;
 186         walk_env_t  *wenv = env;
 187
 188         if (opcode == iro_Proj) {
 189                 pred   = get_Proj_pred(node);
 190                 opcode = get_irn_opcode(pred);
 191
 192                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 193                         ldst_info = get_ldst_info(pred, &wenv->obst);
 194
 195                         wenv->changes |= update_projs(ldst_info, node);
 196
 197                         /*
 198                          * Place the Proj's to the same block as the
 199                          * predecessor Load. This is always ok and prevents
 200                          * "non-SSA" form after optimizations if the Proj
 201                          * is in a wrong block.
 202                          */
 203                         blk      = get_nodes_block(node);
 204                         pred_blk = get_nodes_block(pred);
 205                         if (blk != pred_blk) {
 206                                 wenv->changes |= DF_CHANGED;
 207                                 set_nodes_block(node, pred_blk);
 208                         }
 209                 }
 210         } else if (opcode == iro_Block) {
 211                 int i;
 212
 213                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 214                         ir_node      *pred_block, *proj;
 215                         block_info_t *bl_info;
 216                         int          is_exc = 0;
 217
 218                         pred = proj = get_Block_cfgpred(node, i);
 219
 220                         if (is_Proj(proj)) {
 221                                 pred   = get_Proj_pred(proj);
 222                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 223                         }
 224
 225                         /* ignore Bad predecessors, they will be removed later */
 226                         if (is_Bad(pred))
 227                                 continue;
 228
 229                         pred_block = get_nodes_block(pred);
 230                         bl_info    = get_block_info(pred_block, &wenv->obst);
 231
 232                         if (is_fragile_op(pred) && is_exc)
 233                                 bl_info->flags |= BLOCK_HAS_EXC;
 234                         else if (is_irn_forking(pred))
 235                                 bl_info->flags |= BLOCK_HAS_COND;
 236
 237                         opcode = get_irn_opcode(pred);
 238                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 239                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 240
 241                                 wenv->changes |= update_exc(ldst_info, node, i);
 242                         }
 243                 }
 244         }
 245 }  /* collect_nodes */
 246
 247 /**
 248  * Returns an entity if the address ptr points to a constant one.
 249  *
 250  * @param ptr  the address
 251  *
 252  * @return an entity or NULL
 253  */
 254 static ir_entity *find_constant_entity(ir_node *ptr)
 255 {
 256         for (;;) {
 257                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 258                         return get_SymConst_entity(ptr);
 259                 } else if (is_Sel(ptr)) {
 260                         ir_entity *ent = get_Sel_entity(ptr);
 261                         ir_type   *tp  = get_entity_owner(ent);
 262
 263                         /* Do not fiddle with polymorphism. */
 264                         if (is_Class_type(get_entity_owner(ent)) &&
 265                                 ((get_entity_n_overwrites(ent)    != 0) ||
 266                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 267                                 return NULL;
 268
 269                         if (is_Array_type(tp)) {
 270                                 /* check bounds */
 271                                 int i, n;
 272
 273                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 274                                         ir_node *bound;
 275                                         tarval *tlower, *tupper;
 276                                         ir_node *index = get_Sel_index(ptr, i);
 277                                         tarval *tv     = computed_value(index);
 278
 279                                         /* check if the index is constant */
 280                                         if (tv == tarval_bad)
 281                                                 return NULL;
 282
 283                                         bound  = get_array_lower_bound(tp, i);
 284                                         tlower = computed_value(bound);
 285                                         bound  = get_array_upper_bound(tp, i);
 286                                         tupper = computed_value(bound);
 287
 288                                         if (tlower == tarval_bad || tupper == tarval_bad)
 289                                                 return NULL;
 290
 291                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 292                                                 return NULL;
 293                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 294                                                 return NULL;
 295
 296                                         /* ok, bounds check finished */
 297                                 }
 298                         }
 299
 300                         if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
 301                                 return ent;
 302
 303                         /* try next */
 304                         ptr = get_Sel_ptr(ptr);
 305                 } else if (is_Add(ptr)) {
 306                         ir_node *l = get_Add_left(ptr);
 307                         ir_node *r = get_Add_right(ptr);
 308
 309                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 310                                 ptr = l;
 311                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 312                                 ptr = r;
 313                         else
 314                                 return NULL;
 315
 316                         /* for now, we support only one addition, reassoc should fold all others */
 317                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 318                                 return NULL;
 319                 } else if (is_Sub(ptr)) {
 320                         ir_node *l = get_Sub_left(ptr);
 321                         ir_node *r = get_Sub_right(ptr);
 322
 323                         if (get_irn_mode(l) == get_irn_mode(ptr) &&     is_Const(r))
 324                                 ptr = l;
 325                         else
 326                                 return NULL;
 327                         /* for now, we support only one substraction, reassoc should fold all others */
 328                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 329                                 return NULL;
 330                 } else
 331                         return NULL;
 332         }
 333 }  /* find_constant_entity */
 334
 335 /**
 336  * Return the Selection index of a Sel node from dimension n
 337  */
 338 static long get_Sel_array_index_long(ir_node *n, int dim)
 339 {
 340         ir_node *index = get_Sel_index(n, dim);
 341         assert(is_Const(index));
 342         return get_tarval_long(get_Const_tarval(index));
 343 }  /* get_Sel_array_index_long */
 344
 345 /**
 346  * Returns the accessed component graph path for an
 347  * node computing an address.
 348  *
 349  * @param ptr    the node computing the address
 350  * @param depth  current depth in steps upward from the root
 351  *               of the address
 352  */
 353 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth)
 354 {
 355         compound_graph_path *res = NULL;
 356         ir_entity           *root, *field, *ent;
 357         int                 path_len, pos, idx;
 358         tarval              *tv;
 359         ir_type             *tp;
 360
 361         if (is_SymConst(ptr)) {
 362                 /* a SymConst. If the depth is 0, this is an access to a global
 363                  * entity and we don't need a component path, else we know
 364                  * at least its length.
