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

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