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

   1 /*
   2  * Copyright (C) 1995-2007 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.h"
  46 #include "irhooks.h"
  47 #include "iredges.h"
  48 #include "irtools.h"
  49 #include "opt_polymorphy.h"
  50 #include "irmemory.h"
  51 #include "xmalloc.h"
  52
  53 #ifdef DO_CACHEOPT
  54 #include "cacheopt/cachesim.h"
  55 #endif
  56
  57 #undef IMAX
  58 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  59
  60 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  61
  62 enum changes_t {
  63         DF_CHANGED = 1,       /**< data flow changed */
  64         CF_CHANGED = 2,       /**< control flow changed */
  65 };
  66
  67 /**
  68  * walker environment
  69  */
  70 typedef struct _walk_env_t {
  71         struct obstack obst;          /**< list of all stores */
  72         unsigned changes;             /**< a bitmask of graph changes */
  73 } walk_env_t;
  74
  75 /**
  76  * flags for Load/Store
  77  */
  78 enum ldst_flags_t {
  79         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  80 };
  81
  82 /** A Load/Store info. */
  83 typedef struct _ldst_info_t {
  84         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  85         ir_node  *exc_block;          /**< the exception block if available */
  86         int      exc_idx;             /**< predecessor index in the exception block */
  87         unsigned flags;               /**< flags */
  88         unsigned visited;             /**< visited counter for breaking loops */
  89 } ldst_info_t;
  90
  91 /**
  92  * flags for control flow.
  93  */
  94 enum block_flags_t {
  95         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  96         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  97 };
  98
  99 /**
 100  * a Block info.
 101  */
 102 typedef struct _block_info_t {
 103         unsigned flags;               /**< flags for the block */
 104 } block_info_t;
 105
 106 /** the master visited flag for loop detection. */
 107 static unsigned master_visited = 0;
 108
 109 #define INC_MASTER()       ++master_visited
 110 #define MARK_NODE(info)    (info)->visited = master_visited
 111 #define NODE_VISITED(info) (info)->visited >= master_visited
 112
 113 /**
 114  * get the Load/Store info of a node
 115  */
 116 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 117         ldst_info_t *info = get_irn_link(node);
 118
 119         if (! info) {
 120                 info = obstack_alloc(&env->obst, sizeof(*info));
 121                 memset(info, 0, sizeof(*info));
 122                 set_irn_link(node, info);
 123         }
 124         return info;
 125 }  /* get_ldst_info */
 126
 127 /**
 128  * get the Block info of a node
 129  */
 130 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 131         block_info_t *info = get_irn_link(node);
 132
 133         if (! info) {
 134                 info = obstack_alloc(&env->obst, sizeof(*info));
 135                 memset(info, 0, sizeof(*info));
 136                 set_irn_link(node, info);
 137         }
 138         return info;
 139 }  /* get_block_info */
 140
 141 /**
 142  * update the projection info for a Load/Store
 143  */
 144 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 145 {
 146         long nr = get_Proj_proj(proj);
 147
 148         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 149
 150         if (info->projs[nr]) {
 151                 /* there is already one, do CSE */
 152                 exchange(proj, info->projs[nr]);
 153                 return DF_CHANGED;
 154         }
 155         else {
 156                 info->projs[nr] = proj;
 157                 return 0;
 158         }
 159 }  /* update_projs */
 160
 161 /**
 162  * update the exception block info for a Load/Store node.
 163  *
 164  * @param info   the load/store info struct
 165  * @param block  the exception handler block for this load/store
 166  * @param pos    the control flow input of the block
 167  */
 168 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 169 {
 170         assert(info->exc_block == NULL && "more than one exception block found");
 171
 172         info->exc_block = block;
 173         info->exc_idx   = pos;
 174         return 0;
 175 }  /* update_exc */
 176
 177 /** Return the number of uses of an address node */
 178 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 179
 180 /**
 181  * walker, collects all Load/Store/Proj nodes
 182  *
 183  * walks from Start -> End
 184  */
 185 static void collect_nodes(ir_node *node, void *env)
 186 {
 187         ir_op       *op = get_irn_op(node);
 188         ir_node     *pred, *blk, *pred_blk;
 189         ldst_info_t *ldst_info;
 190         walk_env_t  *wenv = env;
 191
 192         if (op == op_Proj) {
 193                 ir_node *adr;
 194                 ir_op *op;
 195
 196                 pred = get_Proj_pred(node);
 197                 op   = get_irn_op(pred);
 198
 199                 if (op == op_Load) {
 200                         ldst_info = get_ldst_info(pred, wenv);
 201
 202                         wenv->changes |= update_projs(ldst_info, node);
 203
 204                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 205                                 adr = get_Load_ptr(pred);
 206                                 ldst_info->flags |= LDST_VISITED;
 207                         }
 208
 209                         /*
 210                         * Place the Proj's to the same block as the
 211                         * predecessor Load. This is always ok and prevents
 212                         * "non-SSA" form after optimizations if the Proj
 213                         * is in a wrong block.
