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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  * @version $Id$
  25  */
  26 #include "config.h"
  27
  28 #include <string.h>
  29
  30 #include "iroptimize.h"
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irmode_t.h"
  34 #include "iropt_t.h"
  35 #include "ircons_t.h"
  36 #include "irgmod.h"
  37 #include "irgwalk.h"
  38 #include "irvrfy.h"
  39 #include "tv_t.h"
  40 #include "dbginfo_t.h"
  41 #include "iropt_dbg.h"
  42 #include "irflag_t.h"
  43 #include "array_t.h"
  44 #include "irhooks.h"
  45 #include "iredges.h"
  46 #include "irpass.h"
  47 #include "opt_polymorphy.h"
  48 #include "irmemory.h"
  49 #include "irphase_t.h"
  50 #include "irgopt.h"
  51 #include "set.h"
  52 #include "debug.h"
  53
  54 /** The debug handle. */
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  56
  57 #ifdef DO_CACHEOPT
  58 #include "cacheopt/cachesim.h"
  59 #endif
  60
  61 #undef IMAX
  62 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  63
  64 #define MAX_PROJ        IMAX(IMAX(pn_Load_max, pn_Store_max), pn_Call_max)
  65
  66 enum changes_t {
  67         DF_CHANGED = 1,       /**< data flow changed */
  68         CF_CHANGED = 2,       /**< control flow changed */
  69 };
  70
  71 /**
  72  * walker environment
  73  */
  74 typedef struct _walk_env_t {
  75         struct obstack obst;          /**< list of all stores */
  76         unsigned changes;             /**< a bitmask of graph changes */
  77 } walk_env_t;
  78
  79 /** A Load/Store info. */
  80 typedef struct _ldst_info_t {
  81         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  82         ir_node  *exc_block;          /**< the exception block if available */
  83         int      exc_idx;             /**< predecessor index in the exception block */
  84         unsigned visited;             /**< visited counter for breaking loops */
  85 } ldst_info_t;
  86
  87 /**
  88  * flags for control flow.
  89  */
  90 enum block_flags_t {
  91         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  92         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  93 };
  94
  95 /**
  96  * a Block info.
  97  */
  98 typedef struct _block_info_t {
  99         unsigned flags;               /**< flags for the block */
 100 } block_info_t;
 101
 102 /** the master visited flag for loop detection. */
 103 static unsigned master_visited = 0;
 104
 105 #define INC_MASTER()       ++master_visited
 106 #define MARK_NODE(info)    (info)->visited = master_visited
 107 #define NODE_VISITED(info) (info)->visited >= master_visited
 108
 109 /**
 110  * get the Load/Store info of a node
 111  */
 112 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 113         ldst_info_t *info = get_irn_link(node);
 114
 115         if (! info) {
 116                 info = OALLOCZ(obst, ldst_info_t);
 117                 set_irn_link(node, info);
 118         }
 119         return info;
 120 }  /* get_ldst_info */
 121
 122 /**
 123  * get the Block info of a node
 124  */
 125 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
 126         block_info_t *info = get_irn_link(node);
 127
 128         if (! info) {
 129                 info = OALLOCZ(obst, block_info_t);
 130                 set_irn_link(node, info);
 131         }
 132         return info;
 133 }  /* get_block_info */
 134
 135 /**
 136  * update the projection info for a Load/Store
 137  */
 138 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 139 {
 140         long nr = get_Proj_proj(proj);
 141
 142         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 143
 144         if (info->projs[nr]) {
 145                 /* there is already one, do CSE */
 146                 exchange(proj, info->projs[nr]);
 147                 return DF_CHANGED;
 148         }
 149         else {
 150                 info->projs[nr] = proj;
 151                 return 0;
 152         }
 153 }  /* update_projs */
 154
 155 /**
 156  * update the exception block info for a Load/Store node.
 157  *
 158  * @param info   the load/store info struct
 159  * @param block  the exception handler block for this load/store
 160  * @param pos    the control flow input of the block
 161  */
 162 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 163 {
 164         assert(info->exc_block == NULL && "more than one exception block found");
 165
 166         info->exc_block = block;
 167         info->exc_idx   = pos;
 168         return 0;
 169 }  /* update_exc */
 170
 171 /** Return the number of uses of an address node */
 172 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 173
 174 /**
 175  * walker, collects all Load/Store/Proj nodes
 176  *
 177  * walks from Start -> End
 178  */
 179 static void collect_nodes(ir_node *node, void *env)
 180 {
 181         ir_opcode   opcode = get_irn_opcode(node);
 182         ir_node     *pred, *blk, *pred_blk;
 183         ldst_info_t *ldst_info;
 184         walk_env_t  *wenv = env;
 185
 186         if (opcode == iro_Proj) {
 187                 pred   = get_Proj_pred(node);
 188                 opcode = get_irn_opcode(pred);
 189
 190                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 191                         ldst_info = get_ldst_info(pred, &wenv->obst);
 192
 193                         wenv->changes |= update_projs(ldst_info, node);
 194
 195                         /*
 196                          * Place the Proj's to the same block as the
 197                          * predecessor Load. This is always ok and prevents
 198                          * "non-SSA" form after optimizations if the Proj
 199                          * is in a wrong block.
 200                          */
 201                         blk      = get_nodes_block(node);
 202                         pred_blk = get_nodes_block(pred);
 203                         if (blk != pred_blk) {
 204                                 wenv->changes |= DF_CHANGED;
 205                                 set_nodes_block(node, pred_blk);
 206                         }
 207                 }
 208         } else if (opcode == iro_Block) {
 209                 int i;
 210
 211                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 212                         ir_node      *pred_block, *proj;
 213                         block_info_t *bl_info;
 214                         int          is_exc = 0;
 215
 216                         pred = proj = get_Block_cfgpred(node, i);
 217
 218                         if (is_Proj(proj)) {
 219                                 pred   = get_Proj_pred(proj);
 220                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 221                         }
 222
 223                         /* ignore Bad predecessors, they will be removed later */
 224                         if (is_Bad(pred))
 225                                 continue;
 226
 227                         pred_block = get_nodes_block(pred);
 228                         bl_info    = get_block_info(pred_block, &wenv->obst);
 229
 230                         if (is_fragile_op(pred) && is_exc)
 231                                 bl_info->flags |= BLOCK_HAS_EXC;
 232                         else if (is_irn_forking(pred))
 233                                 bl_info->flags |= BLOCK_HAS_COND;
 234
 235                         opcode = get_irn_opcode(pred);
 236                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 237                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 238
 239                                 wenv->changes |= update_exc(ldst_info, node, i);
 240                         }
 241                 }
 242         }
 243 }  /* collect_nodes */
 244
 245 /**
 246  * Returns an entity if the address ptr points to a constant one.
 247  *
 248  * @param ptr  the address
 249  *
 250  * @return an entity or NULL
 251  */
 252 static ir_entity *find_constant_entity(ir_node *ptr)
 253 {
 254         for (;;) {
 255                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 256                         return get_SymConst_entity(ptr);
 257                 } else if (is_Sel(ptr)) {
 258                         ir_entity *ent = get_Sel_entity(ptr);
 259                         ir_type   *tp  = get_entity_owner(ent);
 260
 261                         /* Do not fiddle with polymorphism. */
 262                         if (is_Class_type(get_entity_owner(ent)) &&
 263                                 ((get_entity_n_overwrites(ent)    != 0) ||
 264                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 265                                 return NULL;
 266
 267                         if (is_Array_type(tp)) {
 268                                 /* check bounds */
 269                                 int i, n;
 270
 271                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 272                                         ir_node *bound;
 273                                         tarval *tlower, *tupper;
 274                                         ir_node *index = get_Sel_index(ptr, i);
 275                                         tarval *tv     = computed_value(index);
 276
 277                                         /* check if the index is constant */
 278                                         if (tv == tarval_bad)
 279                                                 return NULL;
 280
 281                                         bound  = get_array_lower_bound(tp, i);
 282                                         tlower = computed_value(bound);
 283                                         bound  = get_array_upper_bound(tp, i);
 284                                         tupper = computed_value(bound);
 285
 286                                         if (tlower == tarval_bad || tupper == tarval_bad)
 287                                                 return NULL;
 288
 289                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 290                                                 return NULL;
 291                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 292                                                 return NULL;
 293
 294                                         /* ok, bounds check finished */
 295                                 }
 296                         }
 297
 298                         if (variability_constant == get_entity_variability(ent))
 299                                 return ent;
 300
 301                         /* try next */
 302                         ptr = get_Sel_ptr(ptr);
 303                 } else if (is_Add(ptr)) {
 304                         ir_node *l = get_Add_left(ptr);
 305                         ir_node *r = get_Add_right(ptr);
 306
 307                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 308                                 ptr = l;
 309                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 310                                 ptr = r;
 311                         else
 312                                 return NULL;
 313
 314                         /* for now, we support only one addition, reassoc should fold all others */
 315                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 316                                 return NULL;
 317                 } else if (is_Sub(ptr)) {
 318                         ir_node *l = get_Sub_left(ptr);
 319                         ir_node *r = get_Sub_right(ptr);
 320
 321                         if (get_irn_mode(l) == get_irn_mode(ptr) &&     is_Const(r))
 322                                 ptr = l;
 323                         else
 324                                 return NULL;
 325                         /* for now, we support only one substraction, reassoc should fold all others */
 326                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 327                                 return NULL;
 328                 } else
 329                         return NULL;
 330         }
 331 }  /* find_constant_entity */
 332
 333 /**
 334  * Return the Selection index of a Sel node from dimension n
 335  */
 336 static long get_Sel_array_index_long(ir_node *n, int dim) {
 337         ir_node *index = get_Sel_index(n, dim);
 338         assert(is_Const(index));
 339         return get_tarval_long(get_Const_tarval(index));
 340 }  /* get_Sel_array_index_long */
 341
 342 /**
 343  * Returns the accessed component graph path for an
 344  * node computing an address.
