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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  * @version $Id$
  25  */
  26 #include "config.h"
  27
  28 #include <string.h>
  29
  30 #include "iroptimize.h"
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irmode_t.h"
  34 #include "iropt_t.h"
  35 #include "ircons_t.h"
  36 #include "irgmod.h"
  37 #include "irgwalk.h"
  38 #include "irvrfy.h"
  39 #include "tv_t.h"
  40 #include "dbginfo_t.h"
  41 #include "iropt_dbg.h"
  42 #include "irflag_t.h"
  43 #include "array_t.h"
  44 #include "irhooks.h"
  45 #include "iredges.h"
  46 #include "irtools.h"
  47 #include "opt_polymorphy.h"
  48 #include "irmemory.h"
  49 #include "irphase_t.h"
  50 #include "irgopt.h"
  51 #include "set.h"
  52 #include "debug.h"
  53
  54 /** The debug handle. */
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  56
  57 #ifdef DO_CACHEOPT
  58 #include "cacheopt/cachesim.h"
  59 #endif
  60
  61 #undef IMAX
  62 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  63
  64 #define MAX_PROJ        IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
  65
  66 enum changes_t {
  67         DF_CHANGED = 1,       /**< data flow changed */
  68         CF_CHANGED = 2,       /**< control flow changed */
  69 };
  70
  71 /**
  72  * walker environment
  73  */
  74 typedef struct _walk_env_t {
  75         struct obstack obst;          /**< list of all stores */
  76         unsigned changes;             /**< a bitmask of graph changes */
  77 } walk_env_t;
  78
  79 /** A Load/Store info. */
  80 typedef struct _ldst_info_t {
  81         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  82         ir_node  *exc_block;          /**< the exception block if available */
  83         int      exc_idx;             /**< predecessor index in the exception block */
  84         unsigned visited;             /**< visited counter for breaking loops */
  85 } ldst_info_t;
  86
  87 /**
  88  * flags for control flow.
  89  */
  90 enum block_flags_t {
  91         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  92         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  93 };
  94
  95 /**
  96  * a Block info.
  97  */
  98 typedef struct _block_info_t {
  99         unsigned flags;               /**< flags for the block */
 100 } block_info_t;
 101
 102 /** the master visited flag for loop detection. */
 103 static unsigned master_visited = 0;
 104
 105 #define INC_MASTER()       ++master_visited
 106 #define MARK_NODE(info)    (info)->visited = master_visited
 107 #define NODE_VISITED(info) (info)->visited >= master_visited
 108
 109 /**
 110  * get the Load/Store info of a node
 111  */
 112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 113         ldst_info_t *info = get_irn_link(node);
 114
 115         if (! info) {
 116                 info = obstack_alloc(obst, sizeof(*info));
 117                 memset(info, 0, sizeof(*info));
 118                 set_irn_link(node, info);
 119         }
 120         return info;
 121 }  /* get_ldst_info */
 122
 123 /**
 124  * get the Block info of a node
 125  */
 126 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
 127         block_info_t *info = get_irn_link(node);
 128
 129         if (! info) {
 130                 info = obstack_alloc(obst, sizeof(*info));
 131                 memset(info, 0, sizeof(*info));
 132                 set_irn_link(node, info);
 133         }
 134         return info;
 135 }  /* get_block_info */
 136
 137 /**
 138  * update the projection info for a Load/Store
 139  */
 140 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 141 {
 142         long nr = get_Proj_proj(proj);
 143
 144         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 145
 146         if (info->projs[nr]) {
 147                 /* there is already one, do CSE */
 148                 exchange(proj, info->projs[nr]);
 149                 return DF_CHANGED;
 150         }
 151         else {
 152                 info->projs[nr] = proj;
 153                 return 0;
 154         }
 155 }  /* update_projs */
 156
 157 /**
 158  * update the exception block info for a Load/Store node.
 159  *
 160  * @param info   the load/store info struct
 161  * @param block  the exception handler block for this load/store
 162  * @param pos    the control flow input of the block
 163  */
 164 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 165 {
 166         assert(info->exc_block == NULL && "more than one exception block found");
 167
 168         info->exc_block = block;
 169         info->exc_idx   = pos;
 170         return 0;
 171 }  /* update_exc */
 172
 173 /** Return the number of uses of an address node */
 174 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 175
 176 /**
 177  * walker, collects all Load/Store/Proj nodes
 178  *
 179  * walks from Start -> End
 180  */
 181 static void collect_nodes(ir_node *node, void *env)
 182 {
 183         ir_opcode   opcode = get_irn_opcode(node);
 184         ir_node     *pred, *blk, *pred_blk;
 185         ldst_info_t *ldst_info;
 186         walk_env_t  *wenv = env;
 187
 188         if (opcode == iro_Proj) {
 189                 pred   = get_Proj_pred(node);
 190                 opcode = get_irn_opcode(pred);
 191
 192                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 193                         ldst_info = get_ldst_info(pred, &wenv->obst);
 194
 195                         wenv->changes |= update_projs(ldst_info, node);
 196
 197                         /*
 198                          * Place the Proj's to the same block as the
 199                          * predecessor Load. This is always ok and prevents
 200                          * "non-SSA" form after optimizations if the Proj
 201                          * is in a wrong block.
 202                          */
 203                         blk      = get_nodes_block(node);
 204                         pred_blk = get_nodes_block(pred);
 205                         if (blk != pred_blk) {
 206                                 wenv->changes |= DF_CHANGED;
 207                                 set_nodes_block(node, pred_blk);
 208                         }
 209                 }
 210         } else if (opcode == iro_Block) {
 211                 int i;
 212
 213                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 214                         ir_node      *pred_block, *proj;
 215                         block_info_t *bl_info;
 216                         int          is_exc = 0;
 217
 218                         pred = proj = get_Block_cfgpred(node, i);
 219
 220                         if (is_Proj(proj)) {
 221                                 pred   = get_Proj_pred(proj);
 222                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 223                         }
 224
 225                         /* ignore Bad predecessors, they will be removed later */
 226                         if (is_Bad(pred))
 227                                 continue;
 228
 229                         pred_block = get_nodes_block(pred);
 230                         bl_info    = get_block_info(pred_block, &wenv->obst);
 231
 232                         if (is_fragile_op(pred) && is_exc)
 233                                 bl_info->flags |= BLOCK_HAS_EXC;
 234                         else if (is_irn_forking(pred))
 235                                 bl_info->flags |= BLOCK_HAS_COND;
 236
 237                         opcode = get_irn_opcode(pred);
 238                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 239                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 240
 241                                 wenv->changes |= update_exc(ldst_info, node, i);
 242                         }
 243                 }
 244         }
 245 }  /* collect_nodes */
 246
 247 /**
 248  * Returns an entity if the address ptr points to a constant one.
 249  *
 250  * @param ptr  the address
 251  *
 252  * @return an entity or NULL
 253  */
 254 static ir_entity *find_constant_entity(ir_node *ptr)
 255 {
 256         for (;;) {
 257                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 258                         return get_SymConst_entity(ptr);
 259                 } else if (is_Sel(ptr)) {
 260                         ir_entity *ent = get_Sel_entity(ptr);
 261                         ir_type   *tp  = get_entity_owner(ent);
 262
 263                         /* Do not fiddle with polymorphism. */
 264                         if (is_Class_type(get_entity_owner(ent)) &&
 265                                 ((get_entity_n_overwrites(ent)    != 0) ||
 266                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 267                                 return NULL;
 268
 269                         if (is_Array_type(tp)) {
 270                                 /* check bounds */
 271                                 int i, n;
 272
 273                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 274                                         ir_node *bound;
 275                                         tarval *tlower, *tupper;
 276                                         ir_node *index = get_Sel_index(ptr, i);
 277                                         tarval *tv     = computed_value(index);
 278
 279                                         /* check if the index is constant */
 280                                         if (tv == tarval_bad)
 281                                                 return NULL;
 282
 283                                         bound  = get_array_lower_bound(tp, i);
 284                                         tlower = computed_value(bound);
 285                                         bound  = get_array_upper_bound(tp, i);
 286                                         tupper = computed_value(bound);
 287
 288                                         if (tlower == tarval_bad || tupper == tarval_bad)
 289                                                 return NULL;
 290
 291                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 292                                                 return NULL;
 293                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 294                                                 return NULL;
 295
 296                                         /* ok, bounds check finished */
 297                                 }
 298                         }
 299
 300                         if (variability_constant == get_entity_variability(ent))
 301                                 return ent;
 302
 303                         /* try next */
 304                         ptr = get_Sel_ptr(ptr);
 305                 } else if (is_Add(ptr)) {
 306                         ir_node *l = get_Add_left(ptr);
 307                         ir_node *r = get_Add_right(ptr);
 308
 309                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 310                                 ptr = l;
 311                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 312                                 ptr = r;
 313                         else
 314                                 return NULL;
 315
 316                         /* for now, we support only one addition, reassoc should fold all others */
 317                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 318                                 return NULL;
 319                 } else if (is_Sub(ptr)) {
 320                         ir_node *l = get_Sub_left(ptr);
 321                         ir_node *r = get_Sub_right(ptr);
 322
 323                         if (get_irn_mode(l) == get_irn_mode(ptr) &&     is_Const(r))
 324                                 ptr = l;
 325                         else
 326                                 return NULL;
 327                         /* for now, we support only one substraction, reassoc should fold all others */
 328                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 329                                 return NULL;
 330                 } else
 331                         return NULL;
 332         }
 333 }  /* find_constant_entity */
 334
 335 /**
 336  * Return the Selection index of a Sel node from dimension n
 337  */
 338 static long get_Sel_array_index_long(ir_node *n, int dim) {
 339         ir_node *index = get_Sel_index(n, dim);
 340         assert(is_Const(index));
 341         return get_tarval_long(get_Const_tarval(index));
 342 }  /* get_Sel_array_index_long */
 343
 344 /**
 345  * Returns the accessed component graph path for an
 346  * node computing an address.