 365                  */
 366                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 367                 root = get_SymConst_entity(ptr);
 368                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 369         } else if (is_Sel(ptr)) {
 370                 /* it's a Sel, go up until we find the root */
 371                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 372                 if (res == NULL)
 373                         return NULL;
 374
 375                 /* fill up the step in the path at the current position */
 376                 field    = get_Sel_entity(ptr);
 377                 path_len = get_compound_graph_path_length(res);
 378                 pos      = path_len - depth - 1;
 379                 set_compound_graph_path_node(res, pos, field);
 380
 381                 if (is_Array_type(get_entity_owner(field))) {
 382                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 383                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 384                 }
 385         } else if (is_Add(ptr)) {
 386                 ir_node *l    = get_Add_left(ptr);
 387                 ir_node *r    = get_Add_right(ptr);
 388                 ir_mode *mode = get_irn_mode(ptr);
 389                 tarval  *tmp;
 390
 391                 if (is_Const(r) && get_irn_mode(l) == mode) {
 392                         ptr = l;
 393                         tv  = get_Const_tarval(r);
 394                 } else {
 395                         ptr = r;
 396                         tv  = get_Const_tarval(l);
 397                 }
 398 ptr_arith:
 399                 mode = get_tarval_mode(tv);
 400                 tmp  = tv;
 401
 402                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 403                 if (is_Sel(ptr)) {
 404                         field = get_Sel_entity(ptr);
 405                 } else {
 406                         field = get_SymConst_entity(ptr);
 407                 }
 408                 idx = 0;
 409                 for (ent = field;;) {
 410                         unsigned size;
 411                         tarval   *sz, *tv_index, *tlower, *tupper;
 412                         ir_node  *bound;
 413
 414                         tp = get_entity_type(ent);
 415                         if (! is_Array_type(tp))
 416                                 break;
 417                         ent = get_array_element_entity(tp);
 418                         size = get_type_size_bytes(get_entity_type(ent));
 419                         sz   = new_tarval_from_long(size, mode);
 420
 421                         tv_index = tarval_div(tmp, sz);
 422                         tmp      = tarval_mod(tmp, sz);
 423
 424                         if (tv_index == tarval_bad || tmp == tarval_bad)
 425                                 return NULL;
 426
 427                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 428                         bound  = get_array_lower_bound(tp, 0);
 429                         tlower = computed_value(bound);
 430                         bound  = get_array_upper_bound(tp, 0);
 431                         tupper = computed_value(bound);
 432
 433                         if (tlower == tarval_bad || tupper == tarval_bad)
 434                                 return NULL;
 435
 436                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 437                                 return NULL;
 438                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 439                                 return NULL;
 440
 441                         /* ok, bounds check finished */
 442                         ++idx;
 443                 }
 444                 if (! tarval_is_null(tmp)) {
 445                         /* access to some struct/union member */
 446                         return NULL;
 447                 }
 448
 449                 /* should be at least ONE array */
 450                 if (idx == 0)
 451                         return NULL;
 452
 453                 res = rec_get_accessed_path(ptr, depth + idx);
 454                 if (res == NULL)
 455                         return NULL;
 456
 457                 path_len = get_compound_graph_path_length(res);
 458                 pos      = path_len - depth - idx;
 459
 460                 for (ent = field;;) {
 461                         unsigned size;
 462                         tarval   *sz, *tv_index;
 463                         long     index;
 464
 465                         tp = get_entity_type(ent);
 466                         if (! is_Array_type(tp))
 467                                 break;
 468                         ent = get_array_element_entity(tp);
 469                         set_compound_graph_path_node(res, pos, ent);
 470
 471                         size = get_type_size_bytes(get_entity_type(ent));
 472                         sz   = new_tarval_from_long(size, mode);
 473
 474                         tv_index = tarval_div(tv, sz);
 475                         tv       = tarval_mod(tv, sz);
 476
 477                         /* worked above, should work again */
 478                         assert(tv_index != tarval_bad && tv != tarval_bad);
 479
 480                         /* bounds already checked above */
 481                         index = get_tarval_long(tv_index);
 482                         set_compound_graph_path_array_index(res, pos, index);
 483                         ++pos;
 484                 }
 485         } else if (is_Sub(ptr)) {
 486                 ir_node *l = get_Sub_left(ptr);
 487                 ir_node *r = get_Sub_right(ptr);
 488
 489                 ptr = l;
 490                 tv  = get_Const_tarval(r);
 491                 tv  = tarval_neg(tv);
 492                 goto ptr_arith;
 493         }
 494         return res;
 495 }  /* rec_get_accessed_path */
 496
 497 /**
 498  * Returns an access path or NULL.  The access path is only
 499  * valid, if the graph is in phase_high and _no_ address computation is used.
 500  */
 501 static compound_graph_path *get_accessed_path(ir_node *ptr)
 502 {
 503         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 504         return gr;
 505 }  /* get_accessed_path */
 506
 507 typedef struct path_entry {
 508         ir_entity         *ent;
 509         struct path_entry *next;
 510         long              index;
 511 } path_entry;
 512
 513 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
 514 {
 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 >= (int) 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 {
 690         return rec_find_compound_ent_value(ptr, NULL);
 691 }
 692
 693 /* forward */
 694 static void reduce_adr_usage(ir_node *ptr);
 695
 696 /**
 697  * Update a Load that may have lost its users.
 698  */
 699 static void handle_load_update(ir_node *load)
 700 {
 701         ldst_info_t *info = get_irn_link(load);
 702
 703         /* do NOT touch volatile loads for now */
 704         if (get_Load_volatility(load) == volatility_is_volatile)
 705                 return;
 706
 707         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 708                 ir_node *ptr = get_Load_ptr(load);
 709                 ir_node *mem = get_Load_mem(load);
 710
 711                 /* a Load whose value is neither used nor exception checked, remove it */
 712                 exchange(info->projs[pn_Load_M], mem);
 713                 if (info->projs[pn_Load_X_regular])
 714                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 715                 kill_node(load);
 716                 reduce_adr_usage(ptr);
 717         }
 718 }  /* handle_load_update */
 719
 720 /**
 721  * A use of an address node has vanished. Check if this was a Proj
 722  * node and update the counters.
 723  */
 724 static void reduce_adr_usage(ir_node *ptr)
 725 {
 726         if (is_Proj(ptr)) {
 727                 if (get_irn_n_edges(ptr) <= 0) {
 728                         /* this Proj is dead now */
 729                         ir_node *pred = get_Proj_pred(ptr);
 730
 731                         if (is_Load(pred)) {
 732                                 ldst_info_t *info = get_irn_link(pred);
 733                                 info->projs[get_Proj_proj(ptr)] = NULL;
 734
 735                                 /* this node lost its result proj, handle that */
 736                                 handle_load_update(pred);
 737                         }
 738                 }
 739         }
 740 }  /* reduce_adr_usage */
 741
 742 /**
 743  * Check, if an already existing value of mode old_mode can be converted
 744  * into the needed one new_mode without loss.
 745  */
 746 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
 747 {
 748         if (old_mode == new_mode)
 749                 return 1;
 750
 751         /* if both modes are two-complement ones, we can always convert the
 752            Stored value into the needed one. */
 753         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 754                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 755                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 756                 return 1;
 757         return 0;
 758 }  /* can_use_stored_value */
 759
 760 /**
 761  * Check whether a Call is at least pure, ie. does only read memory.
 762  */
 763 static unsigned is_Call_pure(ir_node *call)
 764 {
 765         ir_type *call_tp = get_Call_type(call);
 766         unsigned prop = get_method_additional_properties(call_tp);
 767
 768         /* check first the call type */
 769         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 770                 /* try the called entity */
 771                 ir_node *ptr = get_Call_ptr(call);
 772
 773                 if (is_Global(ptr)) {
 774                         ir_entity *ent = get_Global_entity(ptr);
 775
 776                         prop = get_entity_additional_properties(ent);
 777                 }
 778         }
 779         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 780 }  /* is_Call_pure */
 781
 782 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 783 {
 784         ir_mode *mode  = get_irn_mode(ptr);
 785         long    offset = 0;
 786
 787         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 788         for (;;) {
 789                 if (is_Add(ptr)) {
 790                         ir_node *l = get_Add_left(ptr);
 791                         ir_node *r = get_Add_right(ptr);
 792
 793                         if (get_irn_mode(l) != mode || !is_Const(r))
 794                                 break;
 795
 796                         offset += get_tarval_long(get_Const_tarval(r));
 797                         ptr     = l;
 798                 } else if (is_Sub(ptr)) {
 799                         ir_node *l = get_Sub_left(ptr);
 800                         ir_node *r = get_Sub_right(ptr);
 801
 802                         if (get_irn_mode(l) != mode || !is_Const(r))
 803                                 break;
 804
 805                         offset -= get_tarval_long(get_Const_tarval(r));
 806                         ptr     = l;
 807                 } else if (is_Sel(ptr)) {
 808                         ir_entity *ent = get_Sel_entity(ptr);
 809                         ir_type   *tp  = get_entity_owner(ent);
 810
 811                         if (is_Array_type(tp)) {
 812                                 int     size;
 813                                 ir_node *index;
 814
 815                                 /* only one dimensional arrays yet */
 816                                 if (get_Sel_n_indexs(ptr) != 1)
 817                                         break;