 345  *
 346  * @param ptr    the node computing the address
 347  * @param depth  current depth in steps upward from the root
 348  *               of the address
 349  */
 350 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 351         compound_graph_path *res = NULL;
 352         ir_entity           *root, *field, *ent;
 353         int                 path_len, pos, idx;
 354         tarval              *tv;
 355         ir_type             *tp;
 356
 357         if (is_SymConst(ptr)) {
 358                 /* a SymConst. If the depth is 0, this is an access to a global
 359                  * entity and we don't need a component path, else we know
 360                  * at least its length.
 361                  */
 362                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 363                 root = get_SymConst_entity(ptr);
 364                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 365         } else if (is_Sel(ptr)) {
 366                 /* it's a Sel, go up until we find the root */
 367                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 368                 if (res == NULL)
 369                         return NULL;
 370
 371                 /* fill up the step in the path at the current position */
 372                 field    = get_Sel_entity(ptr);
 373                 path_len = get_compound_graph_path_length(res);
 374                 pos      = path_len - depth - 1;
 375                 set_compound_graph_path_node(res, pos, field);
 376
 377                 if (is_Array_type(get_entity_owner(field))) {
 378                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 379                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 380                 }
 381         } else if (is_Add(ptr)) {
 382                 ir_node *l    = get_Add_left(ptr);
 383                 ir_node *r    = get_Add_right(ptr);
 384                 ir_mode *mode = get_irn_mode(ptr);
 385                 tarval  *tmp;
 386
 387                 if (is_Const(r) && get_irn_mode(l) == mode) {
 388                         ptr = l;
 389                         tv  = get_Const_tarval(r);
 390                 } else {
 391                         ptr = r;
 392                         tv  = get_Const_tarval(l);
 393                 }
 394 ptr_arith:
 395                 mode = get_tarval_mode(tv);
 396                 tmp  = tv;
 397
 398                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 399                 if (is_Sel(ptr)) {
 400                         field = get_Sel_entity(ptr);
 401                 } else {
 402                         field = get_SymConst_entity(ptr);
 403                 }
 404                 idx = 0;
 405                 for (ent = field;;) {
 406                         unsigned size;
 407                         tarval   *sz, *tv_index, *tlower, *tupper;
 408                         ir_node  *bound;
 409
 410                         tp = get_entity_type(ent);
 411                         if (! is_Array_type(tp))
 412                                 break;
 413                         ent = get_array_element_entity(tp);
 414                         size = get_type_size_bytes(get_entity_type(ent));
 415                         sz   = new_tarval_from_long(size, mode);
 416
 417                         tv_index = tarval_div(tmp, sz);
 418                         tmp      = tarval_mod(tmp, sz);
 419
 420                         if (tv_index == tarval_bad || tmp == tarval_bad)
 421                                 return NULL;
 422
 423                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 424                         bound  = get_array_lower_bound(tp, 0);
 425                         tlower = computed_value(bound);
 426                         bound  = get_array_upper_bound(tp, 0);
 427                         tupper = computed_value(bound);
 428
 429                         if (tlower == tarval_bad || tupper == tarval_bad)
 430                                 return NULL;
 431
 432                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 433                                 return NULL;
 434                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 435                                 return NULL;
 436
 437                         /* ok, bounds check finished */
 438                         ++idx;
 439                 }
 440                 if (! tarval_is_null(tmp)) {
 441                         /* access to some struct/union member */
 442                         return NULL;
 443                 }
 444
 445                 /* should be at least ONE array */
 446                 if (idx == 0)
 447                         return NULL;
 448
 449                 res = rec_get_accessed_path(ptr, depth + idx);
 450                 if (res == NULL)
 451                         return NULL;
 452
 453                 path_len = get_compound_graph_path_length(res);
 454                 pos      = path_len - depth - idx;
 455
 456                 for (ent = field;;) {
 457                         unsigned size;
 458                         tarval   *sz, *tv_index;
 459                         long     index;
 460
 461                         tp = get_entity_type(ent);
 462                         if (! is_Array_type(tp))
 463                                 break;
 464                         ent = get_array_element_entity(tp);
 465                         set_compound_graph_path_node(res, pos, ent);
 466
 467                         size = get_type_size_bytes(get_entity_type(ent));
 468                         sz   = new_tarval_from_long(size, mode);
 469
 470                         tv_index = tarval_div(tv, sz);
 471                         tv       = tarval_mod(tv, sz);
 472
 473                         /* worked above, should work again */
 474                         assert(tv_index != tarval_bad && tv != tarval_bad);
 475
 476                         /* bounds already checked above */
 477                         index = get_tarval_long(tv_index);
 478                         set_compound_graph_path_array_index(res, pos, index);
 479                         ++pos;
 480                 }
 481         } else if (is_Sub(ptr)) {
 482                 ir_node *l = get_Sub_left(ptr);
 483                 ir_node *r = get_Sub_right(ptr);
 484
 485                 ptr = l;
 486                 tv  = get_Const_tarval(r);
 487                 tv  = tarval_neg(tv);
 488                 goto ptr_arith;
 489         }
 490         return res;
 491 }  /* rec_get_accessed_path */
 492
 493 /**
 494  * Returns an access path or NULL.  The access path is only
 495  * valid, if the graph is in phase_high and _no_ address computation is used.
 496  */
 497 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 498         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 499         return gr;
 500 }  /* get_accessed_path */
 501
 502 typedef struct path_entry {
 503         ir_entity         *ent;
 504         struct path_entry *next;
 505         long              index;
 506 } path_entry;
 507
 508 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
 509         path_entry       entry, *p;
 510         ir_entity        *ent, *field;
 511         ir_initializer_t *initializer;
 512         tarval           *tv;
 513         ir_type          *tp;
 514         unsigned         n;
 515
 516         entry.next = next;
 517         if (is_SymConst(ptr)) {
 518                 /* found the root */
 519                 ent         = get_SymConst_entity(ptr);
 520                 initializer = get_entity_initializer(ent);
 521                 for (p = next; p != NULL;) {
 522                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 523                                 return NULL;
 524                         n  = get_initializer_compound_n_entries(initializer);
 525                         tp = get_entity_type(ent);
 526
 527                         if (is_Array_type(tp)) {
 528                                 ent = get_array_element_entity(tp);
 529                                 if (ent != p->ent) {
 530                                         /* a missing [0] */
 531                                         if (0 >= n)
 532                                                 return NULL;
 533                                         initializer = get_initializer_compound_value(initializer, 0);
 534                                         continue;
 535                                 }
 536                         }
 537                         if (p->index >= (int) n)
 538                                 return NULL;
 539                         initializer = get_initializer_compound_value(initializer, p->index);
 540
 541                         ent = p->ent;
 542                         p   = p->next;
 543                 }
 544                 tp = get_entity_type(ent);
 545                 while (is_Array_type(tp)) {
 546                         ent = get_array_element_entity(tp);
 547                         tp = get_entity_type(ent);
 548                         /* a missing [0] */
 549                         n  = get_initializer_compound_n_entries(initializer);
 550                         if (0 >= n)
 551                                 return NULL;
 552                         initializer = get_initializer_compound_value(initializer, 0);
 553                 }
 554
 555                 switch (initializer->kind) {
 556                 case IR_INITIALIZER_CONST:
 557                         return get_initializer_const_value(initializer);
 558                 case IR_INITIALIZER_TARVAL:
 559                 case IR_INITIALIZER_NULL:
 560                 default:
 561                         return NULL;
 562                 }
 563         } else if (is_Sel(ptr)) {
 564                 entry.ent = field = get_Sel_entity(ptr);
 565                 tp = get_entity_owner(field);
 566                 if (is_Array_type(tp)) {
 567                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 568                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 569                 } else {
 570                         int i, n_members = get_compound_n_members(tp);
 571                         for (i = 0; i < n_members; ++i) {
 572                                 if (get_compound_member(tp, i) == field)
 573                                         break;
 574                         }
 575                         if (i >= n_members) {
 576                                 /* not found: should NOT happen */
 577                                 return NULL;
 578                         }
 579                         entry.index = i;
 580                 }
 581                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 582         }  else if (is_Add(ptr)) {
 583                 ir_node  *l = get_Add_left(ptr);
 584                 ir_node  *r = get_Add_right(ptr);
 585                 ir_mode  *mode;
 586                 unsigned pos;
 587
 588                 if (is_Const(r)) {
 589                         ptr = l;
 590                         tv  = get_Const_tarval(r);
 591                 } else {
 592                         ptr = r;
 593                         tv  = get_Const_tarval(l);
 594                 }
 595 ptr_arith:
 596                 mode = get_tarval_mode(tv);
 597
 598                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 599                 if (is_Sel(ptr)) {
 600                         field = get_Sel_entity(ptr);
 601                 } else {
 602                         field = get_SymConst_entity(ptr);
 603                 }
 604
 605                 /* count needed entries */
 606                 pos = 0;
 607                 for (ent = field;;) {
 608                         tp = get_entity_type(ent);
 609                         if (! is_Array_type(tp))
 610                                 break;
 611                         ent = get_array_element_entity(tp);
 612                         ++pos;
 613                 }
 614                 /* should be at least ONE entry */
 615                 if (pos == 0)
 616                         return NULL;
 617
 618                 /* allocate the right number of entries */
 619                 NEW_ARR_A(path_entry, p, pos);
 620
 621                 /* fill them up */
 622                 pos = 0;
 623                 for (ent = field;;) {
 624                         unsigned size;
 625                         tarval   *sz, *tv_index, *tlower, *tupper;
 626                         long     index;
 627                         ir_node  *bound;
 628
 629                         tp = get_entity_type(ent);
 630                         if (! is_Array_type(tp))
 631                                 break;
 632                         ent = get_array_element_entity(tp);
 633                         p[pos].ent  = ent;
 634                         p[pos].next = &p[pos + 1];
 635
 636                         size = get_type_size_bytes(get_entity_type(ent));
 637                         sz   = new_tarval_from_long(size, mode);
 638
 639                         tv_index = tarval_div(tv, sz);
 640                         tv       = tarval_mod(tv, sz);
 641
 642                         if (tv_index == tarval_bad || tv == tarval_bad)
 643                                 return NULL;
 644
 645                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 646                         bound  = get_array_lower_bound(tp, 0);
 647                         tlower = computed_value(bound);
 648                         bound  = get_array_upper_bound(tp, 0);
 649                         tupper = computed_value(bound);
 650
 651                         if (tlower == tarval_bad || tupper == tarval_bad)
 652                                 return NULL;
 653
 654                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 655                                 return NULL;
 656                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 657                                 return NULL;
 658
 659                         /* ok, bounds check finished */
 660                         index = get_tarval_long(tv_index);
 661                         p[pos].index = index;
 662                         ++pos;
 663                 }
 664                 if (! tarval_is_null(tv)) {
 665                         /* hmm, wrong access */
 666                         return NULL;
 667                 }
 668                 p[pos - 1].next = next;
 669                 return rec_find_compound_ent_value(ptr, p);
 670         } else if (is_Sub(ptr)) {
 671                 ir_node *l = get_Sub_left(ptr);
 672                 ir_node *r = get_Sub_right(ptr);
 673
 674                 ptr = l;
 675                 tv  = get_Const_tarval(r);
 676                 tv  = tarval_neg(tv);
 677                 goto ptr_arith;
 678         }
 679         return NULL;
 680 }
 681
 682 static ir_node *find_compound_ent_value(ir_node *ptr) {
 683         return rec_find_compound_ent_value(ptr, NULL);
 684 }
 685
 686 /* forward */
 687 static void reduce_adr_usage(ir_node *ptr);
 688
 689 /**
 690  * Update a Load that may have lost its users.