 347  *
 348  * @param ptr    the node computing the address
 349  * @param depth  current depth in steps upward from the root
 350  *               of the address
 351  */
 352 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 353         compound_graph_path *res = NULL;
 354         ir_entity           *root, *field, *ent;
 355         int                 path_len, pos, idx;
 356         tarval              *tv;
 357         ir_type             *tp;
 358
 359         if (is_SymConst(ptr)) {
 360                 /* a SymConst. If the depth is 0, this is an access to a global
 361                  * entity and we don't need a component path, else we know
 362                  * at least its length.
 363                  */
 364                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 365                 root = get_SymConst_entity(ptr);
 366                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 367         } else if (is_Sel(ptr)) {
 368                 /* it's a Sel, go up until we find the root */
 369                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 370                 if (res == NULL)
 371                         return NULL;
 372
 373                 /* fill up the step in the path at the current position */
 374                 field    = get_Sel_entity(ptr);
 375                 path_len = get_compound_graph_path_length(res);
 376                 pos      = path_len - depth - 1;
 377                 set_compound_graph_path_node(res, pos, field);
 378
 379                 if (is_Array_type(get_entity_owner(field))) {
 380                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 381                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 382                 }
 383         } else if (is_Add(ptr)) {
 384                 ir_node *l    = get_Add_left(ptr);
 385                 ir_node *r    = get_Add_right(ptr);
 386                 ir_mode *mode = get_irn_mode(ptr);
 387                 tarval  *tmp;
 388
 389                 if (is_Const(r) && get_irn_mode(l) == mode) {
 390                         ptr = l;
 391                         tv  = get_Const_tarval(r);
 392                 } else {
 393                         ptr = r;
 394                         tv  = get_Const_tarval(l);
 395                 }
 396 ptr_arith:
 397                 mode = get_tarval_mode(tv);
 398                 tmp  = tv;
 399
 400                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 401                 if (is_Sel(ptr)) {
 402                         field = get_Sel_entity(ptr);
 403                 } else {
 404                         field = get_SymConst_entity(ptr);
 405                 }
 406                 idx = 0;
 407                 for (ent = field;;) {
 408                         unsigned size;
 409                         tarval   *sz, *tv_index, *tlower, *tupper;
 410                         ir_node  *bound;
 411
 412                         tp = get_entity_type(ent);
 413                         if (! is_Array_type(tp))
 414                                 break;
 415                         ent = get_array_element_entity(tp);
 416                         size = get_type_size_bytes(get_entity_type(ent));
 417                         sz   = new_tarval_from_long(size, mode);
 418
 419                         tv_index = tarval_div(tmp, sz);
 420                         tmp      = tarval_mod(tmp, sz);
 421
 422                         if (tv_index == tarval_bad || tmp == tarval_bad)
 423                                 return NULL;
 424
 425                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 426                         bound  = get_array_lower_bound(tp, 0);
 427                         tlower = computed_value(bound);
 428                         bound  = get_array_upper_bound(tp, 0);
 429                         tupper = computed_value(bound);
 430
 431                         if (tlower == tarval_bad || tupper == tarval_bad)
 432                                 return NULL;
 433
 434                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 435                                 return NULL;
 436                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 437                                 return NULL;
 438
 439                         /* ok, bounds check finished */
 440                         ++idx;
 441                 }
 442                 if (! tarval_is_null(tmp)) {
 443                         /* access to some struct/union member */
 444                         return NULL;
 445                 }
 446
 447                 /* should be at least ONE array */
 448                 if (idx == 0)
 449                         return NULL;
 450
 451                 res = rec_get_accessed_path(ptr, depth + idx);
 452                 if (res == NULL)
 453                         return NULL;
 454
 455                 path_len = get_compound_graph_path_length(res);
 456                 pos      = path_len - depth - idx;
 457
 458                 for (ent = field;;) {
 459                         unsigned size;
 460                         tarval   *sz, *tv_index;
 461                         long     index;
 462
 463                         tp = get_entity_type(ent);
 464                         if (! is_Array_type(tp))
 465                                 break;
 466                         ent = get_array_element_entity(tp);
 467                         set_compound_graph_path_node(res, pos, ent);
 468
 469                         size = get_type_size_bytes(get_entity_type(ent));
 470                         sz   = new_tarval_from_long(size, mode);
 471
 472                         tv_index = tarval_div(tv, sz);
 473                         tv       = tarval_mod(tv, sz);
 474
 475                         /* worked above, should work again */
 476                         assert(tv_index != tarval_bad && tv != tarval_bad);
 477
 478                         /* bounds already checked above */
 479                         index = get_tarval_long(tv_index);
 480                         set_compound_graph_path_array_index(res, pos, index);
 481                         ++pos;
 482                 }
 483         } else if (is_Sub(ptr)) {
 484                 ir_node *l = get_Sub_left(ptr);
 485                 ir_node *r = get_Sub_right(ptr);
 486
 487                 ptr = l;
 488                 tv  = get_Const_tarval(r);
 489                 tv  = tarval_neg(tv);
 490                 goto ptr_arith;
 491         }
 492         return res;
 493 }  /* rec_get_accessed_path */
 494
 495 /**
 496  * Returns an access path or NULL.  The access path is only
 497  * valid, if the graph is in phase_high and _no_ address computation is used.
 498  */
 499 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 500         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 501         return gr;
 502 }  /* get_accessed_path */
 503
 504 typedef struct path_entry {
 505         ir_entity         *ent;
 506         struct path_entry *next;
 507         long              index;
 508 } path_entry;
 509
 510 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
 511         path_entry       entry, *p;
 512         ir_entity        *ent, *field;
 513         ir_initializer_t *initializer;
 514         tarval           *tv;
 515         ir_type          *tp;
 516         unsigned         n;
 517
 518         entry.next = next;
 519         if (is_SymConst(ptr)) {
 520                 /* found the root */
 521                 ent         = get_SymConst_entity(ptr);
 522                 initializer = get_entity_initializer(ent);
 523                 for (p = next; p != NULL;) {
 524                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 525                                 return NULL;
 526                         n  = get_initializer_compound_n_entries(initializer);
 527                         tp = get_entity_type(ent);
 528
 529                         if (is_Array_type(tp)) {
 530                                 ent = get_array_element_entity(tp);
 531                                 if (ent != p->ent) {
 532                                         /* a missing [0] */
 533                                         if (0 >= n)
 534                                                 return NULL;
 535                                         initializer = get_initializer_compound_value(initializer, 0);
 536                                         continue;
 537                                 }
 538                         }
 539                         if (p->index >= (int) n)
 540                                 return NULL;
 541                         initializer = get_initializer_compound_value(initializer, p->index);
 542
 543                         ent = p->ent;
 544                         p   = p->next;
 545                 }
 546                 tp = get_entity_type(ent);
 547                 while (is_Array_type(tp)) {
 548                         ent = get_array_element_entity(tp);
 549                         tp = get_entity_type(ent);
 550                         /* a missing [0] */
 551                         n  = get_initializer_compound_n_entries(initializer);
 552                         if (0 >= n)
 553                                 return NULL;
 554                         initializer = get_initializer_compound_value(initializer, 0);
 555                 }
 556
 557                 switch (initializer->kind) {
 558                 case IR_INITIALIZER_CONST:
 559                         return get_initializer_const_value(initializer);
 560                 case IR_INITIALIZER_TARVAL:
 561                 case IR_INITIALIZER_NULL:
 562                 default:
 563                         return NULL;
 564                 }
 565         } else if (is_Sel(ptr)) {
 566                 entry.ent = field = get_Sel_entity(ptr);
 567                 tp = get_entity_owner(field);
 568                 if (is_Array_type(tp)) {
 569                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 570                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 571                 } else {
 572                         int i, n_members = get_compound_n_members(tp);
 573                         for (i = 0; i < n_members; ++i) {
 574                                 if (get_compound_member(tp, i) == field)
 575                                         break;
 576                         }
 577                         if (i >= n_members) {
 578                                 /* not found: should NOT happen */
 579                                 return NULL;
 580                         }
 581                         entry.index = i;
 582                 }
 583                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 584         }  else if (is_Add(ptr)) {
 585                 ir_node  *l = get_Add_left(ptr);
 586                 ir_node  *r = get_Add_right(ptr);
 587                 ir_mode  *mode;
 588                 unsigned pos;
 589
 590                 if (is_Const(r)) {
 591                         ptr = l;
 592                         tv  = get_Const_tarval(r);
 593                 } else {
 594                         ptr = r;
 595                         tv  = get_Const_tarval(l);
 596                 }
 597 ptr_arith:
 598                 mode = get_tarval_mode(tv);
 599
 600                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 601                 if (is_Sel(ptr)) {
 602                         field = get_Sel_entity(ptr);
 603                 } else {
 604                         field = get_SymConst_entity(ptr);
 605                 }
 606
 607                 /* count needed entries */
 608                 pos = 0;
 609                 for (ent = field;;) {
 610                         tp = get_entity_type(ent);
 611                         if (! is_Array_type(tp))
 612                                 break;
 613                         ent = get_array_element_entity(tp);
 614                         ++pos;
 615                 }
 616                 /* should be at least ONE entry */
 617                 if (pos == 0)
 618                         return NULL;
 619
 620                 /* allocate the right number of entries */
 621                 NEW_ARR_A(path_entry, p, pos);
 622
 623                 /* fill them up */
 624                 pos = 0;
 625                 for (ent = field;;) {
 626                         unsigned size;
 627                         tarval   *sz, *tv_index, *tlower, *tupper;
 628                         long     index;
 629                         ir_node  *bound;
 630
 631                         tp = get_entity_type(ent);
 632                         if (! is_Array_type(tp))
 633                                 break;
 634                         ent = get_array_element_entity(tp);
 635                         p[pos].ent  = ent;
 636                         p[pos].next = &p[pos + 1];
 637
 638                         size = get_type_size_bytes(get_entity_type(ent));
 639                         sz   = new_tarval_from_long(size, mode);
 640
 641                         tv_index = tarval_div(tv, sz);
 642                         tv       = tarval_mod(tv, sz);
 643
 644                         if (tv_index == tarval_bad || tv == tarval_bad)
 645                                 return NULL;
 646
 647                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 648                         bound  = get_array_lower_bound(tp, 0);
 649                         tlower = computed_value(bound);
 650                         bound  = get_array_upper_bound(tp, 0);
 651                         tupper = computed_value(bound);
 652
 653                         if (tlower == tarval_bad || tupper == tarval_bad)
 654                                 return NULL;
 655
 656                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 657                                 return NULL;
 658                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 659                                 return NULL;
 660
 661                         /* ok, bounds check finished */
 662                         index = get_tarval_long(tv_index);
 663                         p[pos].index = index;
 664                         ++pos;
 665                 }
 666                 if (! tarval_is_null(tv)) {
 667                         /* hmm, wrong access */
 668                         return NULL;
 669                 }
 670                 p[pos - 1].next = next;
 671                 return rec_find_compound_ent_value(ptr, p);
 672         } else if (is_Sub(ptr)) {
 673                 ir_node *l = get_Sub_left(ptr);
 674                 ir_node *r = get_Sub_right(ptr);
 675
 676                 ptr = l;
 677                 tv  = get_Const_tarval(r);
 678                 tv  = tarval_neg(tv);
 679                 goto ptr_arith;
 680         }
 681         return NULL;
 682 }
 683
 684 static ir_node *find_compound_ent_value(ir_node *ptr) {
 685         return rec_find_compound_ent_value(ptr, NULL);
 686 }
 687
 688 /* forward */
 689 static void reduce_adr_usage(ir_node *ptr);
 690
 691 /**
 692  * Update a Load that may have lost its users.
