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