 214                         */
 215                         blk      = get_nodes_block(node);
 216                         pred_blk = get_nodes_block(pred);
 217                         if (blk != pred_blk) {
 218                                 wenv->changes |= DF_CHANGED;
 219                                 set_nodes_block(node, pred_blk);
 220                         }
 221                 } else if (op == op_Store) {
 222                         ldst_info = get_ldst_info(pred, wenv);
 223
 224                         wenv->changes |= update_projs(ldst_info, node);
 225
 226                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 227                                 adr = get_Store_ptr(pred);
 228                                 ldst_info->flags |= LDST_VISITED;
 229                         }
 230
 231                         /*
 232                         * Place the Proj's to the same block as the
 233                         * predecessor Store. This is always ok and prevents
 234                         * "non-SSA" form after optimizations if the Proj
 235                         * is in a wrong block.
 236                         */
 237                         blk      = get_nodes_block(node);
 238                         pred_blk = get_nodes_block(pred);
 239                         if (blk != pred_blk) {
 240                                 wenv->changes |= DF_CHANGED;
 241                                 set_nodes_block(node, pred_blk);
 242                         }
 243                 }
 244         } else if (op == op_Block) {
 245                 int i;
 246
 247                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 248                         ir_node      *pred_block;
 249                         block_info_t *bl_info;
 250
 251                         pred = skip_Proj(get_Block_cfgpred(node, i));
 252
 253                         /* ignore Bad predecessors, they will be removed later */
 254                         if (is_Bad(pred))
 255                                 continue;
 256
 257                         pred_block = get_nodes_block(pred);
 258                         bl_info    = get_block_info(pred_block, wenv);
 259
 260                         if (is_fragile_op(pred))
 261                                 bl_info->flags |= BLOCK_HAS_EXC;
 262                         else if (is_irn_forking(pred))
 263                                 bl_info->flags |= BLOCK_HAS_COND;
 264
 265                         if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
 266                                 ldst_info = get_ldst_info(pred, wenv);
 267
 268                                 wenv->changes |= update_exc(ldst_info, node, i);
 269                         }
 270                 }
 271         }
 272 }  /* collect_nodes */
 273
 274 /**
 275  * Returns an entity if the address ptr points to a constant one.
 276  *
 277  * @param ptr  the address
 278  *
 279  * @return an entity or NULL
 280  */
 281 static ir_entity *find_constant_entity(ir_node *ptr)
 282 {
 283         for (;;) {
 284                 ir_op *op = get_irn_op(ptr);
 285
 286                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 287                         ir_entity *ent = get_SymConst_entity(ptr);
 288                         if (variability_constant == get_entity_variability(ent))
 289                                 return ent;
 290                         return NULL;
 291                 } else if (op == op_Sel) {
 292                         ir_entity *ent = get_Sel_entity(ptr);
 293                         ir_type   *tp  = get_entity_owner(ent);
 294
 295                         /* Do not fiddle with polymorphism. */
 296                         if (is_Class_type(get_entity_owner(ent)) &&
 297                                 ((get_entity_n_overwrites(ent)    != 0) ||
 298                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 299                                 return NULL;
 300
 301                         if (is_Array_type(tp)) {
 302                                 /* check bounds */
 303                                 int i, n;
 304
 305                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 306                                         ir_node *bound;
 307                                         tarval *tlower, *tupper;
 308                                         ir_node *index = get_Sel_index(ptr, i);
 309                                         tarval *tv     = computed_value(index);
 310
 311                                         /* check if the index is constant */
 312                                         if (tv == tarval_bad)
 313                                                 return NULL;
 314
 315                                         bound  = get_array_lower_bound(tp, i);
 316                                         tlower = computed_value(bound);
 317                                         bound  = get_array_upper_bound(tp, i);
 318                                         tupper = computed_value(bound);
 319
 320                                         if (tlower == tarval_bad || tupper == tarval_bad)
 321                                                 return NULL;
 322
 323                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 324                                                 return NULL;
 325                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 326                                                 return NULL;
 327
 328                                         /* ok, bounds check finished */
 329                                 }
 330                         }
 331
 332                         if (variability_constant == get_entity_variability(ent))
 333                                 return ent;
 334
 335                         /* try next */
 336                         ptr = get_Sel_ptr(ptr);
 337                 } else
 338                         return NULL;
 339         }
 340 }  /* find_constant_entity */
 341
 342 /**
 343  * Return the Selection index of a Sel node from dimension n
 344  */
 345 static long get_Sel_array_index_long(ir_node *n, int dim) {
 346         ir_node *index = get_Sel_index(n, dim);
 347         assert(get_irn_op(index) == op_Const);
 348         return get_tarval_long(get_Const_tarval(index));
 349 }  /* get_Sel_array_index_long */
 350
 351 /**
 352  * Returns the accessed component graph path for an
 353  * node computing an address.