 693  */
 694 static void handle_load_update(ir_node *load) {
 695         ldst_info_t *info = get_irn_link(load);
 696
 697         /* do NOT touch volatile loads for now */
 698         if (get_Load_volatility(load) == volatility_is_volatile)
 699                 return;
 700
 701         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 702                 ir_node *ptr = get_Load_ptr(load);
 703                 ir_node *mem = get_Load_mem(load);
 704
 705                 /* a Load whose value is neither used nor exception checked, remove it */
 706                 exchange(info->projs[pn_Load_M], mem);
 707                 if (info->projs[pn_Load_X_regular])
 708                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 709                 kill_node(load);
 710                 reduce_adr_usage(ptr);
 711         }
 712 }  /* handle_load_update */
 713
 714 /**
 715  * A use of an address node has vanished. Check if this was a Proj
 716  * node and update the counters.
 717  */
 718 static void reduce_adr_usage(ir_node *ptr) {
 719         if (is_Proj(ptr)) {
 720                 if (get_irn_n_edges(ptr) <= 0) {
 721                         /* this Proj is dead now */
 722                         ir_node *pred = get_Proj_pred(ptr);
 723
 724                         if (is_Load(pred)) {
 725                                 ldst_info_t *info = get_irn_link(pred);
 726                                 info->projs[get_Proj_proj(ptr)] = NULL;
 727
 728                                 /* this node lost its result proj, handle that */
 729                                 handle_load_update(pred);
 730                         }
 731                 }
 732         }
 733 }  /* reduce_adr_usage */
 734
 735 /**
 736  * Check, if an already existing value of mode old_mode can be converted
 737  * into the needed one new_mode without loss.
 738  */
 739 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 740         if (old_mode == new_mode)
 741                 return 1;
 742
 743         /* if both modes are two-complement ones, we can always convert the
 744            Stored value into the needed one. */
 745         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 746                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 747                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 748                 return 1;
 749         return 0;
 750 }  /* can_use_stored_value */
 751
 752 /**
 753  * Check whether a Call is at least pure, ie. does only read memory.
 754  */
 755 static unsigned is_Call_pure(ir_node *call) {
 756         ir_type *call_tp = get_Call_type(call);
 757         unsigned prop = get_method_additional_properties(call_tp);
 758
 759         /* check first the call type */
 760         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 761                 /* try the called entity */
 762                 ir_node *ptr = get_Call_ptr(call);
 763
 764                 if (is_Global(ptr)) {
 765                         ir_entity *ent = get_Global_entity(ptr);
 766
 767                         prop = get_entity_additional_properties(ent);
 768                 }
 769         }
 770         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 771 }  /* is_Call_pure */
 772
 773 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 774 {
 775         ir_mode *mode  = get_irn_mode(ptr);
 776         long    offset = 0;
 777
 778         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 779         for (;;) {
 780                 if (is_Add(ptr)) {
 781                         ir_node *l = get_Add_left(ptr);
 782                         ir_node *r = get_Add_right(ptr);
 783
 784                         if (get_irn_mode(l) != mode || !is_Const(r))
 785                                 break;
 786
 787                         offset += get_tarval_long(get_Const_tarval(r));
 788                         ptr     = l;
 789                 } else if (is_Sub(ptr)) {
 790                         ir_node *l = get_Sub_left(ptr);
 791                         ir_node *r = get_Sub_right(ptr);
 792
 793                         if (get_irn_mode(l) != mode || !is_Const(r))
 794                                 break;
 795
 796                         offset -= get_tarval_long(get_Const_tarval(r));
 797                         ptr     = l;
 798                 } else if (is_Sel(ptr)) {
 799                         ir_entity *ent = get_Sel_entity(ptr);
 800                         ir_type   *tp  = get_entity_owner(ent);
 801
 802                         if (is_Array_type(tp)) {
 803                                 int     size;
 804                                 ir_node *index;
 805
 806                                 /* only one dimensional arrays yet */
 807                                 if (get_Sel_n_indexs(ptr) != 1)
 808                                         break;
 809                                 index = get_Sel_index(ptr, 0);
 810                                 if (! is_Const(index))
 811                                         break;
 812
 813                                 tp = get_entity_type(ent);
 814                                 if (get_type_state(tp) != layout_fixed)
 815                                         break;
 816
 817                                 size    = get_type_size_bytes(tp);
 818                                 offset += size * get_tarval_long(get_Const_tarval(index));
 819                         } else {
 820                                 if (get_type_state(tp) != layout_fixed)
 821                                         break;
 822                                 offset += get_entity_offset(ent);
 823                         }
 824                         ptr = get_Sel_ptr(ptr);
 825                 } else
 826                         break;
 827         }
 828
 829         *pOffset = offset;
 830         return ptr;
 831 }
 832
 833 static int try_load_after_store(ir_node *load,
 834                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 835 {
 836         ldst_info_t *info;
 837         ir_node *store_ptr      = get_Store_ptr(store);
 838         long     store_offset;
 839         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 840         ir_node *store_value;
 841         ir_mode *store_mode;
 842         ir_node *load_ptr;
 843         ir_mode *load_mode;
 844         long     load_mode_len;
 845         long     store_mode_len;
 846         long     delta;
 847         int      res;
 848
 849         if (load_base_ptr != store_base_ptr)
 850                 return 0;
 851
 852         load_mode      = get_Load_mode(load);
 853         load_mode_len  = get_mode_size_bytes(load_mode);
 854         store_mode     = get_irn_mode(get_Store_value(store));
 855         store_mode_len = get_mode_size_bytes(store_mode);
 856         delta          = load_offset - store_offset;
 857         store_value    = get_Store_value(store);
 858
 859         if (delta != 0 || store_mode != load_mode) {
 860                 if (delta < 0 || delta + load_mode_len > store_mode_len)
 861                         return 0;
 862
 863                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 864                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 865                         return 0;
 866
 867
 868                 /* produce a shift to adjust offset delta */
 869                 if (delta > 0) {
 870                         ir_node *cnst;
 871
 872                         /* FIXME: only true for little endian */
 873                         cnst        = new_Const_long(mode_Iu, delta * 8);
 874                         store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
 875                                                                         store_value, cnst, store_mode);
 876                 }
 877
 878                 /* add an convert if needed */
 879                 if (store_mode != load_mode) {
 880                         store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
 881                                                                          store_value, load_mode);
 882                 }
 883         }
 884
 885         DBG_OPT_RAW(load, store_value);
 886
 887         info = get_irn_link(load);
 888         if (info->projs[pn_Load_M])
 889                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 890
 891         res = 0;
 892         /* no exception */
 893         if (info->projs[pn_Load_X_except]) {
 894                 exchange( info->projs[pn_Load_X_except], new_Bad());
 895                 res |= CF_CHANGED;
 896         }
 897         if (info->projs[pn_Load_X_regular]) {
 898                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 899                 res |= CF_CHANGED;
 900         }
 901
 902         if (info->projs[pn_Load_res])
 903                 exchange(info->projs[pn_Load_res], store_value);
 904
 905         load_ptr = get_Load_ptr(load);
 906         kill_node(load);
 907         reduce_adr_usage(load_ptr);
 908         return res | DF_CHANGED;
 909 }
 910
 911 /**
 912  * Follow the memory chain as long as there are only Loads,
 913  * alias free Stores, and constant Calls and try to replace the
 914  * current Load by a previous ones.