 691  */
 692 static void handle_load_update(ir_node *load) {
 693         ldst_info_t *info = get_irn_link(load);
 694
 695         /* do NOT touch volatile loads for now */
 696         if (get_Load_volatility(load) == volatility_is_volatile)
 697                 return;
 698
 699         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 700                 ir_node *ptr = get_Load_ptr(load);
 701                 ir_node *mem = get_Load_mem(load);
 702
 703                 /* a Load whose value is neither used nor exception checked, remove it */
 704                 exchange(info->projs[pn_Load_M], mem);
 705                 if (info->projs[pn_Load_X_regular])
 706                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 707                 kill_node(load);
 708                 reduce_adr_usage(ptr);
 709         }
 710 }  /* handle_load_update */
 711
 712 /**
 713  * A use of an address node has vanished. Check if this was a Proj
 714  * node and update the counters.
 715  */
 716 static void reduce_adr_usage(ir_node *ptr) {
 717         if (is_Proj(ptr)) {
 718                 if (get_irn_n_edges(ptr) <= 0) {
 719                         /* this Proj is dead now */
 720                         ir_node *pred = get_Proj_pred(ptr);
 721
 722                         if (is_Load(pred)) {
 723                                 ldst_info_t *info = get_irn_link(pred);
 724                                 info->projs[get_Proj_proj(ptr)] = NULL;
 725
 726                                 /* this node lost its result proj, handle that */
 727                                 handle_load_update(pred);
 728                         }
 729                 }
 730         }
 731 }  /* reduce_adr_usage */
 732
 733 /**
 734  * Check, if an already existing value of mode old_mode can be converted
 735  * into the needed one new_mode without loss.
 736  */
 737 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 738         if (old_mode == new_mode)
 739                 return 1;
 740
 741         /* if both modes are two-complement ones, we can always convert the
 742            Stored value into the needed one. */
 743         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 744                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 745                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 746                 return 1;
 747         return 0;
 748 }  /* can_use_stored_value */
 749
 750 /**
 751  * Check whether a Call is at least pure, ie. does only read memory.
 752  */
 753 static unsigned is_Call_pure(ir_node *call) {
 754         ir_type *call_tp = get_Call_type(call);
 755         unsigned prop = get_method_additional_properties(call_tp);
 756
 757         /* check first the call type */
 758         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 759                 /* try the called entity */
 760                 ir_node *ptr = get_Call_ptr(call);
 761
 762                 if (is_Global(ptr)) {
 763                         ir_entity *ent = get_Global_entity(ptr);
 764
 765                         prop = get_entity_additional_properties(ent);
 766                 }
 767         }
 768         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 769 }  /* is_Call_pure */
 770
 771 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 772 {
 773         ir_mode *mode  = get_irn_mode(ptr);
 774         long    offset = 0;
 775
 776         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 777         for (;;) {
 778                 if (is_Add(ptr)) {
 779                         ir_node *l = get_Add_left(ptr);
 780                         ir_node *r = get_Add_right(ptr);
 781
 782                         if (get_irn_mode(l) != mode || !is_Const(r))
 783                                 break;
 784
 785                         offset += get_tarval_long(get_Const_tarval(r));
 786                         ptr     = l;
 787                 } else if (is_Sub(ptr)) {
 788                         ir_node *l = get_Sub_left(ptr);
 789                         ir_node *r = get_Sub_right(ptr);
 790
 791                         if (get_irn_mode(l) != mode || !is_Const(r))
 792                                 break;
 793
 794                         offset -= get_tarval_long(get_Const_tarval(r));
 795                         ptr     = l;
 796                 } else if (is_Sel(ptr)) {
 797                         ir_entity *ent = get_Sel_entity(ptr);
 798                         ir_type   *tp  = get_entity_owner(ent);
 799
 800                         if (is_Array_type(tp)) {
 801                                 int     size;
 802                                 ir_node *index;
 803
 804                                 /* only one dimensional arrays yet */
 805                                 if (get_Sel_n_indexs(ptr) != 1)
 806                                         break;
 807                                 index = get_Sel_index(ptr, 0);
 808                                 if (! is_Const(index))
 809                                         break;
 810
 811                                 tp = get_entity_type(ent);
 812                                 if (get_type_state(tp) != layout_fixed)
 813                                         break;
 814
 815                                 size    = get_type_size_bytes(tp);
 816                                 offset += size * get_tarval_long(get_Const_tarval(index));
 817                         } else {
 818                                 if (get_type_state(tp) != layout_fixed)
 819                                         break;
 820                                 offset += get_entity_offset(ent);
 821                         }
 822                         ptr = get_Sel_ptr(ptr);
 823                 } else
 824                         break;
 825         }
 826
 827         *pOffset = offset;
 828         return ptr;
 829 }
 830
 831 static int try_load_after_store(ir_node *load,
 832                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 833 {
 834         ldst_info_t *info;
 835         ir_node *store_ptr      = get_Store_ptr(store);
 836         long     store_offset;
 837         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 838         ir_node *store_value;
 839         ir_mode *store_mode;
 840         ir_node *load_ptr;
 841         ir_mode *load_mode;
 842         long     load_mode_len;
 843         long     store_mode_len;
 844         long     delta;
 845         int      res;
 846
 847         if (load_base_ptr != store_base_ptr)
 848                 return 0;
 849
 850         load_mode      = get_Load_mode(load);
 851         load_mode_len  = get_mode_size_bytes(load_mode);
 852         store_mode     = get_irn_mode(get_Store_value(store));
 853         store_mode_len = get_mode_size_bytes(store_mode);
 854         delta          = load_offset - store_offset;
 855         store_value    = get_Store_value(store);
 856
 857         if (delta != 0 || store_mode != load_mode) {
 858                 if (delta < 0 || delta + load_mode_len > store_mode_len)
 859                         return 0;
 860
 861                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 862                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 863                         return 0;
 864
 865
 866                 /* produce a shift to adjust offset delta */
 867                 if (delta > 0) {
 868                         ir_node *cnst;
 869
 870                         /* FIXME: only true for little endian */
 871                         cnst        = new_Const_long(mode_Iu, delta * 8);
 872                         store_value = new_r_Shr(get_nodes_block(load),
 873                                                                         store_value, cnst, store_mode);
 874                 }
 875
 876                 /* add an convert if needed */
 877                 if (store_mode != load_mode) {
 878                         store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
 879                 }
 880         }
 881
 882         DBG_OPT_RAW(load, store_value);
 883
 884         info = get_irn_link(load);
 885         if (info->projs[pn_Load_M])
 886                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 887
 888         res = 0;
 889         /* no exception */
 890         if (info->projs[pn_Load_X_except]) {
 891                 exchange( info->projs[pn_Load_X_except], new_Bad());
 892                 res |= CF_CHANGED;
 893         }
 894         if (info->projs[pn_Load_X_regular]) {
 895                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 896                 res |= CF_CHANGED;
 897         }
 898
 899         if (info->projs[pn_Load_res])
 900                 exchange(info->projs[pn_Load_res], store_value);
 901
 902         load_ptr = get_Load_ptr(load);
 903         kill_node(load);
 904         reduce_adr_usage(load_ptr);
 905         return res | DF_CHANGED;
 906 }
 907
 908 /**
 909  * Follow the memory chain as long as there are only Loads,
 910  * alias free Stores, and constant Calls and try to replace the
 911  * current Load by a previous ones.