 354  *
 355  * @param ptr    the node computing the address
 356  * @param depth  current depth in steps upward from the root
 357  *               of the address
 358  */
 359 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 360         compound_graph_path *res = NULL;
 361         ir_entity           *root, *field;
 362         int                 path_len, pos;
 363
 364         if (get_irn_op(ptr) == op_SymConst) {
 365                 /* a SymConst. If the depth is 0, this is an access to a global
 366                  * entity and we don't need a component path, else we know
 367                  * at least it's length.
 368                  */
 369                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 370                 root = get_SymConst_entity(ptr);
 371                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 372         } else {
 373                 assert(get_irn_op(ptr) == op_Sel);
 374                 /* it's a Sel, go up until we find the root */
 375                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 376
 377                 /* fill up the step in the path at the current position */
 378                 field    = get_Sel_entity(ptr);
 379                 path_len = get_compound_graph_path_length(res);
 380                 pos      = path_len - depth - 1;
 381                 set_compound_graph_path_node(res, pos, field);
 382
 383                 if (is_Array_type(get_entity_owner(field))) {
 384                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 385                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 386                 }
 387         }
 388         return res;
 389 }  /* rec_get_accessed_path */
 390
 391 /** Returns an access path or NULL.  The access path is only
 392  *  valid, if the graph is in phase_high and _no_ address computation is used.
 393  */
 394 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 395         return rec_get_accessed_path(ptr, 0);
 396 }  /* get_accessed_path */
 397
 398 /* forward */
 399 static void reduce_adr_usage(ir_node *ptr);
 400
 401 /**
 402  * Update a Load that may lost it's usage.
 403  */
 404 static void handle_load_update(ir_node *load) {
 405         ldst_info_t *info = get_irn_link(load);
 406
 407         /* do NOT touch volatile loads for now */
 408         if (get_Load_volatility(load) == volatility_is_volatile)
 409                 return;
 410
 411         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 412                 ir_node *ptr = get_Load_ptr(load);
 413                 ir_node *mem = get_Load_mem(load);
 414
 415                 /* a Load which value is neither used nor exception checked, remove it */
 416                 exchange(info->projs[pn_Load_M], mem);
 417                 exchange(load, new_Bad());
 418                 reduce_adr_usage(ptr);
 419         }
 420 }  /* handle_load_update */
 421
 422 /**
 423  * A Use of an address node is vanished. Check if this was a Proj
 424  * node and update the counters.
 425  */
 426 static void reduce_adr_usage(ir_node *ptr) {
 427         if (is_Proj(ptr)) {
 428                 if (get_irn_n_edges(ptr) <= 0) {
 429                         /* this Proj is dead now */
 430                         ir_node *pred = get_Proj_pred(ptr);
 431
 432                         if (is_Load(pred)) {
 433                                 ldst_info_t *info = get_irn_link(pred);
 434                                 info->projs[get_Proj_proj(ptr)] = NULL;
 435
 436                                 /* this node lost it's result proj, handle that */
 437                                 handle_load_update(pred);
 438                         }
 439                 }
 440         }
 441 }  /* reduce_adr_usage */
 442
 443 /**
 444  * Check, if an already existing value of mode old_mode can be converted
 445  * into the needed one new_mode without loss.
 446  */
 447 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 448         if (old_mode == new_mode)
 449                 return 1;
 450
 451         /* if both modes are two-complement ones, we can always convert the
 452            Stored value into the needed one. */
 453         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 454                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 455                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 456                 return 1;
 457         return 0;
 458 }  /* can_use_stored_value */
 459
 460 /**
 461  * Follow the memory chain as long as there are only Loads
 462  * and alias free Stores and try to replace current Load or Store
 463  * by a previous ones.
 464  * Note that in unreachable loops it might happen that we reach
 465  * load again, as well as we can fall into a cycle.
 466  * We break such cycles using a special visited flag.
 467  *
 468  * INC_MASTER() must be called before dive into
 469  */
 470 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 471         unsigned res = 0;
 472         ldst_info_t *info = get_irn_link(load);
 473         ir_node *pred;
 474         ir_node *ptr       = get_Load_ptr(load);
 475         ir_node *mem       = get_Load_mem(load);
 476         ir_mode *load_mode = get_Load_mode(load);
 477
 478         for (pred = curr; load != pred; ) {
 479                 ldst_info_t *pred_info = get_irn_link(pred);
 480
 481                 /*
 482                  * BEWARE: one might think that checking the modes is useless, because
 483                  * if the pointers are identical, they refer to the same object.