 915  * Note that in unreachable loops it might happen that we reach
 916  * load again, as well as we can fall into a cycle.
 917  * We break such cycles using a special visited flag.
 918  *
 919  * INC_MASTER() must be called before dive into
 920  */
 921 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 922         unsigned    res = 0;
 923         ldst_info_t *info = get_irn_link(load);
 924         ir_node     *pred;
 925         ir_node     *ptr       = get_Load_ptr(load);
 926         ir_node     *mem       = get_Load_mem(load);
 927         ir_mode     *load_mode = get_Load_mode(load);
 928
 929         for (pred = curr; load != pred; ) {
 930                 ldst_info_t *pred_info = get_irn_link(pred);
 931
 932                 /*
 933                  * a Load immediately after a Store -- a read after write.
 934                  * We may remove the Load, if both Load & Store does not have an
 935                  * exception handler OR they are in the same MacroBlock. In the latter
 936                  * case the Load cannot throw an exception when the previous Store was
 937                  * quiet.
 938                  *
 939                  * Why we need to check for Store Exception? If the Store cannot
 940                  * be executed (ROM) the exception handler might simply jump into
 941                  * the load MacroBlock :-(
 942                  * We could make it a little bit better if we would know that the
 943                  * exception handler of the Store jumps directly to the end...
 944                  */
 945                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 946                                 && info->projs[pn_Load_X_except] == NULL)
 947                                 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
 948                 {
 949                         long    load_offset;
 950                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 951                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 952
 953                         if (changes != 0)
 954                                 return res | changes;
 955                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 956                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 957                         /*
 958                          * a Load after a Load -- a read after read.
 959                          * We may remove the second Load, if it does not have an exception handler
 960                          * OR they are in the same MacroBlock. In the later case the Load cannot
 961                          * throw an exception when the previous Load was quiet.
 962                          *
 963                          * Here, there is no need to check if the previous Load has an exception
 964                          * hander because they would have exact the same exception...
 965                          */
 966                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 967                                 ir_node *value;
 968
 969                                 DBG_OPT_RAR(load, pred);
 970
 971                                 /* the result is used */
 972                                 if (info->projs[pn_Load_res]) {
 973                                         if (pred_info->projs[pn_Load_res] == NULL) {
 974                                                 /* create a new Proj again */
 975                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 976                                         }
 977                                         value = pred_info->projs[pn_Load_res];
 978
 979                                         /* add an convert if needed */
 980                                         if (get_Load_mode(pred) != load_mode) {
 981                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 982                                         }
 983
 984                                         exchange(info->projs[pn_Load_res], value);
 985                                 }
 986
 987                                 if (info->projs[pn_Load_M])
 988                                         exchange(info->projs[pn_Load_M], mem);
 989
 990                                 /* no exception */
 991                                 if (info->projs[pn_Load_X_except]) {
 992                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 993                                         res |= CF_CHANGED;
 994                                 }
 995                                 if (info->projs[pn_Load_X_regular]) {
 996                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 997                                         res |= CF_CHANGED;
 998                                 }
 999
1000                                 kill_node(load);
1001                                 reduce_adr_usage(ptr);
1002                                 return res |= DF_CHANGED;
1003                         }
1004                 }
1005
1006                 if (is_Store(pred)) {
1007                         /* check if we can pass through this store */
1008                         ir_alias_relation rel = get_alias_relation(
1009                                 current_ir_graph,
1010                                 get_Store_ptr(pred),
1011                                 get_irn_mode(get_Store_value(pred)),
1012                                 ptr, load_mode);
1013                         /* if the might be an alias, we cannot pass this Store */
1014                         if (rel != ir_no_alias)
1015                                 break;
1016                         pred = skip_Proj(get_Store_mem(pred));
1017                 } else if (is_Load(pred)) {
1018                         pred = skip_Proj(get_Load_mem(pred));
1019                 } else if (is_Call(pred)) {
1020                         if (is_Call_pure(pred)) {
1021                                 /* The called graph is at least pure, so there are no Store's
1022                                    in it. We can handle it like a Load and skip it. */
1023                                 pred = skip_Proj(get_Call_mem(pred));
1024                         } else {
1025                                 /* there might be Store's in the graph, stop here */
1026                                 break;
1027                         }
1028                 } else {
1029                         /* follow only Load chains */
1030                         break;
1031                 }
1032
1033                 /* check for cycles */
1034                 if (NODE_VISITED(pred_info))
1035                         break;
1036                 MARK_NODE(pred_info);
1037         }
1038
1039         if (is_Sync(pred)) {
1040                 int i;
1041
1042                 /* handle all Sync predecessors */
1043                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1044                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1045                         if (res)
1046                                 return res;
1047                 }
1048         }
1049
1050         return res;
1051 }  /* follow_Mem_chain */
1052
1053 /*
1054  * Check if we can replace the load by a given const from
1055  * the const code irg.
1056  */
1057 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1058         ir_mode *c_mode = get_irn_mode(c);
1059         ir_mode *l_mode = get_Load_mode(load);
1060         ir_node *res    = NULL;
1061
1062         if (c_mode != l_mode) {
1063                 /* check, if the mode matches OR can be easily converted info */
1064                 if (is_reinterpret_cast(c_mode, l_mode)) {
1065                         /* we can safely cast */
1066                         dbg_info *dbg   = get_irn_dbg_info(load);
1067                         ir_node  *block = get_nodes_block(load);
1068
1069                         /* copy the value from the const code irg and cast it */
1070                         res = copy_const_value(dbg, c);
1071                         res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
1072                 }
1073         } else {
1074                 /* copy the value from the const code irg */
1075                 res = copy_const_value(get_irn_dbg_info(load), c);
1076         }
1077         return res;
1078 }  /* can_replace_load_by_const */
1079
1080 /**
1081  * optimize a Load
1082  *
1083  * @param load  the Load node
1084  */
1085 static unsigned optimize_load(ir_node *load)
1086 {
1087         ldst_info_t *info = get_irn_link(load);
1088         ir_node     *mem, *ptr, *value;
1089         ir_entity   *ent;
1090         long        dummy;
1091         unsigned    res = 0;
1092
1093         /* do NOT touch volatile loads for now */
1094         if (get_Load_volatility(load) == volatility_is_volatile)
1095                 return 0;
1096
1097         /* the address of the load to be optimized */
1098         ptr = get_Load_ptr(load);
1099
1100         /*
1101          * Check if we can remove the exception from a Load:
1102          * This can be done, if the address is from an Sel(Alloc) and
1103          * the Sel type is a subtype of the allocated type.
1104          *
1105          * This optimizes some often used OO constructs,
1106          * like x = new O; x->t;
1107          */
1108         if (info->projs[pn_Load_X_except]) {
1109                 ir_node *addr = ptr;
1110
1111                 /* find base address */
1112                 while (is_Sel(addr))
1113                         addr = get_Sel_ptr(addr);
1114                 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1115                         /* simple case: a direct load after an Alloc. Firm Alloc throw
1116                          * an exception in case of out-of-memory. So, there is no way for an
1117                          * exception in this load.
1118                          * This code is constructed by the "exception lowering" in the Jack compiler.