 818                                 index = get_Sel_index(ptr, 0);
 819                                 if (! is_Const(index))
 820                                         break;
 821
 822                                 tp = get_entity_type(ent);
 823                                 if (get_type_state(tp) != layout_fixed)
 824                                         break;
 825
 826                                 size    = get_type_size_bytes(tp);
 827                                 offset += size * get_tarval_long(get_Const_tarval(index));
 828                         } else {
 829                                 if (get_type_state(tp) != layout_fixed)
 830                                         break;
 831                                 offset += get_entity_offset(ent);
 832                         }
 833                         ptr = get_Sel_ptr(ptr);
 834                 } else
 835                         break;
 836         }
 837
 838         *pOffset = offset;
 839         return ptr;
 840 }
 841
 842 static int try_load_after_store(ir_node *load,
 843                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 844 {
 845         ldst_info_t *info;
 846         ir_node *store_ptr      = get_Store_ptr(store);
 847         long     store_offset;
 848         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 849         ir_node *store_value;
 850         ir_mode *store_mode;
 851         ir_node *load_ptr;
 852         ir_mode *load_mode;
 853         long     load_mode_len;
 854         long     store_mode_len;
 855         long     delta;
 856         int      res;
 857
 858         if (load_base_ptr != store_base_ptr)
 859                 return 0;
 860
 861         load_mode      = get_Load_mode(load);
 862         load_mode_len  = get_mode_size_bytes(load_mode);
 863         store_mode     = get_irn_mode(get_Store_value(store));
 864         store_mode_len = get_mode_size_bytes(store_mode);
 865         delta          = load_offset - store_offset;
 866         store_value    = get_Store_value(store);
 867
 868         if (delta != 0 || store_mode != load_mode) {
 869                 if (delta < 0 || delta + load_mode_len > store_mode_len)
 870                         return 0;
 871
 872                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 873                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 874                         return 0;
 875
 876
 877                 /* produce a shift to adjust offset delta */
 878                 if (delta > 0) {
 879                         ir_node *cnst;
 880
 881                         /* FIXME: only true for little endian */
 882                         cnst        = new_Const_long(mode_Iu, delta * 8);
 883                         store_value = new_r_Shr(get_nodes_block(load),
 884                                                                         store_value, cnst, store_mode);
 885                 }
 886
 887                 /* add an convert if needed */
 888                 if (store_mode != load_mode) {
 889                         store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
 890                 }
 891         }
 892
 893         DBG_OPT_RAW(load, store_value);
 894
 895         info = get_irn_link(load);
 896         if (info->projs[pn_Load_M])
 897                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 898
 899         res = 0;
 900         /* no exception */
 901         if (info->projs[pn_Load_X_except]) {
 902                 exchange( info->projs[pn_Load_X_except], new_Bad());
 903                 res |= CF_CHANGED;
 904         }
 905         if (info->projs[pn_Load_X_regular]) {
 906                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 907                 res |= CF_CHANGED;
 908         }
 909
 910         if (info->projs[pn_Load_res])
 911                 exchange(info->projs[pn_Load_res], store_value);
 912
 913         load_ptr = get_Load_ptr(load);
 914         kill_node(load);
 915         reduce_adr_usage(load_ptr);
 916         return res | DF_CHANGED;
 917 }
 918
 919 /**
 920  * Follow the memory chain as long as there are only Loads,
 921  * alias free Stores, and constant Calls and try to replace the
 922  * current Load by a previous ones.
 923  * Note that in unreachable loops it might happen that we reach
 924  * load again, as well as we can fall into a cycle.
 925  * We break such cycles using a special visited flag.
 926  *
 927  * INC_MASTER() must be called before dive into
 928  */
 929 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
 930 {
 931         unsigned    res = 0;
 932         ldst_info_t *info = get_irn_link(load);
 933         ir_node     *pred;
 934         ir_node     *ptr       = get_Load_ptr(load);
 935         ir_node     *mem       = get_Load_mem(load);
 936         ir_mode     *load_mode = get_Load_mode(load);
 937
 938         for (pred = curr; load != pred; ) {
 939                 ldst_info_t *pred_info = get_irn_link(pred);
 940
 941                 /*
 942                  * a Load immediately after a Store -- a read after write.
 943                  * We may remove the Load, if both Load & Store does not have an
 944                  * exception handler OR they are in the same MacroBlock. In the latter
 945                  * case the Load cannot throw an exception when the previous Store was
 946                  * quiet.
 947                  *
 948                  * Why we need to check for Store Exception? If the Store cannot
 949                  * be executed (ROM) the exception handler might simply jump into
 950                  * the load MacroBlock :-(
 951                  * We could make it a little bit better if we would know that the
 952                  * exception handler of the Store jumps directly to the end...
 953                  */
 954                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 955                                 && info->projs[pn_Load_X_except] == NULL)
 956                                 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
 957                 {
 958                         long    load_offset;
 959                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 960                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 961
 962                         if (changes != 0)
 963                                 return res | changes;
 964                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 965                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 966                         /*
 967                          * a Load after a Load -- a read after read.
 968                          * We may remove the second Load, if it does not have an exception handler
 969                          * OR they are in the same MacroBlock. In the later case the Load cannot
 970                          * throw an exception when the previous Load was quiet.
 971                          *
 972                          * Here, there is no need to check if the previous Load has an exception
 973                          * hander because they would have exact the same exception...
 974                          */
 975                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 976                                 ir_node *value;
 977
 978                                 DBG_OPT_RAR(load, pred);
 979
 980                                 /* the result is used */
 981                                 if (info->projs[pn_Load_res]) {
 982                                         if (pred_info->projs[pn_Load_res] == NULL) {
 983                                                 /* create a new Proj again */
 984                                                 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
 985                                         }
 986                                         value = pred_info->projs[pn_Load_res];
 987
 988                                         /* add an convert if needed */
 989                                         if (get_Load_mode(pred) != load_mode) {
 990                                                 value = new_r_Conv(get_nodes_block(load), value, load_mode);
 991                                         }
 992
 993                                         exchange(info->projs[pn_Load_res], value);
 994                                 }
 995
 996                                 if (info->projs[pn_Load_M])
 997                                         exchange(info->projs[pn_Load_M], mem);
 998
 999                                 /* no exception */
1000                                 if (info->projs[pn_Load_X_except]) {
1001                                         exchange(info->projs[pn_Load_X_except], new_Bad());
1002                                         res |= CF_CHANGED;
1003                                 }
1004                                 if (info->projs[pn_Load_X_regular]) {
1005                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1006                                         res |= CF_CHANGED;
1007                                 }
1008
1009                                 kill_node(load);
1010                                 reduce_adr_usage(ptr);
1011                                 return res |= DF_CHANGED;
1012                         }
1013                 }
1014
1015                 if (is_Store(pred)) {
1016                         /* check if we can pass through this store */
1017                         ir_alias_relation rel = get_alias_relation(
1018                                 current_ir_graph,
1019                                 get_Store_ptr(pred),
1020                                 get_irn_mode(get_Store_value(pred)),
1021                                 ptr, load_mode);
1022                         /* if the might be an alias, we cannot pass this Store */
1023                         if (rel != ir_no_alias)
1024                                 break;
1025                         pred = skip_Proj(get_Store_mem(pred));
1026                 } else if (is_Load(pred)) {
1027                         pred = skip_Proj(get_Load_mem(pred));
1028                 } else if (is_Call(pred)) {
1029                         if (is_Call_pure(pred)) {
1030                                 /* The called graph is at least pure, so there are no Store's
1031                                    in it. We can handle it like a Load and skip it. */
1032                                 pred = skip_Proj(get_Call_mem(pred));
1033                         } else {
1034                                 /* there might be Store's in the graph, stop here */
1035                                 break;
1036                         }
1037                 } else {
1038                         /* follow only Load chains */
1039                         break;
1040                 }
1041
1042                 /* check for cycles */
1043                 if (NODE_VISITED(pred_info))
1044                         break;
1045                 MARK_NODE(pred_info);
1046         }
1047
1048         if (is_Sync(pred)) {
1049                 int i;
1050
1051                 /* handle all Sync predecessors */
1052                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1053                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1054                         if (res)
1055                                 return res;
1056                 }
1057         }
1058
1059         return res;
1060 }  /* follow_Mem_chain */
1061
1062 /*
1063  * Check if we can replace the load by a given const from
1064  * the const code irg.