 484                  * This is only true in strong typed languages, not in C were the following
 485                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 486                  */
 487                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 488                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 489                         /*
 490                          * a Load immediately after a Store -- a read after write.
 491                          * We may remove the Load, if both Load & Store does not have an exception handler
 492                          * OR they are in the same block. In the latter case the Load cannot
 493                          * throw an exception when the previous Store was quiet.
 494                          *
 495                          * Why we need to check for Store Exception? If the Store cannot
 496                          * be executed (ROM) the exception handler might simply jump into
 497                          * the load block :-(
 498                          * We could make it a little bit better if we would know that the exception
 499                          * handler of the Store jumps directly to the end...
 500                          */
 501                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 502                             get_nodes_block(load) == get_nodes_block(pred)) {
 503                                 ir_node *value = get_Store_value(pred);
 504
 505                                 DBG_OPT_RAW(load, value);
 506
 507                                 /* add an convert if needed */
 508                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 509                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 510                                 }
 511
 512                                 if (info->projs[pn_Load_M])
 513                                         exchange(info->projs[pn_Load_M], mem);
 514
 515                                 /* no exception */
 516                                 if (info->projs[pn_Load_X_except]) {
 517                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 518                                         res |= CF_CHANGED;
 519                                 }
 520
 521                                 if (info->projs[pn_Load_res])
 522                                         exchange(info->projs[pn_Load_res], value);
 523
 524                                 exchange(load, new_Bad());
 525                                 reduce_adr_usage(ptr);
 526                                 return res | DF_CHANGED;
 527                         }
 528                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 529                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 530                         /*
 531                          * a Load after a Load -- a read after read.
 532                          * We may remove the second Load, if it does not have an exception handler
 533                          * OR they are in the same block. In the later case the Load cannot
 534                          * throw an exception when the previous Load was quiet.
 535                          *
 536                          * Here, there is no need to check if the previous Load has an exception
 537                          * hander because they would have exact the same exception...
 538                          */
 539                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
 540                                 ir_node *value;
 541
 542                                 DBG_OPT_RAR(load, pred);
 543
 544                                 /* the result is used */
 545                                 if (info->projs[pn_Load_res]) {
 546                                         if (pred_info->projs[pn_Load_res] == NULL) {
 547                                                 /* create a new Proj again */
 548                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 549                                         }
 550                                         value = pred_info->projs[pn_Load_res];
 551
 552                                         /* add an convert if needed */
 553                                         if (get_Load_mode(pred) != load_mode) {
 554                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 555                                         }
 556
 557                                         exchange(info->projs[pn_Load_res], value);
 558                                 }
 559
 560                                 if (info->projs[pn_Load_M])
 561                                         exchange(info->projs[pn_Load_M], mem);
 562
 563                                 /* no exception */
 564                                 if (info->projs[pn_Load_X_except]) {
 565                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 566                                         res |= CF_CHANGED;
 567                                 }
 568
 569                                 exchange(load, new_Bad());
 570                                 reduce_adr_usage(ptr);
 571                                 return res |= DF_CHANGED;
 572                         }
 573                 }
 574
 575                 if (get_irn_op(pred) == op_Store) {
 576                         /* check if we can pass through this store */
 577                         ir_alias_relation rel = get_alias_relation(
 578                                 current_ir_graph,
 579                                 get_Store_ptr(pred),
 580                                 get_irn_mode(get_Store_value(pred)),
 581                                 ptr, load_mode);
 582                         /* if the might be an alias, we cannot pass this Store */
 583                         if (rel != no_alias)
 584                                 break;
 585                         pred = skip_Proj(get_Store_mem(pred));
 586                 } else if (get_irn_op(pred) == op_Load) {
 587                         pred = skip_Proj(get_Load_mem(pred));
 588                 } else {
 589                         /* follow only Load chains */
 590                         break;
 591                 }
 592
 593                 /* check for cycles */
 594                 if (NODE_VISITED(pred_info))
 595                         break;
 596                 MARK_NODE(pred_info);
 597         }
 598
 599         if (get_irn_op(pred) == op_Sync) {
 600                 int i;
 601
 602                 /* handle all Sync predecessors */
 603                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 604                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 605                         if (res)
 606                                 break;
 607                 }
 608         }
 609
 610         return res;
 611 }  /* follow_Mem_chain */
 612
 613 /**
 614  * optimize a Load
 615  *
 616  * @param load  the Load node
 617  */
 618 static unsigned optimize_load(ir_node *load)
 619 {
 620         ldst_info_t *info = get_irn_link(load);
 621         ir_node *mem, *ptr, *new_node;
 622         ir_entity *ent;
 623         unsigned res = 0;
 624
 625         /* do NOT touch volatile loads for now */
 626         if (get_Load_volatility(load) == volatility_is_volatile)
 627                 return 0;
 628
 629         /* the address of the load to be optimized */
 630         ptr = get_Load_ptr(load);
 631
 632         /*
 633          * Check if we can remove the exception from a Load:
 634          * This can be done, if the address is from an Sel(Alloc) and
 635          * the Sel type is a subtype of the allocated type.