 912  * Note that in unreachable loops it might happen that we reach
 913  * load again, as well as we can fall into a cycle.
 914  * We break such cycles using a special visited flag.
 915  *
 916  * INC_MASTER() must be called before dive into
 917  */
 918 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 919         unsigned    res = 0;
 920         ldst_info_t *info = get_irn_link(load);
 921         ir_node     *pred;
 922         ir_node     *ptr       = get_Load_ptr(load);
 923         ir_node     *mem       = get_Load_mem(load);
 924         ir_mode     *load_mode = get_Load_mode(load);
 925
 926         for (pred = curr; load != pred; ) {
 927                 ldst_info_t *pred_info = get_irn_link(pred);
 928
 929                 /*
 930                  * a Load immediately after a Store -- a read after write.
 931                  * We may remove the Load, if both Load & Store does not have an
 932                  * exception handler OR they are in the same MacroBlock. In the latter
 933                  * case the Load cannot throw an exception when the previous Store was
 934                  * quiet.
 935                  *
 936                  * Why we need to check for Store Exception? If the Store cannot
 937                  * be executed (ROM) the exception handler might simply jump into
 938                  * the load MacroBlock :-(
 939                  * We could make it a little bit better if we would know that the
 940                  * exception handler of the Store jumps directly to the end...
 941                  */
 942                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 943                                 && info->projs[pn_Load_X_except] == NULL)
 944                                 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
 945                 {
 946                         long    load_offset;
 947                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 948                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 949
 950                         if (changes != 0)
 951                                 return res | changes;
 952                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 953                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 954                         /*
 955                          * a Load after a Load -- a read after read.
 956                          * We may remove the second Load, if it does not have an exception handler
 957                          * OR they are in the same MacroBlock. In the later case the Load cannot
 958                          * throw an exception when the previous Load was quiet.
 959                          *
 960                          * Here, there is no need to check if the previous Load has an exception
 961                          * hander because they would have exact the same exception...
 962                          */
 963                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 964                                 ir_node *value;
 965
 966                                 DBG_OPT_RAR(load, pred);
 967
 968                                 /* the result is used */
 969                                 if (info->projs[pn_Load_res]) {
 970                                         if (pred_info->projs[pn_Load_res] == NULL) {
 971                                                 /* create a new Proj again */
 972                                                 pred_info->projs[pn_Load_res] = new_r_Proj(get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 973                                         }
 974                                         value = pred_info->projs[pn_Load_res];
 975
 976                                         /* add an convert if needed */
 977                                         if (get_Load_mode(pred) != load_mode) {
 978                                                 value = new_r_Conv(get_nodes_block(load), value, load_mode);
 979                                         }
 980
 981                                         exchange(info->projs[pn_Load_res], value);
 982                                 }
 983
 984                                 if (info->projs[pn_Load_M])
 985                                         exchange(info->projs[pn_Load_M], mem);
 986
 987                                 /* no exception */
 988                                 if (info->projs[pn_Load_X_except]) {
 989                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 990                                         res |= CF_CHANGED;
 991                                 }
 992                                 if (info->projs[pn_Load_X_regular]) {
 993                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 994                                         res |= CF_CHANGED;
 995                                 }
 996
 997                                 kill_node(load);
 998                                 reduce_adr_usage(ptr);
 999                                 return res |= DF_CHANGED;
1000                         }
1001                 }
1002
1003                 if (is_Store(pred)) {
1004                         /* check if we can pass through this store */
1005                         ir_alias_relation rel = get_alias_relation(
1006                                 current_ir_graph,
1007                                 get_Store_ptr(pred),
1008                                 get_irn_mode(get_Store_value(pred)),
1009                                 ptr, load_mode);
1010                         /* if the might be an alias, we cannot pass this Store */
1011                         if (rel != ir_no_alias)
1012                                 break;
1013                         pred = skip_Proj(get_Store_mem(pred));
1014                 } else if (is_Load(pred)) {
1015                         pred = skip_Proj(get_Load_mem(pred));
1016                 } else if (is_Call(pred)) {
1017                         if (is_Call_pure(pred)) {
1018                                 /* The called graph is at least pure, so there are no Store's
1019                                    in it. We can handle it like a Load and skip it. */
1020                                 pred = skip_Proj(get_Call_mem(pred));
1021                         } else {
1022                                 /* there might be Store's in the graph, stop here */
1023                                 break;
1024                         }
1025                 } else {
1026                         /* follow only Load chains */
1027                         break;
1028                 }
1029
1030                 /* check for cycles */
1031                 if (NODE_VISITED(pred_info))
1032                         break;
1033                 MARK_NODE(pred_info);
1034         }
1035
1036         if (is_Sync(pred)) {
1037                 int i;
1038
1039                 /* handle all Sync predecessors */
1040                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1041                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1042                         if (res)
1043                                 return res;
1044                 }
1045         }
1046
1047         return res;
1048 }  /* follow_Mem_chain */
1049
1050 /*
1051  * Check if we can replace the load by a given const from
1052  * the const code irg.
1053  */
1054 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1055         ir_mode *c_mode = get_irn_mode(c);
1056         ir_mode *l_mode = get_Load_mode(load);
1057         ir_node *res    = NULL;
1058
1059         if (c_mode != l_mode) {
1060                 /* check, if the mode matches OR can be easily converted info */
1061                 if (is_reinterpret_cast(c_mode, l_mode)) {
1062                         /* we can safely cast */
1063                         dbg_info *dbg   = get_irn_dbg_info(load);
1064                         ir_node  *block = get_nodes_block(load);
1065
1066                         /* copy the value from the const code irg and cast it */
1067                         res = copy_const_value(dbg, c);
1068                         res = new_rd_Conv(dbg, block, res, l_mode);
1069                 }
1070         } else {
1071                 /* copy the value from the const code irg */
1072                 res = copy_const_value(get_irn_dbg_info(load), c);
1073         }
1074         return res;
1075 }  /* can_replace_load_by_const */
1076
1077 /**
1078  * optimize a Load
1079  *
1080  * @param load  the Load node
1081  */
1082 static unsigned optimize_load(ir_node *load)
1083 {
1084         ldst_info_t *info = get_irn_link(load);
1085         ir_node     *mem, *ptr, *value;
1086         ir_entity   *ent;
1087         long        dummy;
1088         unsigned    res = 0;
1089
1090         /* do NOT touch volatile loads for now */
1091         if (get_Load_volatility(load) == volatility_is_volatile)
1092                 return 0;
1093
1094         /* the address of the load to be optimized */
1095         ptr = get_Load_ptr(load);
1096
1097         /*
1098          * Check if we can remove the exception from a Load:
1099          * This can be done, if the address is from an Sel(Alloc) and
1100          * the Sel type is a subtype of the allocated type.
1101          *
1102          * This optimizes some often used OO constructs,
1103          * like x = new O; x->t;
1104          */
1105         if (info->projs[pn_Load_X_except]) {
1106                 ir_node *addr = ptr;
1107
1108                 /* find base address */
1109                 while (is_Sel(addr))
1110                         addr = get_Sel_ptr(addr);
1111                 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1112                         /* simple case: a direct load after an Alloc. Firm Alloc throw
1113                          * an exception in case of out-of-memory. So, there is no way for an
1114                          * exception in this load.
1115                          * This code is constructed by the "exception lowering" in the Jack compiler.