1065  */
1066 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1067 {
1068         ir_mode *c_mode = get_irn_mode(c);
1069         ir_mode *l_mode = get_Load_mode(load);
1070         ir_node *res    = NULL;
1071
1072         if (c_mode != l_mode) {
1073                 /* check, if the mode matches OR can be easily converted info */
1074                 if (is_reinterpret_cast(c_mode, l_mode)) {
1075                         /* we can safely cast */
1076                         dbg_info *dbg   = get_irn_dbg_info(load);
1077                         ir_node  *block = get_nodes_block(load);
1078
1079                         /* copy the value from the const code irg and cast it */
1080                         res = copy_const_value(dbg, c);
1081                         res = new_rd_Conv(dbg, block, res, l_mode);
1082                 }
1083         } else {
1084                 /* copy the value from the const code irg */
1085                 res = copy_const_value(get_irn_dbg_info(load), c);
1086         }
1087         return res;
1088 }  /* can_replace_load_by_const */
1089
1090 /**
1091  * optimize a Load
1092  *
1093  * @param load  the Load node
1094  */
1095 static unsigned optimize_load(ir_node *load)
1096 {
1097         ldst_info_t *info = get_irn_link(load);
1098         ir_node     *mem, *ptr, *value;
1099         ir_entity   *ent;
1100         long        dummy;
1101         unsigned    res = 0;
1102
1103         /* do NOT touch volatile loads for now */
1104         if (get_Load_volatility(load) == volatility_is_volatile)
1105                 return 0;
1106
1107         /* the address of the load to be optimized */
1108         ptr = get_Load_ptr(load);
1109
1110         /* The mem of the Load. Must still be returned after optimization. */
1111         mem = get_Load_mem(load);
1112
1113         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1114                 /* a Load which value is neither used nor exception checked, remove it */
1115                 exchange(info->projs[pn_Load_M], mem);
1116
1117                 if (info->projs[pn_Load_X_regular]) {
1118                         /* should not happen, but if it does, remove it */
1119                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1120                         res |= CF_CHANGED;
1121                 }
1122                 kill_node(load);
1123                 reduce_adr_usage(ptr);
1124                 return res | DF_CHANGED;
1125         }
1126
1127         /* Load from a constant polymorphic field, where we can resolve
1128            polymorphism. */
1129         value = transform_polymorph_Load(load);
1130         if (value == load) {
1131                 value = NULL;
1132                 /* check if we can determine the entity that will be loaded */
1133                 ent = find_constant_entity(ptr);
1134                 if (ent != NULL
1135                                 && get_entity_visibility(ent) != ir_visibility_external) {
1136                         /* a static allocation that is not external: there should be NO
1137                          * exception when loading even if we cannot replace the load itself.
1138                          */
1139
1140                         /* no exception, clear the info field as it might be checked later again */
1141                         if (info->projs[pn_Load_X_except]) {
1142                                 exchange(info->projs[pn_Load_X_except], new_Bad());
1143                                 info->projs[pn_Load_X_except] = NULL;
1144                                 res |= CF_CHANGED;
1145                         }
1146                         if (info->projs[pn_Load_X_regular]) {
1147                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1148                                 info->projs[pn_Load_X_regular] = NULL;
1149                                 res |= CF_CHANGED;
1150                         }
1151
1152                         if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1153                                 if (ent->initializer != NULL) {
1154                                         /* new style initializer */
1155                                         value = find_compound_ent_value(ptr);
1156                                 } else if (entity_has_compound_ent_values(ent)) {
1157                                         /* old style initializer */
1158                                         compound_graph_path *path = get_accessed_path(ptr);
1159
1160                                         if (path != NULL) {
1161                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1162
1163                                                 value = get_compound_ent_value_by_path(ent, path);
1164                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1165                                                 free_compound_graph_path(path);
1166                                         }
1167                                 }
1168                                 if (value != NULL)
1169                                         value = can_replace_load_by_const(load, value);
1170                         }
1171                 }
1172         }
1173         if (value != NULL) {
1174                 /* we completely replace the load by this value */
1175                 if (info->projs[pn_Load_X_except]) {
1176                         exchange(info->projs[pn_Load_X_except], new_Bad());
1177                         info->projs[pn_Load_X_except] = NULL;
1178                         res |= CF_CHANGED;
1179                 }
1180                 if (info->projs[pn_Load_X_regular]) {
1181                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1182                         info->projs[pn_Load_X_regular] = NULL;
1183                         res |= CF_CHANGED;
1184                 }
1185                 if (info->projs[pn_Load_M]) {
1186                         exchange(info->projs[pn_Load_M], mem);
1187                         res |= DF_CHANGED;
1188                 }
1189                 if (info->projs[pn_Load_res]) {
1190                         exchange(info->projs[pn_Load_res], value);
1191                         res |= DF_CHANGED;
1192                 }
1193                 kill_node(load);
1194                 reduce_adr_usage(ptr);
1195                 return res;
1196         }
1197
1198         /* Check, if the address of this load is used more than once.
1199          * If not, more load cannot be removed in any case. */
1200         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1201                 return res;
1202
1203         /*
1204          * follow the memory chain as long as there are only Loads
1205          * and try to replace current Load or Store by a previous one.
1206          * Note that in unreachable loops it might happen that we reach
1207          * load again, as well as we can fall into a cycle.
1208          * We break such cycles using a special visited flag.
1209          */
1210         INC_MASTER();
1211         res = follow_Mem_chain(load, skip_Proj(mem));
1212         return res;
1213 }  /* optimize_load */
1214
1215 /**
1216  * Check whether a value of mode new_mode would completely overwrite a value
1217  * of mode old_mode in memory.
1218  */
1219 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1220 {
1221         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1222 }  /* is_completely_overwritten */
1223
1224 /**
1225  * Check whether small is a part of large (starting at same address).
1226  */
1227 static int is_partially_same(ir_node *small, ir_node *large)
1228 {
1229         ir_mode *sm = get_irn_mode(small);
1230         ir_mode *lm = get_irn_mode(large);
1231
1232         /* FIXME: Check endianness */
1233         return is_Conv(small) && get_Conv_op(small) == large
1234             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1235             && get_mode_arithmetic(sm) == irma_twos_complement
1236             && get_mode_arithmetic(lm) == irma_twos_complement;
1237 }  /* is_partially_same */
1238
1239 /**
1240  * follow the memory chain as long as there are only Loads and alias free Stores.
1241  *
1242  * INC_MASTER() must be called before dive into
1243  */
1244 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1245 {
1246         unsigned res = 0;
1247         ldst_info_t *info = get_irn_link(store);
1248         ir_node *pred;
1249         ir_node *ptr = get_Store_ptr(store);
1250         ir_node *mem = get_Store_mem(store);
1251         ir_node *value = get_Store_value(store);
1252         ir_mode *mode  = get_irn_mode(value);
1253         ir_node *block = get_nodes_block(store);
1254         ir_node *mblk  = get_Block_MacroBlock(block);
1255
1256         for (pred = curr; pred != store;) {
1257                 ldst_info_t *pred_info = get_irn_link(pred);
1258
1259                 /*
1260                  * BEWARE: one might think that checking the modes is useless, because
1261                  * if the pointers are identical, they refer to the same object.
1262                  * This is only true in strong typed languages, not is C were the following
1263                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1264                  * However, if the size of the mode that is written is bigger or equal the
1265                  * size of the old one, the old value is completely overwritten and can be
1266                  * killed ...
1267                  */
1268                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1269             get_nodes_MacroBlock(pred) == mblk) {
1270                         /*
1271                          * a Store after a Store in the same MacroBlock -- a write after write.
1272                          */
1273
1274                         /*
1275                          * We may remove the first Store, if the old value is completely
1276                          * overwritten or the old value is a part of the new value,
1277                          * and if it does not have an exception handler.
1278                          *
1279                          * TODO: What, if both have the same exception handler ???
1280                          */
1281                         if (get_Store_volatility(pred) != volatility_is_volatile
1282                                 && !pred_info->projs[pn_Store_X_except]) {
1283                                 ir_node *predvalue = get_Store_value(pred);
1284                                 ir_mode *predmode  = get_irn_mode(predvalue);
1285
1286                                 if (is_completely_overwritten(predmode, mode)
1287                                         || is_partially_same(predvalue, value)) {
1288                                         DBG_OPT_WAW(pred, store);
1289                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1290                                         kill_node(pred);
1291                                         reduce_adr_usage(ptr);
1292                                         return DF_CHANGED;
1293                                 }
1294                         }
1295
1296                         /*
1297                          * We may remove the Store, if the old value already contains
1298                          * the new value, and if it does not have an exception handler.