1119                          */
1120                         exchange(info->projs[pn_Load_X_except], new_Bad());
1121                         info->projs[pn_Load_X_except] = NULL;
1122                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1123                         info->projs[pn_Load_X_regular] = NULL;
1124                         res |= CF_CHANGED;
1125                 }
1126         }
1127
1128         /* The mem of the Load. Must still be returned after optimization. */
1129         mem = get_Load_mem(load);
1130
1131         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1132                 /* a Load which value is neither used nor exception checked, remove it */
1133                 exchange(info->projs[pn_Load_M], mem);
1134
1135                 if (info->projs[pn_Load_X_regular]) {
1136                         /* should not happen, but if it does, remove it */
1137                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1138                         res |= CF_CHANGED;
1139                 }
1140                 kill_node(load);
1141                 reduce_adr_usage(ptr);
1142                 return res | DF_CHANGED;
1143         }
1144
1145         /* Load from a constant polymorphic field, where we can resolve
1146            polymorphism. */
1147         value = transform_polymorph_Load(load);
1148         if (value == load) {
1149                 value = NULL;
1150                 /* check if we can determine the entity that will be loaded */
1151                 ent = find_constant_entity(ptr);
1152                 if (ent != NULL                                     &&
1153                     allocation_static == get_entity_allocation(ent) &&
1154                     visibility_external_allocated != get_entity_visibility(ent)) {
1155                         /* a static allocation that is not external: there should be NO exception
1156                          * when loading even if we cannot replace the load itself. */
1157
1158                         /* no exception, clear the info field as it might be checked later again */
1159                         if (info->projs[pn_Load_X_except]) {
1160                                 exchange(info->projs[pn_Load_X_except], new_Bad());
1161                                 info->projs[pn_Load_X_except] = NULL;
1162                                 res |= CF_CHANGED;
1163                         }
1164                         if (info->projs[pn_Load_X_regular]) {
1165                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1166                                 info->projs[pn_Load_X_regular] = NULL;
1167                                 res |= CF_CHANGED;
1168                         }
1169
1170                         if (variability_constant == get_entity_variability(ent)) {
1171                                 if (is_atomic_entity(ent)) {
1172                                         /* Might not be atomic after lowering of Sels.  In this case we
1173                                          * could also load, but it's more complicated. */
1174                                         /* more simpler case: we load the content of a constant value:
1175                                          * replace it by the constant itself */
1176                                         value = get_atomic_ent_value(ent);
1177                                 } else if (ent->has_initializer) {
1178                                         /* new style initializer */
1179                                         value = find_compound_ent_value(ptr);
1180                                 } else {
1181                                         /* old style initializer */
1182                                         compound_graph_path *path = get_accessed_path(ptr);
1183
1184                                         if (path != NULL) {
1185                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1186
1187                                                 value = get_compound_ent_value_by_path(ent, path);
1188                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1189                                                 free_compound_graph_path(path);
1190                                         }
1191                                 }
1192                                 if (value != NULL)
1193                                         value = can_replace_load_by_const(load, value);
1194                         }
1195                 }
1196         }
1197         if (value != NULL) {
1198                 /* we completely replace the load by this value */
1199                 if (info->projs[pn_Load_X_except]) {
1200                         exchange(info->projs[pn_Load_X_except], new_Bad());
1201                         info->projs[pn_Load_X_except] = NULL;
1202                         res |= CF_CHANGED;
1203                 }
1204                 if (info->projs[pn_Load_X_regular]) {
1205                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1206                         info->projs[pn_Load_X_regular] = NULL;
1207                         res |= CF_CHANGED;
1208                 }
1209                 if (info->projs[pn_Load_M]) {
1210                         exchange(info->projs[pn_Load_M], mem);
1211                         res |= DF_CHANGED;
1212                 }
1213                 if (info->projs[pn_Load_res]) {
1214                         exchange(info->projs[pn_Load_res], value);
1215                         res |= DF_CHANGED;
1216                 }
1217                 kill_node(load);
1218                 reduce_adr_usage(ptr);
1219                 return res;
1220         }
1221
1222         /* Check, if the address of this load is used more than once.
1223          * If not, more load cannot be removed in any case. */
1224         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1225                 return res;
1226
1227         /*
1228          * follow the memory chain as long as there are only Loads
1229          * and try to replace current Load or Store by a previous one.
1230          * Note that in unreachable loops it might happen that we reach
1231          * load again, as well as we can fall into a cycle.
1232          * We break such cycles using a special visited flag.
1233          */
1234         INC_MASTER();
1235         res = follow_Mem_chain(load, skip_Proj(mem));
1236         return res;
1237 }  /* optimize_load */
1238
1239 /**
1240  * Check whether a value of mode new_mode would completely overwrite a value
1241  * of mode old_mode in memory.
1242  */
1243 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1244 {
1245         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1246 }  /* is_completely_overwritten */
1247
1248 /**
1249  * Check whether small is a part of large (starting at same address).
1250  */
1251 static int is_partially_same(ir_node *small, ir_node *large)
1252 {
1253         ir_mode *sm = get_irn_mode(small);
1254         ir_mode *lm = get_irn_mode(large);
1255
1256         /* FIXME: Check endianness */
1257         return is_Conv(small) && get_Conv_op(small) == large
1258             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1259             && get_mode_arithmetic(sm) == irma_twos_complement
1260             && get_mode_arithmetic(lm) == irma_twos_complement;
1261 }  /* is_partially_same */
1262
1263 /**
1264  * follow the memory chain as long as there are only Loads and alias free Stores.
1265  *
1266  * INC_MASTER() must be called before dive into
1267  */
1268 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1269         unsigned res = 0;
1270         ldst_info_t *info = get_irn_link(store);
1271         ir_node *pred;
1272         ir_node *ptr = get_Store_ptr(store);
1273         ir_node *mem = get_Store_mem(store);
1274         ir_node *value = get_Store_value(store);
1275         ir_mode *mode  = get_irn_mode(value);
1276         ir_node *block = get_nodes_block(store);
1277         ir_node *mblk  = get_Block_MacroBlock(block);
1278
1279         for (pred = curr; pred != store;) {
1280                 ldst_info_t *pred_info = get_irn_link(pred);
1281
1282                 /*
1283                  * BEWARE: one might think that checking the modes is useless, because
1284                  * if the pointers are identical, they refer to the same object.
1285                  * This is only true in strong typed languages, not is C were the following
1286                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1287                  * However, if the size of the mode that is written is bigger or equal the
1288                  * size of the old one, the old value is completely overwritten and can be
1289                  * killed ...
1290                  */
1291                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1292             get_nodes_MacroBlock(pred) == mblk) {
1293                         /*
1294                          * a Store after a Store in the same MacroBlock -- a write after write.
1295                          */
1296
1297                         /*
1298                          * We may remove the first Store, if the old value is completely
1299                          * overwritten or the old value is a part of the new value,
1300                          * 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(pred) != volatility_is_volatile
1305                                 && !pred_info->projs[pn_Store_X_except]) {
1306                                 ir_node *predvalue = get_Store_value(pred);
1307                                 ir_mode *predmode  = get_irn_mode(predvalue);
1308
1309                                 if(is_completely_overwritten(predmode, mode)
1310                                         || is_partially_same(predvalue, value)) {
1311                                         DBG_OPT_WAW(pred, store);
1312                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1313                                         kill_node(pred);
1314                                         reduce_adr_usage(ptr);
1315                                         return DF_CHANGED;
1316                                 }
1317                         }
1318
1319                         /*
1320                          * We may remove the Store, if the old value already contains
1321                          * the new value, and if it does not have an exception handler.
1322                          *
1323                          * TODO: What, if both have the same exception handler ???
1324                          */
1325                         if (get_Store_volatility(store) != volatility_is_volatile
1326                                 && !info->projs[pn_Store_X_except]) {
1327                                 ir_node *predvalue = get_Store_value(pred);
1328
1329                                 if(is_partially_same(value, predvalue)) {
1330                                         DBG_OPT_WAW(pred, store);
1331                                         exchange(info->projs[pn_Store_M], mem);
1332                                         kill_node(store);
1333                                         reduce_adr_usage(ptr);
1334                                         return DF_CHANGED;
1335                                 }
1336                         }
1337                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1338                            value == pred_info->projs[pn_Load_res]) {
1339                         /*
1340                          * a Store of a value just loaded from the same address
1341                          * -- a write after read.
1342                          * We may remove the Store, if it does not have an exception
1343                          * handler.
1344                          */
1345                         if (! info->projs[pn_Store_X_except]) {
1346                                 DBG_OPT_WAR(store, pred);
1347                                 exchange(info->projs[pn_Store_M], mem);
1348                                 kill_node(store);
1349                                 reduce_adr_usage(ptr);
1350                                 return DF_CHANGED;
1351                         }
1352                 }
1353
1354                 if (is_Store(pred)) {
1355                         /* check if we can pass through this store */
1356                         ir_alias_relation rel = get_alias_relation(
1357                                 current_ir_graph,
1358                                 get_Store_ptr(pred),
1359                                 get_irn_mode(get_Store_value(pred)),
1360                                 ptr, mode);
1361                         /* if the might be an alias, we cannot pass this Store */
1362                         if (rel != ir_no_alias)
1363                                 break;
1364                         pred = skip_Proj(get_Store_mem(pred));
1365                 } else if (is_Load(pred)) {
1366                         ir_alias_relation rel = get_alias_relation(
1367                                 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1368                                 ptr, mode);
1369                         if (rel != ir_no_alias)
1370                                 break;
1371
1372                         pred = skip_Proj(get_Load_mem(pred));
1373                 } else {
1374                         /* follow only Load chains */
1375                         break;
1376                 }
1377
1378                 /* check for cycles */
1379                 if (NODE_VISITED(pred_info))
1380                         break;
1381                 MARK_NODE(pred_info);
1382         }
1383
1384         if (is_Sync(pred)) {
1385                 int i;
1386
1387                 /* handle all Sync predecessors */
1388                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1389                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1390                         if (res)
1391                                 break;
1392                 }
1393         }
1394         return res;
1395 }  /* follow_Mem_chain_for_Store */
1396
1397 /** find entity used as base for an address calculation */
1398 static ir_entity *find_entity(ir_node *ptr)
1399 {
1400         switch(get_irn_opcode(ptr)) {
1401         case iro_SymConst:
1402                 return get_SymConst_entity(ptr);
1403         case iro_Sel: {
1404                 ir_node *pred = get_Sel_ptr(ptr);
1405                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1406                         return get_Sel_entity(ptr);
1407
1408                 return find_entity(pred);
1409         }
1410         case iro_Sub:
1411         case iro_Add: {
1412                 ir_node *left = get_binop_left(ptr);
1413                 ir_node *right;
1414                 if (mode_is_reference(get_irn_mode(left)))
1415                         return find_entity(left);
1416                 right = get_binop_right(ptr);
1417                 if (mode_is_reference(get_irn_mode(right)))
1418                         return find_entity(right);
1419                 return NULL;
1420         }
1421         default:
1422                 return NULL;
1423         }
1424 }
1425
1426 /**
1427  * optimize a Store
1428  *
1429  * @param store  the Store node
1430  */
1431 static unsigned optimize_store(ir_node *store) {
1432         ir_node   *ptr;
1433         ir_node   *mem;
1434         ir_entity *entity;
1435
1436         if (get_Store_volatility(store) == volatility_is_volatile)
1437                 return 0;
1438
1439         ptr    = get_Store_ptr(store);
1440         entity = find_entity(ptr);
1441
1442         /* a store to an entity which is never read is unnecessary */
1443         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1444                 ldst_info_t *info = get_irn_link(store);
1445                 if (info->projs[pn_Store_X_except] == NULL) {
1446                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1447                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1448                         kill_node(store);
1449                         reduce_adr_usage(ptr);
1450                         return DF_CHANGED;
1451                 }
1452         }
1453
1454         /* Check, if the address of this Store is used more than once.