1116                          */
1117                         exchange(info->projs[pn_Load_X_except], new_Bad());
1118                         info->projs[pn_Load_X_except] = NULL;
1119                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1120                         info->projs[pn_Load_X_regular] = NULL;
1121                         res |= CF_CHANGED;
1122                 }
1123         }
1124
1125         /* The mem of the Load. Must still be returned after optimization. */
1126         mem = get_Load_mem(load);
1127
1128         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1129                 /* a Load which value is neither used nor exception checked, remove it */
1130                 exchange(info->projs[pn_Load_M], mem);
1131
1132                 if (info->projs[pn_Load_X_regular]) {
1133                         /* should not happen, but if it does, remove it */
1134                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1135                         res |= CF_CHANGED;
1136                 }
1137                 kill_node(load);
1138                 reduce_adr_usage(ptr);
1139                 return res | DF_CHANGED;
1140         }
1141
1142         /* Load from a constant polymorphic field, where we can resolve
1143            polymorphism. */
1144         value = transform_polymorph_Load(load);
1145         if (value == load) {
1146                 value = NULL;
1147                 /* check if we can determine the entity that will be loaded */
1148                 ent = find_constant_entity(ptr);
1149                 if (ent != NULL                                     &&
1150                     allocation_static == get_entity_allocation(ent) &&
1151                     visibility_external_allocated != get_entity_visibility(ent)) {
1152                         /* a static allocation that is not external: there should be NO exception
1153                          * when loading even if we cannot replace the load itself. */
1154
1155                         /* no exception, clear the info field as it might be checked later again */
1156                         if (info->projs[pn_Load_X_except]) {
1157                                 exchange(info->projs[pn_Load_X_except], new_Bad());
1158                                 info->projs[pn_Load_X_except] = NULL;
1159                                 res |= CF_CHANGED;
1160                         }
1161                         if (info->projs[pn_Load_X_regular]) {
1162                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1163                                 info->projs[pn_Load_X_regular] = NULL;
1164                                 res |= CF_CHANGED;
1165                         }
1166
1167                         if (variability_constant == get_entity_variability(ent)) {
1168                                 if (is_atomic_entity(ent)) {
1169                                         /* Might not be atomic after lowering of Sels.  In this case we
1170                                          * could also load, but it's more complicated. */
1171                                         /* more simpler case: we load the content of a constant value:
1172                                          * replace it by the constant itself */
1173                                         value = get_atomic_ent_value(ent);
1174                                 } else if (ent->has_initializer) {
1175                                         /* new style initializer */
1176                                         value = find_compound_ent_value(ptr);
1177                                 } else {
1178                                         /* old style initializer */
1179                                         compound_graph_path *path = get_accessed_path(ptr);
1180
1181                                         if (path != NULL) {
1182                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1183
1184                                                 value = get_compound_ent_value_by_path(ent, path);
1185                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1186                                                 free_compound_graph_path(path);
1187                                         }
1188                                 }
1189                                 if (value != NULL)
1190                                         value = can_replace_load_by_const(load, value);
1191                         }
1192                 }
1193         }
1194         if (value != NULL) {
1195                 /* we completely replace the load by this value */
1196                 if (info->projs[pn_Load_X_except]) {
1197                         exchange(info->projs[pn_Load_X_except], new_Bad());
1198                         info->projs[pn_Load_X_except] = NULL;
1199                         res |= CF_CHANGED;
1200                 }
1201                 if (info->projs[pn_Load_X_regular]) {
1202                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1203                         info->projs[pn_Load_X_regular] = NULL;
1204                         res |= CF_CHANGED;
1205                 }
1206                 if (info->projs[pn_Load_M]) {
1207                         exchange(info->projs[pn_Load_M], mem);
1208                         res |= DF_CHANGED;
1209                 }
1210                 if (info->projs[pn_Load_res]) {
1211                         exchange(info->projs[pn_Load_res], value);
1212                         res |= DF_CHANGED;
1213                 }
1214                 kill_node(load);
1215                 reduce_adr_usage(ptr);
1216                 return res;
1217         }
1218
1219         /* Check, if the address of this load is used more than once.
1220          * If not, more load cannot be removed in any case. */
1221         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1222                 return res;
1223
1224         /*
1225          * follow the memory chain as long as there are only Loads
1226          * and try to replace current Load or Store by a previous one.
1227          * Note that in unreachable loops it might happen that we reach
1228          * load again, as well as we can fall into a cycle.
1229          * We break such cycles using a special visited flag.
1230          */
1231         INC_MASTER();
1232         res = follow_Mem_chain(load, skip_Proj(mem));
1233         return res;
1234 }  /* optimize_load */
1235
1236 /**
1237  * Check whether a value of mode new_mode would completely overwrite a value
1238  * of mode old_mode in memory.
1239  */
1240 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1241 {
1242         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1243 }  /* is_completely_overwritten */
1244
1245 /**
1246  * Check whether small is a part of large (starting at same address).
1247  */
1248 static int is_partially_same(ir_node *small, ir_node *large)
1249 {
1250         ir_mode *sm = get_irn_mode(small);
1251         ir_mode *lm = get_irn_mode(large);
1252
1253         /* FIXME: Check endianness */
1254         return is_Conv(small) && get_Conv_op(small) == large
1255             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1256             && get_mode_arithmetic(sm) == irma_twos_complement
1257             && get_mode_arithmetic(lm) == irma_twos_complement;
1258 }  /* is_partially_same */
1259
1260 /**
1261  * follow the memory chain as long as there are only Loads and alias free Stores.
1262  *
1263  * INC_MASTER() must be called before dive into
1264  */
1265 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1266         unsigned res = 0;
1267         ldst_info_t *info = get_irn_link(store);
1268         ir_node *pred;
1269         ir_node *ptr = get_Store_ptr(store);
1270         ir_node *mem = get_Store_mem(store);
1271         ir_node *value = get_Store_value(store);
1272         ir_mode *mode  = get_irn_mode(value);
1273         ir_node *block = get_nodes_block(store);
1274         ir_node *mblk  = get_Block_MacroBlock(block);
1275
1276         for (pred = curr; pred != store;) {
1277                 ldst_info_t *pred_info = get_irn_link(pred);
1278
1279                 /*
1280                  * BEWARE: one might think that checking the modes is useless, because
1281                  * if the pointers are identical, they refer to the same object.
1282                  * This is only true in strong typed languages, not is C were the following
1283                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1284                  * However, if the size of the mode that is written is bigger or equal the
1285                  * size of the old one, the old value is completely overwritten and can be
1286                  * killed ...
1287                  */
1288                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1289             get_nodes_MacroBlock(pred) == mblk) {
1290                         /*
1291                          * a Store after a Store in the same MacroBlock -- a write after write.
1292                          */
1293
1294                         /*
1295                          * We may remove the first Store, if the old value is completely
1296                          * overwritten or the old value is a part of the new value,
1297                          * and if it does not have an exception handler.
1298                          *
1299                          * TODO: What, if both have the same exception handler ???
1300                          */
1301                         if (get_Store_volatility(pred) != volatility_is_volatile
1302                                 && !pred_info->projs[pn_Store_X_except]) {
1303                                 ir_node *predvalue = get_Store_value(pred);
1304                                 ir_mode *predmode  = get_irn_mode(predvalue);
1305
1306                                 if(is_completely_overwritten(predmode, mode)
1307                                         || is_partially_same(predvalue, value)) {
1308                                         DBG_OPT_WAW(pred, store);
1309                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1310                                         kill_node(pred);
1311                                         reduce_adr_usage(ptr);
1312                                         return DF_CHANGED;
1313                                 }
1314                         }
1315
1316                         /*
1317                          * We may remove the Store, if the old value already contains
1318                          * the new value, and if it does not have an exception handler.
1319                          *
1320                          * TODO: What, if both have the same exception handler ???
1321                          */
1322                         if (get_Store_volatility(store) != volatility_is_volatile
1323                                 && !info->projs[pn_Store_X_except]) {
1324                                 ir_node *predvalue = get_Store_value(pred);
1325
1326                                 if(is_partially_same(value, predvalue)) {
1327                                         DBG_OPT_WAW(pred, store);
1328                                         exchange(info->projs[pn_Store_M], mem);
1329                                         kill_node(store);
1330                                         reduce_adr_usage(ptr);
1331                                         return DF_CHANGED;
1332                                 }
1333                         }
1334                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1335                            value == pred_info->projs[pn_Load_res]) {
1336                         /*
1337                          * a Store of a value just loaded from the same address
1338                          * -- a write after read.
1339                          * We may remove the Store, if it does not have an exception
1340                          * handler.
1341                          */
1342                         if (! info->projs[pn_Store_X_except]) {
1343                                 DBG_OPT_WAR(store, pred);
1344                                 exchange(info->projs[pn_Store_M], mem);
1345                                 kill_node(store);
1346                                 reduce_adr_usage(ptr);
1347                                 return DF_CHANGED;
1348                         }
1349                 }
1350
1351                 if (is_Store(pred)) {
1352                         /* check if we can pass through this store */
1353                         ir_alias_relation rel = get_alias_relation(
1354                                 current_ir_graph,
1355                                 get_Store_ptr(pred),
1356                                 get_irn_mode(get_Store_value(pred)),
1357                                 ptr, mode);
1358                         /* if the might be an alias, we cannot pass this Store */
1359                         if (rel != ir_no_alias)
1360                                 break;
1361                         pred = skip_Proj(get_Store_mem(pred));
1362                 } else if (is_Load(pred)) {
1363                         ir_alias_relation rel = get_alias_relation(
1364                                 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1365                                 ptr, mode);
1366                         if (rel != ir_no_alias)
1367                                 break;
1368
1369                         pred = skip_Proj(get_Load_mem(pred));
1370                 } else {
1371                         /* follow only Load chains */
1372                         break;
1373                 }
1374
1375                 /* check for cycles */
1376                 if (NODE_VISITED(pred_info))
1377                         break;
1378                 MARK_NODE(pred_info);
1379         }
1380
1381         if (is_Sync(pred)) {
1382                 int i;
1383
1384                 /* handle all Sync predecessors */
1385                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1386                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1387                         if (res)
1388                                 break;
1389                 }
1390         }
1391         return res;
1392 }  /* follow_Mem_chain_for_Store */
1393
1394 /** find entity used as base for an address calculation */
1395 static ir_entity *find_entity(ir_node *ptr)
1396 {
1397         switch(get_irn_opcode(ptr)) {
1398         case iro_SymConst:
1399                 return get_SymConst_entity(ptr);
1400         case iro_Sel: {
1401                 ir_node *pred = get_Sel_ptr(ptr);
1402                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1403                         return get_Sel_entity(ptr);
1404
1405                 return find_entity(pred);
1406         }
1407         case iro_Sub:
1408         case iro_Add: {
1409                 ir_node *left = get_binop_left(ptr);
1410                 ir_node *right;
1411                 if (mode_is_reference(get_irn_mode(left)))
1412                         return find_entity(left);
1413                 right = get_binop_right(ptr);
1414                 if (mode_is_reference(get_irn_mode(right)))
1415                         return find_entity(right);
1416                 return NULL;
1417         }
1418         default:
1419                 return NULL;
1420         }
1421 }
1422
1423 /**
1424  * optimize a Store
1425  *
1426  * @param store  the Store node
1427  */
1428 static unsigned optimize_store(ir_node *store) {
1429         ir_node   *ptr;
1430         ir_node   *mem;
1431         ir_entity *entity;
1432
1433         if (get_Store_volatility(store) == volatility_is_volatile)
1434                 return 0;
1435
1436         ptr    = get_Store_ptr(store);
1437         entity = find_entity(ptr);
1438
1439         /* a store to an entity which is never read is unnecessary */
1440         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1441                 ldst_info_t *info = get_irn_link(store);
1442                 if (info->projs[pn_Store_X_except] == NULL) {
1443                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1444                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1445                         kill_node(store);
1446                         reduce_adr_usage(ptr);
1447                         return DF_CHANGED;
1448                 }
1449         }
1450
1451         /* Check, if the address of this Store is used more than once.