1299                          *
1300                          * TODO: What, if both have the same exception handler ???
1301                          */
1302                         if (get_Store_volatility(store) != volatility_is_volatile
1303                                 && !info->projs[pn_Store_X_except]) {
1304                                 ir_node *predvalue = get_Store_value(pred);
1305
1306                                 if (is_partially_same(value, predvalue)) {
1307                                         DBG_OPT_WAW(pred, store);
1308                                         exchange(info->projs[pn_Store_M], mem);
1309                                         kill_node(store);
1310                                         reduce_adr_usage(ptr);
1311                                         return DF_CHANGED;
1312                                 }
1313                         }
1314                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1315                            value == pred_info->projs[pn_Load_res]) {
1316                         /*
1317                          * a Store of a value just loaded from the same address
1318                          * -- a write after read.
1319                          * We may remove the Store, if it does not have an exception
1320                          * handler.
1321                          */
1322                         if (! info->projs[pn_Store_X_except]) {
1323                                 DBG_OPT_WAR(store, pred);
1324                                 exchange(info->projs[pn_Store_M], mem);
1325                                 kill_node(store);
1326                                 reduce_adr_usage(ptr);
1327                                 return DF_CHANGED;
1328                         }
1329                 }
1330
1331                 if (is_Store(pred)) {
1332                         /* check if we can pass through this store */
1333                         ir_alias_relation rel = get_alias_relation(
1334                                 current_ir_graph,
1335                                 get_Store_ptr(pred),
1336                                 get_irn_mode(get_Store_value(pred)),
1337                                 ptr, mode);
1338                         /* if the might be an alias, we cannot pass this Store */
1339                         if (rel != ir_no_alias)
1340                                 break;
1341                         pred = skip_Proj(get_Store_mem(pred));
1342                 } else if (is_Load(pred)) {
1343                         ir_alias_relation rel = get_alias_relation(
1344                                 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1345                                 ptr, mode);
1346                         if (rel != ir_no_alias)
1347                                 break;
1348
1349                         pred = skip_Proj(get_Load_mem(pred));
1350                 } else {
1351                         /* follow only Load chains */
1352                         break;
1353                 }
1354
1355                 /* check for cycles */
1356                 if (NODE_VISITED(pred_info))
1357                         break;
1358                 MARK_NODE(pred_info);
1359         }
1360
1361         if (is_Sync(pred)) {
1362                 int i;
1363
1364                 /* handle all Sync predecessors */
1365                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1366                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1367                         if (res)
1368                                 break;
1369                 }
1370         }
1371         return res;
1372 }  /* follow_Mem_chain_for_Store */
1373
1374 /** find entity used as base for an address calculation */
1375 static ir_entity *find_entity(ir_node *ptr)
1376 {
1377         switch (get_irn_opcode(ptr)) {
1378         case iro_SymConst:
1379                 return get_SymConst_entity(ptr);
1380         case iro_Sel: {
1381                 ir_node *pred = get_Sel_ptr(ptr);
1382                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1383                         return get_Sel_entity(ptr);
1384
1385                 return find_entity(pred);
1386         }
1387         case iro_Sub:
1388         case iro_Add: {
1389                 ir_node *left = get_binop_left(ptr);
1390                 ir_node *right;
1391                 if (mode_is_reference(get_irn_mode(left)))
1392                         return find_entity(left);
1393                 right = get_binop_right(ptr);
1394                 if (mode_is_reference(get_irn_mode(right)))
1395                         return find_entity(right);
1396                 return NULL;
1397         }
1398         default:
1399                 return NULL;
1400         }
1401 }
1402
1403 /**
1404  * optimize a Store
1405  *
1406  * @param store  the Store node
1407  */
1408 static unsigned optimize_store(ir_node *store)
1409 {
1410         ir_node   *ptr;
1411         ir_node   *mem;
1412         ir_entity *entity;
1413
1414         if (get_Store_volatility(store) == volatility_is_volatile)
1415                 return 0;
1416
1417         ptr    = get_Store_ptr(store);
1418         entity = find_entity(ptr);
1419
1420         /* a store to an entity which is never read is unnecessary */
1421         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1422                 ldst_info_t *info = get_irn_link(store);
1423                 if (info->projs[pn_Store_X_except] == NULL) {
1424                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1425                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1426                         kill_node(store);
1427                         reduce_adr_usage(ptr);
1428                         return DF_CHANGED;
1429                 }
1430         }
1431
1432         /* Check, if the address of this Store is used more than once.
1433          * If not, this Store cannot be removed in any case. */
1434         if (get_irn_n_uses(ptr) <= 1)
1435                 return 0;
1436
1437         mem = get_Store_mem(store);
1438
1439         /* follow the memory chain as long as there are only Loads */
1440         INC_MASTER();
1441
1442         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1443 }  /* optimize_store */
1444
1445 /**
1446  * walker, optimizes Phi after Stores to identical places:
1447  * Does the following optimization:
1448  * @verbatim
1449  *
1450  *   val1   val2   val3          val1  val2  val3
1451  *    |      |      |               \    |    /
1452  *  Store  Store  Store              \   |   /
1453  *      \    |    /                   PhiData
1454  *       \   |   /                       |
1455  *        \  |  /                      Store
1456  *          PhiM
1457  *
1458  * @endverbatim
1459  * This reduces the number of stores and allows for predicated execution.
1460  * Moves Stores back to the end of a function which may be bad.
1461  *
1462  * This is only possible if the predecessor blocks have only one successor.
1463  */
1464 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1465 {
1466         int i, n;
1467         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1468         ir_mode *mode;
1469         ir_node **inM, **inD, **projMs;
1470         int *idx;
1471         dbg_info *db = NULL;
1472         ldst_info_t *info;
1473         block_info_t *bl_info;
1474         unsigned res = 0;
1475
1476         /* Must be a memory Phi */
1477         if (get_irn_mode(phi) != mode_M)
1478                 return 0;
1479
1480         n = get_Phi_n_preds(phi);
1481         if (n <= 0)
1482                 return 0;
1483
1484         /* must be only one user */
1485         projM = get_Phi_pred(phi, 0);
1486         if (get_irn_n_edges(projM) != 1)
1487                 return 0;
1488
1489         store = skip_Proj(projM);
1490         old_store = store;
1491         if (!is_Store(store))
1492                 return 0;
1493
1494         block = get_nodes_block(store);
1495
1496         /* abort on dead blocks */
1497         if (is_Block_dead(block))
1498                 return 0;
1499
1500         /* check if the block is post dominated by Phi-block
1501            and has no exception exit */
1502         bl_info = get_irn_link(block);
1503         if (bl_info->flags & BLOCK_HAS_EXC)
1504                 return 0;
1505
1506         phi_block = get_nodes_block(phi);
1507         if (! block_strictly_postdominates(phi_block, block))
1508                 return 0;
1509
1510         /* this is the address of the store */
1511         ptr  = get_Store_ptr(store);
1512         mode = get_irn_mode(get_Store_value(store));
1513         info = get_irn_link(store);
1514         exc  = info->exc_block;
1515
1516         for (i = 1; i < n; ++i) {
1517                 ir_node *pred = get_Phi_pred(phi, i);
1518
1519                 if (get_irn_n_edges(pred) != 1)
1520                         return 0;
1521
1522                 pred = skip_Proj(pred);
1523                 if (!is_Store(pred))
1524                         return 0;
1525
1526                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1527                         return 0;
1528
1529                 info = get_irn_link(pred);
1530
1531                 /* check, if all stores have the same exception flow */
1532                 if (exc != info->exc_block)
1533                         return 0;
1534
1535                 /* abort on dead blocks */
1536                 block = get_nodes_block(pred);
1537                 if (is_Block_dead(block))
1538                         return 0;
1539
1540                 /* check if the block is post dominated by Phi-block
1541                    and has no exception exit. Note that block must be different from
1542                    Phi-block, else we would move a Store from end End of a block to its
1543                    Start... */
1544                 bl_info = get_irn_link(block);
1545                 if (bl_info->flags & BLOCK_HAS_EXC)
1546                         return 0;
1547                 if (block == phi_block || ! block_postdominates(phi_block, block))
1548                         return 0;
1549         }
1550
1551         /*
1552          * ok, when we are here, we found all predecessors of a Phi that
1553          * are Stores to the same address and size. That means whatever
1554          * we do before we enter the block of the Phi, we do a Store.