 636          *
 637          * This optimizes some often used OO constructs,
 638          * like x = new O; x->t;
 639          */
 640         if (info->projs[pn_Load_X_except]) {
 641                 if (is_Sel(ptr)) {
 642                         ir_node *mem = get_Sel_mem(ptr);
 643
 644                         /* FIXME: works with the current FE, but better use the base */
 645                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 646                                 /* ok, check the types */
 647                                 ir_entity *ent    = get_Sel_entity(ptr);
 648                                 ir_type   *s_type = get_entity_type(ent);
 649                                 ir_type   *a_type = get_Alloc_type(mem);
 650
 651                                 if (is_SubClass_of(s_type, a_type)) {
 652                                         /* ok, condition met: there can't be an exception because
 653                                         * Alloc guarantees that enough memory was allocated */
 654
 655                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 656                                         info->projs[pn_Load_X_except] = NULL;
 657                                         res |= CF_CHANGED;
 658                                 }
 659                         }
 660                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 661                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 662                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 663                                  * an exception in case of out-of-memory. So, there is no way for an
 664                                  * exception in this load.
 665                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 666                                  */
 667                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 668                                 info->projs[pn_Load_X_except] = NULL;
 669                                 res |= CF_CHANGED;
 670                 }
 671         }
 672
 673         /* The mem of the Load. Must still be returned after optimization. */
 674         mem  = get_Load_mem(load);
 675
 676         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 677                 /* a Load which value is neither used nor exception checked, remove it */
 678                 exchange(info->projs[pn_Load_M], mem);
 679
 680                 exchange(load, new_Bad());
 681                 reduce_adr_usage(ptr);
 682                 return res | DF_CHANGED;
 683         }
 684
 685         /* Load from a constant polymorphic field, where we can resolve
 686            polymorphism. */
 687         new_node = transform_node_Load(load);
 688         if (new_node != load) {
 689                 if (info->projs[pn_Load_M]) {
 690                         exchange(info->projs[pn_Load_M], mem);
 691                         info->projs[pn_Load_M] = NULL;
 692                 }
 693                 if (info->projs[pn_Load_X_except]) {
 694                         exchange(info->projs[pn_Load_X_except], new_Bad());
 695                         info->projs[pn_Load_X_except] = NULL;
 696                 }
 697                 if (info->projs[pn_Load_res])
 698                         exchange(info->projs[pn_Load_res], new_node);
 699
 700                 exchange(load, new_Bad());
 701                 reduce_adr_usage(ptr);
 702                 return res | DF_CHANGED;
 703         }
 704
 705         /* check if we can determine the entity that will be loaded */
 706         ent = find_constant_entity(ptr);
 707         if (ent) {
 708                 if ((allocation_static == get_entity_allocation(ent)) &&
 709                         (visibility_external_allocated != get_entity_visibility(ent))) {
 710                         /* a static allocation that is not external: there should be NO exception
 711                          * when loading. */
 712
 713                         /* no exception, clear the info field as it might be checked later again */
 714                         if (info->projs[pn_Load_X_except]) {
 715                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 716                                 info->projs[pn_Load_X_except] = NULL;
 717                                 res |= CF_CHANGED;
 718                         }
 719
 720                         if (variability_constant == get_entity_variability(ent)
 721                                 && is_atomic_entity(ent)) {
 722                                 /* Might not be atomic after
 723                                    lowering of Sels.  In this
 724                                    case we could also load, but
 725                                    it's more complicated. */
 726                                 /* more simpler case: we load the content of a constant value:
 727                                  * replace it by the constant itself
 728                                  */
 729
 730                                 /* no memory */
 731                                 if (info->projs[pn_Load_M]) {
 732                                         exchange(info->projs[pn_Load_M], mem);
 733                                         res |= DF_CHANGED;
 734                                 }
 735                                 /* no result :-) */
 736                                 if (info->projs[pn_Load_res]) {
 737                                         if (is_atomic_entity(ent)) {
 738                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 739
 740                                                 DBG_OPT_RC(load, c);
 741                                                 exchange(info->projs[pn_Load_res], c);
 742                                                 res |= DF_CHANGED;
 743                                         }
 744                                 }
 745                                 exchange(load, new_Bad());
 746                                 reduce_adr_usage(ptr);
 747                                 return res;
 748                         } else if (variability_constant == get_entity_variability(ent)) {
 749                                 compound_graph_path *path = get_accessed_path(ptr);
 750
 751                                 if (path) {
 752                                         ir_node *c;
 753
 754                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 755                                         /*
 756                                         {
 757                                                 int j;
 758                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 759                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 760                                                         fprintf(stdout, ".%s", get_entity_name(node));
 761                                                         if (is_Array_type(get_entity_owner(node)))
 762                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 763                                                 }
 764                                                 printf("\n");
 765                                         }
 766                                         */
 767
 768                                         c = get_compound_ent_value_by_path(ent, path);
 769                                         free_compound_graph_path(path);
 770
 771                                         /* printf("  cons: "); DDMN(c); */
 772
 773                                         if (info->projs[pn_Load_M]) {
 774                                                 exchange(info->projs[pn_Load_M], mem);
 775                                                 res |= DF_CHANGED;
 776                                         }
 777                                         if (info->projs[pn_Load_res]) {
 778                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 779                                                 res |= DF_CHANGED;
 780                                         }
 781                                         exchange(load, new_Bad());
 782                                         reduce_adr_usage(ptr);
 783                                         return res;
 784                                 } else {
 785                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 786                                         a byte from an object to generate an exception.   Happens in test program