1452          * If not, this Store cannot be removed in any case. */
1453         if (get_irn_n_uses(ptr) <= 1)
1454                 return 0;
1455
1456         mem = get_Store_mem(store);
1457
1458         /* follow the memory chain as long as there are only Loads */
1459         INC_MASTER();
1460
1461         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1462 }  /* optimize_store */
1463
1464 /**
1465  * walker, optimizes Phi after Stores to identical places:
1466  * Does the following optimization:
1467  * @verbatim
1468  *
1469  *   val1   val2   val3          val1  val2  val3
1470  *    |      |      |               \    |    /
1471  *  Store  Store  Store              \   |   /
1472  *      \    |    /                   PhiData
1473  *       \   |   /                       |
1474  *        \  |  /                      Store
1475  *          PhiM
1476  *
1477  * @endverbatim
1478  * This reduces the number of stores and allows for predicated execution.
1479  * Moves Stores back to the end of a function which may be bad.
1480  *
1481  * This is only possible if the predecessor blocks have only one successor.
1482  */
1483 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1484 {
1485         int i, n;
1486         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1487         ir_mode *mode;
1488         ir_node **inM, **inD, **projMs;
1489         int *idx;
1490         dbg_info *db = NULL;
1491         ldst_info_t *info;
1492         block_info_t *bl_info;
1493         unsigned res = 0;
1494
1495         /* Must be a memory Phi */
1496         if (get_irn_mode(phi) != mode_M)
1497                 return 0;
1498
1499         n = get_Phi_n_preds(phi);
1500         if (n <= 0)
1501                 return 0;
1502
1503         /* must be only one user */
1504         projM = get_Phi_pred(phi, 0);
1505         if (get_irn_n_edges(projM) != 1)
1506                 return 0;
1507
1508         store = skip_Proj(projM);
1509         old_store = store;
1510         if (!is_Store(store))
1511                 return 0;
1512
1513         block = get_nodes_block(store);
1514
1515         /* abort on dead blocks */
1516         if (is_Block_dead(block))
1517                 return 0;
1518
1519         /* check if the block is post dominated by Phi-block
1520            and has no exception exit */
1521         bl_info = get_irn_link(block);
1522         if (bl_info->flags & BLOCK_HAS_EXC)
1523                 return 0;
1524
1525         phi_block = get_nodes_block(phi);
1526         if (! block_strictly_postdominates(phi_block, block))
1527                 return 0;
1528
1529         /* this is the address of the store */
1530         ptr  = get_Store_ptr(store);
1531         mode = get_irn_mode(get_Store_value(store));
1532         info = get_irn_link(store);
1533         exc  = info->exc_block;
1534
1535         for (i = 1; i < n; ++i) {
1536                 ir_node *pred = get_Phi_pred(phi, i);
1537
1538                 if (get_irn_n_edges(pred) != 1)
1539                         return 0;
1540
1541                 pred = skip_Proj(pred);
1542                 if (!is_Store(pred))
1543                         return 0;
1544
1545                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1546                         return 0;
1547
1548                 info = get_irn_link(pred);
1549
1550                 /* check, if all stores have the same exception flow */
1551                 if (exc != info->exc_block)
1552                         return 0;
1553
1554                 /* abort on dead blocks */
1555                 block = get_nodes_block(pred);
1556                 if (is_Block_dead(block))
1557                         return 0;
1558
1559                 /* check if the block is post dominated by Phi-block
1560                    and has no exception exit. Note that block must be different from
1561                    Phi-block, else we would move a Store from end End of a block to its
1562                    Start... */
1563                 bl_info = get_irn_link(block);
1564                 if (bl_info->flags & BLOCK_HAS_EXC)
1565                         return 0;
1566                 if (block == phi_block || ! block_postdominates(phi_block, block))
1567                         return 0;
1568         }
1569
1570         /*
1571          * ok, when we are here, we found all predecessors of a Phi that
1572          * are Stores to the same address and size. That means whatever
1573          * we do before we enter the block of the Phi, we do a Store.
1574          * So, we can move the Store to the current block:
1575          *
1576          *   val1    val2    val3          val1  val2  val3
1577          *    |       |       |               \    |    /
1578          * | Str | | Str | | Str |             \   |   /
1579          *      \     |     /                   PhiData
1580          *       \    |    /                       |
1581          *        \   |   /                       Str
1582          *           PhiM
1583          *
1584          * Is only allowed if the predecessor blocks have only one successor.
1585          */
1586
1587         NEW_ARR_A(ir_node *, projMs, n);
1588         NEW_ARR_A(ir_node *, inM, n);
1589         NEW_ARR_A(ir_node *, inD, n);
1590         NEW_ARR_A(int, idx, n);
1591
1592         /* Prepare: Collect all Store nodes.  We must do this
1593            first because we otherwise may loose a store when exchanging its
1594            memory Proj.
1595          */
1596         for (i = n - 1; i >= 0; --i) {
1597                 ir_node *store;
1598
1599                 projMs[i] = get_Phi_pred(phi, i);
1600                 assert(is_Proj(projMs[i]));
1601
1602                 store = get_Proj_pred(projMs[i]);
1603                 info  = get_irn_link(store);
1604
1605                 inM[i] = get_Store_mem(store);
1606                 inD[i] = get_Store_value(store);
1607                 idx[i] = info->exc_idx;
1608         }
1609         block = get_nodes_block(phi);
1610
1611         /* second step: create a new memory Phi */
1612         phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1613
1614         /* third step: create a new data Phi */
1615         phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1616
1617         /* rewire memory and kill the node */
1618         for (i = n - 1; i >= 0; --i) {
1619                 ir_node *proj  = projMs[i];
1620
1621                 if(is_Proj(proj)) {
1622                         ir_node *store = get_Proj_pred(proj);
1623                         exchange(proj, inM[i]);
1624                         kill_node(store);
1625                 }
1626         }
1627
1628         /* fourth step: create the Store */
1629         store = new_rd_Store(db, block, phiM, ptr, phiD, 0);
1630 #ifdef DO_CACHEOPT
1631         co_set_irn_name(store, co_get_irn_ident(old_store));
1632 #endif
1633
1634         projM = new_rd_Proj(NULL, block, store, mode_M, pn_Store_M);
1635
1636         info = get_ldst_info(store, &wenv->obst);
1637         info->projs[pn_Store_M] = projM;
1638
1639         /* fifths step: repair exception flow */
1640         if (exc) {
1641                 ir_node *projX = new_rd_Proj(NULL, block, store, mode_X, pn_Store_X_except);
1642
1643                 info->projs[pn_Store_X_except] = projX;
1644                 info->exc_block                = exc;
1645                 info->exc_idx                  = idx[0];
1646
1647                 for (i = 0; i < n; ++i) {
1648                         set_Block_cfgpred(exc, idx[i], projX);
1649                 }
1650
1651                 if (n > 1) {
1652                         /* the exception block should be optimized as some inputs are identical now */
1653                 }
1654
1655                 res |= CF_CHANGED;
1656         }
1657
1658         /* sixth step: replace old Phi */
1659         exchange(phi, projM);
1660
1661         return res | DF_CHANGED;
1662 }  /* optimize_phi */
1663
1664 /**
1665  * walker, do the optimizations
1666  */
1667 static void do_load_store_optimize(ir_node *n, void *env) {
1668         walk_env_t *wenv = env;
1669
1670         switch (get_irn_opcode(n)) {
1671
1672         case iro_Load:
1673                 wenv->changes |= optimize_load(n);
1674                 break;
1675
1676         case iro_Store:
1677                 wenv->changes |= optimize_store(n);
1678                 break;
1679
1680         case iro_Phi:
1681                 wenv->changes |= optimize_phi(n, wenv);
1682                 break;
1683
1684         default:
1685                 ;
1686         }
1687 }  /* do_load_store_optimize */
1688
1689 /** A scc. */
1690 typedef struct scc {
1691         ir_node *head;          /**< the head of the list */
1692 } scc;
1693
1694 /** A node entry. */
1695 typedef struct node_entry {
1696         unsigned DFSnum;    /**< the DFS number of this node */
1697         unsigned low;       /**< the low number of this node */
1698         int      in_stack;  /**< flag, set if the node is on the stack */
1699         ir_node  *next;     /**< link to the next node the the same scc */
1700         scc      *pscc;     /**< the scc of this node */
1701         unsigned POnum;     /**< the post order number for blocks */
1702 } node_entry;
1703
1704 /** A loop entry. */
1705 typedef struct loop_env {
1706         ir_phase ph;           /**< the phase object */
1707         ir_node  **stack;      /**< the node stack */
1708         int      tos;          /**< tos index */
1709         unsigned nextDFSnum;   /**< the current DFS number */
1710         unsigned POnum;        /**< current post order number */
1711
1712         unsigned changes;      /**< a bitmask of graph changes */
1713 } loop_env;
1714
1715 /**
1716 * Gets the node_entry of a node
1717 */
1718 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1719         ir_phase   *ph = &env->ph;
1720         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1721
1722         if (! e) {
1723                 e = phase_alloc(ph, sizeof(*e));
1724                 memset(e, 0, sizeof(*e));
1725                 phase_set_irn_data(ph, irn, e);
1726         }
1727         return e;
1728 }  /* get_irn_ne */
1729
1730 /**
1731  * Push a node onto the stack.