1555          * So, we can move the Store to the current block:
1556          *
1557          *   val1    val2    val3          val1  val2  val3
1558          *    |       |       |               \    |    /
1559          * | Str | | Str | | Str |             \   |   /
1560          *      \     |     /                   PhiData
1561          *       \    |    /                       |
1562          *        \   |   /                       Str
1563          *           PhiM
1564          *
1565          * Is only allowed if the predecessor blocks have only one successor.
1566          */
1567
1568         NEW_ARR_A(ir_node *, projMs, n);
1569         NEW_ARR_A(ir_node *, inM, n);
1570         NEW_ARR_A(ir_node *, inD, n);
1571         NEW_ARR_A(int, idx, n);
1572
1573         /* Prepare: Collect all Store nodes.  We must do this
1574            first because we otherwise may loose a store when exchanging its
1575            memory Proj.
1576          */
1577         for (i = n - 1; i >= 0; --i) {
1578                 ir_node *store;
1579
1580                 projMs[i] = get_Phi_pred(phi, i);
1581                 assert(is_Proj(projMs[i]));
1582
1583                 store = get_Proj_pred(projMs[i]);
1584                 info  = get_irn_link(store);
1585
1586                 inM[i] = get_Store_mem(store);
1587                 inD[i] = get_Store_value(store);
1588                 idx[i] = info->exc_idx;
1589         }
1590         block = get_nodes_block(phi);
1591
1592         /* second step: create a new memory Phi */
1593         phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1594
1595         /* third step: create a new data Phi */
1596         phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1597
1598         /* rewire memory and kill the node */
1599         for (i = n - 1; i >= 0; --i) {
1600                 ir_node *proj  = projMs[i];
1601
1602                 if (is_Proj(proj)) {
1603                         ir_node *store = get_Proj_pred(proj);
1604                         exchange(proj, inM[i]);
1605                         kill_node(store);
1606                 }
1607         }
1608
1609         /* fourth step: create the Store */
1610         store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1611 #ifdef DO_CACHEOPT
1612         co_set_irn_name(store, co_get_irn_ident(old_store));
1613 #endif
1614
1615         projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1616
1617         info = get_ldst_info(store, &wenv->obst);
1618         info->projs[pn_Store_M] = projM;
1619
1620         /* fifths step: repair exception flow */
1621         if (exc) {
1622                 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1623
1624                 info->projs[pn_Store_X_except] = projX;
1625                 info->exc_block                = exc;
1626                 info->exc_idx                  = idx[0];
1627
1628                 for (i = 0; i < n; ++i) {
1629                         set_Block_cfgpred(exc, idx[i], projX);
1630                 }
1631
1632                 if (n > 1) {
1633                         /* the exception block should be optimized as some inputs are identical now */
1634                 }
1635
1636                 res |= CF_CHANGED;
1637         }
1638
1639         /* sixth step: replace old Phi */
1640         exchange(phi, projM);
1641
1642         return res | DF_CHANGED;
1643 }  /* optimize_phi */
1644
1645 /**
1646  * walker, do the optimizations
1647  */
1648 static void do_load_store_optimize(ir_node *n, void *env)
1649 {
1650         walk_env_t *wenv = env;
1651
1652         switch (get_irn_opcode(n)) {
1653
1654         case iro_Load:
1655                 wenv->changes |= optimize_load(n);
1656                 break;
1657
1658         case iro_Store:
1659                 wenv->changes |= optimize_store(n);
1660                 break;
1661
1662         case iro_Phi:
1663                 wenv->changes |= optimize_phi(n, wenv);
1664                 break;
1665
1666         default:
1667                 ;
1668         }
1669 }  /* do_load_store_optimize */
1670
1671 /** A scc. */
1672 typedef struct scc {
1673         ir_node *head;          /**< the head of the list */
1674 } scc;
1675
1676 /** A node entry. */
1677 typedef struct node_entry {
1678         unsigned DFSnum;    /**< the DFS number of this node */
1679         unsigned low;       /**< the low number of this node */
1680         int      in_stack;  /**< flag, set if the node is on the stack */
1681         ir_node  *next;     /**< link to the next node the the same scc */
1682         scc      *pscc;     /**< the scc of this node */
1683         unsigned POnum;     /**< the post order number for blocks */
1684 } node_entry;
1685
1686 /** A loop entry. */
1687 typedef struct loop_env {
1688         ir_phase ph;           /**< the phase object */
1689         ir_node  **stack;      /**< the node stack */
1690         int      tos;          /**< tos index */
1691         unsigned nextDFSnum;   /**< the current DFS number */
1692         unsigned POnum;        /**< current post order number */
1693
1694         unsigned changes;      /**< a bitmask of graph changes */
1695 } loop_env;
1696
1697 /**
1698 * Gets the node_entry of a node
1699 */
1700 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1701 {
1702         ir_phase   *ph = &env->ph;
1703         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1704
1705         if (! e) {
1706                 e = phase_alloc(ph, sizeof(*e));
1707                 memset(e, 0, sizeof(*e));
1708                 phase_set_irn_data(ph, irn, e);
1709         }
1710         return e;
1711 }  /* get_irn_ne */
1712
1713 /**
1714  * Push a node onto the stack.
1715  *
1716  * @param env   the loop environment
1717  * @param n     the node to push
1718  */
1719 static void push(loop_env *env, ir_node *n)
1720 {
1721         node_entry *e;
1722
1723         if (env->tos == ARR_LEN(env->stack)) {
1724                 int nlen = ARR_LEN(env->stack) * 2;
1725                 ARR_RESIZE(ir_node *, env->stack, nlen);
1726         }
1727         env->stack[env->tos++] = n;
1728         e = get_irn_ne(n, env);
1729         e->in_stack = 1;
1730 }  /* push */
1731
1732 /**
1733  * pop a node from the stack
1734  *
1735  * @param env   the loop environment
1736  *
1737  * @return  The topmost node
1738  */
1739 static ir_node *pop(loop_env *env)
1740 {
1741         ir_node *n = env->stack[--env->tos];
1742         node_entry *e = get_irn_ne(n, env);
1743
1744         e->in_stack = 0;
1745         return n;
1746 }  /* pop */
1747
1748 /**
1749  * Check if irn is a region constant.
1750  * The block or irn must strictly dominate the header block.
1751  *
1752  * @param irn           the node to check
1753  * @param header_block  the header block of the induction variable
1754  */
1755 static int is_rc(ir_node *irn, ir_node *header_block)
1756 {
1757         ir_node *block = get_nodes_block(irn);
1758
1759         return (block != header_block) && block_dominates(block, header_block);
1760 }  /* is_rc */
1761
1762 typedef struct phi_entry phi_entry;
1763 struct phi_entry {
1764         ir_node   *phi;    /**< A phi with a region const memory. */
1765         int       pos;     /**< The position of the region const memory */
1766         ir_node   *load;   /**< the newly created load for this phi */
1767         phi_entry *next;
1768 };
1769
1770 /**
1771  * An entry in the avail set.
1772  */
1773 typedef struct avail_entry_t {
1774         ir_node *ptr;   /**< the address pointer */
1775         ir_mode *mode;  /**< the load mode */
1776         ir_node *load;  /**< the associated Load */
1777 } avail_entry_t;
1778
1779 /**
1780  * Compare two avail entries.
1781  */
1782 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1783 {
1784         const avail_entry_t *a = elt;
1785         const avail_entry_t *b = key;
1786         (void) size;
1787
1788         return a->ptr != b->ptr || a->mode != b->mode;
1789 }  /* cmp_avail_entry */
1790
1791 /**
1792  * Calculate the hash value of an avail entry.