1455          * If not, this Store cannot be removed in any case. */
1456         if (get_irn_n_uses(ptr) <= 1)
1457                 return 0;
1458
1459         mem = get_Store_mem(store);
1460
1461         /* follow the memory chain as long as there are only Loads */
1462         INC_MASTER();
1463
1464         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1465 }  /* optimize_store */
1466
1467 /**
1468  * walker, optimizes Phi after Stores to identical places:
1469  * Does the following optimization:
1470  * @verbatim
1471  *
1472  *   val1   val2   val3          val1  val2  val3
1473  *    |      |      |               \    |    /
1474  *  Store  Store  Store              \   |   /
1475  *      \    |    /                   PhiData
1476  *       \   |   /                       |
1477  *        \  |  /                      Store
1478  *          PhiM
1479  *
1480  * @endverbatim
1481  * This reduces the number of stores and allows for predicated execution.
1482  * Moves Stores back to the end of a function which may be bad.
1483  *
1484  * This is only possible if the predecessor blocks have only one successor.
1485  */
1486 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1487 {
1488         int i, n;
1489         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1490         ir_mode *mode;
1491         ir_node **inM, **inD, **projMs;
1492         int *idx;
1493         dbg_info *db = NULL;
1494         ldst_info_t *info;
1495         block_info_t *bl_info;
1496         unsigned res = 0;
1497
1498         /* Must be a memory Phi */
1499         if (get_irn_mode(phi) != mode_M)
1500                 return 0;
1501
1502         n = get_Phi_n_preds(phi);
1503         if (n <= 0)
1504                 return 0;
1505
1506         /* must be only one user */
1507         projM = get_Phi_pred(phi, 0);
1508         if (get_irn_n_edges(projM) != 1)
1509                 return 0;
1510
1511         store = skip_Proj(projM);
1512         old_store = store;
1513         if (!is_Store(store))
1514                 return 0;
1515
1516         block = get_nodes_block(store);
1517
1518         /* abort on dead blocks */
1519         if (is_Block_dead(block))
1520                 return 0;
1521
1522         /* check if the block is post dominated by Phi-block
1523            and has no exception exit */
1524         bl_info = get_irn_link(block);
1525         if (bl_info->flags & BLOCK_HAS_EXC)
1526                 return 0;
1527
1528         phi_block = get_nodes_block(phi);
1529         if (! block_strictly_postdominates(phi_block, block))
1530                 return 0;
1531
1532         /* this is the address of the store */
1533         ptr  = get_Store_ptr(store);
1534         mode = get_irn_mode(get_Store_value(store));
1535         info = get_irn_link(store);
1536         exc  = info->exc_block;
1537
1538         for (i = 1; i < n; ++i) {
1539                 ir_node *pred = get_Phi_pred(phi, i);
1540
1541                 if (get_irn_n_edges(pred) != 1)
1542                         return 0;
1543
1544                 pred = skip_Proj(pred);
1545                 if (!is_Store(pred))
1546                         return 0;
1547
1548                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1549                         return 0;
1550
1551                 info = get_irn_link(pred);
1552
1553                 /* check, if all stores have the same exception flow */
1554                 if (exc != info->exc_block)
1555                         return 0;
1556
1557                 /* abort on dead blocks */
1558                 block = get_nodes_block(pred);
1559                 if (is_Block_dead(block))
1560                         return 0;
1561
1562                 /* check if the block is post dominated by Phi-block
1563                    and has no exception exit. Note that block must be different from
1564                    Phi-block, else we would move a Store from end End of a block to its
1565                    Start... */
1566                 bl_info = get_irn_link(block);
1567                 if (bl_info->flags & BLOCK_HAS_EXC)
1568                         return 0;
1569                 if (block == phi_block || ! block_postdominates(phi_block, block))
1570                         return 0;
1571         }
1572
1573         /*
1574          * ok, when we are here, we found all predecessors of a Phi that
1575          * are Stores to the same address and size. That means whatever
1576          * we do before we enter the block of the Phi, we do a Store.
1577          * So, we can move the Store to the current block:
1578          *
1579          *   val1    val2    val3          val1  val2  val3
1580          *    |       |       |               \    |    /
1581          * | Str | | Str | | Str |             \   |   /
1582          *      \     |     /                   PhiData
1583          *       \    |    /                       |
1584          *        \   |   /                       Str
1585          *           PhiM
1586          *
1587          * Is only allowed if the predecessor blocks have only one successor.
1588          */
1589
1590         NEW_ARR_A(ir_node *, projMs, n);
1591         NEW_ARR_A(ir_node *, inM, n);
1592         NEW_ARR_A(ir_node *, inD, n);
1593         NEW_ARR_A(int, idx, n);
1594
1595         /* Prepare: Collect all Store nodes.  We must do this
1596            first because we otherwise may loose a store when exchanging its
1597            memory Proj.
1598          */
1599         for (i = n - 1; i >= 0; --i) {
1600                 ir_node *store;
1601
1602                 projMs[i] = get_Phi_pred(phi, i);
1603                 assert(is_Proj(projMs[i]));
1604
1605                 store = get_Proj_pred(projMs[i]);
1606                 info  = get_irn_link(store);
1607
1608                 inM[i] = get_Store_mem(store);
1609                 inD[i] = get_Store_value(store);
1610                 idx[i] = info->exc_idx;
1611         }
1612         block = get_nodes_block(phi);
1613
1614         /* second step: create a new memory Phi */
1615         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1616
1617         /* third step: create a new data Phi */
1618         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1619
1620         /* rewire memory and kill the node */
1621         for (i = n - 1; i >= 0; --i) {
1622                 ir_node *proj  = projMs[i];
1623
1624                 if(is_Proj(proj)) {
1625                         ir_node *store = get_Proj_pred(proj);
1626                         exchange(proj, inM[i]);
1627                         kill_node(store);
1628                 }
1629         }
1630
1631         /* fourth step: create the Store */
1632         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD, 0);
1633 #ifdef DO_CACHEOPT
1634         co_set_irn_name(store, co_get_irn_ident(old_store));
1635 #endif
1636
1637         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1638
1639         info = get_ldst_info(store, &wenv->obst);
1640         info->projs[pn_Store_M] = projM;
1641
1642         /* fifths step: repair exception flow */
1643         if (exc) {
1644                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1645
1646                 info->projs[pn_Store_X_except] = projX;
1647                 info->exc_block                = exc;
1648                 info->exc_idx                  = idx[0];
1649
1650                 for (i = 0; i < n; ++i) {
1651                         set_Block_cfgpred(exc, idx[i], projX);
1652                 }
1653
1654                 if (n > 1) {
1655                         /* the exception block should be optimized as some inputs are identical now */
1656                 }
1657
1658                 res |= CF_CHANGED;
1659         }
1660
1661         /* sixth step: replace old Phi */
1662         exchange(phi, projM);
1663
1664         return res | DF_CHANGED;
1665 }  /* optimize_phi */
1666
1667 /**
1668  * walker, do the optimizations
1669  */
1670 static void do_load_store_optimize(ir_node *n, void *env) {
1671         walk_env_t *wenv = env;
1672
1673         switch (get_irn_opcode(n)) {
1674
1675         case iro_Load:
1676                 wenv->changes |= optimize_load(n);
1677                 break;
1678
1679         case iro_Store:
1680                 wenv->changes |= optimize_store(n);
1681                 break;
1682
1683         case iro_Phi:
1684                 wenv->changes |= optimize_phi(n, wenv);
1685                 break;
1686
1687         default:
1688                 ;
1689         }
1690 }  /* do_load_store_optimize */
1691
1692 /** A scc. */
1693 typedef struct scc {
1694         ir_node *head;          /**< the head of the list */
1695 } scc;
1696
1697 /** A node entry. */
1698 typedef struct node_entry {
1699         unsigned DFSnum;    /**< the DFS number of this node */
1700         unsigned low;       /**< the low number of this node */
1701         int      in_stack;  /**< flag, set if the node is on the stack */
1702         ir_node  *next;     /**< link to the next node the the same scc */
1703         scc      *pscc;     /**< the scc of this node */
1704         unsigned POnum;     /**< the post order number for blocks */
1705 } node_entry;
1706
1707 /** A loop entry. */
1708 typedef struct loop_env {
1709         ir_phase ph;           /**< the phase object */
1710         ir_node  **stack;      /**< the node stack */
1711         int      tos;          /**< tos index */
1712         unsigned nextDFSnum;   /**< the current DFS number */
1713         unsigned POnum;        /**< current post order number */
1714
1715         unsigned changes;      /**< a bitmask of graph changes */
1716 } loop_env;
1717
1718 /**
1719 * Gets the node_entry of a node
1720 */
1721 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1722         ir_phase   *ph = &env->ph;
1723         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1724
1725         if (! e) {
1726                 e = phase_alloc(ph, sizeof(*e));
1727                 memset(e, 0, sizeof(*e));
1728                 phase_set_irn_data(ph, irn, e);
1729         }
1730         return e;
1731 }  /* get_irn_ne */
1732
1733 /**
1734  * Push a node onto the stack.