1732  *
1733  * @param env   the loop environment
1734  * @param n     the node to push
1735  */
1736 static void push(loop_env *env, ir_node *n) {
1737         node_entry *e;
1738
1739         if (env->tos == ARR_LEN(env->stack)) {
1740                 int nlen = ARR_LEN(env->stack) * 2;
1741                 ARR_RESIZE(ir_node *, env->stack, nlen);
1742         }
1743         env->stack[env->tos++] = n;
1744         e = get_irn_ne(n, env);
1745         e->in_stack = 1;
1746 }  /* push */
1747
1748 /**
1749  * pop a node from the stack
1750  *
1751  * @param env   the loop environment
1752  *
1753  * @return  The topmost node
1754  */
1755 static ir_node *pop(loop_env *env) {
1756         ir_node *n = env->stack[--env->tos];
1757         node_entry *e = get_irn_ne(n, env);
1758
1759         e->in_stack = 0;
1760         return n;
1761 }  /* pop */
1762
1763 /**
1764  * Check if irn is a region constant.
1765  * The block or irn must strictly dominate the header block.
1766  *
1767  * @param irn           the node to check
1768  * @param header_block  the header block of the induction variable
1769  */
1770 static int is_rc(ir_node *irn, ir_node *header_block) {
1771         ir_node *block = get_nodes_block(irn);
1772
1773         return (block != header_block) && block_dominates(block, header_block);
1774 }  /* is_rc */
1775
1776 typedef struct phi_entry phi_entry;
1777 struct phi_entry {
1778         ir_node   *phi;    /**< A phi with a region const memory. */
1779         int       pos;     /**< The position of the region const memory */
1780         ir_node   *load;   /**< the newly created load for this phi */
1781         phi_entry *next;
1782 };
1783
1784 /**
1785  * An entry in the avail set.
1786  */
1787 typedef struct avail_entry_t {
1788         ir_node *ptr;   /**< the address pointer */
1789         ir_mode *mode;  /**< the load mode */
1790         ir_node *load;  /**< the associated Load */
1791 } avail_entry_t;
1792
1793 /**
1794  * Compare two avail entries.
1795  */
1796 static int cmp_avail_entry(const void *elt, const void *key, size_t size) {
1797         const avail_entry_t *a = elt;
1798         const avail_entry_t *b = key;
1799         (void) size;
1800
1801         return a->ptr != b->ptr || a->mode != b->mode;
1802 }  /* cmp_avail_entry */
1803
1804 /**
1805  * Calculate the hash value of an avail entry.
1806  */
1807 static unsigned hash_cache_entry(const avail_entry_t *entry) {
1808         return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1809 }  /* hash_cache_entry */
1810
1811 /**
1812  * Move loops out of loops if possible.
1813  *
1814  * @param pscc   the loop described by an SCC
1815  * @param env    the loop environment
1816  */
1817 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1818         ir_node   *phi, *load, *next, *other, *next_other;
1819         ir_entity *ent;
1820         int       j;
1821         phi_entry *phi_list = NULL;
1822         set       *avail;
1823
1824         avail = new_set(cmp_avail_entry, 8);
1825
1826         /* collect all outer memories */
1827         for (phi = pscc->head; phi != NULL; phi = next) {
1828                 node_entry *ne = get_irn_ne(phi, env);
1829                 next = ne->next;
1830
1831                 /* check all memory Phi's */
1832                 if (! is_Phi(phi))
1833                         continue;
1834
1835                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1836
1837                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1838                         ir_node    *pred = get_irn_n(phi, j);
1839                         node_entry *pe   = get_irn_ne(pred, env);
1840
1841                         if (pe->pscc != ne->pscc) {
1842                                 /* not in the same SCC, is region const */
1843                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1844
1845                                 pe->phi  = phi;
1846                                 pe->pos  = j;
1847                                 pe->next = phi_list;
1848                                 phi_list = pe;
1849                         }
1850                 }
1851         }
1852         /* no Phis no fun */
1853         assert(phi_list != NULL && "DFS found a loop without Phi");
1854
1855         /* for now, we cannot handle more than one input (only reducible cf) */
1856         if (phi_list->next != NULL)
1857                 return;
1858
1859         for (load = pscc->head; load; load = next) {
1860                 ir_mode *load_mode;
1861                 node_entry *ne = get_irn_ne(load, env);
1862                 next = ne->next;
1863
1864                 if (is_Load(load)) {
1865                         ldst_info_t *info = get_irn_link(load);
1866                         ir_node     *ptr = get_Load_ptr(load);
1867
1868                         /* for now, we cannot handle Loads with exceptions */
1869                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1870                                 continue;
1871
1872                         /* for now, we can only move Load(Global) */
1873                         if (! is_Global(ptr))
1874                                 continue;
1875                         ent       = get_Global_entity(ptr);
1876                         load_mode = get_Load_mode(load);
1877                         for (other = pscc->head; other != NULL; other = next_other) {
1878                                 node_entry *ne = get_irn_ne(other, env);
1879                                 next_other = ne->next;
1880
1881                                 if (is_Store(other)) {
1882                                         ir_alias_relation rel = get_alias_relation(
1883                                                 current_ir_graph,
1884                                                 get_Store_ptr(other),
1885                                                 get_irn_mode(get_Store_value(other)),
1886                                                 ptr, load_mode);
1887                                         /* if the might be an alias, we cannot pass this Store */
1888                                         if (rel != ir_no_alias)
1889                                                 break;
1890                                 }
1891                                 /* only Phis and pure Calls are allowed here, so ignore them */
1892                         }
1893                         if (other == NULL) {
1894                                 ldst_info_t *ninfo;
1895                                 phi_entry   *pe;
1896                                 dbg_info    *db;
1897
1898                                 /* yep, no aliasing Store found, Load can be moved */
1899                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1900
1901                                 db   = get_irn_dbg_info(load);
1902                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1903                                         int     pos   = pe->pos;
1904                                         ir_node *phi  = pe->phi;
1905                                         ir_node *blk  = get_nodes_block(phi);
1906                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1907                                         ir_node *irn, *mem;
1908                                         avail_entry_t entry, *res;
1909
1910                                         entry.ptr  = ptr;
1911                                         entry.mode = load_mode;
1912                                         res = set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1913                                         if (res != NULL) {
1914                                                 irn = res->load;
1915                                         } else {
1916                                                 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, 0);
1917                                                 entry.load = irn;
1918                                                 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1919                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1920                                         }
1921                                         pe->load = irn;
1922                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1923
1924                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(pred, irn, mode_M, pn_Load_M);
1925                                         set_Phi_pred(phi, pos, mem);
1926
1927                                         ninfo->projs[pn_Load_res] = new_r_Proj(pred, irn, load_mode, pn_Load_res);
1928                                 }
1929
1930                                 /* now kill the old Load */
1931                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1932                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1933
1934                                 env->changes |= DF_CHANGED;
1935                         }
1936                 }
1937         }
1938         del_set(avail);
1939 }  /* move_loads_out_of_loops */
1940
1941 /**
1942  * Process a loop SCC.
1943  *
1944  * @param pscc  the SCC
1945  * @param env   the loop environment
1946  */
1947 static void process_loop(scc *pscc, loop_env *env) {
1948         ir_node *irn, *next, *header = NULL;
1949         node_entry *b, *h = NULL;
1950         int j, only_phi, num_outside, process = 0;
1951         ir_node *out_rc;
1952
1953         /* find the header block for this scc */
1954         for (irn = pscc->head; irn; irn = next) {
1955                 node_entry *e = get_irn_ne(irn, env);
1956                 ir_node *block = get_nodes_block(irn);
1957
1958                 next = e->next;
1959                 b = get_irn_ne(block, env);
1960
1961                 if (header != NULL) {
1962                         if (h->POnum < b->POnum) {
1963                                 header = block;
1964                                 h      = b;
1965                         }
1966                 } else {
1967                         header = block;
1968                         h      = b;
1969                 }
1970         }
1971
1972         /* check if this scc contains only Phi, Loads or Stores nodes */
1973         only_phi    = 1;
1974         num_outside = 0;
1975         out_rc      = NULL;
1976         for (irn = pscc->head; irn; irn = next) {
1977                 node_entry *e = get_irn_ne(irn, env);
1978
1979                 next = e->next;
1980                 switch (get_irn_opcode(irn)) {
1981                 case iro_Call:
1982                         if (is_Call_pure(irn)) {
1983                                 /* pure calls can be treated like loads */
1984                                 only_phi = 0;
1985                                 break;
1986                         }
1987                         /* non-pure calls must be handle like may-alias Stores */
1988                         goto fail;
1989                 case iro_CopyB:
1990                         /* cannot handle CopyB yet */
1991                         goto fail;
1992                 case iro_Load:
1993                         process = 1;
1994                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1995                                 /* cannot handle loops with volatile Loads */
1996                                 goto fail;
1997                         }
1998                         only_phi = 0;
1999                         break;
2000                 case iro_Store:
2001                         if (get_Store_volatility(irn) == volatility_is_volatile) {
2002                                 /* cannot handle loops with volatile Stores */
2003                                 goto fail;
2004                         }
2005                         only_phi = 0;
2006                         break;
2007                 default:
2008                         only_phi = 0;
2009                         break;
2010                 case iro_Phi:
2011                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2012                                 ir_node *pred  = get_irn_n(irn, j);
2013                                 node_entry *pe = get_irn_ne(pred, env);
2014
2015                                 if (pe->pscc != e->pscc) {
2016                                         /* not in the same SCC, must be a region const */
2017                                         if (! is_rc(pred, header)) {
2018                                                 /* not a memory loop */
2019                                                 goto fail;
2020                                         }
2021                                         if (out_rc == NULL) {
2022                                                 /* first region constant */
2023                                                 out_rc = pred;
2024                                                 ++num_outside;
2025                                         } else if (out_rc != pred) {
2026                                                 /* another region constant */
2027                                                 ++num_outside;
2028                                         }
2029                                 }
2030                         }
2031                         break;
2032                 }
2033         }
2034         if (! process)
2035                 goto fail;
2036
2037         /* found a memory loop */
2038         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
2039         if (only_phi && num_outside == 1) {
2040                 /* a phi cycle with only one real predecessor can be collapsed */
2041                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
2042
2043                 for (irn = pscc->head; irn; irn = next) {
2044                         node_entry *e = get_irn_ne(irn, env);
2045                         next = e->next;
2046                         exchange(irn, out_rc);
2047                 }
2048                 env->changes |= DF_CHANGED;
2049                 return;
2050         }
2051
2052 #ifdef DEBUG_libfirm
2053         for (irn = pscc->head; irn; irn = next) {
2054                 node_entry *e = get_irn_ne(irn, env);
2055                 next = e->next;
2056                 DB((dbg, LEVEL_2, " %+F,", irn));
2057         }
2058         DB((dbg, LEVEL_2, "\n"));
2059 #endif
2060         move_loads_out_of_loops(pscc, env);
2061
2062 fail:
2063         ;
2064 }  /* process_loop */
2065
2066 /**
2067  * Process a SCC.