 787                                         Reflectiontest.
 788                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 789                                         get_entity_name(get_irg_entity(current_ir_graph)));
 790                                         printf("  load: "); DDMN(load);
 791                                         printf("  ptr:  "); DDMN(ptr);
 792                                         */
 793                                 }
 794                         }
 795                 }
 796         }
 797
 798         /* Check, if the address of this load is used more than once.
 799          * If not, this load cannot be removed in any case. */
 800         if (get_irn_n_uses(ptr) <= 1)
 801                 return res;
 802
 803         /*
 804          * follow the memory chain as long as there are only Loads
 805          * and try to replace current Load or Store by a previous one.
 806          * Note that in unreachable loops it might happen that we reach
 807          * load again, as well as we can fall into a cycle.
 808          * We break such cycles using a special visited flag.
 809          */
 810         INC_MASTER();
 811         res = follow_Mem_chain(load, skip_Proj(mem));
 812         return res;
 813 }  /* optimize_load */
 814
 815 /**
 816  * Check whether a value of mode new_mode would completely overwrite a value
 817  * of mode old_mode in memory.
 818  */
 819 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 820 {
 821         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 822 }  /* is_completely_overwritten */
 823
 824 /**
 825  * follow the memory chain as long as there are only Loads and alias free Stores.
 826  *
 827  * INC_MASTER() must be called before dive into
 828  */
 829 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 830         unsigned res = 0;
 831         ldst_info_t *info = get_irn_link(store);
 832         ir_node *pred;
 833         ir_node *ptr = get_Store_ptr(store);
 834         ir_node *mem = get_Store_mem(store);
 835         ir_node *value = get_Store_value(store);
 836         ir_mode *mode  = get_irn_mode(value);
 837         ir_node *block = get_nodes_block(store);
 838
 839         for (pred = curr; pred != store;) {
 840                 ldst_info_t *pred_info = get_irn_link(pred);
 841
 842                 /*
 843                  * BEWARE: one might think that checking the modes is useless, because
 844                  * if the pointers are identical, they refer to the same object.
 845                  * This is only true in strong typed languages, not is C were the following
 846                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 847                  * However, if the mode that is written have a bigger  or equal size the the old
 848                  * one, the old value is completely overwritten and can be killed ...
 849                  */
 850                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 851                     get_nodes_block(pred) == block &&
 852                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 853                         /*
 854                          * a Store after a Store in the same block -- a write after write.
 855                          * We may remove the first Store, if it does not have an exception handler.
 856                          *
 857                          * TODO: What, if both have the same exception handler ???
 858                          */
 859                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 860                                 DBG_OPT_WAW(pred, store);
 861                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 862                                 exchange(pred, new_Bad());
 863                                 reduce_adr_usage(ptr);
 864                                 return DF_CHANGED;
 865                         }
 866                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 867                            value == pred_info->projs[pn_Load_res]) {
 868                         /*
 869                          * a Store of a value after a Load -- a write after read.
 870                          * We may remove the second Store, if it does not have an exception handler.