1793  */
1794 static unsigned hash_cache_entry(const avail_entry_t *entry)
1795 {
1796         return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1797 }  /* hash_cache_entry */
1798
1799 /**
1800  * Move loops out of loops if possible.
1801  *
1802  * @param pscc   the loop described by an SCC
1803  * @param env    the loop environment
1804  */
1805 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1806 {
1807         ir_node   *phi, *load, *next, *other, *next_other;
1808         ir_entity *ent;
1809         int       j;
1810         phi_entry *phi_list = NULL;
1811         set       *avail;
1812
1813         avail = new_set(cmp_avail_entry, 8);
1814
1815         /* collect all outer memories */
1816         for (phi = pscc->head; phi != NULL; phi = next) {
1817                 node_entry *ne = get_irn_ne(phi, env);
1818                 next = ne->next;
1819
1820                 /* check all memory Phi's */
1821                 if (! is_Phi(phi))
1822                         continue;
1823
1824                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1825
1826                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1827                         ir_node    *pred = get_irn_n(phi, j);
1828                         node_entry *pe   = get_irn_ne(pred, env);
1829
1830                         if (pe->pscc != ne->pscc) {
1831                                 /* not in the same SCC, is region const */
1832                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1833
1834                                 pe->phi  = phi;
1835                                 pe->pos  = j;
1836                                 pe->next = phi_list;
1837                                 phi_list = pe;
1838                         }
1839                 }
1840         }
1841         /* no Phis no fun */
1842         assert(phi_list != NULL && "DFS found a loop without Phi");
1843
1844         /* for now, we cannot handle more than one input (only reducible cf) */
1845         if (phi_list->next != NULL)
1846                 return;
1847
1848         for (load = pscc->head; load; load = next) {
1849                 ir_mode *load_mode;
1850                 node_entry *ne = get_irn_ne(load, env);
1851                 next = ne->next;
1852
1853                 if (is_Load(load)) {
1854                         ldst_info_t *info = get_irn_link(load);
1855                         ir_node     *ptr = get_Load_ptr(load);
1856
1857                         /* for now, we cannot handle Loads with exceptions */
1858                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1859                                 continue;
1860
1861                         /* for now, we can only move Load(Global) */
1862                         if (! is_Global(ptr))
1863                                 continue;
1864                         ent       = get_Global_entity(ptr);
1865                         load_mode = get_Load_mode(load);
1866                         for (other = pscc->head; other != NULL; other = next_other) {
1867                                 node_entry *ne = get_irn_ne(other, env);
1868                                 next_other = ne->next;
1869
1870                                 if (is_Store(other)) {
1871                                         ir_alias_relation rel = get_alias_relation(
1872                                                 current_ir_graph,
1873                                                 get_Store_ptr(other),
1874                                                 get_irn_mode(get_Store_value(other)),
1875                                                 ptr, load_mode);
1876                                         /* if the might be an alias, we cannot pass this Store */
1877                                         if (rel != ir_no_alias)
1878                                                 break;
1879                                 }
1880                                 /* only Phis and pure Calls are allowed here, so ignore them */
1881                         }
1882                         if (other == NULL) {
1883                                 ldst_info_t *ninfo;
1884                                 phi_entry   *pe;
1885                                 dbg_info    *db;
1886
1887                                 /* yep, no aliasing Store found, Load can be moved */
1888                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1889
1890                                 db   = get_irn_dbg_info(load);
1891                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1892                                         int     pos   = pe->pos;
1893                                         ir_node *phi  = pe->phi;
1894                                         ir_node *blk  = get_nodes_block(phi);
1895                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1896                                         ir_node *irn, *mem;
1897                                         avail_entry_t entry, *res;
1898
1899                                         entry.ptr  = ptr;
1900                                         entry.mode = load_mode;
1901                                         res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1902                                         if (res != NULL) {
1903                                                 irn = res->load;
1904                                         } else {
1905                                                 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1906                                                 entry.load = irn;
1907                                                 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1908                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1909                                         }
1910                                         pe->load = irn;
1911                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1912
1913                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1914                                         set_Phi_pred(phi, pos, mem);
1915
1916                                         ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1917                                 }
1918
1919                                 /* now kill the old Load */
1920                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1921                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1922
1923                                 env->changes |= DF_CHANGED;
1924                         }
1925                 }
1926         }
1927         del_set(avail);
1928 }  /* move_loads_out_of_loops */
1929
1930 /**
1931  * Process a loop SCC.
1932  *
1933  * @param pscc  the SCC
1934  * @param env   the loop environment
1935  */
1936 static void process_loop(scc *pscc, loop_env *env)
1937 {
1938         ir_node *irn, *next, *header = NULL;
1939         node_entry *b, *h = NULL;
1940         int j, only_phi, num_outside, process = 0;
1941         ir_node *out_rc;
1942
1943         /* find the header block for this scc */
1944         for (irn = pscc->head; irn; irn = next) {
1945                 node_entry *e = get_irn_ne(irn, env);
1946                 ir_node *block = get_nodes_block(irn);
1947
1948                 next = e->next;
1949                 b = get_irn_ne(block, env);
1950
1951                 if (header != NULL) {
1952                         if (h->POnum < b->POnum) {
1953                                 header = block;
1954                                 h      = b;
1955                         }
1956                 } else {
1957                         header = block;
1958                         h      = b;
1959                 }
1960         }
1961
1962         /* check if this scc contains only Phi, Loads or Stores nodes */
1963         only_phi    = 1;
1964         num_outside = 0;
1965         out_rc      = NULL;
1966         for (irn = pscc->head; irn; irn = next) {
1967                 node_entry *e = get_irn_ne(irn, env);
1968
1969                 next = e->next;
1970                 switch (get_irn_opcode(irn)) {
1971                 case iro_Call:
1972                         if (is_Call_pure(irn)) {
1973                                 /* pure calls can be treated like loads */
1974                                 only_phi = 0;
1975                                 break;
1976                         }
1977                         /* non-pure calls must be handle like may-alias Stores */
1978                         goto fail;
1979                 case iro_CopyB:
1980                         /* cannot handle CopyB yet */
1981                         goto fail;
1982                 case iro_Load:
1983                         process = 1;
1984                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1985                                 /* cannot handle loops with volatile Loads */
1986                                 goto fail;
1987                         }
1988                         only_phi = 0;
1989                         break;
1990                 case iro_Store:
1991                         if (get_Store_volatility(irn) == volatility_is_volatile) {
1992                                 /* cannot handle loops with volatile Stores */
1993                                 goto fail;
1994                         }
1995                         only_phi = 0;
1996                         break;
1997                 default:
1998                         only_phi = 0;
1999                         break;
2000                 case iro_Phi:
2001                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2002                                 ir_node *pred  = get_irn_n(irn, j);
2003                                 node_entry *pe = get_irn_ne(pred, env);
2004
2005                                 if (pe->pscc != e->pscc) {
2006                                         /* not in the same SCC, must be a region const */
2007                                         if (! is_rc(pred, header)) {
2008                                                 /* not a memory loop */
2009                                                 goto fail;
2010                                         }
2011                                         if (out_rc == NULL) {
2012                                                 /* first region constant */
2013                                                 out_rc = pred;
2014                                                 ++num_outside;
2015                                         } else if (out_rc != pred) {
2016                                                 /* another region constant */
2017                                                 ++num_outside;
2018                                         }
2019                                 }
2020                         }
2021                         break;
2022                 }
2023         }
2024         if (! process)
2025                 goto fail;
2026
2027         /* found a memory loop */
2028         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
2029         if (only_phi && num_outside == 1) {
2030                 /* a phi cycle with only one real predecessor can be collapsed */
2031                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
2032
2033                 for (irn = pscc->head; irn; irn = next) {
2034                         node_entry *e = get_irn_ne(irn, env);
2035                         next = e->next;
2036                         exchange(irn, out_rc);
2037                 }
2038                 env->changes |= DF_CHANGED;
2039                 return;
2040         }
2041
2042 #ifdef DEBUG_libfirm
2043         for (irn = pscc->head; irn; irn = next) {
2044                 node_entry *e = get_irn_ne(irn, env);
2045                 next = e->next;
2046                 DB((dbg, LEVEL_2, " %+F,", irn));
2047         }
2048         DB((dbg, LEVEL_2, "\n"));
2049 #endif
2050         move_loads_out_of_loops(pscc, env);
2051
2052 fail:
2053         ;
2054 }  /* process_loop */
2055
2056 /**
2057  * Process a SCC.