2068  *
2069  * @param pscc  the SCC
2070  * @param env   the loop environment
2071  */
2072 static void process_scc(scc *pscc, loop_env *env) {
2073         ir_node *head = pscc->head;
2074         node_entry *e = get_irn_ne(head, env);
2075
2076 #ifdef DEBUG_libfirm
2077         {
2078                 ir_node *irn, *next;
2079
2080                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2081                 for (irn = pscc->head; irn; irn = next) {
2082                         node_entry *e = get_irn_ne(irn, env);
2083
2084                         next = e->next;
2085
2086                         DB((dbg, LEVEL_4, " %+F,", irn));
2087                 }
2088                 DB((dbg, LEVEL_4, "\n"));
2089         }
2090 #endif
2091
2092         if (e->next != NULL) {
2093                 /* this SCC has more than one member */
2094                 process_loop(pscc, env);
2095         }
2096 }  /* process_scc */
2097
2098 /**
2099  * Do Tarjan's SCC algorithm and drive load/store optimization.
2100  *
2101  * @param irn  start at this node
2102  * @param env  the loop environment
2103  */
2104 static void dfs(ir_node *irn, loop_env *env)
2105 {
2106         int i, n;
2107         node_entry *node = get_irn_ne(irn, env);
2108
2109         mark_irn_visited(irn);
2110
2111         node->DFSnum = env->nextDFSnum++;
2112         node->low    = node->DFSnum;
2113         push(env, irn);
2114
2115         /* handle preds */
2116         if (is_Phi(irn) || is_Sync(irn)) {
2117                 n = get_irn_arity(irn);
2118                 for (i = 0; i < n; ++i) {
2119                         ir_node *pred = get_irn_n(irn, i);
2120                         node_entry *o = get_irn_ne(pred, env);
2121
2122                         if (!irn_visited(pred)) {
2123                                 dfs(pred, env);
2124                                 node->low = MIN(node->low, o->low);
2125                         }
2126                         if (o->DFSnum < node->DFSnum && o->in_stack)
2127                                 node->low = MIN(o->DFSnum, node->low);
2128                 }
2129         } else if (is_fragile_op(irn)) {
2130                 ir_node *pred = get_fragile_op_mem(irn);
2131                 node_entry *o = get_irn_ne(pred, env);
2132
2133                 if (!irn_visited(pred)) {
2134                         dfs(pred, env);
2135                         node->low = MIN(node->low, o->low);
2136                 }
2137                 if (o->DFSnum < node->DFSnum && o->in_stack)
2138                         node->low = MIN(o->DFSnum, node->low);
2139         } else if (is_Proj(irn)) {
2140                 ir_node *pred = get_Proj_pred(irn);
2141                 node_entry *o = get_irn_ne(pred, env);
2142
2143                 if (!irn_visited(pred)) {
2144                         dfs(pred, env);
2145                         node->low = MIN(node->low, o->low);
2146                 }
2147                 if (o->DFSnum < node->DFSnum && o->in_stack)
2148                         node->low = MIN(o->DFSnum, node->low);
2149         }
2150         else {
2151                  /* IGNORE predecessors */
2152         }
2153
2154         if (node->low == node->DFSnum) {
2155                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2156                 ir_node *x;
2157
2158                 pscc->head = NULL;
2159                 do {
2160                         node_entry *e;
2161
2162                         x = pop(env);
2163                         e = get_irn_ne(x, env);
2164                         e->pscc    = pscc;
2165                         e->next    = pscc->head;
2166                         pscc->head = x;
2167                 } while (x != irn);
2168
2169                 process_scc(pscc, env);
2170         }
2171 }  /* dfs */
2172
2173 /**
2174  * Do the DFS on the memory edges a graph.
2175  *
2176  * @param irg  the graph to process
2177  * @param env  the loop environment
2178  */
2179 static void do_dfs(ir_graph *irg, loop_env *env) {
2180         ir_graph *rem = current_ir_graph;
2181         ir_node  *endblk, *end;
2182         int      i;
2183
2184         current_ir_graph = irg;
2185         inc_irg_visited(irg);
2186
2187         /* visit all memory nodes */
2188         endblk = get_irg_end_block(irg);
2189         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2190                 ir_node *pred = get_Block_cfgpred(endblk, i);
2191
2192                 pred = skip_Proj(pred);
2193                 if (is_Return(pred))
2194                         dfs(get_Return_mem(pred), env);
2195                 else if (is_Raise(pred))
2196                         dfs(get_Raise_mem(pred), env);
2197                 else if (is_fragile_op(pred))
2198                         dfs(get_fragile_op_mem(pred), env);
2199                 else {
2200                         assert(0 && "Unknown EndBlock predecessor");
2201                 }
2202         }
2203
2204         /* visit the keep-alives */
2205         end = get_irg_end(irg);
2206         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2207                 ir_node *ka = get_End_keepalive(end, i);
2208
2209                 if (is_Phi(ka) && !irn_visited(ka))
2210                         dfs(ka, env);
2211         }
2212         current_ir_graph = rem;
2213 }  /* do_dfs */
2214
2215 /**
2216  * Initialize new phase data. We do this always explicit, so return NULL here
2217  */
2218 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2219         (void)ph;
2220         (void)irn;
2221         (void)data;
2222         return NULL;
2223 }  /* init_loop_data */
2224
2225 /**
2226  * Optimize Loads/Stores in loops.
2227  *
2228  * @param irg  the graph
2229  */
2230 static int optimize_loops(ir_graph *irg) {
2231         loop_env env;
2232
2233         env.stack         = NEW_ARR_F(ir_node *, 128);
2234         env.tos           = 0;
2235         env.nextDFSnum    = 0;
2236         env.POnum         = 0;
2237         env.changes       = 0;
2238         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2239
2240         /* calculate the SCC's and drive loop optimization. */
2241         do_dfs(irg, &env);
2242
2243         DEL_ARR_F(env.stack);
2244         phase_free(&env.ph);
2245
2246         return env.changes;
2247 }  /* optimize_loops */
2248
2249 /*
2250  * do the load store optimization
2251  */
2252 int optimize_load_store(ir_graph *irg) {
2253         walk_env_t env;
2254
2255         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2256
2257         assert(get_irg_phase_state(irg) != phase_building);
2258         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2259                 "LoadStore optimization needs pinned graph");
2260
2261         /* we need landing pads */
2262         remove_critical_cf_edges(irg);
2263
2264         edges_assure(irg);
2265
2266         /* for Phi optimization post-dominators are needed ... */
2267         assure_postdoms(irg);
2268
2269         if (get_opt_alias_analysis()) {
2270                 assure_irg_entity_usage_computed(irg);
2271                 assure_irp_globals_entity_usage_computed();
2272         }
2273
2274         obstack_init(&env.obst);
2275         env.changes = 0;
2276
2277         /* init the links, then collect Loads/Stores/Proj's in lists */
2278         master_visited = 0;
2279         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2280
2281         /* now we have collected enough information, optimize */
2282         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2283
2284         env.changes |= optimize_loops(irg);
2285
2286         obstack_free(&env.obst, NULL);
2287
2288         /* Handle graph state */
2289         if (env.changes) {
2290                 set_irg_outs_inconsistent(irg);
2291                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2292         }
2293
2294         if (env.changes & CF_CHANGED) {
2295                 /* is this really needed: Yes, control flow changed, block might
2296                 have Bad() predecessors. */
2297                 set_irg_doms_inconsistent(irg);
2298         }
2299         return env.changes != 0;
2300 }  /* optimize_load_store */
2301
2302 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2303 {
2304         return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2305 }  /* optimize_load_store_pass */