 871                          */
 872                         if (! info->projs[pn_Store_X_except]) {
 873                                 DBG_OPT_WAR(store, pred);
 874                                 exchange( info->projs[pn_Store_M], mem );
 875                                 exchange(store, new_Bad());
 876                                 reduce_adr_usage(ptr);
 877                                 return DF_CHANGED;
 878                         }
 879                 }
 880
 881                 if (get_irn_op(pred) == op_Store) {
 882                         /* check if we can pass thru this store */
 883                         ir_alias_relation rel = get_alias_relation(
 884                                 current_ir_graph,
 885                                 get_Store_ptr(pred),
 886                                 get_irn_mode(get_Store_value(pred)),
 887                                 ptr, mode);
 888                         /* if the might be an alias, we cannot pass this Store */
 889                         if (rel != no_alias)
 890                                 break;
 891                         pred = skip_Proj(get_Store_mem(pred));
 892                 } else if (get_irn_op(pred) == op_Load) {
 893                         pred = skip_Proj(get_Load_mem(pred));
 894                 } else {
 895                         /* follow only Load chains */
 896                         break;
 897                 }
 898
 899                 /* check for cycles */
 900                 if (NODE_VISITED(pred_info))
 901                         break;
 902                 MARK_NODE(pred_info);
 903         }
 904
 905         if (get_irn_op(pred) == op_Sync) {
 906                 int i;
 907
 908                 /* handle all Sync predecessors */
 909                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 910                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 911                         if (res)
 912                                 break;
 913                 }
 914         }
 915         return res;
 916 }  /* follow_Mem_chain_for_Store */
 917
 918 /**
 919  * optimize a Store
 920  *
 921  * @param store  the Store node
 922  */
 923 static unsigned optimize_store(ir_node *store) {
 924         ir_node *ptr, *mem;
 925
 926         if (get_Store_volatility(store) == volatility_is_volatile)
 927                 return 0;
 928
 929         ptr = get_Store_ptr(store);
 930
 931         /* Check, if the address of this Store is used more than once.
 932          * If not, this Store cannot be removed in any case. */
 933         if (get_irn_n_uses(ptr) <= 1)
 934                 return 0;
 935
 936         mem = get_Store_mem(store);
 937
 938         /* follow the memory chain as long as there are only Loads */
 939         INC_MASTER();
 940         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 941 }  /* optimize_store */
 942
 943 /**
 944  * walker, optimizes Phi after Stores to identical places:
 945  * Does the following optimization:
 946  * @verbatim
 947  *
 948  *   val1   val2   val3          val1  val2  val3
 949  *    |      |      |               \    |    /
 950  *  Store  Store  Store              \   |   /
 951  *      \    |    /                   PhiData
 952  *       \   |   /                       |
 953  *        \  |  /                      Store
 954  *          PhiM
 955  *
 956  * @endverbatim
 957  * This reduces the number of stores and allows for predicated execution.
 958  * Moves Stores back to the end of a function which may be bad.
 959  *
 960  * This is only possible if the predecessor blocks have only one successor.
 961  */
 962 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 963 {
 964         int i, n;
 965         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 966         ir_mode *mode;
 967         ir_node **inM, **inD, **stores;
 968         int *idx;
 969         dbg_info *db = NULL;
 970         ldst_info_t *info;
 971         block_info_t *bl_info;
 972         unsigned res = 0;
 973
 974         /* Must be a memory Phi */
 975         if (get_irn_mode(phi) != mode_M)
 976                 return 0;
 977
 978         n = get_Phi_n_preds(phi);
 979         if (n <= 0)
 980                 return 0;
 981
 982         store = skip_Proj(get_Phi_pred(phi, 0));
 983         old_store = store;
 984         if (get_irn_op(store) != op_Store)
 985                 return 0;
 986
 987         block = get_nodes_block(store);
 988
 989         /* abort on dead blocks */
 990         if (is_Block_dead(block))
 991                 return 0;
 992
 993         /* check if the block is post dominated by Phi-block
 994            and has no exception exit */
 995         bl_info = get_irn_link(block);
 996         if (bl_info->flags & BLOCK_HAS_EXC)
 997                 return 0;
 998
 999         phi_block = get_nodes_block(phi);
1000         if (! block_postdominates(phi_block, block))
1001                 return 0;
1002
1003         /* this is the address of the store */
1004         ptr  = get_Store_ptr(store);
1005         mode = get_irn_mode(get_Store_value(store));
1006         info = get_irn_link(store);
1007         exc  = info->exc_block;
1008
1009         for (i = 1; i < n; ++i) {
1010                 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
1011
1012                 if (get_irn_op(pred) != op_Store)
1013                         return 0;
1014
1015                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1016                         return 0;
1017
1018                 info = get_irn_link(pred);
1019
1020                 /* check, if all stores have the same exception flow */
1021                 if (exc != info->exc_block)
1022                         return 0;
1023
1024                 /* abort on dead blocks */
1025                 block = get_nodes_block(pred);
1026                 if (is_Block_dead(block))
1027                         return 0;
1028
1029                 /* check if the block is post dominated by Phi-block
1030                    and has no exception exit. Note that block must be different from
1031                    Phi-block, else we would move a Store from end End of a block to its
1032                    Start... */
1033                 bl_info = get_irn_link(block);
1034                 if (bl_info->flags & BLOCK_HAS_EXC)
1035                         return 0;
1036                 if (block == phi_block || ! block_postdominates(phi_block, block))
1037                         return 0;
1038         }
1039
1040         /*
1041          * ok, when we are here, we found all predecessors of a Phi that
1042          * are Stores to the same address and size. That means whatever
1043          * we do before we enter the block of the Phi, we do a Store.