1735  *
1736  * @param env   the loop environment
1737  * @param n     the node to push
1738  */
1739 static void push(loop_env *env, ir_node *n) {
1740         node_entry *e;
1741
1742         if (env->tos == ARR_LEN(env->stack)) {
1743                 int nlen = ARR_LEN(env->stack) * 2;
1744                 ARR_RESIZE(ir_node *, env->stack, nlen);
1745         }
1746         env->stack[env->tos++] = n;
1747         e = get_irn_ne(n, env);
1748         e->in_stack = 1;
1749 }  /* push */
1750
1751 /**
1752  * pop a node from the stack
1753  *
1754  * @param env   the loop environment
1755  *
1756  * @return  The topmost node
1757  */
1758 static ir_node *pop(loop_env *env) {
1759         ir_node *n = env->stack[--env->tos];
1760         node_entry *e = get_irn_ne(n, env);
1761
1762         e->in_stack = 0;
1763         return n;
1764 }  /* pop */
1765
1766 /**
1767  * Check if irn is a region constant.
1768  * The block or irn must strictly dominate the header block.
1769  *
1770  * @param irn           the node to check
1771  * @param header_block  the header block of the induction variable
1772  */
1773 static int is_rc(ir_node *irn, ir_node *header_block) {
1774         ir_node *block = get_nodes_block(irn);
1775
1776         return (block != header_block) && block_dominates(block, header_block);
1777 }  /* is_rc */
1778
1779 typedef struct phi_entry phi_entry;
1780 struct phi_entry {
1781         ir_node   *phi;    /**< A phi with a region const memory. */
1782         int       pos;     /**< The position of the region const memory */
1783         ir_node   *load;   /**< the newly created load for this phi */
1784         phi_entry *next;
1785 };
1786
1787 /**
1788  * An entry in the avail set.
1789  */
1790 typedef struct avail_entry_t {
1791         ir_node *ptr;   /**< the address pointer */
1792         ir_mode *mode;  /**< the load mode */
1793         ir_node *load;  /**< the associated Load */
1794 } avail_entry_t;
1795
1796 /**
1797  * Compare two avail entries.
1798  */
1799 static int cmp_avail_entry(const void *elt, const void *key, size_t size) {
1800         (void) size;
1801
1802         const avail_entry_t *a = elt;
1803         const avail_entry_t *b = key;
1804
1805         return a->ptr != b->ptr || a->mode != b->mode;
1806 }  /* cmp_avail_entry */
1807
1808 /**
1809  * Calculate the hash value of an avail entry.
1810  */
1811 static unsigned hash_cache_entry(const avail_entry_t *entry) {
1812         return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1813 }  /* hash_cache_entry */
1814
1815 /**
1816  * Move loops out of loops if possible.
1817  *
1818  * @param pscc   the loop described by an SCC
1819  * @param env    the loop environment
1820  */
1821 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1822         ir_node   *phi, *load, *next, *other, *next_other;
1823         ir_entity *ent;
1824         int       j;
1825         phi_entry *phi_list = NULL;
1826         set       *avail;
1827
1828         avail = new_set(cmp_avail_entry, 8);
1829
1830         /* collect all outer memories */
1831         for (phi = pscc->head; phi != NULL; phi = next) {
1832                 node_entry *ne = get_irn_ne(phi, env);
1833                 next = ne->next;
1834
1835                 /* check all memory Phi's */
1836                 if (! is_Phi(phi))
1837                         continue;
1838
1839                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1840
1841                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1842                         ir_node    *pred = get_irn_n(phi, j);
1843                         node_entry *pe   = get_irn_ne(pred, env);
1844
1845                         if (pe->pscc != ne->pscc) {
1846                                 /* not in the same SCC, is region const */
1847                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1848
1849                                 pe->phi  = phi;
1850                                 pe->pos  = j;
1851                                 pe->next = phi_list;
1852                                 phi_list = pe;
1853                         }
1854                 }
1855         }
1856         /* no Phis no fun */
1857         assert(phi_list != NULL && "DFS found a loop without Phi");
1858
1859         /* for now, we cannot handle more than one input (only reducible cf) */
1860         if (phi_list->next != NULL)
1861                 return;
1862
1863         for (load = pscc->head; load; load = next) {
1864                 ir_mode *load_mode;
1865                 node_entry *ne = get_irn_ne(load, env);
1866                 next = ne->next;
1867
1868                 if (is_Load(load)) {
1869                         ldst_info_t *info = get_irn_link(load);
1870                         ir_node     *ptr = get_Load_ptr(load);
1871
1872                         /* for now, we cannot handle Loads with exceptions */
1873                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1874                                 continue;
1875
1876                         /* for now, we can only move Load(Global) */
1877                         if (! is_Global(ptr))
1878                                 continue;
1879                         ent       = get_Global_entity(ptr);
1880                         load_mode = get_Load_mode(load);
1881                         for (other = pscc->head; other != NULL; other = next_other) {
1882                                 node_entry *ne = get_irn_ne(other, env);
1883                                 next_other = ne->next;
1884
1885                                 if (is_Store(other)) {
1886                                         ir_alias_relation rel = get_alias_relation(
1887                                                 current_ir_graph,
1888                                                 get_Store_ptr(other),
1889                                                 get_irn_mode(get_Store_value(other)),
1890                                                 ptr, load_mode);
1891                                         /* if the might be an alias, we cannot pass this Store */
1892                                         if (rel != ir_no_alias)
1893                                                 break;
1894                                 }
1895                                 /* only Phis and pure Calls are allowed here, so ignore them */
1896                         }
1897                         if (other == NULL) {
1898                                 ldst_info_t *ninfo;
1899                                 phi_entry   *pe;
1900                                 dbg_info    *db;
1901
1902                                 /* yep, no aliasing Store found, Load can be moved */
1903                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1904
1905                                 db   = get_irn_dbg_info(load);
1906                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1907                                         int     pos   = pe->pos;
1908                                         ir_node *phi  = pe->phi;
1909                                         ir_node *blk  = get_nodes_block(phi);
1910                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1911                                         ir_node *irn, *mem;
1912                                         avail_entry_t entry, *res;
1913
1914                                         entry.ptr  = ptr;
1915                                         entry.mode = load_mode;
1916                                         res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1917                                         if (res != NULL) {
1918                                                 irn = res->load;
1919                                         } else {
1920                                                 irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1921                                                 entry.load = irn;
1922                                                 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1923                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1924                                         }
1925                                         pe->load = irn;
1926                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1927
1928                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1929                                         set_Phi_pred(phi, pos, mem);
1930
1931                                         ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1932                                 }
1933
1934                                 /* now kill the old Load */
1935                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1936                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1937
1938                                 env->changes |= DF_CHANGED;
1939                         }
1940                 }
1941         }
1942         del_set(avail);
1943 }  /* move_loads_out_of_loops */
1944
1945 /**
1946  * Process a loop SCC.
1947  *
1948  * @param pscc  the SCC
1949  * @param env   the loop environment
1950  */
1951 static void process_loop(scc *pscc, loop_env *env) {
1952         ir_node *irn, *next, *header = NULL;
1953         node_entry *b, *h = NULL;
1954         int j, only_phi, num_outside, process = 0;
1955         ir_node *out_rc;
1956
1957         /* find the header block for this scc */
1958         for (irn = pscc->head; irn; irn = next) {
1959                 node_entry *e = get_irn_ne(irn, env);
1960                 ir_node *block = get_nodes_block(irn);
1961
1962                 next = e->next;
1963                 b = get_irn_ne(block, env);
1964
1965                 if (header != NULL) {
1966                         if (h->POnum < b->POnum) {
1967                                 header = block;
1968                                 h      = b;
1969                         }
1970                 } else {
1971                         header = block;
1972                         h      = b;
1973                 }
1974         }
1975
1976         /* check if this scc contains only Phi, Loads or Stores nodes */
1977         only_phi    = 1;
1978         num_outside = 0;
1979         out_rc      = NULL;
1980         for (irn = pscc->head; irn; irn = next) {
1981                 node_entry *e = get_irn_ne(irn, env);
1982
1983                 next = e->next;
1984                 switch (get_irn_opcode(irn)) {
1985                 case iro_Call:
1986                         if (is_Call_pure(irn)) {
1987                                 /* pure calls can be treated like loads */
1988                                 only_phi = 0;
1989                                 break;
1990                         }
1991                         /* non-pure calls must be handle like may-alias Stores */
1992                         goto fail;
1993                 case iro_CopyB:
1994                         /* cannot handle CopyB yet */
1995                         goto fail;
1996                 case iro_Load:
1997                         process = 1;
1998                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1999                                 /* cannot handle loops with volatile Loads */
2000                                 goto fail;
2001                         }
2002                         only_phi = 0;
2003                         break;
2004                 case iro_Store:
2005                         if (get_Store_volatility(irn) == volatility_is_volatile) {
2006                                 /* cannot handle loops with volatile Stores */
2007                                 goto fail;
2008                         }
2009                         only_phi = 0;
2010                         break;
2011                 default:
2012                         only_phi = 0;
2013                         break;
2014                 case iro_Phi:
2015                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2016                                 ir_node *pred  = get_irn_n(irn, j);
2017                                 node_entry *pe = get_irn_ne(pred, env);
2018
2019                                 if (pe->pscc != e->pscc) {
2020                                         /* not in the same SCC, must be a region const */
2021                                         if (! is_rc(pred, header)) {
2022                                                 /* not a memory loop */
2023                                                 goto fail;
2024                                         }
2025                                         if (out_rc == NULL) {
2026                                                 /* first region constant */
2027                                                 out_rc = pred;
2028                                                 ++num_outside;
2029                                         } else if (out_rc != pred) {
2030                                                 /* another region constant */
2031                                                 ++num_outside;
2032                                         }
2033                                 }
2034                         }
2035                         break;
2036                 }
2037         }
2038         if (! process)
2039                 goto fail;
2040
2041         /* found a memory loop */
2042         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
2043         if (only_phi && num_outside == 1) {
2044                 /* a phi cycle with only one real predecessor can be collapsed */
2045                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
2046
2047                 for (irn = pscc->head; irn; irn = next) {
2048                         node_entry *e = get_irn_ne(irn, env);
2049                         next = e->next;
2050                         exchange(irn, out_rc);
2051                 }
2052                 env->changes |= DF_CHANGED;
2053                 return;
2054         }
2055
2056 #ifdef DEBUG_libfirm
2057         for (irn = pscc->head; irn; irn = next) {
2058                 node_entry *e = get_irn_ne(irn, env);
2059                 next = e->next;
2060                 DB((dbg, LEVEL_2, " %+F,", irn));
2061         }
2062         DB((dbg, LEVEL_2, "\n"));
2063 #endif
2064         move_loads_out_of_loops(pscc, env);
2065
2066 fail:
2067         ;
2068 }  /* process_loop */
2069
2070 /**
2071  * Process a SCC.