2058  *
2059  * @param pscc  the SCC
2060  * @param env   the loop environment
2061  */
2062 static void process_scc(scc *pscc, loop_env *env)
2063 {
2064         ir_node *head = pscc->head;
2065         node_entry *e = get_irn_ne(head, env);
2066
2067 #ifdef DEBUG_libfirm
2068         {
2069                 ir_node *irn, *next;
2070
2071                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2072                 for (irn = pscc->head; irn; irn = next) {
2073                         node_entry *e = get_irn_ne(irn, env);
2074
2075                         next = e->next;
2076
2077                         DB((dbg, LEVEL_4, " %+F,", irn));
2078                 }
2079                 DB((dbg, LEVEL_4, "\n"));
2080         }
2081 #endif
2082
2083         if (e->next != NULL) {
2084                 /* this SCC has more than one member */
2085                 process_loop(pscc, env);
2086         }
2087 }  /* process_scc */
2088
2089 /**
2090  * Do Tarjan's SCC algorithm and drive load/store optimization.
2091  *
2092  * @param irn  start at this node
2093  * @param env  the loop environment
2094  */
2095 static void dfs(ir_node *irn, loop_env *env)
2096 {
2097         int i, n;
2098         node_entry *node = get_irn_ne(irn, env);
2099
2100         mark_irn_visited(irn);
2101
2102         node->DFSnum = env->nextDFSnum++;
2103         node->low    = node->DFSnum;
2104         push(env, irn);
2105
2106         /* handle preds */
2107         if (is_Phi(irn) || is_Sync(irn)) {
2108                 n = get_irn_arity(irn);
2109                 for (i = 0; i < n; ++i) {
2110                         ir_node *pred = get_irn_n(irn, i);
2111                         node_entry *o = get_irn_ne(pred, env);
2112
2113                         if (!irn_visited(pred)) {
2114                                 dfs(pred, env);
2115                                 node->low = MIN(node->low, o->low);
2116                         }
2117                         if (o->DFSnum < node->DFSnum && o->in_stack)
2118                                 node->low = MIN(o->DFSnum, node->low);
2119                 }
2120         } else if (is_fragile_op(irn)) {
2121                 ir_node *pred = get_fragile_op_mem(irn);
2122                 node_entry *o = get_irn_ne(pred, env);
2123
2124                 if (!irn_visited(pred)) {
2125                         dfs(pred, env);
2126                         node->low = MIN(node->low, o->low);
2127                 }
2128                 if (o->DFSnum < node->DFSnum && o->in_stack)
2129                         node->low = MIN(o->DFSnum, node->low);
2130         } else if (is_Proj(irn)) {
2131                 ir_node *pred = get_Proj_pred(irn);
2132                 node_entry *o = get_irn_ne(pred, env);
2133
2134                 if (!irn_visited(pred)) {
2135                         dfs(pred, env);
2136                         node->low = MIN(node->low, o->low);
2137                 }
2138                 if (o->DFSnum < node->DFSnum && o->in_stack)
2139                         node->low = MIN(o->DFSnum, node->low);
2140         }
2141         else {
2142                  /* IGNORE predecessors */
2143         }
2144
2145         if (node->low == node->DFSnum) {
2146                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2147                 ir_node *x;
2148
2149                 pscc->head = NULL;
2150                 do {
2151                         node_entry *e;
2152
2153                         x = pop(env);
2154                         e = get_irn_ne(x, env);
2155                         e->pscc    = pscc;
2156                         e->next    = pscc->head;
2157                         pscc->head = x;
2158                 } while (x != irn);
2159
2160                 process_scc(pscc, env);
2161         }
2162 }  /* dfs */
2163
2164 /**
2165  * Do the DFS on the memory edges a graph.
2166  *
2167  * @param irg  the graph to process
2168  * @param env  the loop environment
2169  */
2170 static void do_dfs(ir_graph *irg, loop_env *env)
2171 {
2172         ir_graph *rem = current_ir_graph;
2173         ir_node  *endblk, *end;
2174         int      i;
2175
2176         current_ir_graph = irg;
2177         inc_irg_visited(irg);
2178
2179         /* visit all memory nodes */
2180         endblk = get_irg_end_block(irg);
2181         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2182                 ir_node *pred = get_Block_cfgpred(endblk, i);
2183
2184                 pred = skip_Proj(pred);
2185                 if (is_Return(pred))
2186                         dfs(get_Return_mem(pred), env);
2187                 else if (is_Raise(pred))
2188                         dfs(get_Raise_mem(pred), env);
2189                 else if (is_fragile_op(pred))
2190                         dfs(get_fragile_op_mem(pred), env);
2191                 else {
2192                         assert(0 && "Unknown EndBlock predecessor");
2193                 }
2194         }
2195
2196         /* visit the keep-alives */
2197         end = get_irg_end(irg);
2198         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2199                 ir_node *ka = get_End_keepalive(end, i);
2200
2201                 if (is_Phi(ka) && !irn_visited(ka))
2202                         dfs(ka, env);
2203         }
2204         current_ir_graph = rem;
2205 }  /* do_dfs */
2206
2207 /**
2208  * Initialize new phase data. We do this always explicit, so return NULL here
2209  */
2210 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data)
2211 {
2212         (void)ph;
2213         (void)irn;
2214         (void)data;
2215         return NULL;
2216 }  /* init_loop_data */
2217
2218 /**
2219  * Optimize Loads/Stores in loops.
2220  *
2221  * @param irg  the graph
2222  */
2223 static int optimize_loops(ir_graph *irg)
2224 {
2225         loop_env env;
2226
2227         env.stack         = NEW_ARR_F(ir_node *, 128);
2228         env.tos           = 0;
2229         env.nextDFSnum    = 0;
2230         env.POnum         = 0;
2231         env.changes       = 0;
2232         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2233
2234         /* calculate the SCC's and drive loop optimization. */
2235         do_dfs(irg, &env);
2236
2237         DEL_ARR_F(env.stack);
2238         phase_free(&env.ph);
2239
2240         return env.changes;
2241 }  /* optimize_loops */
2242
2243 /*
2244  * do the load store optimization
2245  */
2246 int optimize_load_store(ir_graph *irg)
2247 {
2248         walk_env_t env;
2249
2250         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2251
2252         assert(get_irg_phase_state(irg) != phase_building);
2253         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2254                 "LoadStore optimization needs pinned graph");
2255
2256         /* we need landing pads */
2257         remove_critical_cf_edges(irg);
2258
2259         edges_assure(irg);
2260
2261         /* for Phi optimization post-dominators are needed ... */
2262         assure_postdoms(irg);
2263
2264         if (get_opt_alias_analysis()) {
2265                 assure_irg_entity_usage_computed(irg);
2266                 assure_irp_globals_entity_usage_computed();
2267         }
2268
2269         obstack_init(&env.obst);
2270         env.changes = 0;
2271
2272         /* init the links, then collect Loads/Stores/Proj's in lists */
2273         master_visited = 0;
2274         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2275
2276         /* now we have collected enough information, optimize */
2277         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2278
2279         env.changes |= optimize_loops(irg);
2280
2281         obstack_free(&env.obst, NULL);
2282
2283         /* Handle graph state */
2284         if (env.changes) {
2285                 set_irg_outs_inconsistent(irg);
2286                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2287         }
2288
2289         if (env.changes & CF_CHANGED) {
2290                 /* is this really needed: Yes, control flow changed, block might
2291                 have Bad() predecessors. */
2292                 set_irg_doms_inconsistent(irg);
2293         }
2294         return env.changes != 0;
2295 }  /* optimize_load_store */
2296
2297 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2298 {
2299         return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2300 }  /* optimize_load_store_pass */