1044          * So, we can move the Store to the current block:
1045          *
1046          *   val1    val2    val3          val1  val2  val3
1047          *    |       |       |               \    |    /
1048          * | Str | | Str | | Str |             \   |   /
1049          *      \     |     /                   PhiData
1050          *       \    |    /                       |
1051          *        \   |   /                       Str
1052          *           PhiM
1053          *
1054          * Is only allowed if the predecessor blocks have only one successor.
1055          */
1056
1057         NEW_ARR_A(ir_node *, stores, n);
1058         NEW_ARR_A(ir_node *, inM, n);
1059         NEW_ARR_A(ir_node *, inD, n);
1060         NEW_ARR_A(int, idx, n);
1061
1062         /* Prepare: Collect all Store nodes.  We must do this
1063            first because we otherwise may loose a store when exchanging its
1064            memory Proj.
1065          */
1066         for (i = 0; i < n; ++i)
1067                 stores[i] = skip_Proj(get_Phi_pred(phi, i));
1068
1069         /* first step: collect all inputs */
1070         for (i = 0; i < n; ++i) {
1071                 ir_node *store = stores[i];
1072                 info = get_irn_link(store);
1073
1074                 inM[i] = get_Store_mem(store);
1075                 inD[i] = get_Store_value(store);
1076                 idx[i] = info->exc_idx;
1077
1078                 kill_node(store);
1079         }
1080         block = get_nodes_block(phi);
1081
1082         /* second step: create a new memory Phi */
1083         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1084
1085         /* third step: create a new data Phi */
1086         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1087
1088         /* fourth step: create the Store */
1089         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1090 #ifdef DO_CACHEOPT
1091         co_set_irn_name(store, co_get_irn_ident(old_store));
1092 #endif
1093
1094         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1095
1096         info = get_ldst_info(store, wenv);
1097         info->projs[pn_Store_M] = projM;
1098
1099         /* fifths step: repair exception flow */
1100         if (exc) {
1101                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1102
1103                 info->projs[pn_Store_X_except] = projX;
1104                 info->exc_block                = exc;
1105                 info->exc_idx                  = idx[0];
1106
1107                 for (i = 0; i < n; ++i) {
1108                         set_Block_cfgpred(exc, idx[i], projX);
1109                 }
1110
1111                 if (n > 1) {
1112                         /* the exception block should be optimized as some inputs are identical now */
1113                 }
1114
1115                 res |= CF_CHANGED;
1116         }
1117
1118         /* sixth step: replace old Phi */
1119         exchange(phi, projM);
1120
1121         return res | DF_CHANGED;
1122 }  /* optimize_phi */
1123
1124 /**
1125  * walker, do the optimizations
1126  */
1127 static void do_load_store_optimize(ir_node *n, void *env) {
1128         walk_env_t *wenv = env;
1129
1130         switch (get_irn_opcode(n)) {
1131
1132         case iro_Load:
1133                 wenv->changes |= optimize_load(n);
1134                 break;
1135
1136         case iro_Store:
1137                 wenv->changes |= optimize_store(n);
1138                 break;
1139
1140         case iro_Phi:
1141                 wenv->changes |= optimize_phi(n, wenv);
1142
1143         default:
1144                 ;
1145         }
1146 }  /* do_load_store_optimize */
1147
1148 /*
1149  * do the load store optimization
1150  */
1151 void optimize_load_store(ir_graph *irg) {
1152         walk_env_t env;
1153
1154         assert(get_irg_phase_state(irg) != phase_building);
1155         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1156                 "LoadStore optimization needs pinned graph");
1157
1158         if (! get_opt_redundant_loadstore())
1159                 return;
1160
1161         edges_assure(irg);
1162
1163         /* for Phi optimization post-dominators are needed ... */
1164         assure_postdoms(irg);
1165
1166         if (get_opt_alias_analysis()) {
1167                 assure_irg_address_taken_computed(irg);
1168                 assure_irp_globals_address_taken_computed();
1169         }
1170
1171         obstack_init(&env.obst);
1172         env.changes = 0;
1173
1174         /* init the links, then collect Loads/Stores/Proj's in lists */
1175         master_visited = 0;
1176         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1177
1178         /* now we have collected enough information, optimize */
1179         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1180
1181         obstack_free(&env.obst, NULL);
1182
1183         /* Handle graph state */
1184         if (env.changes) {
1185                 if (get_irg_outs_state(irg) == outs_consistent)
1186                         set_irg_outs_inconsistent(irg);
1187         }
1188
1189         if (env.changes & CF_CHANGED) {
1190                 /* is this really needed: Yes, control flow changed, block might
1191                 have Bad() predecessors. */
1192                 set_irg_doms_inconsistent(irg);
1193         }
1194 }  /* optimize_load_store */