2072  *
2073  * @param pscc  the SCC
2074  * @param env   the loop environment
2075  */
2076 static void process_scc(scc *pscc, loop_env *env) {
2077         ir_node *head = pscc->head;
2078         node_entry *e = get_irn_ne(head, env);
2079
2080 #ifdef DEBUG_libfirm
2081         {
2082                 ir_node *irn, *next;
2083
2084                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2085                 for (irn = pscc->head; irn; irn = next) {
2086                         node_entry *e = get_irn_ne(irn, env);
2087
2088                         next = e->next;
2089
2090                         DB((dbg, LEVEL_4, " %+F,", irn));
2091                 }
2092                 DB((dbg, LEVEL_4, "\n"));
2093         }
2094 #endif
2095
2096         if (e->next != NULL) {
2097                 /* this SCC has more than one member */
2098                 process_loop(pscc, env);
2099         }
2100 }  /* process_scc */
2101
2102 /**
2103  * Do Tarjan's SCC algorithm and drive load/store optimization.
2104  *
2105  * @param irn  start at this node
2106  * @param env  the loop environment
2107  */
2108 static void dfs(ir_node *irn, loop_env *env)
2109 {
2110         int i, n;
2111         node_entry *node = get_irn_ne(irn, env);
2112
2113         mark_irn_visited(irn);
2114
2115         node->DFSnum = env->nextDFSnum++;
2116         node->low    = node->DFSnum;
2117         push(env, irn);
2118
2119         /* handle preds */
2120         if (is_Phi(irn) || is_Sync(irn)) {
2121                 n = get_irn_arity(irn);
2122                 for (i = 0; i < n; ++i) {
2123                         ir_node *pred = get_irn_n(irn, i);
2124                         node_entry *o = get_irn_ne(pred, env);
2125
2126                         if (!irn_visited(pred)) {
2127                                 dfs(pred, env);
2128                                 node->low = MIN(node->low, o->low);
2129                         }
2130                         if (o->DFSnum < node->DFSnum && o->in_stack)
2131                                 node->low = MIN(o->DFSnum, node->low);
2132                 }
2133         } else if (is_fragile_op(irn)) {
2134                 ir_node *pred = get_fragile_op_mem(irn);
2135                 node_entry *o = get_irn_ne(pred, env);
2136
2137                 if (!irn_visited(pred)) {
2138                         dfs(pred, env);
2139                         node->low = MIN(node->low, o->low);
2140                 }
2141                 if (o->DFSnum < node->DFSnum && o->in_stack)
2142                         node->low = MIN(o->DFSnum, node->low);
2143         } else if (is_Proj(irn)) {
2144                 ir_node *pred = get_Proj_pred(irn);
2145                 node_entry *o = get_irn_ne(pred, env);
2146
2147                 if (!irn_visited(pred)) {
2148                         dfs(pred, env);
2149                         node->low = MIN(node->low, o->low);
2150                 }
2151                 if (o->DFSnum < node->DFSnum && o->in_stack)
2152                         node->low = MIN(o->DFSnum, node->low);
2153         }
2154         else {
2155                  /* IGNORE predecessors */
2156         }
2157
2158         if (node->low == node->DFSnum) {
2159                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2160                 ir_node *x;
2161
2162                 pscc->head = NULL;
2163                 do {
2164                         node_entry *e;
2165
2166                         x = pop(env);
2167                         e = get_irn_ne(x, env);
2168                         e->pscc    = pscc;
2169                         e->next    = pscc->head;
2170                         pscc->head = x;
2171                 } while (x != irn);
2172
2173                 process_scc(pscc, env);
2174         }
2175 }  /* dfs */
2176
2177 /**
2178  * Do the DFS on the memory edges a graph.
2179  *
2180  * @param irg  the graph to process
2181  * @param env  the loop environment
2182  */
2183 static void do_dfs(ir_graph *irg, loop_env *env) {
2184         ir_graph *rem = current_ir_graph;
2185         ir_node  *endblk, *end;
2186         int      i;
2187
2188         current_ir_graph = irg;
2189         inc_irg_visited(irg);
2190
2191         /* visit all memory nodes */
2192         endblk = get_irg_end_block(irg);
2193         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2194                 ir_node *pred = get_Block_cfgpred(endblk, i);
2195
2196                 pred = skip_Proj(pred);
2197                 if (is_Return(pred))
2198                         dfs(get_Return_mem(pred), env);
2199                 else if (is_Raise(pred))
2200                         dfs(get_Raise_mem(pred), env);
2201                 else if (is_fragile_op(pred))
2202                         dfs(get_fragile_op_mem(pred), env);
2203                 else {
2204                         assert(0 && "Unknown EndBlock predecessor");
2205                 }
2206         }
2207
2208         /* visit the keep-alives */
2209         end = get_irg_end(irg);
2210         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2211                 ir_node *ka = get_End_keepalive(end, i);
2212
2213                 if (is_Phi(ka) && !irn_visited(ka))
2214                         dfs(ka, env);
2215         }
2216         current_ir_graph = rem;
2217 }  /* do_dfs */
2218
2219 /**
2220  * Initialize new phase data. We do this always explicit, so return NULL here
2221  */
2222 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2223         (void)ph;
2224         (void)irn;
2225         (void)data;
2226         return NULL;
2227 }  /* init_loop_data */
2228
2229 /**
2230  * Optimize Loads/Stores in loops.
2231  *
2232  * @param irg  the graph
2233  */
2234 static int optimize_loops(ir_graph *irg) {
2235         loop_env env;
2236
2237         env.stack         = NEW_ARR_F(ir_node *, 128);
2238         env.tos           = 0;
2239         env.nextDFSnum    = 0;
2240         env.POnum         = 0;
2241         env.changes       = 0;
2242         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2243
2244         /* calculate the SCC's and drive loop optimization. */
2245         do_dfs(irg, &env);
2246
2247         DEL_ARR_F(env.stack);
2248         phase_free(&env.ph);
2249
2250         return env.changes;
2251 }  /* optimize_loops */
2252
2253 /*
2254  * do the load store optimization
2255  */
2256 int optimize_load_store(ir_graph *irg) {
2257         walk_env_t env;
2258
2259         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2260
2261         assert(get_irg_phase_state(irg) != phase_building);
2262         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2263                 "LoadStore optimization needs pinned graph");
2264
2265         /* we need landing pads */
2266         remove_critical_cf_edges(irg);
2267
2268         edges_assure(irg);
2269
2270         /* for Phi optimization post-dominators are needed ... */
2271         assure_postdoms(irg);
2272
2273         if (get_opt_alias_analysis()) {
2274                 assure_irg_entity_usage_computed(irg);
2275                 assure_irp_globals_entity_usage_computed();
2276         }
2277
2278         obstack_init(&env.obst);
2279         env.changes = 0;
2280
2281         /* init the links, then collect Loads/Stores/Proj's in lists */
2282         master_visited = 0;
2283         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2284
2285         /* now we have collected enough information, optimize */
2286         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2287
2288         env.changes |= optimize_loops(irg);
2289
2290         obstack_free(&env.obst, NULL);
2291
2292         /* Handle graph state */
2293         if (env.changes) {
2294                 set_irg_outs_inconsistent(irg);
2295                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2296         }
2297
2298         if (env.changes & CF_CHANGED) {
2299                 /* is this really needed: Yes, control flow changed, block might
2300                 have Bad() predecessors. */
2301                 set_irg_doms_inconsistent(irg);
2302         }
2303         return env.changes != 0;
2304 }  /* optimize_load_store */